Machine learning in automated text categorization

Machine Learning in Automated Text Categorization
FABRIZIO SEBASTIANI
Consiglio Nazionale delle Ricerche, Italy

The automated categorization (or classification) of texts into predefined categories has
witnessed a booming interest in the last 10 years, due to the increased availability of
documents in digital form and the ensuing need to organize them. In the research
community the dominant approach to this problem is based on machine learning
techniques: a general inductive process automatically builds a classifier by learning,
from a set of preclassified documents, the characteristics of the categories. The
advantages of this approach over the knowledge engineering approach (consisting in
the manual definition of a classifier by domain experts) are a very good effectiveness,
considerable savings in terms of expert labor power, and straightforward portability to
different domains. This survey discusses the main approaches to text categorization
that fall within the machine learning paradigm. We will discuss in detail issues
pertaining to three different problems, namely, document representation, classifier
construction, and classifier evaluation.
Categories and Subject Descriptors: H.3.1 [Information Storage and Retrieval]:
Content Analysis and Indexing—Indexing methods; H.3.3 [Information Storage and
Retrieval]: Information Search and Retrieval—Information filtering; H.3.4
[Information Storage and Retrieval]: Systems and Software—Performance
evaluation (efficiency and effectiveness); I.2.6 [Artificial Intelligence]: Learning—
Induction
General Terms: Algorithms, Experimentation, Theory
Additional Key Words and Phrases: Machine learning, text categorization, text
classification

1. INTRODUCTION

In the last 10 years content-based document management tasks (collectively
known as information retrieval—IR) have
gained a prominent status in the information systems field, due to the increased
availability of documents in digital form
and the ensuing need to access them in
flexible ways. Text categorization (TC—
a.k.a. text classification, or topic spotting),
the activity of labeling natural language

texts with thematic categories from a predefined set, is one such task. TC dates
back to the early ’60s, but only in the early
’90s did it become a major subfield of the
information systems discipline, thanks to
increased applicative interest and to the
availability of more powerful hardware.
TC is now being applied in many contexts,
ranging from document indexing based
on a controlled vocabulary, to document
filtering, automated metadata generation,
word sense disambiguation, population of

Author’s address: Istituto di Elaborazione dell’Informazione, Consiglio Nazionale delle Ricerche, Via G.
Moruzzi 1, 56124 Pisa, Italy; e-mail: fabrizio@iei.pi.cnr.it.
Permission to make digital/hard copy of part or all of this work for personal or classroom use is granted
without fee provided that the copies are not made or distributed for profit or commercial advantage, the
copyright notice, the title of the publication, and its date appear, and notice is given that copying is by
permission of the ACM, Inc. To copy otherwise, to republish, to post on servers, or to redistribute to lists,
requires prior specific permission and/or a fee.
c 2002 ACM 0360-0300/02/0300-0001 $5.00

ACM Computing Surveys, Vol. 34, No. 1, March 2002, pp. 1–47.

2
hierarchical catalogues of Web resources,
and in general any application requiring
document organization or selective and
adaptive document dispatching.
Until the late ’80s the most popular approach to TC, at least in the “operational”
(i.e., real-world applications) community,
was a knowledge engineering (KE) one,
consisting in manually defining a set of
rules encoding expert knowledge on how
to classify documents under the given categories. In the ’90s this approach has increasingly lost popularity (especially in
the research community) in favor of the
machine learning (ML) paradigm, according to which a general inductive process
automatically builds an automatic text
classifier by learning, from a set of preclassified documents, the characteristics of the
categories of interest. The advantages of
this approach are an accuracy comparable
to that achieved by human experts, and
a considerable savings in terms of expert
labor power, since no intervention from either knowledge engineers or domain experts is needed for the construction of the
classifier or for its porting to a different set
of categories. It is the ML approach to TC
that this paper concentrates on.
Current-day TC is thus a discipline at
the crossroads of ML and IR, and as
such it shares a number of characteristics with other tasks such as information/
knowledge extraction from texts and text
mining [Knight 1999; Pazienza 1997].
There is still considerable debate on where
the exact border between these disciplines
lies, and the terminology is still evolving.
“Text mining” is increasingly being used
to denote all the tasks that, by analyzing large quantities of text and detecting usage patterns, try to extract probably
useful (although only probably correct)
information. According to this view, TC is
an instance of text mining. TC enjoys quite
a rich literature now, but this is still fairly
scattered.1 Although two international
journals have devoted special issues to

Sebastiani
this topic [Joachims and Sebastiani 2002;
Lewis and Hayes 1994], there are no systematic treatments of the subject: there
are neither textbooks nor journals entirely devoted to TC yet, and Manning
¨
and Schutze [1999, Chapter 16] is the only
chapter-length treatment of the subject.
As a note, we should warn the reader
that the term “automatic text classification” has sometimes been used in the literature to mean things quite different from
the ones discussed here. Aside from (i) the
automatic assignment of documents to a
predefined set of categories, which is the
main topic of this paper, the term has also
been used to mean (ii) the automatic identification of such a set of categories (e.g.,
Borko and Bernick [1963]), or (iii) the automatic identification of such a set of categories and the grouping of documents
under them (e.g., Merkl [1998]), a task
usually called text clustering, or (iv) any
activity of placing text items into groups,
a task that has thus both TC and text clustering as particular instances [Manning
¨
and Schutze 1999].
This paper is organized as follows. In
Section 2 we formally define TC and its
various subcases, and in Section 3 we
review its most important applications.
Section 4 describes the main ideas underlying the ML approach to classification.
Our discussion of text classification starts
in Section 5 by introducing text indexing, that is, the transformation of textual
documents into a form that can be interpreted by a classifier-building algorithm
and by the classifier eventually built by it.
Section 6 tackles the inductive construction of a text classifier from a “training”
set of preclassified documents. Section 7
discusses the evaluation of text classifiers. Section 8 concludes, discussing open
issues and possible avenues of further
research for TC.
2. TEXT CATEGORIZATION
2.1. A Definition of Text Categorization

1

A fully searchable bibliography on TC created and
maintained by this author is available at http://
liinwww.ira.uka.de/bibliography/Ai/automated.text.
categorization.html.

Text categorization is the task of assigning
a Boolean value to each pair d j , ci ∈ D ×
C, where D is a domain of documents and
ACM Computing Surveys, Vol. 34, No. 1, March 2002.

C = {c1 , . . . , c|C| } is a set of predefined categories. A value of T assigned to d j , ci
indicates a decision to file d j under ci ,
while a value of F indicates a decision
not to file d j under ci . More formally, the
task is to approximate the unknown target function ˘ : D × C → {T, F } (that describes how documents ought to be classified) by means of a function : D × C →
{T, F } called the classifier (aka rule, or
hypothesis, or model) such that ˘ and
“coincide as much as possible.” How to precisely define and measure this coincidence
(called effectiveness) will be discussed in
Section 7.1. From now on we will assume
that:
—The categories are just symbolic labels, and no additional knowledge (of
a procedural or declarative nature) of
their meaning is available.
—No exogenous knowledge (i.e., data provided for classification purposes by an
external source) is available; therefore,
classification must be accomplished on
the basis of endogenous knowledge only
(i.e., knowledge extracted from the documents). In particular, this means that
metadata such as, for example, publication date, document type, publication source, etc., is not assumed to be
available.
The TC methods we will discuss are
thus completely general, and do not depend on the availability of special-purpose
resources that might be unavailable or
costly to develop. Of course, these assumptions need not be verified in operational settings, where it is legitimate to
use any source of information that might
be available or deemed worth developing
[Dáz Esteban et al. 1998; Junker and
ı
Abecker 1997]. Relying only on endogenous knowledge means classifying a document based solely on its semantics, and
given that the semantics of a document
is a subjective notion, it follows that the
membership of a document in a category (pretty much as the relevance of a
document to an information need in IR
[Saracevic 1975]) cannot be decided deterministically. This is exemplified by the

3

phenomenon of inter-indexer inconsistency
[Cleverdon 1984]: when two human experts decide whether to classify document
d j under category ci , they may disagree,
and this in fact happens with relatively
high frequency. A news article on Clinton
attending Dizzy Gillespie’s funeral could
be filed under Politics, or under Jazz, or under both, or even under neither, depending
on the subjective judgment of the expert.
2.2. Single-Label Versus Multilabel
Text Categorization

Different constraints may be enforced on
the TC task, depending on the application. For instance we might need that, for
a given integer k, exactly k (or ≤ k, or ≥ k)
elements of C be assigned to each d j ∈ D.
The case in which exactly one category
must be assigned to each d j ∈ D is often
called the single-label (a.k.a. nonoverlapping categories) case, while the case in
which any number of categories from 0
to |C| may be assigned to the same d j ∈ D
is dubbed the multilabel (aka overlapping
categories) case. A special case of singlelabel TC is binary TC, in which each d j ∈ D
must be assigned either to category ci or
to its complement ci .
¯
From a theoretical point of view, the
binary case (hence, the single-label case,
too) is more general than the multilabel,
since an algorithm for binary classification can also be used for multilabel classification: one needs only transform the
problem of multilabel classification under
{c1 , . . . , c|C| } into |C| independent problems
¯
of binary classification under {ci , ci }, for
i = 1, . . . , |C|. However, this requires that
categories be stochastically independent
of each other, that is, for any c , c , the
value of ˘ (d j , c ) does not depend on
the value of ˘ (d j , c ) and vice versa;
this is usually assumed to be the case
(applications in which this is not the case
are discussed in Section 3.5). The converse
is not true: an algorithm for multilabel
classification cannot be used for either binary or single-label classification. In fact,
given a document d j to classify, (i) the classifier might attribute k > 1 categories to
d j , and it might not be obvious how to

4
choose a “most appropriate” category from
them; or (ii) the classifier might attribute
to d j no category at all, and it might not
be obvious how to choose a “least inappropriate” category from C.
In the rest of the paper, unless explicitly
mentioned, we will deal with the binary
case. There are various reasons for this:
—The binary case is important in itself
because important TC applications, including filtering (see Section 3.3), consist of binary classification problems
(e.g., deciding whether d j is about Jazz
or not). In TC, most binary classification
problems feature unevenly populated
categories (e.g., much fewer documents
are about Jazz than are not) and unevenly characterized categories (e.g.,
what is about Jazz can be characterized
much better than what is not).
—Solving the binary case also means solving the multilabel case, which is also
representative of important TC applications, including automated indexing for
Boolean systems (see Section 3.1).
—Most of the TC literature is couched in
terms of the binary case.
—Most techniques for binary classification are just special cases of existing
techniques for the single-label case, and
are simpler to illustrate than these
latter.
This ultimately means that we will view
classification under C = {c1 , . . . , c|C| } as
consisting of |C| independent problems of
classifying the documents in D under a
given category ci , for i = 1, . . . , |C|. A classifier for ci is then a function i : D →
{T, F } that approximates an unknown target function ˘ i : D → {T, F }.
2.3. Category-Pivoted Versus
Document-Pivoted Text Categorization

There are two different ways of using
a text classifier. Given d j ∈ D, we might
want to find all the ci ∈ C under which it
should be filed (document-pivoted categorization—DPC); alternatively, given ci ∈ C,
we might want to find all the d j ∈ D that
should be filed under it (category-pivoted

Sebastiani
categorization—CPC). This distinction is
more pragmatic than conceptual, but is
important since the sets C and D might not
be available in their entirety right from
the start. It is also relevant to the choice
of the classifier-building method, as some
of these methods (see Section 6.9) allow
the construction of classifiers with a definite slant toward one or the other style.
DPC is thus suitable when documents
become available at different moments in
time, e.g., in filtering e-mail. CPC is instead suitable when (i) a new category
c|C|+1 may be added to an existing set
C = {c1 , . . . , c|C| } after a number of documents have already been classified under
C, and (ii) these documents need to be reconsidered for classification under c|C|+1
(e.g., Larkey [1999]). DPC is used more often than CPC, as the former situation is
more common than the latter.
Although some specific techniques apply to one style and not to the other (e.g.,
the proportional thresholding method discussed in Section 6.1 applies only to CPC),
this is more the exception than the rule:
most of the techniques we will discuss allow the construction of classifiers capable
of working in either mode.
2.4. “Hard” Categorization Versus
Ranking Categorization

While a complete automation of the
TC task requires a T or F decision
for each pair d j , ci , a partial automation of this process might have different
requirements.
For instance, given d j ∈ D a system
might simply rank the categories in
C = {c1 , . . . , c|C| } according to their estimated appropriateness to d j , without taking any “hard” decision on any of them.
Such a ranked list would be of great
help to a human expert in charge of
taking the final categorization decision,
since she could thus restrict the choice
to the category (or categories) at the top
of the list, rather than having to examine
the entire set. Alternatively, given ci ∈ C
a system might simply rank the documents in D according to their estimated
appropriateness to ci ; symmetrically, for

classification under ci a human expert
would just examine the top-ranked documents instead of the entire document
set. These two modalities are sometimes
called category-ranking TC and documentranking TC [Yang 1999], respectively,
and are the obvious counterparts of DPC
and CPC.
Semiautomated, “interactive” classification systems [Larkey and Croft 1996] are
useful especially in critical applications
in which the effectiveness of a fully automated system may be expected to be
significantly lower than that of a human
expert. This may be the case when the
quality of the training data (see Section 4)
is low, or when the training documents
cannot be trusted to be a representative
sample of the unseen documents that are
to come, so that the results of a completely
automatic classifier could not be trusted
completely.
In the rest of the paper, unless explicitly
mentioned, we will deal with “hard” classification; however, many of the algorithms
we will discuss naturally lend themselves
to ranking TC too (more details on this in
Section 6.1).
3. APPLICATIONS OF TEXT
CATEGORIZATION

TC goes back to Maron’s [1961] seminal work on probabilistic text classification. Since then, it has been used for a
number of different applications, of which
we here briefly review the most important ones. Note that the borders between
the different classes of applications listed
here are fuzzy and somehow artificial, and
some of these may be considered special
cases of others. Other applications we do
not explicitly discuss are speech categorization by means of a combination of
speech recognition and TC [Myers et al.
2000; Schapire and Singer 2000], multimedia document categorization through
the analysis of textual captions [Sable
and Hatzivassiloglou 2000], author identification for literary texts of unknown or
disputed authorship [Forsyth 1999], language identification for texts of unknown
language [Cavnar and Trenkle 1994],

5

automated identification of text genre
[Kessler et al. 1997], and automated essay
grading [Larkey 1998].
3.1. Automatic Indexing for Boolean
Information Retrieval Systems

The application that has spawned most
of the early research in the field [Borko
and Bernick 1963; Field 1975; Gray and
Harley 1971; Heaps 1973; Maron 1961]
is that of automatic document indexing
for IR systems relying on a controlled
dictionary, the most prominent example
of which is Boolean systems. In these
latter each document is assigned one or
more key words or key phrases describing its content, where these key words and
key phrases belong to a finite set called
controlled dictionary, often consisting of
a thematic hierarchical thesaurus (e.g.,
the NASA thesaurus for the aerospace
discipline, or the MESH thesaurus for
medicine). Usually, this assignment is
done by trained human indexers, and is
thus a costly activity.
If the entries in the controlled vocabulary are viewed as categories, text indexing is an instance of TC, and may
thus be addressed by the automatic techniques described in this paper. Recalling Section 2.2, note that this application may typically require that k1 ≤ x ≤ k2
key words are assigned to each document, for given k1 , k2 . Document-pivoted
TC is probably the best option, so that
new documents may be classified as they
become available. Various text classifiers
explicitly conceived for document indexing have been described in the literature;
see, for example, Fuhr and Knorz [1984],
Robertson and Harding [1984], and Tzeras
and Hartmann [1993].
Automatic indexing with controlled dictionaries is closely related to automated
metadata generation. In digital libraries,
one is usually interested in tagging documents by metadata that describes them
under a variety of aspects (e.g., creation
date, document type or format, availability, etc.). Some of this metadata is thematic, that is, its role is to describe the
semantics of the document by means of

6
bibliographic codes, key words or key
phrases. The generation of this metadata
may thus be viewed as a problem of document indexing with controlled dictionary, and thus tackled by means of TC
techniques.
3.2. Document Organization

Indexing with a controlled vocabulary is
an instance of the general problem of document base organization. In general, many
other issues pertaining to document organization and filing, be it for purposes
of personal organization or structuring of
a corporate document base, may be addressed by TC techniques. For instance,
at the offices of a newspaper incoming
“classified” ads must be, prior to publication, categorized under categories such
as Personals, Cars for Sale, Real Estate,
etc. Newspapers dealing with a high volume of classified ads would benefit from an
automatic system that chooses the most
suitable category for a given ad. Other
possible applications are the organization of patents into categories for making their search easier [Larkey 1999], the
automatic filing of newspaper articles under the appropriate sections (e.g., Politics,
Home News, Lifestyles, etc.), or the automatic grouping of conference papers into
sessions.
3.3. Text Filtering

Text filtering is the activity of classifying a stream of incoming documents dispatched in an asynchronous way by an
information producer to an information
consumer [Belkin and Croft 1992]. A typical case is a newsfeed, where the producer is a news agency and the consumer
is a newspaper [Hayes et al. 1990]. In
this case, the filtering system should block
the delivery of the documents the consumer is likely not interested in (e.g., all
news not concerning sports, in the case
of a sports newspaper). Filtering can be
seen as a case of single-label TC, that
is, the classification of incoming documents into two disjoint categories, the
relevant and the irrelevant. Additionally,

Sebastiani
a filtering system may also further classify the documents deemed relevant to
the consumer into thematic categories;
in the example above, all articles about
sports should be further classified according to which sport they deal with, so as
to allow journalists specialized in individual sports to access only documents of
prospective interest for them. Similarly,
an e-mail filter might be trained to discard
“junk” mail [Androutsopoulos et al. 2000;
Drucker et al. 1999] and further classify
nonjunk mail into topical categories of interest to the user.
A filtering system may be installed at
the producer end, in which case it must
route the documents to the interested consumers only, or at the consumer end, in
which case it must block the delivery of
documents deemed uninteresting to the
consumer. In the former case, the system
builds and updates a “profile” for each consumer [Liddy et al. 1994], while in the latter case (which is the more common, and
to which we will refer in the rest of this
section) a single profile is needed.
A profile may be initially specified by
the user, thereby resembling a standing
IR query, and is updated by the system
by using feedback information provided
(either implicitly or explicitly) by the user
on the relevance or nonrelevance of the delivered messages. In the TREC community
[Lewis 1995c], this is called adaptive filtering, while the case in which no userspecified profile is available is called either routing or batch filtering, depending
on whether documents have to be ranked
in decreasing order of estimated relevance
or just accepted/rejected. Batch filtering
thus coincides with single-label TC under |C| = 2 categories; since this latter is
a completely general TC task, some authors [Hull 1994; Hull et al. 1996; Schapire
¨
et al. 1998; Schutze et al. 1995], somewhat confusingly, use the term “filtering”
in place of the more appropriate term
“categorization.”
In information science, document filtering has a tradition dating back to the
’60s, when, addressed by systems of various degrees of automation and dealing
with the multiconsumer case discussed

above, it was called selective dissemination of information or current awareness
(see Korfhage [1997, Chapter 6]). The explosion in the availability of digital information has boosted the importance of such
systems, which are nowadays being used
in contexts such as the creation of personalized Web newspapers, junk e-mail blocking, and Usenet news selection.
Information filtering by ML techniques
is widely discussed in the literature: see
Amati and Crestani [1999], Iyer et al.
[2000], Kim et al. [2000], Tauritz et al.
[2000], and Yu and Lam [1998].

3.4. Word Sense Disambiguation

Word sense disambiguation (WSD) is the
activity of finding, given the occurrence in
a text of an ambiguous (i.e., polysemous
or homonymous) word, the sense of this
particular word occurrence. For instance,
bank may have (at least) two different
senses in English, as in the Bank of
England (a financial institution) or the
bank of river Thames (a hydraulic engineering artifact). It is thus a WSD task
to decide which of the above senses the occurrence of bank in Last week I borrowed
some money from the bank has. WSD is
very important for many applications, including natural language processing, and
indexing documents by word senses rather
than by words for IR purposes. WSD may
be seen as a TC task (see Gale et al.
[1993]; Escudero et al. [2000]) once we
view word occurrence contexts as documents and word senses as categories.
Quite obviously, this is a single-label TC
case, and one in which document-pivoted
TC is usually the right choice.
WSD is just an example of the more general issue of resolving natural language
ambiguities, one of the most important
problems in computational linguistics.
Other examples, which may all be tackled
by means of TC techniques along the lines
discussed for WSD, are context-sensitive
spelling correction, prepositional phrase
attachment, part of speech tagging, and
word choice selection in machine translation; see Roth [1998] for an introduction.

7

3.5. Hierarchical Categorization
of Web Pages

TC has recently aroused a lot of interest
also for its possible application to automatically classifying Web pages, or sites,
under the hierarchical catalogues hosted
by popular Internet portals. When Web
documents are catalogued in this way,
rather than issuing a query to a generalpurpose Web search engine a searcher
may find it easier to first navigate in
the hierarchy of categories and then restrict her search to a particular category
of interest.
Classifying Web pages automatically
has obvious advantages, since the manual categorization of a large enough subset of the Web is infeasible. Unlike in the
previous applications, it is typically the
case that each category must be populated
by a set of k1 ≤ x ≤ k2 documents. CPC
should be chosen so as to allow new categories to be added and obsolete ones to be
deleted.
With respect to previously discussed TC
applications, automatic Web page categorization has two essential peculiarities:
(1) The hypertextual nature of the documents: Links are a rich source of
information, as they may be understood as stating the relevance of the
linked page to the linking page. Techniques exploiting this intuition in a
TC context have been presented by
Attardi et al. [1998], Chakrabarti et al.
¨
[1998b], Furnkranz [1999], G¨ vert
o
et al. [1999], and Oh et al. [2000]
and experimentally compared by Yang
et al. [2002].
(2) The hierarchical structure of the category set: This may be used, for example,
by decomposing the classification problem into a number of smaller classification problems, each corresponding to a
branching decision at an internal node.
Techniques exploiting this intuition in
a TC context have been presented by
Dumais and Chen [2000], Chakrabarti
et al. [1998a], Koller and Sahami
[1997], McCallum et al. [1998], Ruiz
and Srinivasan [1999], and Weigend
et al. [1999].

8

Sebastiani
if

((wheat & farm)
(wheat & commodity)
(bushels & export)
(wheat & tonnes)
(wheat & winter & ¬ soft))

or
or
or
or
then WHEAT else ¬ WHEAT

Fig. 1. Rule-based classifier for the WHEAT category; key words
are indicated in italic, categories are indicated in SMALL CAPS (from
Apt´ et al. [1994]).
e

4. THE MACHINE LEARNING APPROACH
TO TEXT CATEGORIZATION

In the ’80s, the most popular approach
(at least in operational settings) for the
creation of automatic document classifiers
consisted in manually building, by means
of knowledge engineering (KE) techniques,
an expert system capable of taking TC decisions. Such an expert system would typically consist of a set of manually defined
logical rules, one per category, of type
if DNF formula then category .
A DNF (“disjunctive normal form”) formula is a disjunction of conjunctive
clauses; the document is classified under
category iff it satisfies the formula, that
is, iff it satisfies at least one of the clauses.
The most famous example of this approach
is the CONSTRUE system [Hayes et al. 1990],
built by Carnegie Group for the Reuters
news agency. A sample rule of the type
used in CONSTRUE is illustrated in Figure 1.
The drawback of this approach is
the knowledge acquisition bottleneck well
known from the expert systems literature.
That is, the rules must be manually defined by a knowledge engineer with the
aid of a domain expert (in this case, an
expert in the membership of documents in
the chosen set of categories): if the set of
categories is updated, then these two professionals must intervene again, and if the
classifier is ported to a completely different domain (i.e., set of categories), a different domain expert needs to intervene and
the work has to be repeated from scratch.
On the other hand, it was originally
suggested that this approach can give very
good effectiveness results: Hayes et al.
[1990] reported a .90 “breakeven” result
(see Section 7) on a subset of the Reuters
test collection, a figure that outperforms

even the best classifiers built in the late
’90s by state-of-the-art ML techniques.
However, no other classifier has been
tested on the same dataset as CONSTRUE,
and it is not clear whether this was a
randomly chosen or a favorable subset of
the entire Reuters collection. As argued
by Yang [1999], the results above do not
allow us to state that these effectiveness
results may be obtained in general.
Since the early ’90s, the ML approach
to TC has gained popularity and has
eventually become the dominant one, at
least in the research community (see
Mitchell [1996] for a comprehensive introduction to ML). In this approach, a general
inductive process (also called the learner)
automatically builds a classifier for a category ci by observing the characteristics
of a set of documents manually classified
¯
under ci or ci by a domain expert; from
these characteristics, the inductive process gleans the characteristics that a new
unseen document should have in order to
be classified under ci . In ML terminology,
the classification problem is an activity
of supervised learning, since the learning
process is “supervised” by the knowledge
of the categories and of the training instances that belong to them.2
The advantages of the ML approach
over the KE approach are evident. The engineering effort goes toward the construction not of a classifier, but of an automatic
builder of classifiers (the learner). This
means that if a learner is (as it often is)
available off-the-shelf, all that is needed
is the inductive, automatic construction of
a classifier from a set of manually classified documents. The same happens if a
2

Within the area of content-based document management tasks, an example of an unsupervised learning activity is document clustering (see Section 1).


classifier already exists and the original
set of categories is updated, or if the classifier is ported to a completely different
domain.
In the ML approach, the preclassified
documents are then the key resource.
In the most favorable case, they are already available; this typically happens for
organizations which have previously carried out the same categorization activity
manually and decide to automate the process. The less favorable case is when no
manually classified documents are available; this typically happens for organizations that start a categorization activity and opt for an automated modality
straightaway. The ML approach is more
convenient than the KE approach also in
this latter case. In fact, it is easier to manually classify a set of documents than to
build and tune a set of rules, since it is
easier to characterize a concept extensionally (i.e., to select instances of it) than intensionally (i.e., to describe the concept in
words, or to describe a procedure for recognizing its instances).
Classifiers built by means of ML techniques nowadays achieve impressive levels of effectiveness (see Section 7), making
automatic classification a qualitatively
(and not only economically) viable alternative to manual classification.
4.1. Training Set, Test Set, and
Validation Set

The ML approach relies on the availability of an initial corpus = {d 1 , . . . , d | | } ⊂
D of documents preclassified under C =
{c1 , . . . , c|C| }. That is, the values of the total
function ˘ : D × C → {T, F } are known for
every pair d j , ci ∈ × C. A document d j
is a positive example of ci if ˘ (d j , ci ) = T ,
a negative example of ci if ˘ (d j , ci ) = F .
In research settings (and in most operational settings too), once a classifier has
been built it is desirable to evaluate its effectiveness. In this case, prior to classifier
construction the initial corpus is split in
two sets, not necessarily of equal size:
—a training(-and-validation) set T V =
{d 1 , . . . , d |T V | }. The classifier for categories C = {c1 , . . . , c|C| } is inductively

9

built by observing the characteristics of
these documents;
—a test set Te = {d |T V |+1 , . . . , d | | }, used
for testing the effectiveness of the classifiers. Each d j ∈ Te is fed to the classifier, and the classifier decisions (d j , ci )
are compared with the expert decisions
˘ (d j , ci ). A measure of classification
effectiveness is based on how often
the (d j , ci ) values match the ˘ (d j , ci )
values.
The documents in T e cannot participate
in any way in the inductive construction of the classifiers; if this condition
were not satisfied, the experimental results obtained would likely be unrealistically good, and the evaluation would
thus have no scientific character [Mitchell
1996, page 129]. In an operational setting,
after evaluation has been performed one
would typically retrain the classifier on
the entire initial corpus, in order to boost
effectiveness. In this case, the results of
the previous evaluation would be a pessimistic estimate of the real performance,
since the final classifier has been trained
on more data than the classifier evaluated.
This is called the train-and-test approach. An alternative is the k-fold crossvalidation approach (see Mitchell [1996],
page 146), in which k different classifiers 1 , . . . , k are built by partitioning the initial corpus into k disjoint sets
T e1 , . . . , T ek and then iteratively applying the train-and-test approach on pairs
T Vi = −Tei , Tei . The final effectiveness
figure is obtained by individually computing the effectiveness of 1 , . . . , k , and
then averaging the individual results in
some way.
In both approaches, it is often the case
that the internal parameters of the classifiers must be tuned by testing which
values of the parameters yield the best
effectiveness. In order to make this optimization possible, in the train-and-test
approach the set {d 1 , . . . , d |T V | } is further
split into a training set Tr = {d 1 , . . . , d |Tr| },
from which the classifier is built, and a validation set Va = {d |Tr|+1 , . . . , d |T V | } (sometimes called a hold-out set), on which
the repeated tests of the classifier aimed

10

Sebastiani

at parameter optimization are performed;
the obvious variant may be used in the
k-fold cross-validation case. Note that, for
the same reason why we do not test a classifier on the documents it has been trained
on, we do not test it on the documents it
has been optimized on: test set and validation set must be kept separate.3
Given a corpus , one may define the
generality g (ci ) of a category ci as the
percentage of documents that belong to ci ,
that is:
g (ci ) =

|{d j ∈

| ˘ (d j , ci ) = T }|
.
| |

The training set generality g Tr (ci ), validation set generality g Va (ci ), and test set generality g Te (ci ) of ci may be defined in the
obvious way.
4.2. Information Retrieval Techniques
and Text Categorization

Text categorization heavily relies on the
basic machinery of IR. The reason is that
TC is a content-based document management task, and as such it shares many
characteristics with other IR tasks such
as text search.
IR techniques are used in three phases
of the text classifier life cycle:
(1) IR-style indexing is always performed
on the documents of the initial corpus
and on those to be classified during the
operational phase;
(2) IR-style techniques (such as document-request matching, query reformulation, . . .) are often used in the inductive construction of the classifiers;
(3) IR-style evaluation of the effectiveness
of the classifiers is performed.
The various approaches to classification
differ mostly for how they tackle (2),
although in a few cases nonstandard
3 From now on, we will take the freedom to use the
expression “test document” to denote any document
not in the training set and validation set. This includes thus any document submitted to the classifier
in the operational phase.

approaches to (1) and (3) are also used. Indexing, induction, and evaluation are the
themes of Sections 5, 6 and 7, respectively.
5. DOCUMENT INDEXING AND
DIMENSIONALITY REDUCTION
5.1. Document Indexing

Texts cannot be directly interpreted by a
classifier or by a classifier-building algorithm. Because of this, an indexing procedure that maps a text d j into a compact
representation of its content needs to be
uniformly applied to training, validation,
and test documents. The choice of a representation for text depends on what one
regards as the meaningful units of text
(the problem of lexical semantics) and the
meaningful natural language rules for the
combination of these units (the problem
of compositional semantics). Similarly to
what happens in IR, in TC this latter problem is usually disregarded,4 and a text
d j is usually represented as a vector of
term weights d j = w1 j , . . . , w|T | j , where
T is the set of terms (sometimes called
features) that occur at least once in at least
one document of Tr, and 0 ≤ wk j ≤ 1 represents, loosely speaking, how much term
tk contributes to the semantics of document d j . Differences among approaches
are accounted for by
(1) different ways to understand what a
term is;
(2) different ways to compute term
weights.
A typical choice for (1) is to identify terms
with words. This is often called either the
set of words or the bag of words approach
to document representation, depending on
whether weights are binary or not.
In a number of experiments [Apt´
e
et al. 1994; Dumais et al. 1998; Lewis
1992a], it has been found that representations more sophisticated than this do
not yield significantly better effectiveness,
thereby confirming similar results from IR
4

An exception to this is represented by learning approaches based on hidden Markov models [Denoyer
et al. 2001; Frasconi et al. 2002].


[Salton and Buckley 1988]. In particular,
some authors have used phrases, rather
than individual words, as indexing terms
¨
[Fuhr et al. 1991; Schutze et al. 1995;
Tzeras and Hartmann 1993], but the experimental results found to date have
not been uniformly encouraging, irrespectively of whether the notion of “phrase” is
motivated
—syntactically, that is, the phrase is such
according to a grammar of the language
(see Lewis [1992a]); or
—statistically, that is, the phrase is
not grammatically such, but is composed of a set/sequence of words whose
patterns of contiguous occurrence in the
collection are statistically significant
(see Caropreso et al. [2001]).
Lewis [1992a] argued that the likely reason for the discouraging results is that,
although indexing languages based on
phrases have superior semantic qualities,
they have inferior statistical qualities
with respect to word-only indexing languages: a phrase-only indexing language
has “more terms, more synonymous or
nearly synonymous terms, lower consistency of assignment (since synonymous
terms are not assigned to the same documents), and lower document frequency for
terms” [Lewis 1992a, page 40]. Although
his remarks are about syntactically motivated phrases, they also apply to statistically motivated ones, although perhaps to
a smaller degree. A combination of the two
approaches is probably the best way to
go: Tzeras and Hartmann [1993] obtained
significant improvements by using noun
phrases obtained through a combination
of syntactic and statistical criteria, where
a “crude” syntactic method was complemented by a statistical filter (only those
syntactic phrases that occurred at least
three times in the positive examples of a
category ci were retained). It is likely that
the final word on the usefulness of phrase
indexing in TC has still to be told, and
investigations in this direction are still
being actively pursued [Caropreso et al.
2001; Mladeni´ and Grobelnik 1998].
c
As for issue (2), weights usually
range between 0 and 1 (an exception is

11

Lewis et al. [1996]), and for ease of exposition we will assume they always do. As a
special case, binary weights may be used
(1 denoting presence and 0 absence of the
term in the document); whether binary or
nonbinary weights are used depends on
the classifier learning algorithm used. In
the case of nonbinary indexing, for determining the weight wkj of term tk in document d j any IR-style indexing technique
that represents a document as a vector of
weighted terms may be used. Most of the
times, the standard tfidf function is used
(see Salton and Buckley [1988]), defined as
tfidf (tk , d j ) = #(tk , d j ) · log

|Tr|
, (1)
#Tr (tk )

where #(tk , d j ) denotes the number of
times tk occurs in d j , and #Tr (tk ) denotes
the document frequency of term tk , that
is, the number of documents in Tr in
which tk occurs. This function embodies
the intuitions that (i) the more often a
term occurs in a document, the more it
is representative of its content, and (ii)
the more documents a term occurs in,
the less discriminating it is.5 Note that
this formula (as most other indexing
formulae) weights the importance of a
term to a document in terms of occurrence
considerations only, thereby deeming of
null importance the order in which the
terms occur in the document and the syntactic role they play. In other words, the
semantics of a document is reduced to the
collective lexical semantics of the terms
that occur in it, thereby disregarding the
issue of compositional semantics (an exception are the representation techniques
used for FOIL [Cohen 1995a] and SLEEPING
EXPERTS [Cohen and Singer 1999]).
In order for the weights to fall in the
[0,1] interval and for the documents to
be represented by vectors of equal length,
the weights resulting from tfidf are often
5 There exist many variants of tfidf, that differ from
each other in terms of logarithms, normalization or
other correction factors. Formula 1 is just one of
the possible instances of this class; see Salton and
Buckley [1988] and Singhal et al. [1996] for variations on this theme.

12

Sebastiani

normalized by cosine normalization, given
by
wkj =

tfidf (tk , d j )
|T |
s=1 (tfidf (ts ,

.

(2)

d j ))2

Although normalized tfidf is the most
popular one, other indexing functions
have also been used, including probabilistic techniques [G¨ vert et al. 1999] or
o
techniques for indexing structured documents [Larkey and Croft 1996]. Functions
different from tfidf are especially needed
when Tr is not available in its entirety
from the start and #Tr (tk ) cannot thus be
computed, as in adaptive filtering; in this
case, approximations of tfidf are usually
employed [Dagan et al. 1997, Section 4.3].
Before indexing, the removal of function
words (i.e., topic-neutral words such as articles, prepositions, conjunctions, etc.) is
almost always performed (exceptions include Lewis et al. [1996], Nigam et al.
[2000], and Riloff [1995]).6 Concerning
stemming (i.e., grouping words that share
the same morphological root), its suitability to TC is controversial. Although, similarly to unsupervised term clustering (see
Section 5.5.1) of which it is an instance,
stemming has sometimes been reported
to hurt effectiveness (e.g., Baker and
McCallum [1998]), the recent tendency is
to adopt it, as it reduces both the dimensionality of the term space (see Section 5.3)
and the stochastic dependence between
terms (see Section 6.2).
Depending on the application, either
the full text of the document or selected
parts of it are indexed. While the former
option is the rule, exceptions exist. For
instance, in a patent categorization application Larkey [1999] indexed only the
title, the abstract, the first 20 lines of
the summary, and the section containing
6

One application of TC in which it would be inappropriate to remove function words is author identification for documents of disputed paternity. In fact,
¨
as noted in Manning and Schutze [1999], page 589,
“it is often the ‘little’ words that give an author away
(for example, the relative frequencies of words like
because or though).”

the claims of novelty of the described invention. This approach was made possible by the fact that documents describing
patents are structured. Similarly, when a
document title is available, one can pay
extra importance to the words it contains
[Apt´ et al. 1994; Cohen and Singer 1999;
e
Weiss et al. 1999]. When documents are
flat, identifying the most relevant part of
a document is instead a nonobvious task.
5.2. The Darmstadt Indexing Approach

The AIR/X system [Fuhr et al. 1991] occupies a special place in the literature on
indexing for TC. This system is the final
result of the AIR project, one of the most
important efforts in the history of TC:
spanning a duration of more than 10 years
[Knorz 1982; Tzeras and Hartmann 1993],
it has produced a system operatively employed since 1985 in the classification of
corpora of scientific literature of O(105 )
documents and O(104 ) categories, and has
had important theoretical spin-offs in the
field of probabilistic indexing [Fuhr 1989;
Fuhr and Buckely 1991].7
The approach to indexing taken in
AIR/X is known as the Darmstadt Indexing Approach (DIA) [Fuhr 1985].
Here, “indexing” is used in the sense of
Section 3.1, that is, as using terms from
a controlled vocabulary, and is thus a
synonym of TC (the DIA was later extended to indexing with free terms [Fuhr
and Buckley 1991]). The idea that underlies the DIA is the use of a much wider
set of “features” than described in Section 5.1. All other approaches mentioned
in this paper view terms as the dimensions of the learning space, where terms
may be single words, stems, phrases, or
(see Sections 5.5.1 and 5.5.2) combinations of any of these. In contrast, the DIA
considers properties (of terms, documents,
7 The AIR/X system, its applications (including the
AIR/PHYS system [Biebricher et al. 1988], an application of AIR/X to indexing physics literature), and
its experiments have also been richly documented
in a series of papers and doctoral theses written in
German. The interested reader may consult Fuhr
et al. [1991] for a detailed bibliography.


categories, or pairwise relationships among these) as basic dimensions of the
learning space. Examples of these are
—properties of a term tk : e.g. the idf of tk ;
—properties of the relationship between a
term tk and a document d j : for example,
the t f of tk in d j ; or the location (e.g., in
the title, or in the abstract) of tk within
dj;
—properties of a document d j : for example, the length of d j ;
—properties of a category ci : for example,
the training set generality of ci .
For each possible document-category pair,
the values of these features are collected
in a so-called relevance description vector rd(d j , ci ). The size of this vector is
determined by the number of properties
considered, and is thus independent of
specific terms, categories, or documents
(for multivalued features, appropriate aggregation functions are applied in order
to yield a single value to be included in
rd(d j , ci )); in this way an abstraction from
specific terms, categories, or documents is
achieved.
The main advantage of this approach
is the possibility to consider additional
features that can hardly be accounted for
in the usual term-based approaches, for
example, the location of a term within a
document, or the certainty with which a
phrase was identified in a document. The
term-category relationship is described by
estimates, derived from the training set, of
the probability P (ci | tk ) that a document
belongs to category ci , given that it contains term tk (the DIA association factor).8
Relevance description vectors rd (d j , ci )
are then the final representations that
are used for the classification of document
d j under category ci .
The essential ideas of the DIA—
transforming the classification space by
means of abstraction and using a more detailed text representation than the standard bag-of-words approach—have not
8

Association factors are called adhesion coefficients
in many early papers on TC; see Field [1975];
Robertson and Harding [1984].


13

been taken up by other researchers so
far. For new TC applications dealing with
structured documents or categorization of
Web pages, these ideas may become of increasing importance.
5.3. Dimensionality Reduction

Unlike in text retrieval, in TC the high
dimensionality of the term space (i.e.,
the large value of |T |) may be problematic. In fact, while typical algorithms used
in text retrieval (such as cosine matching) can scale to high values of |T |, the
same does not hold of many sophisticated
learning algorithms used for classifier induction (e.g., the LLSF algorithm of Yang
and Chute [1994]). Because of this, before classifier induction one often applies
a pass of dimensionality reduction (DR),
whose effect is to reduce the size of the
|T |; the set
vector space from |T | to |T |
T is called the reduced term set.
DR is also beneficial since it tends to reduce overfitting, that is, the phenomenon
by which a classifier is tuned also to
the contingent characteristics of the training data rather than just the constitutive characteristics of the categories. Classifiers that overfit the training data are
good at reclassifying the data they have
been trained on, but much worse at classifying previously unseen data. Experiments have shown that, in order to avoid
overfitting a number of training examples roughly proportional to the number
of terms used is needed; Fuhr and Buckley
[1991, page 235] have suggested that 50–
100 training examples per term may be
needed in TC tasks. This means that, if DR
is performed, overfitting may be avoided
even if a smaller amount of training examples is used. However, in removing terms
the risk is to remove potentially useful
information on the meaning of the documents. It is then clear that, in order to
obtain optimal (cost-)effectiveness, the reduction process must be performed with
care. Various DR methods have been proposed, either from the information theory
or from the linear algebra literature, and
their relative merits have been tested by
experimentally evaluating the variation

14
in effectiveness that a given classifier
undergoes after application of the function
to the term space.
There are two distinct ways of viewing DR, depending on whether the task is
performed locally (i.e., for each individual
category) or globally:
—local DR: for each category ci , a set Ti of
|T |, is chosen for clasterms, with |Ti |
e
sification under ci (see Apt´ et al. [1994];
Lewis and Ringuette [1994]; Li and
Jain [1998]; Ng et al. [1997]; Sable and
¨
Hatzivassiloglou [2000]; Schutze et al.
[1995], Wiener et al. [1995]). This means
that different subsets of d j are used
when working with the different categories. Typical values are 10 ≤ |Ti | ≤ 50.
—global DR: a set T of terms, with
|T |, is chosen for the classifica|T |
tion under all categories C = {c1 , . . . , c|C| }
(see Caropreso et al. [2001]; Mladeni´
c
[1998]; Yang [1999]; Yang and Pedersen
[1997]).
This distinction usually does not impact
on the choice of DR technique, since
most such techniques can be used (and
have been used) for local and global
DR alike (supervised DR techniques—see
Section 5.5.1—are exceptions to this rule).
In the rest of this section, we will assume
that the global approach is used, although
everything we will say also applies to the
local approach.
A second, orthogonal distinction may be
drawn in terms of the nature of the resulting terms:
—DR by term selection: T is a subset
of T ;
—DR by term extraction: the terms in
T are not of the same type of the
terms in T (e.g., if the terms in T are
words, the terms in T may not be words
at all), but are obtained by combinations or transformations of the original
ones.
Unlike in the previous distinction, these
two ways of doing DR are tackled by very
different techniques; we will address them
separately in the next two sections.

Sebastiani
by Term Selection

Given a predetermined integer r, techniques for term selection (also called term
space reduction—TSR) attempt to select,
from the original set T , the set T of
|T |) that, when used
terms (with |T |
for document indexing, yields the highest
effectiveness. Yang and Pedersen [1997]
have shown that TSR may even result in
a moderate (≤5%) increase in effectiveness, depending on the classifier, on the
|T
aggressivity |T || of the reduction, and on
the TSR technique used.
Moulinier et al. [1996] have used a socalled wrapper approach, that is, one in
which T is identified by means of the
same learning method that will be used for
building the classifier [John et al. 1994].
Starting from an initial term set, a new
term set is generated by either adding
or removing a term. When a new term
set is generated, a classifier based on it
is built and then tested on a validation
set. The term set that results in the best
effectiveness is chosen. This approach has
the advantage of being tuned to the learning algorithm being used; moreover, if local DR is performed, different numbers of
terms for different categories may be chosen, depending on whether a category is
or is not easily separable from the others.
However, the sheer size of the space of different term sets makes its cost-prohibitive
for standard TC applications.
A computationally easier alternative is
the filtering approach [John et al. 1994],
|T | terms that
that is, keeping the |T |
receive the highest score according to a
function that measures the “importance”
of the term for the TC task. We will explore
this solution in the rest of this section.
5.4.1. Document Frequency. A simple and
effective global TSR function is the document frequency #Tr (tk ) of a term tk , that is,
only the terms that occur in the highest
number of documents are retained. In a
series of experiments Yang and Pedersen
[1997] have shown that with #Tr (tk ) it is
possible to reduce the dimensionality by a
factor of 10 with no loss in effectiveness (a

reduction by a factor of 100 bringing about
just a small loss).
This seems to indicate that the terms
occurring most frequently in the collection
are the most valuable for TC. As such, this
would seem to contradict a well-known
“law” of IR, according to which the terms
with low-to-medium document frequency
are the most informative ones [Salton and
Buckley 1988]. But these two results do
not contradict each other, since it is well
known (see Salton et al. [1975]) that the
large majority of the words occurring in
a corpus have a very low document frequency; this means that by reducing the
term set by a factor of 10 using document
frequency, only such words are removed,
while the words from low-to-medium to
high document frequency are preserved.
Of course, stop words need to be removed
in advance, lest only topic-neutral words
are retained [Mladeni´ 1998].
c
Finally, note that a slightly more empirical form of TSR by document frequency
is adopted by many authors, who remove
all terms occurring in at most x training documents (popular values for x range
from 1 to 3), either as the only form of DR
[Maron 1961; Ittner et al. 1995] or before
applying another more sophisticated form
[Dumais et al. 1998; Li and Jain 1998]. A
variant of this policy is removing all terms
that occur at most x times in the training set (e.g., Dagan et al. [1997]; Joachims
[1997]), with popular values for x ranging from 1 (e.g., Baker and McCallum
[1998]) to 5 (e.g., Apt´ et al. [1994]; Cohen
e
[1995a]).
5.4.2. Other Information-Theoretic Term
Selection Functions. Other more sophis-

ticated information-theoretic functions
have been used in the literature, among
them the DIA association factor [Fuhr
et al. 1991], chi-square [Caropreso et al.
¨
2001; Galavotti et al. 2000; Schutze et al.
1995; Sebastiani et al. 2000; Yang and
Pedersen 1997; Yang and Liu 1999],
NGL coefficient [Ng et al. 1997; Ruiz
and Srinivasan 1999], information gain
[Caropreso et al. 2001; Larkey 1998;
Lewis 1992a; Lewis and Ringuette 1994;
Mladeni´ 1998; Moulinier and Ganascia
c

15

1996; Yang and Pedersen 1997, Yang and
Liu 1999], mutual information [Dumais
et al. 1998; Lam et al. 1997; Larkey
and Croft 1996; Lewis and Ringuette
1994; Li and Jain 1998; Moulinier et al.
1996; Ruiz and Srinivasan 1999; Taira
and Haruno 1999; Yang and Pedersen
1997], odds ratio [Caropreso et al. 2001;
Mladeni´ 1998; Ruiz and Srinivasan
c
1999], relevancy score [Wiener et al.
1995], and GSS coefficient [Galavotti
et al. 2000]. The mathematical definitions
of these measures are summarized for
convenience in Table I.9 Here, probabilities are interpreted on an event space
¯
of documents (e.g., P (t k , ci ) denotes the
probability that, for a random document
x, term tk does not occur in x and x
belongs to category ci ), and are estimated
by counting occurrences in the training
set. All functions are specified “locally” to
a specific category ci ; in order to assess the
value of a term tk in a “global,” categoryindependent sense, either the sum
|C|
f sum (tk ) = i=1 f (tk , ci ), or the weighted
|C|
sum f wsum (tk ) = i=1 P (ci ) f (tk , ci ), or the
|C|
maximum f max (tk ) = maxi=1 f (tk , ci ) of
their category-specific values f (tk , ci ) are
usually computed.
These functions try to capture the intuition that the best terms for ci are the
ones distributed most differently in the
sets of positive and negative examples of
ci . However, interpretations of this principle vary across different functions. For
instance, in the experimental sciences χ 2
is used to measure how the results of an
observation differ (i.e., are independent)
from the results expected according to an
initial hypothesis (lower values indicate
lower dependence). In DR we measure how
independent tk and ci are. The terms tk
9 For better uniformity Table I views all the TSR
functions of this section in terms of subjective probability. In some cases such as χ 2 (tk , ci ) this is slightly
artificial, since this function is not usually viewed in
probabilistic terms. The formulae refer to the “local”
(i.e., category-specific) forms of the functions, which
again is slightly artificial in some cases. Note that
the NGL and GSS coefficients are here named after
their authors, since they had originally been given
names that might generate some confusion if used
here.

16

Sebastiani
Table I. Main Functions Used for Term Space Reduction Purposes. (Information gain is also known as
expected mutual information, and is used under this name by Lewis [1992a, page 44] and
Larkey [1998]. In the RS (t k , c i ) formula, d is a constant damping factor.)
Function
Denoted by
Mathematical form
DIA association factor
Information gain

P (ci | tk )

z(tk , ci )
IG(tk , ci )

P (t, c) · log
¯
c∈{ci , ci } t∈{tk , t k }
¯

P (tk , ci )
P (tk ) · P (ci )

Mutual information

MI(tk , ci )

Chi-square

χ 2 (tk , ci )

NGL coefficient

NGL(tk , ci )

Relevancy score

RS(tk , ci )

Odds ratio

OR(tk , ci )

P (tk | ci ) · (1 − P (tk | ci ))
¯
(1 − P (tk | ci )) · P (tk | ci )
¯

GSS coefficient

GSS(tk , ci )

¯ ¯
¯
P (tk , ci ) · P (t k , ci ) − P (tk , ci ) · P (t k , ci )
¯

with the lowest value for χ 2 (tk , ci ) are thus
the most independent from ci ; since we
are interested in the terms which are not,
we select the terms for which χ 2 (tk , ci ) is
highest.
While each TSR function has its own
rationale, the ultimate word on its value
is the effectiveness it brings about. Various experimental comparisons of TSR
functions have thus been carried out
[Caropreso et al. 2001; Galavotti et al.
2000; Mladeni´ 1998; Yang and Pedersen
c
1997]. In these experiments most functions listed in Table I (with the possible
exception of MI) have improved on the results of document frequency. For instance,
Yang and Pedersen [1997] have shown
that, with various classifiers and various
initial corpora, sophisticated techniques
2
such as IGsum (tk , ci ) or χmax (tk , ci ) can reduce the dimensionality of the term space
by a factor of 100 with no loss (or even
with a small increase) of effectiveness.
Collectively, the experiments reported in
the above-mentioned papers seem to indicate that {ORsum , NGLsum , GSSmax } >
2
2
{χmax , IGsum } > {χwavg } {MImax , MIwsum },
where “>” means “performs better than.”

log

P (t, c)
P (t) · P (c)

¯ ¯
¯
|Tr| · [P (tk , ci ) · P (t k , ci ) − P (tk , ci ) · P (t k , ci )]2
¯
¯
P (tk ) · P (t k ) · P (ci ) · P (ci )
¯
¯ ¯
¯
|Tr| · [P (tk , ci ) · P (t k , ci ) − P (tk , ci ) · P (t k , ci )]
¯
¯
P (tk ) · P (t k ) · P (ci ) · P (ci )
¯
log

P (tk | ci ) + d
¯
P (t k | ci ) + d
¯

However, it should be noted that these
results are just indicative, and that more
general statements on the relative merits of these functions could be made only
as a result of comparative experiments
performed in thoroughly controlled conditions and on a variety of different situations (e.g., different classifiers, different
initial corpora, . . . ).
by Term Extraction

Given a predetermined |T | |T |, term extraction attempts to generate, from the
original set T , a set T of “synthetic”
terms that maximize effectiveness. The
rationale for using synthetic (rather than
naturally occurring) terms is that, due
to the pervasive problems of polysemy,
homonymy, and synonymy, the original
terms may not be optimal dimensions
for document content representation.
Methods for term extraction try to solve
these problems by creating artificial terms
that do not suffer from them. Any term extraction method consists in (i) a method
for extracting the new terms from the

old ones, and (ii) a method for converting the original document representations into new representations based on
the newly synthesized dimensions. Two
term extraction methods have been experimented with in TC, namely term clustering and latent semantic indexing.

5.5.1. Term Clustering. Term clustering
tries to group words with a high degree of
pairwise semantic relatedness, so that the
groups (or their centroids, or a representative of them) may be used instead of the
terms as dimensions of the vector space.
Term clustering is different from term selection, since the former tends to address
terms synonymous (or near-synonymous)
with other terms, while the latter targets
noninformative terms.10
Lewis [1992a] was the first to investigate the use of term clustering in TC.
The method he employed, called reciprocal nearest neighbor clustering, consists
in creating clusters of two terms that are
one the most similar to the other according to some measure of similarity. His results were inferior to those obtained by
single-word indexing, possibly due to a disappointing performance by the clustering
method: as Lewis [1992a, page 48] said,
“The relationships captured in the clusters are mostly accidental, rather than the
systematic relationships that were hoped
for.”
Li and Jain [1998] viewed semantic
relatedness between words in terms of
their co-occurrence and co-absence within
training documents. By using this technique in the context of a hierarchical
clustering algorithm, they witnessed only
a marginal effectiveness improvement;
however, the small size of their experiment
(see Section 6.11) hardly allows any definitive conclusion to be reached.
Both Lewis [1992a] and Li and Jain
[1998] are examples of unsupervised clustering, since clustering is not affected by
the category labels attached to the docu10

Some term selection methods, such as wrapper
methods, also address the problem of redundancy.


17

ments. Baker and McCallum [1998] provided instead an example of supervised
clustering, as the distributional clustering
method they employed clusters together
those terms that tend to indicate the presence of the same category, or group of categories. Their experiments, carried out in
the context of a Na¨ve Bayes classifier (see
ı
Section 6.2), showed only a 2% effectiveness loss with an aggressivity of 1,000,
and even showed some effectiveness improvement with less aggressive levels of
reduction. Later experiments by Slonim
and Tishby [2001] have confirmed the potential of supervised clustering methods
for term extraction.
5.5.2. Latent Semantic Indexing. Latent semantic indexing (LSI—[Deerwester et al.
1990]) is a DR technique developed in IR
in order to address the problems deriving from the use of synonymous, nearsynonymous, and polysemous words as
dimensions of document representations.
This technique compresses document vectors into vectors of a lower-dimensional
space whose dimensions are obtained
as combinations of the original dimensions by looking at their patterns of cooccurrence. In practice, LSI infers the
dependence among the original terms
from a corpus and “wires” this dependence
into the newly obtained, independent dimensions. The function mapping original
vectors into new vectors is obtained by applying a singular value decomposition to
the matrix formed by the original document vectors. In TC this technique is applied by deriving the mapping function
from the training set and then applying
it to training and test documents alike.
One characteristic of LSI is that the
newly obtained dimensions are not, unlike
in term selection and term clustering,
intuitively interpretable. However, they
work well in bringing out the “latent”
semantic structure of the vocabulary
¨
used in the corpus. For instance, Schutze
et al. [1995, page 235] discussed the classification under category Demographic
shifts in the U.S. with economic impact of
a document that was indeed a positive

18
test instance for the category, and that
contained, among others, the quite revealing sentence The nation grew to 249.6
million people in the 1980s as more
Americans left the industrial and agricultural heartlands for the South
and West. The classifier decision was incorrect when local DR had been performed
by χ 2 -based term selection retaining the
top original 200 terms, but was correct
when the same task was tackled by
means of LSI. This well exemplifies
how LSI works: the above sentence does
not contain any of the 200 terms most
relevant to the category selected by χ 2 ,
but quite possibly the words contained in
it have concurred to produce one or more
of the LSI higher-order terms that generate the document space of the category.
¨
As Schutze et al. [1995, page 230] put it,
“if there is a great number of terms which
all contribute a small amount of critical
information, then the combination of evidence is a major problem for a term-based
classifier.” A drawback of LSI, though, is
that if some original term is particularly
good in itself at discriminating a category,
that discrimination power may be lost in
the new vector space.
Wiener et al. [1995] used LSI in two
alternative ways: (i) for local DR, thus
creating several category-specific LSI
representations, and (ii) for global DR,
thus creating a single LSI representation for the entire category set. Their
experiments showed the former approach
to perform better than the latter, and
both approaches to perform better than
simple TSR based on Relevancy Score
(see Table I).
¨
Schutze et al. [1995] experimentally
compared LSI-based term extraction with
χ 2 -based TSR using three different classifier learning techniques (namely, linear
discriminant analysis, logistic regression,
and neural networks). Their experiments
showed LSI to be far more effective than
χ 2 for the first two techniques, while both
methods performed equally well for the
neural network classifier.
For other TC works that have used
LSI or similar term extraction techniques,
see Hull [1994], Li and Jain [1998],

Sebastiani
¨
Schutze [1998], Weigend et al. [1999], and
Yang [1995].

6. INDUCTIVE CONSTRUCTION
OF TEXT CLASSIFIERS

The inductive construction of text classifiers has been tackled in a variety of
ways. Here we will deal only with the
methods that have been most popular
in TC, but we will also briefly mention
the existence of alternative, less standard
approaches.
We start by discussing the general
form that a text classifier has. Let us
recall from Section 2.4 that there are
two alternative ways of viewing classification: “hard” (fully automated) classification and ranking (semiautomated)
classification.
The inductive construction of a ranking
classifier for category ci ∈ C usually consists in the definition of a function CSVi :
D → [0, 1] that, given a document d j , returns a categorization status value for it,
that is, a number between 0 and 1 which,
roughly speaking, represents the evidence
for the fact that d j ∈ ci . Documents are
then ranked according to their CSVi value.
This works for “document-ranking TC”;
“category-ranking TC” is usually tackled
by ranking, for a given document d j , its
CSVi scores for the different categories in
C = {c1 , . . . , c|C| }.
The CSVi function takes up different meanings according to the learning method used: for instance, in the
“Na¨ve Bayes” approach of Section 6.2
ı
CSVi (d j ) is defined in terms of a probability, whereas in the “Rocchio” approach
discussed in Section 6.7 CSVi (d j ) is a measure of vector closeness in |T |-dimensional
space.
The construction of a “hard” classifier may follow two alternative paths.
The former consists in the definition of
a function CSVi : D → {T, F }. The latter consists instead in the definition of
a function CSVi : D → [0, 1], analogous
to the one used for ranking classification,
followed by the definition of a threshold
τi such that CSVi (d j ) ≥ τi is interpreted

as T while CSVi (d j ) < τi is interpreted
as F .11
The definition of thresholds will be the
topic of Section 6.1. In Sections 6.2 to 6.12
we will instead concentrate on the definition of CSVi , discussing a number of approaches that have been applied in the TC
literature. In general we will assume we
are dealing with “hard” classification; it
will be evident from the context how and
whether the approaches can be adapted to
ranking classification. The presentation of
the algorithms will be mostly qualitative
rather than quantitative, that is, will focus on the methods for classifier learning
rather than on the effectiveness and efficiency of the classifiers built by means
of them; this will instead be the focus of
Section 7.
6.1. Determining Thresholds

There are various policies for determining the threshold τi , also depending on the
constraints imposed by the application.
The most important distinction is whether
the threshold is derived analytically or
experimentally.
The former method is possible only in
the presence of a theoretical result that indicates how to compute the threshold that
maximizes the expected value of the effectiveness function [Lewis 1995a]. This is
typical of classifiers that output probability estimates of the membership of d j in ci
(see Section 6.2) and whose effectiveness is
computed by decision-theoretic measures
such as utility (see Section 7.1.3); we thus
defer the discussion of this policy (which
is called probability thresholding in Lewis
[1995a]) to Section 7.1.3.
When such a theoretical result is not
known, one has to revert to the latter
method, which consists in testing different
values for τi on a validation set and choosing the value which maximizes effectiveness. We call this policy CSV thresholding
11

Alternative methods are possible, such as training a classifier for which some standard, predefined
value such as 0 is the threshold. For ease of exposition we will not discuss them.


19

[Cohen and Singer 1999; Schapire et al.
1998; Wiener et al. 1995]; it is also called
Scut in Yang [1999]. Different τi ’s are typically chosen for the different ci ’s.
A second, popular experimental policy is proportional thresholding [Iwayama
and Tokunaga 1995; Larkey 1998; Lewis
1992a; Lewis and Ringuette 1994; Wiener
et al. 1995], also called Pcut in Yang
[1999]. This policy consists in choosing
the value of τi for which g Va (ci ) is closest to g Tr (ci ), and embodies the principle
that the same percentage of documents of
both training and test set should be classified under ci . For obvious reasons, this
policy does not lend itself to documentpivoted TC.
Sometimes, depending on the application, a fixed thresholding policy (a.k.a.
“k-per-doc” thresholding [Lewis 1992a] or
Rcut [Yang 1999]) is applied, whereby it is
stipulated that a fixed number k of categories, equal for all d j ’s, are to be assigned
to each document d j . This is often used,
for instance, in applications of TC to automated document indexing [Field 1975;
Lam et al. 1999]. Strictly speaking, however, this is not a thresholding policy in the
sense defined at the beginning of Section 6,
as it might happen that d is classified under ci , d is not, and CSVi (d ) < CSVi (d ).
Quite clearly, this policy is mostly at home
with document-pivoted TC. However, it
suffers from a certain coarseness, as the
fact that k is equal for all documents (nor
could this be otherwise) allows no finetuning.
In his experiments Lewis [1992a] found
the proportional policy to be superior to
probability thresholding when microaveraged effectiveness was tested but slightly
inferior with macroaveraging (see Section
7.1.1). Yang [1999] found instead CSV
thresholding to be superior to proportional
thresholding (possibly due to her categoryspecific optimization on a validation set),
and found fixed thresholding to be consistently inferior to the other two policies. The fact that these latter results have
been obtained across different classifiers
no doubt reinforces them.
In general, aside from the considerations above, the choice of the thresholding

20

Sebastiani

policy may also be influenced by the
application; for instance, in applying a
text classifier to document indexing for
Boolean systems a fixed thresholding policy might be chosen, while a proportional
or CSV thresholding method might be chosen for Web page classification under hierarchical catalogues.
6.2. Probabilistic Classifiers

Probabilistic classifiers (see Lewis [1998]
for a thorough discussion) view CSVi (d j )
in terms of P (ci |d j ), that is, the probability that a document represented by a
vector d j = w1 j , . . . , w|T | j of (binary or
weighted) terms belongs to ci , and compute this probability by an application of
Bayes’ theorem, given by
P (ci | d j ) =

P (ci )P (d j | ci )
P (d j )

.

(3)

In (3) the event space is the space of documents: P (d j ) is thus the probability that a
randomly picked document has vector d j
as its representation, and P (ci ) the probability that a randomly picked document
belongs to ci .
The estimation of P (d j | ci ) in (3) is
problematic, since the number of possible
vectors d j is too high (the same holds for
P (d j ), but for reasons that will be clear
shortly this will not concern us). In order to alleviate this problem it is common to make the assumption that any two
coordinates of the document vector are,
when viewed as random variables, statistically independent of each other; this independence assumption is encoded by the
equation

Sahami [1997]; Larkey and Croft [1996];
Lewis [1992a]; Lewis and Gale [1994];
Li and Jain [1998]; Robertson and
Harding [1984]). The “na¨ve” character of
ı
the classifier is due to the fact that usually this assumption is, quite obviously,
not verified in practice.
One of the best-known Na¨ve Bayes apı
proaches is the binary independence classifier [Robertson and Sparck Jones 1976],
which results from using binary-valued
vector representations for documents. In
this case, if we write pki as short for
P (wkx = 1 | ci ), the P (wk j | ci ) factors of
(4) may be written as
w

P (wk j | ci ) = pkik j (1 − pki )1−wk j
wk j
pki
=
(1 − pki ). (5)
1 − pki
We may further observe that in TC the
document space is partitioned into two
¯
categories,12 ci and its complement ci , such
that P (¯ i | d j ) = 1 − P (ci | d j ). If we plug
c
in (4) and (5) into (3) and take logs we
obtain
log P (ci | d j )
|T |

= log P (ci ) +
k=1

+

log(1 − pki ) − log P (d j )

(6)

k=1

log(1 − P (ci | d j ))
|T |

= log(1 − P (ci )) +

wk j log
k=1

pki
¯
1 − pki¯

|T |

+
P (wk j | ci ).

pki
1 − pki

|T |

|T |

P (d j | ci ) =

wk j log

log(1 − pki¯) − log P (d j ),

(7)

k=1

(4)

k=1
12

Probabilistic classifiers that use this assumption are called Na¨ve Bayes clası
sifiers, and account for most of the
probabilistic approaches to TC in the literature (see Joachims [1998]; Koller and

Cooper [1995] has pointed out that in this case
the full independence assumption of (4) is not actually made in the Na¨ve Bayes classifier; the ası
sumption needed here is instead the weaker linked
dependence assumption, which may be written as
P (d j | ci )
P (d j | ci )
¯

=

|T | P (wk j | ci )
.
¯
k=1 P (wk j | ci )


where we write pki as short for
¯
¯
P (wkx = 1 | ci ). We may convert (6) and (7)
into a single equation by subtracting componentwise (7) from (6), thus obtaining
log

P (ci | d j )
1 − P (ci | d j )

= log

P (ci )
+
1 − P (ci )
|T |

+

log
k=1

Note that

|T |

wk j log
k=1

pki (1 − pki )
¯
pki (1 − pki )
¯

1 − pki
.
1 − pki
¯

P (ci | d j )
1−P (ci | d j )

(8)

is an increasing mono-

tonic function of P (ci | d j ), and may thus
be used directly as CSVi (d j ). Note also
|T |
1− pki
P (ci )
that log 1−P (ci ) and
k=1 log 1− pki are
¯
constant for all documents, and may
thus be disregarded.13 Defining a classifier for category ci thus basically requires estimating the 2|T | parameters
{ p1i , p1i , . . . , p|T |i , p|T |i } from the training
¯
¯
data, which may be done in the obvious
way. Note that in general the classification of a given document does not require one to compute a sum of |T | factors,
|T |
pki (1− pki )
¯
as the presence of
k=1 wk j log pki (1− pki )
¯
would imply; in fact, all the factors for
which wk j = 0 may be disregarded, and
this accounts for the vast majority of them,
since document vectors are usually very
sparse.
The method we have illustrated is just
one of the many variants of the Na¨ve
ı
Bayes approach, the common denominator of which is (4). A recent paper by Lewis
[1998] is an excellent roadmap on the
various directions that research on Na¨ve
ı
Bayes classifiers has taken; among these
are the ones aiming
—to relax the constraint that document
vectors should be binary-valued. This
13

This is not true, however, if the “fixed thresholding” method of Section 6.1 is adopted. In fact, for a
fixed document d j the first and third factor in the formula above are different for different categories, and
may therefore influence the choice of the categories
under which to file d j .


21

looks natural, given that weighted indexing techniques (see Fuhr [1989];
Salton and Buckley [1988]) accounting
for the “importance” of tk for d j play a
key role in IR.
—to introduce document length normalP (ci | d j )
ization. The value of log 1−P (c | d ) tends
i
j
to be more extreme (i.e., very high
or very low) for long documents (i.e.,
documents such that wk j = 1 for many
values of k), irrespectively of their
semantic relatedness to ci , thus calling for length normalization. Taking
length into account is easy in nonprobabilistic approaches to classification (see Section 6.7), but is problematic
in probabilistic ones (see Lewis [1998],
Section 5). One possible answer is to
switch from an interpretation of Na¨ve
ı
Bayes in which documents are events to
one in which terms are events [Baker
and McCallum 1998; McCallum et al.
1998; Chakrabarti et al. 1998a; Guthrie
et al. 1994]. This accounts for document
length naturally but, as noted by Lewis
[1998], has the drawback that different occurrences of the same word within
the same document are viewed as independent, an assumption even more
implausible than (4).
—to relax the independence assumption.
This may be the hardest route to follow,
since this produces classifiers of higher
computational cost and characterized
by harder parameter estimation problems [Koller and Sahami 1997]. Earlier
efforts in this direction within probabilistic text search (e.g., vanRijsbergen
[1977]) have not shown the performance improvements that were hoped
for. Recently, the fact that the binary independence assumption seldom harms
effectiveness has also been given some
theoretical justification [Domingos and
Pazzani 1997].
The quotation of text search in the last
paragraph is not casual. Unlike other
types of classifiers, the literature on probabilistic classifiers is inextricably intertwined with that on probabilistic search
systems (see Crestani et al. [1998] for a

22

Sebastiani

Fig. 2. A decision tree equivalent to the DNF rule of Figure 1. Edges are labeled
by terms and leaves are labeled by categories (underlining denotes negation).

review), since these latter attempt to determine the probability that a document
falls in the category denoted by the query,
and since they are the only search systems
that take relevance feedback, a notion essentially involving supervised learning, as
central.
6.3. Decision Tree Classifiers

Probabilistic methods are quantitative
(i.e., numeric) in nature, and as such
have sometimes been criticized since, effective as they may be, they are not easily interpretable by humans. A class of
algorithms that do not suffer from this
problem are symbolic (i.e., nonnumeric)
algorithms, among which inductive rule
learners (which we will discuss in Section 6.4) and decision tree learners are the
most important examples.
A decision tree (DT) text classifier (see
Mitchell [1996], Chapter 3) is a tree in
which internal nodes are labeled by terms,
branches departing from them are labeled
by tests on the weight that the term has in
the test document, and leafs are labeled by
categories. Such a classifier categorizes a
test document d j by recursively testing for

the weights that the terms labeling the internal nodes have in vector d j , until a leaf
node is reached; the label of this node is
then assigned to d j . Most such classifiers
use binary document representations, and
thus consist of binary trees. An example
DT is illustrated in Figure 2.
There are a number of standard packages for DT learning, and most DT approaches to TC have made use of one such
package. Among the most popular ones are
ID3 (used by Fuhr et al. [1991]), C4.5 (used
by Cohen and Hirsh [1998], Cohen and
Singer [1999], Joachims [1998], and Lewis
and Catlett [1994]), and C5 (used by Li
and Jain [1998]). TC efforts based on experimental DT packages include Dumais
et al. [1998], Lewis and Ringuette [1994],
and Weiss et al. [1999].
A possible method for learning a DT
for category ci consists in a “divide and
conquer” strategy of (i) checking whether
all the training examples have the same
¯
label (either ci or ci ); (ii) if not, selecting a term tk , partitioning Tr into classes
of documents that have the same value
for tk , and placing each such class in a
separate subtree. The process is recursively repeated on the subtrees until each

leaf of the tree so generated contains training examples assigned to the same category ci , which is then chosen as the label
for the leaf. The key step is the choice of
the term tk on which to operate the partition, a choice which is generally made according to an information gain or entropy
criterion. However, such a “fully grown”
tree may be prone to overfitting, as some
branches may be too specific to the training data. Most DT learning methods thus
include a method for growing the tree and
one for pruning it, that is, for removing
the overly specific branches. Variations on
this basic schema for DT learning abound
[Mitchell 1996, Section 3].
DT text classifiers have been used either
as the main classification tool [Fuhr et al.
1991; Lewis and Catlett 1994; Lewis and
Ringuette 1994], or as baseline classifiers
[Cohen and Singer 1999; Joachims 1998],
or as members of classifier committees [Li
and Jain 1998; Schapire and Singer 2000;
Weiss et al. 1999].
6.4. Decision Rule Classifiers

A classifier for category ci built by an
inductive rule learning method consists
of a DNF rule, that is, of a conditional
rule with a premise in disjunctive normal
form (DNF), of the type illustrated in
Figure 1.14 The literals (i.e., possibly
negated keywords) in the premise denote
the presence (nonnegated keyword) or absence (negated keyword) of the keyword
in the test document d j , while the clause
head denotes the decision to classify d j
under ci . DNF rules are similar to DTs
in that they can encode any Boolean function. However, an advantage of DNF rule
learners is that they tend to generate more
compact classifiers than DT learners.
Rule learning methods usually attempt
to select from all the possible covering
rules (i.e., rules that correctly classify
all the training examples) the “best” one

23

according to some minimality criterion.
While DTs are typically built by a topdown, “divide-and-conquer” strategy, DNF
rules are often built in a bottom-up fashion. Initially, every training example d j is
viewed as a clause η1 , . . . , ηn → γi , where
η1 , . . . , ηn are the terms contained in d j
and γi equals ci or ci according to whether
¯
d j is a positive or negative example of ci .
This set of clauses is already a DNF classifier for ci , but obviously scores high in
terms of overfitting. The learner applies
then a process of generalization in which
the rule is simplified through a series
of modifications (e.g., removing premises
from clauses, or merging clauses) that
maximize its compactness while at the
same time not affecting the “covering”
property of the classifier. At the end of
this process, a “pruning” phase similar in
spirit to that employed in DTs is applied,
where the ability to correctly classify all
the training examples is traded for more
generality.
DNF rule learners vary widely in terms
of the methods, heuristics and criteria
employed for generalization and pruning. Among the DNF rule learners that
have been applied to TC are CHARADE
[Moulinier and Ganascia 1996], DL-ESC
[Li and Yamanishi 1999], RIPPER [Cohen
1995a; Cohen and Hirsh 1998; Cohen and
Singer 1999], SCAR [Moulinier et al. 1996],
e
and SWAP-1 [Apt´ 1994].
While the methods above use rules
of propositional logic (PL), research has
also been carried out using rules of firstorder logic (FOL), obtainable through
the use of inductive logic programming
methods. Cohen [1995a] has extensively
compared PL and FOL learning in TC
(for instance, comparing the PL learner
RIPPER with its FOL version FLIPPER), and
has found that the additional representational power of FOL brings about only
modest benefits.
6.5. Regression Methods

14

Many inductive rule learning algorithms build
decision lists (i.e., arbitrarily nested if-then-else
clauses) instead of DNF rules; since the former may
always be rewritten as the latter, we will disregard
the issue.


Various TC efforts have used regression
models (see Fuhr and Pfeifer [1994]; Ittner
et al. [1995]; Lewis and Gale [1994];
¨
Schutze et al. [1995]). Regression denotes

24
the approximation of a real-valued (instead than binary, as in the case of classification) function ˘ by means of a function that fits the training data [Mitchell
1996, page 236]. Here we will describe one
such model, the Linear Least-Squares Fit
(LLSF) applied to TC by Yang and Chute
[1994]. In LLSF, each document d j has
two vectors associated to it: an input vector I (d j ) of |T | weighted terms, and an
output vector O(d j ) of |C| weights representing the categories (the weights for
this latter vector are binary for training
documents, and are nonbinary CSV s for
test documents). Classification may thus
be seen as the task of determining an output vector O(d j ) for test document d j ,
given its input vector I (d j ); hence, building a classifier boils down to computing
ˆ
ˆ
a |C| × |T | matrix M such that MI(d j ) =
O(d j ).
LLSF computes the matrix from the
training data by computing a linear leastsquares fit that minimizes the error on the
ˆ
training set according to the formula M =
arg min M MI − O F , where arg min M (x)
stands as usual for the M for which x is
def
|C|
|T |
2
minimum, V F =
i=1
j =1 vi j represents the so-called Frobenius norm of a
|C| × |T | matrix, I is the |T | × |Tr| matrix
whose columns are the input vectors of the
training documents, and O is the |C| × |Tr|
matrix whose columns are the output vecˆ
tors of the training documents. The M matrix is usually computed by performing a
singular value decomposition on the training set, and its generic entry mik repreˆ
sents the degree of association between
category ci and term tk .
The experiments of Yang and Chute
[1994] and Yang and Liu [1999] indicate
that LLSF is one of the most effective text
classifiers known to date. One of its disadvantages, though, is that the cost of comˆ
puting the M matrix is much higher than
that of many other competitors in the TC
arena.
6.6. On-Line Methods

A linear classifier for category ci is a vector ci = w1i , . . . , w|T |i belonging to the

Sebastiani
same |T |-dimensional space in which documents are also represented, and such
that CSVi (d j ) corresponds to the dot
|T |
product
k=1 wki wk j of d j and ci . Note
that when both classifier and document
weights are cosine-normalized (see (2)),
the dot product between the two vectors corresponds to their cosine similarity,
that is:
S(ci , d j ) = cos(α)
=

|T |
k=1
|T |
k=1

wki · wk j

2
wki ·

|T |
k=1

,
2
wk j

which represents the cosine of the angle
α that separates the two vectors. This is
the similarity measure between query and
document computed by standard vectorspace IR engines, which means in turn
that once a linear classifier has been built,
classification can be performed by invoking such an engine. Practically all search
engines have a dot product flavor to them,
and can therefore be adapted to doing TC
with a linear classifier.
Methods for learning linear classifiers
are often partitioned in two broad classes,
batch methods and on-line methods.
Batch methods build a classifier by analyzing the training set all at once. Within
the TC literature, one example of a batch
method is linear discriminant analysis,
a model of the stochastic dependence between terms that relies on the covariance matrices of the categories [Hull 1994;
¨
Schutze et al. 1995]. However, the foremost example of a batch method is the
Rocchio method; because of its importance
in the TC literature, this will be discussed
separately in Section 6.7. In this section
we will instead concentrate on on-line
methods.
On-line (a.k.a. incremental) methods
build a classifier soon after examining
the first training document, and incrementally refine it as they examine new
ones. This may be an advantage in the
applications in which Tr is not available in its entirety from the start, or in
which the “meaning” of the category may
change in time, as for example, in adaptive

filtering. This is also apt to applications
(e.g., semiautomated classification, adaptive filtering) in which we may expect the
user of a classifier to provide feedback on
how test documents have been classified,
as in this case further training may be performed during the operating phase by exploiting user feedback.
A simple on-line method is the perceptron algorithm, first applied to TC by
¨
Schutze et al. [1995] and Wiener et al.
[1995], and subsequently used by Dagan
et al. [1997] and Ng et al. [1997]. In this algorithm, the classifier for ci is first initialized by setting all weights wki to the same
positive value. When a training example
d j (represented by a vector d j of binary
weights) is examined, the classifier built
so far classifies it. If the result of the classification is correct, nothing is done, while
if it is wrong, the weights of the classifier
are modified: if d j was a positive example of ci , then the weights wki of “active
terms” (i.e., the terms tk such that wkj = 1)
are “promoted” by increasing them by a
fixed quantity α > 0 (called learning rate),
while if d j was a negative example of ci
then the same weights are “demoted” by
decreasing them by α. Note that when the
classifier has reached a reasonable level of
effectiveness, the fact that a weight wki is
very low means that tk has negatively contributed to the classification process so far,
and may thus be discarded from the representation. We may then see the perceptron
algorithm (as all other incremental learning methods) as allowing for a sort of “onthe-fly term space reduction” [Dagan et al.
1997, Section 4.4]. The perceptron classifier has shown a good effectiveness in all
the experiments quoted above.
The perceptron is an additive weightupdating algorithm. A multiplicative
variant of it is POSITIVE WINNOW [Dagan
et al. 1997], which differs from perceptron
because two different constants α1 > 1 and
0 < α2 < 1 are used for promoting and demoting weights, respectively, and because
promotion and demotion are achieved by
multiplying, instead of adding, by α1 and
α2 . BALANCED WINNOW [Dagan et al. 1997]
is a further variant of POSITIVE WINNOW, in
which the classifier consists of two weights

25

+
−
wki and wki for each term tk ; the final
weight wki used in computing the dot prod+
−
uct is the difference wki − wki . Following
the misclassification of a positive in+
stance, active terms have their wki weight
−
promoted and their wki weight demoted,
whereas in the case of a negative instance
+
−
it is wki that gets demoted while wki gets
promoted (for the rest, promotions and
demotions are as in POSITIVE WINNOW).
BALANCED WINNOW allows negative wki
weights, while in the perceptron and in
POSITIVE WINNOW the wki weights are always positive. In experiments conducted
by Dagan et al. [1997], POSITIVE WINNOW
showed a better effectiveness than perceptron but was in turn outperformed by
(Dagan et al.’s own version of) BALANCED
WINNOW.
Other on-line methods for building text
classifiers are WIDROW-HOFF, a refinement
of it called EXPONENTIATED GRADIENT (both
applied for the first time to TC in [Lewis
et al. 1996]) and SLEEPING EXPERTS [Cohen
and Singer 1999], a version of BALANCED
WINNOW. While the first is an additive
weight-updating algorithm, the second
and third are multiplicative. Key differences with the previously described algorithms are that these three algorithms
(i) update the classifier not only after misclassifying a training example, but also after classifying it correctly, and (ii) update
the weights corresponding to all terms (instead of just active ones).
Linear classifiers lend themselves to
both category-pivoted and documentpivoted TC. For the former the classifier
ci is used, in a standard search engine,
as a query against the set of test documents, while for the latter the vector d j
representing the test document is used
as a query against the set of classifiers
{c1 , . . . , c|C| }.

6.7. The Rocchio Method

Some linear classifiers consist of an explicit profile (or prototypical document)
of the category. This has obvious advantages in terms of interpretability, as such
a profile is more readily understandable
by a human than, say, a neural network

26

Sebastiani

classifier. Learning a linear classifier is often preceded by local TSR; in this case, a
profile of ci is a weighted list of the terms
whose presence or absence is most useful
for discriminating ci .
The Rocchio method is used for inducing linear, profile-style classifiers. It relies on an adaptation to TC of the wellknown Rocchio’s formula for relevance
feedback in the vector-space model, and
it is perhaps the only TC method rooted
in the IR tradition rather than in the
ML one. This adaptation was first proposed by Hull [1994], and has been used
by many authors since then, either as
an object of research in its own right
[Ittner et al. 1995; Joachims 1997; Sable
and Hatzivassiloglou 2000; Schapire et al.
1998; Singhal et al. 1997], or as a baseline classifier [Cohen and Singer 1999;
Galavotti et al. 2000; Joachims 1998;
Lewis et al. 1996; Schapire and Singer
¨
2000; Schutze et al. 1995], or as a member of a classifier committee [Larkey and
Croft 1996] (see Section 6.11).
Rocchio’s method computes a classifier ci = w1i , . . . , w|T |i for category ci by
means of the formula
wki = β ·
{d j ∈POSi }

γ ·
{d j ∈NEGi }

wkj
−
|POSi |
wkj
,
|NEGi |

where wkj is the weight of tk in document
d j , POSi = {d j ∈ Tr | ˘ (d j , ci ) = T }, and
NEGi = {d j ∈ Tr | ˘ (d j , ci ) = F }. In this
formula, β and γ are control parameters
that allow setting the relative importance
of positive and negative examples. For
instance, if β is set to 1 and γ to 0 (as
in Dumais et al. [1998]; Hull [1994];
¨
Joachims [1998]; Schutze et al. [1995]),
the profile of ci is the centroid of its positive training examples. A classifier built
by means of the Rocchio method rewards
the closeness of a test document to the
centroid of the positive training examples,
and its distance from the centroid of the
negative training examples. The role of
negative examples is usually deempha-

sized, by setting β to a high value and γ to
a low one (e.g., Cohen and Singer [1999],
Ittner et al. [1995], and Joachims [1997]
use β = 16 and γ = 4).
This method is quite easy to implement,
and is also quite efficient, since learning
a classifier basically comes down to averaging weights. In terms of effectiveness,
instead, a drawback is that if the documents in the category tend to occur in
disjoint clusters (e.g., a set of newspaper
articles lebeled with the Sports category
and dealing with either boxing or rockclimbing), such a classifier may miss most
of them, as the centroid of these documents may fall outside all of these clusters
(see Figure 3(a)). More generally, a classifier built by the Rocchio method, as all linear classifiers, has the disadvantage that
it divides the space of documents linearly.
This situation is graphically depicted in
Figure 3(a), where documents are classified within ci if and only if they fall within
the circle. Note that even most of the positive training examples would not be classified correctly by the classifier.
6.7.1. Enhancements to the Basic Rocchio
Framework. One issue in the application of

the Rocchio formula to profile extraction
is whether the set NEGi should be considered in its entirety, or whether a wellchosen sample of it, such as the set NPOSi
of near-positives (defined as “the most positive among the negative training examples”), should be selected from it, yielding
wki = β ·
{d j ∈POSi }

wkj
−
|POSi |

γ ·
{d j ∈NPOSi }

wkj
.
|NPOSi |

w

kj
The {d j ∈NPOSi } |NPOSi | factor is more sigwkj
nificant than {d j ∈NEGi } |NEGi | , since nearpositives are the most difficult documents
to tell apart from the positives. Using
near-positives corresponds to the query
zoning method proposed for IR by Singhal
et al. [1997]. This method originates from
the observation that, when the original



27

Fig. 3. A comparison between the TC behavior of (a) the Rocchio classifier, and
(b) the k-NN classifier. Small crosses and circles denote positive and negative
training instances, respectively. The big circles denote the “influence area” of
the classifier. Note that, for ease of illustration, document similarities are here
viewed in terms of Euclidean distance rather than, as is more common, in terms
of dot product or cosine.

Rocchio formula is used for relevance
feedback in IR, near-positives tend to
be used rather than generic negatives, as
the documents on which user judgments
are available are among the ones that
had scored highest in the previous ranking. Early applications of the Rocchio formula to TC (e.g., Hull [1994]; Ittner et al.
[1995]) generally did not make a distinction between near-positives and generic
negatives. In order to select the nearpositives Schapire et al. [1998] issue a
query, consisting of the centroid of the positive training examples, against a document base consisting of the negative training examples; the top-ranked ones are the
most similar to this centroid, and are then
the near-positives. Wiener et al. [1995] instead equate the near-positives of ci to
the positive examples of the sibling categories of ci , as in the application they work
on (TC with hierarchically organized category sets) the notion of a “sibling category of ci ” is well defined. A similar policy
is also adopted by Ng et al. [1997], Ruiz
and Srinivasan [1999], and Weigend et al.
[1999].
By using query zoning plus other enhancements (TSR, statistical phrases, and
a method called dynamic feedback optimization), Schapire et al. [1998] have
found that a Rocchio classifier can achieve

an effectiveness comparable to that of
a state-of-the-art ML method such as
“boosting” (see Section 6.11.1) while being
60 times quicker to train. These recent
results will no doubt bring about a renewed interest for the Rocchio classifier,
previously considered an underperformer
[Cohen and Singer 1999; Joachims 1998;
¨
Lewis et al. 1996; Schutze et al. 1995; Yang
1999].
6.8. Neural Networks

A neural network (NN) text classifier is a
network of units, where the input units
represent terms, the output unit(s) represent the category or categories of interest,
and the weights on the edges connecting
units represent dependence relations. For
classifying a test document d j , its term
weights wkj are loaded into the input units;
the activation of these units is propagated forward through the network, and
the value of the output unit(s) determines
the categorization decision(s). A typical
way of training NNs is backpropagation,
whereby the term weights of a training
document are loaded into the input units,
and if a misclassification occurs the error
is “backpropagated” so as to change the parameters of the network and eliminate or
minimize the error.

28
The simplest type of NN classifier is
the perceptron [Dagan et al. 1997; Ng
et al. 1997], which is a linear classifier and
as such has been extensively discussed
in Section 6.6. Other types of linear NN
classifiers implementing a form of logistic regression have also been proposed
¨
and tested by Schutze et al. [1995] and
Wiener et al. [1995], yielding very good
effectiveness.
A nonlinear NN [Lam and Lee 1999;
¨
Ruiz and Srinivasan 1999; Schutze et al.
1995; Weigend et al. 1999; Wiener et al.
1995; Yang and Liu 1999] is instead a network with one or more additional “layers”
of units, which in TC usually represent
higher-order interactions between terms
that the network is able to learn. When
comparative experiments relating nonlinear NNs to their linear counterparts have
been performed, the former have yielded
¨
either no improvement [Schutze et al.
1995] or very small improvements [Wiener
et al. 1995] over the latter.
6.9. Example-Based Classifiers

Example-based classifiers do not build an
explicit, declarative representation of the
category ci , but rely on the category labels attached to the training documents
similar to the test document. These methods have thus been called lazy learners,
since “they defer the decision on how to
generalize beyond the training data until
each new query instance is encountered”
[Mitchell 1996, page 244].
The first application of example-based
methods (a.k.a. memory-based reasoning methods) to TC is due to Creecy,
Masand and colleagues [Creecy et al.
1992; Masand et al. 1992]; other examples
include Joachims [1998], Lam et al. [1999],
Larkey [1998], Larkey [1999], Li and Jain
[1998], Yang and Pedersen [1997], and
Yang and Liu [1999]. Our presentation of
the example-based approach will be based
on the k-NN (for “k nearest neighbors”)
algorithm used by Yang [1994]. For deciding whether d j ∈ ci , k-NN looks at whether
the k training documents most similar to
d j also are in ci ; if the answer is positive for a large enough proportion of them,

Sebastiani
a positive decision is taken, and a negative decision is taken otherwise. Actually,
Yang’s is a distance-weighted version of
k-NN (see [Mitchell 1996, Section 8.2.1]),
since the fact that a most similar document is in ci is weighted by its similarity with the test document. Classifying d j
by means of k-NN thus comes down to
computing
CSVi (d j )
=

RSV(d j , d z ) · [[ ˘ (d z , ci )]],
d z ∈ Trk (d j )

(9)
where Trk (d j ) is the set of the k documents
d z which maximize RSV(d j , d z ) and
[[α]] =

1 if α = T
.
0 if α = F

The thresholding methods of Section 6.1
can then be used to convert the realvalued CSVi ’s into binary categorization
decisions. In (9), RSV(d j , d z ) represents
some measure or semantic relatedness between a test document d j and a training
document d z ; any matching function, be it
probabilistic (as used by Larkey and Croft
[1996]) or vector-based (as used by Yang
[1994]), from a ranked IR system may be
used for this purpose. The construction of
a k-NN classifier also involves determining (experimentally, on a validation set) a
threshold k that indicates how many topranked training documents have to be considered for computing CSVi (d j ). Larkey
and Croft [1996] used k = 20, while Yang
[1994, 1999] has found 30 ≤ k ≤ 45 to yield
the best effectiveness. Anyhow, various experiments have shown that increasing the
value of k does not significantly degrade
the performance.
Note that k-NN, unlike linear classifiers, does not divide the document space
linearly, and hence does not suffer from
the problem discussed at the end of
Section 6.7. This is graphically depicted
in Figure 3(b), where the more “local”
character of k-NN with respect to Rocchio
can be appreciated.

This method is naturally geared toward
document-pivoted TC, since ranking the
training documents for their similarity
with the test document can be done once
for all categories. For category-pivoted TC,
one would need to store the document
ranks for each test document, which is obviously clumsy; DPC is thus de facto the
only reasonable way to use k-NN.
A number of different experiments (see
Section 7.3) have shown k-NN to be quite
effective. However, its most important
drawback is its inefficiency at classification time: while, for example, with a linear classifier only a dot product needs to
be computed to classify a test document,
k-NN requires the entire training set to
be ranked for similarity with the test document, which is much more expensive. This
is a drawback of “lazy” learning methods,
since they do not have a true training
phase and thus defer all the computation
to classification time.
6.9.1. Other

Example-Based

—clustering the training set, thus obtaining a set of clusters K i = {ki1 , . . . ,
ki|K i | };
—building a profile G(kiz ) (“generalized
instance”) from the documents belonging to cluster kiz by means of some algorithm for learning linear classifiers (e.g.,
Rocchio, WIDROW-HOFF);
—applying k-NN with profiles in place of
training documents, that is, computing
def

CSVi (d j ) =

RSV(d j , G(kiz )) ·
kiz ∈K i

|{d j ∈ kiz | ˘ (d j , ci ) = T }|
·
|{d j ∈ kiz }|
|{d j ∈ kiz }|
|Tr|
RSV(d j , G(kiz )) ·

=

CSVi (d j )

kiz ∈K i

(1 − RSV(d j , d z ))[[

˘ (d z ,ci )]]

d z ∈Trk (d j )

as an alternative to (9), obtaining a small
but statistically significant improvement
over a version of WHIRL using (9). In
their experiments this technique outperformed a number of other classifiers, such
as a C4.5 decision tree classifier and the
RIPPER CNF rule-based classifier.
A variant of the basic k-NN approach was proposed by Galavotti et al.
[2000], who reinterpreted (9) by redefining
[[α]] as
[[α]] =

The difference from the original k-NN approach is that if a training document d z
similar to the test document d j does not
belong to ci , this information is not discarded but weights negatively in the decision to classify d j under ci .
A combination of profile- and examplebased methods was presented in Lam and
Ho [1998]. In this work a k-NN system was
fed generalized instances (GIs) in place of
training documents. This approach may be
seen as the result of

Techniques.

Various example-based techniques have
been used in the TC literature. For example, Cohen and Hirsh [1998] implemented
an example-based classifier by extending
standard relational DBMS technology
with “similarity-based soft joins.” In
their WHIRL system they used the scoring
function

=1−

29

1 if α = T
.
−1 if α = F


|{d j ∈ kiz | ˘ (d j , ci ) = T }|
,
|Tr|

(10)

|{d ∈k | ˘ (d ,c )=T }|

j
iz
j i
represents the
where
|{d j ∈kiz }|
“degree” to which G(kiz ) is a positive in|{d j ∈k }|
stance of ci , and |T r|iz represents its
weight within the entire process.

This exploits the superior effectiveness
(see Figure 3) of k-NN over linear classifiers while at the same time avoiding
the sensitivity of k-NN to the presence of
“outliers” (i.e., positive instances of ci that
“lie out” of the region where most other
positive instances of ci are located) in the
training set.

30

Sebastiani

Fig. 4. Learning support vector classifiers.
The small crosses and circles represent positive and negative training examples, respectively, whereas lines represent decision surfaces. Decision surface σi (indicated by the
thicker line) is, among those shown, the best
possible one, as it is the middle element of
the widest set of parallel decision surfaces
(i.e., its minimum distance to any training
example is maximum). Small boxes indicate
the support vectors.

6.10. Building Classifiers by Support
Vector Machines

The support vector machine (SVM) method
has been introduced in TC by Joachims
[1998, 1999] and subsequently used by
Drucker et al. [1999], Dumais et al. [1998],
Dumais and Chen [2000], Klinkenberg
and Joachims [2000], Taira and Haruno
[1999], and Yang and Liu [1999]. In geometrical terms, it may be seen as the
attempt to find, among all the surfaces
σ1 , σ2 , . . . in |T |-dimensional space that
separate the positive from the negative
training examples (decision surfaces), the
σi that separates the positives from the
negatives by the widest possible margin,
that is, such that the separation property
is invariant with respect to the widest possible traslation of σi .
This idea is best understood in the case
in which the positives and the negatives
are linearly separable, in which case the
decision surfaces are (|T |−1)-hyperplanes.
In the two-dimensional case of Figure 4,
various lines may be chosen as decision
surfaces. The SVM method chooses the

middle element from the “widest” set of
parallel lines, that is, from the set in which
the maximum distance between two elements in the set is highest. It is noteworthy that this “best” decision surface is determined by only a small set of training
examples, called the support vectors.
The method described is applicable also
to the case in which the positives and the
negatives are not linearly separable. Yang
and Liu [1999] experimentally compared
the linear case (namely, when the assumption is made that the categories are linearly separable) with the nonlinear case
on a standard benchmark, and obtained
slightly better results in the former case.
As argued by Joachims [1998], SVMs
offer two important advantages for TC:
—term selection is often not needed, as
SVMs tend to be fairly robust to overfitting and can scale up to considerable
dimensionalities;
—no human and machine effort in parameter tuning on a validation set is needed,
as there is a theoretically motivated,
“default” choice of parameter settings,
which has also been shown to provide
the best effectiveness.
Dumais et al. [1998] have applied a
novel algorithm for training SVMs which
brings about training speeds comparable
to computationally easy learners such as
Rocchio.

6.11. Classifier Committees

Classifier committees (a.k.a. ensembles)
are based on the idea that, given a task
that requires expert knowledge to perform, k experts may be better than one if
their individual judgments are appropriately combined. In TC, the idea is to apply k different classifiers 1 , . . . , k to the
same task of deciding whether d j ∈ ci , and
then combine their outcome appropriately.
A classifier committee is then characterized by (i) a choice of k classifiers, and (ii)
a choice of a combination function.
Concerning Issue (i), it is known from
the ML literature that, in order to guarantee good effectiveness, the classifiers

forming the committee should be as independent as possible [Tumer and Ghosh
1996]. The classifiers may differ for the indexing approach used, or for the inductive
method, or both. Within TC, the avenue
which has been explored most is the latter
(to our knowledge the only example of the
former is Scott and Matwin [1999]).
Concerning Issue (ii), various rules have
been tested. The simplest one is majority
voting (MV), whereby the binary outputs
of the k classifiers are pooled together, and
the classification decision that reaches the
majority of k+1 votes is taken (k obviously
2
needs to be an odd number) [Li and Jain
1998; Liere and Tadepalli 1997]. This
method is particularly suited to the case
in which the committee includes classifiers characterized by a binary decision
function CSVi : D → {T, F }. A second rule
is weighted linear combination (WLC),
whereby a weighted sum of the CSVi ’s produced by the k classifiers yields the final
CSVi . The weights w j reflect the expected
relative effectiveness of classifiers j , and
are usually optimized on a validation set
[Larkey and Croft 1996]. Another policy
is dynamic classifier selection (DCS),
whereby among committee { 1 , . . . , k }
the classifier t most effective on the l
validation examples most similar to d j
is selected, and its judgment adopted by
the committee [Li and Jain 1998]. A still
different policy, somehow intermediate
between WLC and DCS, is adaptive
classifier combination (ACC), whereby the
judgments of all the classifiers in the committee are summed together, but their individual contribution is weighted by their
effectiveness on the l validation examples
most similar to d j [Li and Jain 1998].
Classifier committees have had mixed
results in TC so far. Larkey and Croft
[1996] have used combinations of Rocchio,
Na¨ve Bayes, and k-NN, all together or in
ı
pairwise combinations, using a WLC rule.
In their experiments the combination of
any two classifiers outperformed the best
individual classifier (k-NN), and the combination of the three classifiers improved
an all three pairwise combinations. These
results would seem to give strong support to the idea that classifier committees

31

can somehow profit from the complementary strengths of their individual members. However, the small size of the test set
used (187 documents) suggests that more
experimentation is needed before conclusions can be drawn.
Li and Jain [1998] have tested a committee formed of (various combinations of) a
Na¨ve Bayes classifier, an example-based
ı
classifier, a decision tree classifier, and a
classifier built by means of their own “subspace method”; the combination rules they
have worked with are MV, DCS, and ACC.
Only in the case of a committee formed
by Na¨ve Bayes and the subspace classiı
fier combined by means of ACC has the
committee outperformed, and by a narrow margin, the best individual classifier
(for every attempted classifier combination ACC gave better results than MV and
DCS). This seems discouraging, especially
in light of the fact that the committee approach is computationally expensive (its
cost trivially amounts to the sum of the
costs of the individual classifiers plus
the cost incurred for the computation of
the combination rule). Again, it has to be
remarked that the small size of their experiment (two test sets of less than 700
documents each were used) does not allow
us to draw definitive conclusions on the
approaches adopted.
6.11.1. Boosting. The boosting method
[Schapire et al. 1998; Schapire and Singer
2000] occupies a special place in the classifier committees literature, since the k classifiers 1 , . . . , k forming the committee
are obtained by the same learning method
(here called the weak learner). The key
intuition of boosting is that the k classifiers should be trained not in a conceptually parallel and independent way,
as in the committees described above,
but sequentially. In this way, in training classifier i one may take into account how classifiers 1 , . . . , i−1 perform
on the training examples, and concentrate
on getting right those examples on which
1 , . . . , i−1 have performed worst.
Specifically, for learning classifier t
each d j , ci pair is given an “importance
t
1
weight” hi j (where hi j is set to be equal for

32
all d j , ci pairs15 ), which represents how
hard to get a correct decision for this
pair was for classifiers 1 , . . . , t−1 . These
weights are exploited in learning
t,
which will be specially tuned to correctly
solve the pairs with higher weight. Classifier t is then applied to the training
t
documents, and as a result weights hi j
t+1
are updated to hi j ; in this update operation, pairs correctly classified by t will
have their weight decreased, while pairs
misclassified by t will have their weight
increased. After all the k classifiers have
been built, a weighted linear combination
rule is applied to yield the final committee.
In the BOOSTEXTER system [Schapire and
Singer 2000], two different boosting algorithms are tested, using a one-level
decision tree weak learner. The former
algorithm (ADABOOST.MH, simply called
ADABOOST in Schapire et al. [1998]) is explicitly geared toward the maximization of
microaveraged effectiveness, whereas the
latter (ADABOOST.MR) is aimed at minimizing ranking loss (i.e., at getting a correct category ranking for each individual
document). In experiments conducted over
three different test collections, Schapire
et al. [1998] have shown ADABOOST to
outperform SLEEPING EXPERTS, a classifier
that had proven quite effective in the experiments of Cohen and Singer [1999].
Further experiments by Schapire and
Singer [2000] showed ADABOOST to outperform, aside from SLEEPING EXPERTS, a
Na¨ve Bayes classifier, a standard (nonenı
hanced) Rocchio classifier, and Joachims’
[1997] PRTFIDF classifier.
A boosting algorithm based on a “committee of classifier subcommittees” that
improves on the effectiveness and (especially) the efficiency of ADABOOST.MH was
presented in Sebastiani et al. [2000]. An
approach similar to boosting was also employed by Weiss et al. [1999], who experimented with committees of decision trees
each having an average of 16 leaves (and
hence much more complex than the sim-

15

Schapire et al. [1998] also showed that a simple
modification of this policy allows optimization of the
classifier based on “utility” (see Section 7.1.3).

Sebastiani
ple “decision stumps” used in Schapire
and Singer [2000]), eventually combined
by using the simple MV rule as a combination rule; similarly to boosting, a mechanism for emphasising documents that
have been misclassified by previous decision trees is used. Boosting-based approaches have also been employed in
Escudero et al. [2000], Iyer et al. [2000],
Kim et al. [2000], Li and Jain [1998], and
Myers et al. [2000].
6.12. Other Methods

Although in the previous sections we
have tried to give an overview as complete as possible of the learning approaches proposed in the TC literature, it
is hardly possible to be exhaustive. Some
of the learning approaches adopted do
not fall squarely under one or the other
class of algorithms, or have remained
somehow isolated attempts. Among these,
the most noteworthy are the ones based
on Bayesian inference networks [Dumais
et al. 1998; Lam et al. 1997; Tzeras
and Hartmann 1993], genetic algorithms
[Clack et al. 1997; Masand 1994], and
maximum entropy modelling [Manning
¨
and Schutze 1999].

7. EVALUATION OF TEXT CLASSIFIERS

As for text search systems, the evaluation of document classifiers is typically conducted experimentally, rather
than analytically. The reason is that, in
order to evaluate a system analytically
(e.g., proving that the system is correct
and complete), we would need a formal
specification of the problem that the system is trying to solve (e.g., with respect
to what correctness and completeness are
defined), and the central notion of TC
(namely, that of membership of a document in a category) is, due to its subjective
character, inherently nonformalizable.
The experimental evaluation of a classifier usually measures its effectiveness
(rather than its efficiency), that is, its
ability to take the right classification
decisions.

Table II. The Contingency Table for Category c i
Category
Expert judgments
ci
YES
NO
Classifier
TPi
FPi
YES
Judgments
FNi
TNi
NO

7.1. Measures of Text
Categorization Effectiveness

33

Table III. The Global Contingency Table
Category set
Expert judgments
C = {c1 , . . . , c|C| }
YES
NO
|C|

Classifier

YES

TP =

|C|

TPi

FP =

i=1
|C|

Judgments NO

FN =

FNi TN =
i=1

7.1.1. Precision and Recall. Classification
effectiveness is usually measured in terms
of the classic IR notions of precision (π )
and recall (ρ), adapted to the case of
TC. Precision wrt ci (πi ) is defined as
the conditional probability P ( ˘ (d x , ci ) =
T | (d x , ci ) = T ), that is, as the probability that if a random document d x is
classified under ci , this decision is correct.
Analogously, recall wrt ci (ρi ) is defined
as P ( (d x , ci ) = T | ˘ (d x , ci ) = T ), that
is, as the probability that, if a random
document d x ought to be classified under
ci , this decision is taken. These categoryrelative values may be averaged, in a way
to be discussed shortly, to obtain π and ρ,
that is, values global to the entire category
set. Borrowing terminology from logic, π
may be viewed as the “degree of soundness” of the classifier wrt C, while ρ may
be viewed as its “degree of completeness”
wrt C. As defined here, πi and ρi are to
be understood as subjective probabilities,
that is, as measuring the expectation of
the user that the system will behave correctly when classifying an unseen document under ci . These probabilities may
be estimated in terms of the contingency
table for ci on a given test set (see Table II).
Here, FPi (false positives wrt ci , a.k.a.
errors of commission) is the number of
test documents incorrectly classified under ci ; TNi (true negatives wrt ci ), TPi (true
positives wrt ci ), and FNi (false negatives
wrt ci , a.k.a. errors of omission) are defined accordingly. Estimates (indicated by
carets) of precision wrt ci and recall wrt ci
may thus be obtained as

πi =
ˆ

TPi
,
TPi + FPi

ρi =
ˆ

TPi
.
TPi + FNi

For obtaining estimates of π and ρ, two
different methods may be adopted:

FPi
i=1
|C|

TNi
i=1

—microaveraging: π and ρ are obtained by
summing over all individual decisions:
πµ =
ˆ
ρµ =
ˆ

|C|
i=1

TPi
,
|C|
i=1 (TPi + FPi )
|C|
i=1 TPi
,
|C|
i=1 (TPi + FNi )

TP
=
TP + FP
TP
=
TP + FN

where “µ” indicates microaveraging. The “global” contingency table
(Table III) is thus obtained by summing over category-specific contingency tables;
—macroaveraging: precision and recall
are first evaluated “locally” for each
category, and then “globally” by averaging over the results of the different
categories:
πM =
ˆ

|C|
i=1

πi
ˆ

,

ρM =
ˆ

|C|
i=1

ρi
ˆ

,
|C|
|C|
where “M ” indicates macroaveraging.
These two methods may give quite different results, especially if the different
categories have very different generality.
For instance, the ability of a classifier to
behave well also on categories with low
generality (i.e., categories with few positive training instances) will be emphasized by macroaveraging and much less
so by microaveraging. Whether one or the
other should be used obviously depends on
the application requirements. From now
on, we will assume that microaveraging is
used; everything we will say in the rest of
Section 7 may be adapted to the case of
macroaveraging in the obvious way.
7.1.2. Other

Measures

of

Effectiveness.

Measures alternative to π and ρ and
commonly used in the ML literature, such as accuracy (estimated as

34
TP+TN
ˆ
A = TP+TN+FP+FN ) and error (estimated
FP+FN
ˆ
ˆ
as E = TP+TN+FP+FN = 1 − A), are not
widely used in TC. The reason is that, as
Yang [1999] pointed out, the large value
that their denominator typically has in
TC makes them much more insensitive to
variations in the number of correct decisions (TP + TN) than π and ρ. Besides, if
A is the adopted evaluation measure, in
the frequent case of a very low average
generality the trivial rejector (i.e., the
classifier
such that (d j , ci ) = F for
all d j and ci ) tends to outperform all
nontrivial classifiers (see also Cohen
[1995a], Section 2.3). If A is adopted,
parameter tuning on a validation set may
thus result in parameter choices that
make the classifier behave very much like
the trivial rejector.
A nonstandard effectiveness measure was proposed by Sable and
Hatzivassiloglou [2000, Section 7], who
suggested basing π and ρ not on “absolute” values of success and failure (i.e., 1
if (d j , ci ) = ˘ (d j , ci ) and 0 if (d j , ci ) =
˘ (d j , ci )), but on values of relative success (i.e., CSVi (d j ) if ˘ (d j , ci ) = T and
1 − CSVi (d j ) if ˘ (d j , ci ) = F ). This means
that for a correct (respectively wrong)
decision the classifier is rewarded (respectively penalized) proportionally to its
confidence in the decision. This proposed
measure does not reward the choice of a
good thresholding policy, and is thus unfit
for autonomous (“hard”) classification
systems. However, it might be appropriate for interactive (“ranking”) classifiers
of the type used in Larkey [1999], where
the confidence that the classifier has
in its own decision influences category
ranking and, as a consequence, the overall
usefulness of the system.

7.1.3. Measures Alternative to Effectiveness.

In general, criteria different from effectiveness are seldom used in classifier evaluation. For instance, efficiency, although
very important for applicative purposes,
is seldom used as the sole yardstick, due
to the volatility of the parameters on
which the evaluation rests. However, efficiency may be useful for choosing among

Sebastiani
Table IV. The Utility Matrix
Category set
Expert judgments
C = {c1 , . . . , c|C| }
YES
NO
Classifier
YES
uTP
uFP
Judgments
NO
uFN
uTN

classifiers with similar effectiveness. An
interesting evaluation has been carried
out by Dumais et al. [1998], who have
compared five different learning methods
along three different dimensions, namely,
effectiveness, training efficiency (i.e., the
average time it takes to build a classifier
for category ci from a training set Tr), and
classification efficiency (i.e., the average
time it takes to classify a new document
d j under category ci ).
An important alternative to effectiveness is utility, a class of measures from
decision theory that extend effectiveness
by economic criteria such as gain or loss.
Utility is based on a utility matrix such
as that of Table IV, where the numeric
values uTP , uFP , uFN and uTN represent
the gain brought about by a true positive,
false positive, false negative, and true negative, respectively; both uTP and uTN are
greater than both uFP and uFN . “Standard”
effectiveness is a special case of utility,
i.e., the one in which uTP = uTN > uFP =
uFN . Less trivial cases are those in
which uTP = uTN and/or uFP = uFN ; this
is appropriate, for example, in spam filtering, where failing to discard a piece
of junk mail (FP) is a less serious mistake than discarding a legitimate message (FN) [Androutsopoulos et al. 2000].
If the classifier outputs probability estimates of the membership of d j in ci , then
decision theory provides analytical methods to determine thresholds τi , thus avoiding the need to determine them experimentally (as discussed in Section 6.1).
Specifically, as Lewis [1995a] reminds us,
the expected value of utility is maximized
when
τi =

(uFP − uTN )
,
(uFN − uTP ) + (uFP − uTN )

which, in the case of “standard” effectiveness, is equal to 1 .
2

Table V. Trivial Cases in TC
Precision
Recall
TP
TP
TP + FP
TP + FN

35
C-precision
TN
FP + TN

C-recall
TN
TN + FN

Trivial rejector

TP = FP = 0

Undefined

0
=0
FN

TN
=1
TN

TN
TN + FN

Trivial acceptor

FN = TN = 0

TP
TP + FP

TP
=1
TP

0
=0
FP

Undefined

Trivial “Yes” collection

FP = TN = 0

TP
=1
TP

TP
TP + FN

Undefined

0
=0
FN

Trivial “No” collection

TP = FN = 0

0
=0
FP

Undefined

TN
FP + TN

TN
=1
TN

The use of utility in TC is discussed
in detail by Lewis [1955a]. Other works
where utility is employed are Amati and
Crestani [1999], Cohen and Singer [1999],
Hull et al. [1996], Lewis and Catlett
[1994], and Schapire et al. [1998]. Utility
has become popular within the text filtering community, and the TREC “filtering
track” evaluations have been using it for
a while [Lewis 1995c]. The values of the
utility matrix are extremely applicationdependent. This means that if utility is
used instead of “pure” effectiveness, there
is a further element of difficulty in the
cross-comparison of classification systems
(see Section 7.3), since for two classifiers
to be experimentally comparable also the
two utility matrices must be the same.
Other effectiveness measures different
from the ones discussed here have occasionally been used in the literature; these
include adjacent score [Larkey 1998],
coverage [Schapire and Singer 2000], oneerror [Schapire and Singer 2000], Pearson product-moment correlation [Larkey
1998], recall at n [Larkey and Croft 1996],
top candidate [Larkey and Croft 1996],
and top n [Larkey and Croft 1996]. We
will not attempt to discuss them in detail.
However, their use shows that, although
the TC community is making consistent
efforts at standardizing experimentation
protocols, we are still far from universal
agreement on evaluation issues and, as
a consequence, from understanding precisely the relative merits of the various
methods.

7.1.4. Combined

Effectiveness

Measures.

Neither precision nor recall makes sense
in isolation from each other. In fact the
classifier such that (d j , ci ) = T for all
d j and ci (the trivial acceptor) has ρ = 1.
When the CSVi function has values in
[0, 1], one only needs to set every threshold τi to 0 to obtain the trivial acceptor.
In this case, π would usually be very low
(more precisely, equal to the average test
|C|

g T e (ci )

).16 Conversely, it
set generality i=1|C|
is well known from everyday IR practice
that higher levels of π may be obtained at
the price of low values of ρ.
In practice, by tuning τi a function
CSVi : D → {T, F } is tuned to be, in the
words of Riloff and Lehnert [1994], more
liberal (i.e., improving ρi to the detriment
of πi ) or more conservative (improving πi to

16

From this, one might be tempted to infer, by symmetry, that the trivial rejector always has π = 1.
This is false, as π is undefined (the denominator is
zero) for the trivial rejector (see Table V). In fact,
it is clear from its definition (π = TPTP ) that π
+FP
depends only on how the positives (TP + FP ) are
split between true positives TP and the false positives FP , and does not depend at all on the cardinality of the positives. There is a breakup of “symmetry” between π and ρ here because, from the point of
view of classifier judgment (positives vs. negatives;
this is the dichotomy of interest in trivial acceptor vs.
trivial rejector), the “symmetric” of ρ ( TPTP ) is not
+FN
π ( TPTP ) but C-precision (π c = FPTN ), the “con+FP
+TN
trapositive” of π . In fact, while ρ = 1 and π c = 0 for
the trivial acceptor, π c = 1 and ρ = 0 for the trivial
rejector.

36
the detriment of ρi ).17 A classifier should
thus be evaluated by means of a measure which combines π and ρ.18 Various such measures have been proposed,
among which the most frequent are:
(1) Eleven-point average precision: threshold τi is repeatedly tuned so as to allow
ρi to take up values of 0.0, .1, . . . , .9,
1.0; πi is computed for these 11 different values of τi , and averaged over the
11 resulting values. This is analogous
to the standard evaluation methodology for ranked IR systems, and may be
used
(a) with categories in place of IR
queries. This is most frequently
used for document-ranking clas¨
sifiers (see Schutze et al. [1995];
Yang [1994]; Yang [1999]; Yang and
Pedersen [1997]);
(b) with test documents in place of
IR queries and categories in place
of documents. This is most frequently used for category-ranking
classifiers (see Lam et al. [1999];
Larkey and Croft [1996]; Schapire
and Singer [2000]; Wiener et al.
[1995]). In this case, if macroaveraging is used, it needs to be redefined on a per-document, rather
than per-category, basis.
This measure does not make sense for
binary-valued CSVi functions, since in
this case ρi may not be varied at will.
(2) The breakeven point, that is, the
value at which π equals ρ (e.g., Apt´
e
et al. [1994]; Cohen and Singer [1999];
Dagan et al. [1997]; Joachims [1998];
17 While ρ can always be increased at will by lowi
ering τi , usually at the cost of decreasing πi , πi can
usually be increased at will by raising τi , always at
the cost of decreasing ρi . This kind of tuning is only
possible for CSVi functions with values in [0, 1]; for
binary-valued CSVi functions tuning is not always
possible, or is anyway more difficult (see Weiss et al.
[1999], page 66).
18 An exception is single-label TC, in which π and ρ
are not independent of each other: if a document d j
has been classified under a wrong category cs (thus
decreasing πs ), this also means that it has not been
classified under the right category ct (thus decreasing ρt ). In this case either π or ρ can be used as a
measure of effectiveness.

Sebastiani
Joachims [1999]; Lewis [1992a]; Lewis
and Ringuette [1994]; Moulinier and
Ganascia [1996]; Ng et al. [1997]; Yang
[1999]). This is obtained by a process
analogous to the one used for 11-point
average precision: a plot of π as a function of ρ is computed by repeatedly
varying the thresholds τi ; breakeven
is the value of ρ (or π ) for which the
plot intersects the ρ = π line. This idea
relies on the fact that, by decreasing
the τi ’s from 1 to 0, ρ always increases
monotonically from 0 to 1 and π usually decreases monotonically from a
|C|
1
value near 1 to |C| i=1 g Te (ci ). If for
no values of the τi ’s π and ρ are exactly equal, the τi ’s are set to the value
for which π and ρ are closest, and an
interpolated breakeven is computed as
the average of the values of π and ρ.19
(3) The Fβ function [van Rijsbergen 1979,
Chapter 7], for some 0 ≤ β ≤ + ∞
(e.g., Cohen [1995a]; Cohen and Singer
[1999]; Lewis and Gale [1994]; Lewis
[1995a]; Moulinier et al. [1996]; Ruiz
and Srinivassan [1999]), where
(β 2 + 1)πρ
β 2π + ρ
Here β may be seen as the relative degree of importance attributed to π and
ρ. If β = 0 then Fβ coincides with π,
whereas if β = +∞ then Fβ coincides
with ρ. Usually, a value β = 1 is used,
which attributes equal importance to
π and ρ. As shown in Moulinier et al.
[1996] and Yang [1999], the breakeven
of a classifier is always less or equal
than its F1 value.
Fβ =

19

Breakeven, first proposed by Lewis [1992a, 1992b],
has been recently criticized. Lewis himself (see his
message of 11 Sep 1997 10:49:01 to the DDLBETA
text categorization mailing list—quoted with permission of the author) has pointed out that breakeven is
not a good effectiveness measure, since (i) there may
be no parameter setting that yields the breakeven; in
this case the final breakeven value, obtained by interpolation, is artificial; (ii) to have ρ equal π is not
necessarily desirable, and it is not clear that a system
that achieves high breakeven can be tuned to score
high on other effectiveness measures. Yang [1999]
also noted that when for no value of the parameters π
and ρ are close enough, interpolated breakeven may
not be a reliable indicator of effectiveness.


Once an effectiveness measure is chosen,
a classifier can be tuned (e.g., thresholds and other parameters can be set)
so that the resulting effectiveness is the
best achievable by that classifier. Tuning a parameter p (be it a threshold or
other) is normally done experimentally.
This means performing repeated experiments on the validation set with the values of the other parameters pk fixed (at
a default value, in the case of a yet-tobe-tuned parameter pk , or at the chosen
value, if the parameter pk has already
been tuned) and with different values for
parameter p. The value that has yielded
the best effectiveness is chosen for p.
7.2. Benchmarks for Text Categorization

Standard benchmark collections that can
be used as initial corpora for TC are publically available for experimental purposes.
The most widely used is the Reuters collection, consisting of a set of newswire
stories classified under categories related
to economics. The Reuters collection accounts for most of the experimental work
in TC so far. Unfortunately, this does not
always translate into reliable comparative
results, in the sense that many of these experiments have been carried out in subtly
different conditions.
In general, different sets of experiments
may be used for cross-classifier comparison only if the experiments have been
performed
(1) on exactly the same collection (i.e.,
same documents and same categories);
(2) with the same “split” between training
set and test set;
(3) with the same evaluation measure
and, whenever this measure depends
on some parameters (e.g., the utility
matrix chosen), with the same parameter values.
Unfortunately, a lot of experimentation,
both on Reuters and on other collections, has not been performed with these
caveats in mind: by testing three different classifiers on five popular versions
of Reuters, Yang [1999] has shown that

37

a lack of compliance with these three
conditions may make the experimental
results hardly comparable among each
other. Table VI lists the results of all
experiments known to us performed on
five major versions of the Reuters benchmark: Reuters-22173 “ModLewis” (column
#1), Reuters-22173 “ModApte” (column #2),
´
Reuters-22173 “ModWiener” (column #3),
Reuters-21578 “ModApte” (column #4),
´
and Reuters-21578[10] “ModApte” (column
´
#5).20 Only experiments that have computed either a breakeven or F1 have been
listed, since other less popular effectiveness measures do not readily compare
with these.
Note that only results belonging to the
same column are directly comparable.
In particular, Yang [1999] showed that
experiments carried out on Reuters-22173
“ModLewis” (column #1) are not directly
comparable with those using the other
three versions, since the former strangely
includes a significant percentage (58%) of
“unlabeled” test documents which, being
negative examples of all categories, tend
to depress effectiveness. Also, experiments performed on Reuters-21578[10]
“ModApte” (column #5) are not comparable
´
with the others, since this collection is the
restriction of Reuters-21578 “ModApte” to
´
the 10 categories with the highest generality, and is thus an obviously “easier”
collection.
Other test collections that have been
frequently used are
—the OHSUMED collection, set up
by Hersh et al. [1994] and used by
Joachims [1998], Lam and Ho [1998],
Lam et al. [1999], Lewis et al. [1996],
Ruiz and Srinivasan [1999], and Yang

20 The Reuters-21578 collection may be freely downloaded for experimentation purposes from http://
www.research.att.com/~lewis/reuters21578
.html.
A new corpus, called Reuters Corpus Volume 1 and
consisting of roughly 800,000 documents, has
recently been made available by Reuters for
TC experiments (see http://about.reuters.com/
researchandstandards/corpus/). This will likely
replace Reuters-21578 as the “standard” Reuters
benchmark for TC.

38

Sebastiani

Table VI. Comparative Results Among Different Classiﬁers Obtained on Five Different Versions of Reuters.
(Unless otherwise noted, entries indicate the microaveraged breakeven point; within parentheses, “M”
indicates macroaveraging and “F 1 ” indicates use of the F 1 measure; boldface indicates the best
performer on the collection.)
#1
#2
#3
#4
#5
# of documents
21,450
14,347 13,272 12,902 12,902
# of training documents
14,704
10,667 9,610 9,603 9,603
# of test documents
6,746
3,680 3,662 3,299 3,299
# of categories
135
93
92
90
10
System
Type
Results reported by
WORD
(non-learning)
Yang [1999]
.150
.310
.290
probabilistic
[Dumais et al. 1998]
.752 .815
probabilistic
[Joachims 1998]
.720
probabilistic
[Lam et al. 1997]
.443 (MF1 )
PROPBAYES
probabilistic
[Lewis 1992a]
.650
BIM
probabilistic
[Li and Yamanishi 1999]
.747
probabilistic
.773
NB
probabilistic
[Yang and Liu 1999]
.795
decision trees
.884
C4.5
decision trees
[Joachims 1998]
.794
IND
decision trees [Lewis and Ringuette 1994]
.670
SWAP-1
decision rules
[Apt´ et al. 1994]
e
.805
RIPPER
decision rules
[Cohen and Singer 1999]
.683
.811
.820
SLEEPINGEXPERTS decision rules
.753
.759
.827
DL-ESC
decision rules
.820
CHARADE
decision rules [Moulinier and Ganascia 1996]
.738
CHARADE
decision rules
[Moulinier et al. 1996]
.783 (F1 )
LLSF
regression
[Yang 1999]
.855
.810
LLSF
regression
[Yang and Liu 1999]
.849
BALANCEDWINNOW on-line linear
[Dagan et al. 1997]
.747 (M) .833 (M)
WIDROW-HOFF
on-line linear
[Lam and Ho 1998]
.822
ROCCHIO
batch linear
.660
.748
.776
FINDSIM
batch linear
.617 .646
ROCCHIO
batch linear
[Joachims 1998]
.799
ROCCHIO
batch linear
[Lam and Ho 1998]
.781
ROCCHIO
batch linear
.625
CLASSI
neural network
[Ng et al. 1997]
.802
NNET
neural network
Yang and Liu 1999]
.838
neural network
[Wiener et al. 1995]
.820
GIS-W
example-based
[Lam and Ho 1998]
.860
k-NN
example-based
[Joachims 1998]
.823
k-NN
example-based
[Lam and Ho 1998]
.820
k-NN
example-based
[Yang 1999]
.690
.852
.820
k-NN
example-based
[Yang and Liu 1999]
.856
SVM
.870 .920
SVMLIGHT
SVM
[Joachims 1998]
.864
SVMLIGHT
SVM
[Li Yamanishi 1999]
.841
SVMLIGHT
SVM
[Yang and Liu 1999]
.859
ADABOOST.MH
committee
[Schapire and Singer 2000]
.860
committee
[Weiss et al. 1999]
.878
Bayesian net
.800 .850
Bayesian net
[Lam et al. 1997]
.542 (MF1 )

and Pedersen [1997].21 The documents
are titles or title-plus-abstracts from
medical journals (OHSUMED is actually
a subset of the Medline document base);
21

The OHSUMED collection may be freely downloaded for experimentation purposes from ftp://
medir.ohsu.edu/pub/ohsumed.

the categories are the “postable terms”
of the MESH thesaurus.
—the 20 Newsgroups collection, set up
by Lang [1995] and used by Baker
and McCallum [1998], Joachims
[1997], McCallum and Nigam [1998],
McCallum et al. [1998], Nigam et al.

[2000], and Schapire and Singer [2000].
The documents are messages posted to
Usenet newsgroups, and the categories
are the newsgroups themselves.
—the AP collection, used by Cohen [1995a,
1995b], Cohen and Singer [1999], Lewis
and Catlett [1994], Lewis and Gale
[1994], Lewis et al. [1996], Schapire
and Singer [2000], and Schapire et al.
[1998].
We will not cover the experiments performed on these collections for the same
reasons as those illustrated in footnote 20,
that is, because in no case have a significant enough number of authors used the
same collection in the same experimental conditions, thus making comparisons
difficult.

7.3. Which Text Classifier Is Best?

The published experimental results, and
especially those listed in Table VI, allow
us to attempt some considerations on the
comparative performance of the TC methods discussed. However, we have to bear in
mind that comparisons are reliable only
when based on experiments performed
by the same author under carefully controlled conditions. They are instead more
problematic when they involve different
experiments performed by different authors. In this case various “background
conditions,” often extraneous to the learning algorithm itself, may influence the results. These may include, among others,
different choices in preprocessing (stemming, etc.), indexing, dimensionality reduction, classifier parameter values, etc.,
but also different standards of compliance
with safe scientific practice (such as tuning parameters on the test set rather than
on a separate validation set), which often
are not discussed in the published papers.
Two different methods may thus be
applied for comparing classifiers [Yang
1999]:
and
—direct comparison: classifiers
may be compared when they have been
tested on the same collection , usually
by the same researchers and with the

39

same background conditions. This is the
more reliable method.
—indirect comparison: classifiers
and
may be compared when
(1) they have been tested on collections
and
, respectively, typically
by different researchers and hence
with possibly different background
conditions;
(2) one or more “baseline” classifiers
¯ 1 , . . . , ¯ m have been tested on both
and
by the direct comparison
method.
Test 2 gives an indication on the relaand ; using this
tive “hardness” of
and the results from Test 1, we may
obtain an indication on the relative efand
. For the reafectiveness of
sons discussed above, this method is less
reliable.
A number of interesting conclusions can be
drawn from Table VI by using these two
methods. Concerning the relative “hard>
ness” of the five collections, if by
we indicate that
is a harder collection
, there seems to be enough evithan
dence that Reuters-22173 “ModLewis”
Reuters-22173 “ModWiener” > Reuters22173 “ModApte” ≈ Reuters-21578 “Mod´
Apte” > Reuters-21578[10] “ModApte.”
´
´
These facts are unsurprising; in particular, the first and the last inequalities are a
direct consequence of the peculiar characteristics of Reuters-22173 “ModLewis” and
Reuters-21578[10] “ModApte” discussed in
´
Section 7.2.
Concerning the relative performance of
the classifiers, remembering the considerations above we may attempt a few
conclusions:
—Boosting-based classifier committees,
support vector machines, examplebased methods, and regression methods
deliver top-notch performance. There
seems to be no sufficient evidence to
decidedly opt for either method; efficiency considerations or applicationdependent issues might play a role in
breaking the tie.
—Neural networks and on-line linear classifiers work very well, although slightly

40
worse than the previously mentioned
methods.
—Batch linear classifiers (Rocchio) and
probabilistic Na¨ve Bayes classifiers
ı
look the worst of the learning-based
classifiers. For Rocchio, these results
¨
confirm earlier results by Schutze
et al. [1995], who had found three classifiers based on linear discriminant analysis, linear regression, and neural networks to perform about 15% better
than Rocchio. However, recent results
by Schapire et al. [1998] ranked Rocchio
along the best performers once nearpositives are used in training.
—The data in Table VI is hardly sufficient to say anything about decision
trees. However, the work by Dumais
et al. [1998], in which a decision tree
classifier was shown to perform nearly
as well as their top performing system
(a SVM classifier), will probably renew
the interest in decision trees, an interest
that had dwindled after the unimpressive results reported in earlier literature [Cohen and Singer 1999; Joachims
1998; Lewis and Catlett 1994; Lewis
and Ringuette 1994].
—By far the lowest performance is
displayed by WORD, a classifier implemented by Yang [1999] and not
including any learning component.22
Concerning WORD and no-learning classifiers, for completeness we should recall
that one of the highest effectiveness values
reported in the literature for the Reuters
collection (a .90 breakeven) belongs to
CONSTRUE, a manually constructed classifier. However, this classifier has never
been tested on the standard variants of
Reuters mentioned in Table VI, and it is
not clear [Yang 1999] whether the (small)
test set of Reuters-22173 “ModHayes” on
22 WORD is based on the comparison between documents and category names, each treated as a vector of
weighted terms in the vector space model. WORD was
implemented by Yang with the only purpose of determining the difference in effectiveness that adding
a learning component to a classifier brings about.
WORD is actually called STR in [Yang 1994; Yang and
Chute 1994]. Another no-learning classifier was proposed in Wong et al. [1996].

Sebastiani
which the .90 breakeven value was obtained was chosen randomly, as safe scientific practice would demand. Therefore,
the fact that this figure is indicative of the
performance of CONSTRUE, and of the manual approach it represents, has been convincingly questioned [Yang 1999].
It is important to bear in mind that
the considerations above are not absolute statements (if there may be any)
on the comparative effectiveness of these
TC methods. One of the reasons is that
a particular applicative context may exhibit very different characteristics from
the ones to be found in Reuters, and different classifiers may respond differently
to these characteristics. An experimental study by Joachims [1998] involving
support vector machines, k-NN, decision
trees, Rocchio, and Na¨ve Bayes, showed
ı
all these classifiers to have similar effectiveness on categories with ≥ 300 positive training examples each. The fact
that this experiment involved the methods which have scored best (support vector machines, k-NN) and worst (Rocchio
and Na¨ve Bayes) according to Table VI
ı
shows that applicative contexts different
from Reuters may well invalidate conclusions drawn on this latter.
Finally, a note about the worth of statistical significance testing. Few authors
have gone to the trouble of validating their
results by means of such tests. These tests
are useful for verifying how strongly the
experimental results support the claim
is better than anthat a given system
other system , or for verifying how much
a difference in the experimental setup affects the measured effectiveness of a sys¨
tem . Hull [1994] and Schutze et al.
[1995] have been among the first to work
in this direction, validating their results
by means of the ANOVA test and the Friedman test; the former is aimed at determining the significance of the difference in effectiveness between two methods in terms
of the ratio between this difference and the
effectiveness variability across categories,
while the latter conducts a similar test by
using instead the rank positions of each
method within a category. Yang and Liu
[1999] defined a full suite of significance

tests, some of which apply to microaveraged and some to macroaveraged effectiveness. They applied them systematically
to the comparison between five different
classifiers, and were thus able to infer finegrained conclusions about their relative
effectiveness. For other examples of significance testing in TC, see Cohen [1995a,
1995b]; Cohen and Hirsh [1998], Joachims
[1997], Koller and Sahami [1997], Lewis
et al. [1996], and Wiener et al. [1995].
8. CONCLUSION

Automated TC is now a major research
area within the information systems discipline, thanks to a number of factors:
—Its domains of application are numerous and important, and given the proliferation of documents in digital form
they are bound to increase dramatically
in both number and importance.
—It is indispensable in many applications in which the sheer number of
the documents to be classified and the
short response time required by the application make the manual alternative
implausible.
—It can improve the productivity of
human classifiers in applications in
which no classification decision can be
taken without a final human judgment
[Larkey and Croft 1996], by providing
tools that quickly “suggest” plausible
decisions.
—It has reached effectiveness levels comparable to those of trained professionals. The effectiveness of manual TC is
not 100% anyway [Cleverdon 1984] and,
more importantly, it is unlikely to be
improved substantially by the progress
of research. The levels of effectiveness
of automated TC are instead growing
at a steady pace, and even if they will
likely reach a plateau well below the
100% level, this plateau will probably
be higher than the effectiveness levels
of manual TC.
One of the reasons why from the early
’90s the effectiveness of text classifiers
has dramatically improved is the arrival

41

in the TC arena of ML methods that
are backed by strong theoretical motivations. Examples of these are multiplicative weight updating (e.g., the WINNOW
family, WIDROW-HOFF, etc.), adaptive resampling (e.g., boosting), and support vector machines, which provide a sharp contrast with relatively unsophisticated and
weak methods such as Rocchio. In TC,
ML researchers have found a challenging application, since datasets consisting
of hundreds of thousands of documents
and characterized by tens of thousands of
terms are widely available. This means
that TC is a good benchmark for checking
whether a given learning technique can
scale up to substantial sizes. In turn, this
probably means that the active involvement of the ML community in TC is bound
to grow.
The success story of automated TC is
also going to encourage an extension of
its methods and techniques to neighboring fields of application. Techniques typical of automated TC have already been
extended successfully to the categorization of documents expressed in slightly different media; for instance:
—very noisy text resulting from optical character recognition [Ittner et al.
1995; Junker and Hoch 1998]. In their
experiments Ittner et al. [1995] have
found that, by employing noisy texts
also in the training phase (i.e. texts affected by the same source of noise that
is also at work in the test documents),
effectiveness levels comparable to those
obtainable in the case of standard text
can be achieved.
—speech transcripts [Myers et al.
2000; Schapire and Singer 2000].
For instance, Schapire and Singer
[2000] classified answers given to a
phone operator’s request “How may I
help you?” so as to be able to route the
call to a specialized operator according
to call type.
Concerning other more radically different media, the situation is not as bright
(however, see Lim [1999] for an interesting attempt at image categorization based

42

Sebastiani

on a textual metaphor). The reason for
this is that capturing real semantic content of nontextual media by automatic indexing is still an open problem. While
there are systems that attempt to detect
content, for example, in images by recognizing shapes, color distributions, and
texture, the general problem of image semantics is still unsolved. The main reason
is that natural language, the language of
the text medium, admits far fewer variations than the “languages” employed by
the other media. For instance, while the
concept of a house can be “triggered” by
relatively few natural language expressions such as house, houses, home, housing,
inhabiting, etc., it can be triggered by far
more images: the images of all the different houses that exist, of all possible colors
and shapes, viewed from all possible perspectives, from all possible distances, etc.
If we had solved the multimedia indexing
problem in a satisfactory way, the general
methodology that we have discussed in
this paper for text would also apply to automated multimedia categorization, and
there are reasons to believe that the effectiveness levels could be as high. This
only adds to the common sentiment that
more research in automated contentbased indexing for multimedia documents
is needed.
ACKNOWLEDGMENTS
This paper owes a lot to the suggestions and constructive criticism of Norbert Fuhr and David Lewis.
Thanks also to Umberto Straccia for comments on
an earlier draft, to Evgeniy Gabrilovich, Daniela
Giorgetti, and Alessandro Moschitti for spotting mistakes in an earlier draft, and to Alessandro Sperduti
for many fruitful discussions.

REFERENCES
AMATI, G. AND CRESTANI, F. 1999. Probabilistic
learning for selective dissemination of information. Inform. Process. Man. 35, 5, 633–654.
ANDROUTSOPOULOS, I., KOUTSIAS, J., CHANDRINOS, K. V.,
AND SPYROPOULOS, C. D. 2000. An experimental comparison of naive Bayesian and keywordbased anti-spam filtering with personal e-mail
messages. In Proceedings of SIGIR-00, 23rd
ACM International Conference on Research and
Development in Information Retrieval (Athens,
Greece, 2000), 160–167.

´
APTE, C., DAMERAU, F. J., AND WEISS, S. M. 1994.
Automated learning of decision rules for text
categorization. ACM Trans. on Inform. Syst. 12,
3, 233–251.
ATTARDI, G., DI MARCO, S., AND SALVI, D. 1998. Categorization by context. J. Univers. Comput. Sci.
4, 9, 719–736.
BAKER, L. D. AND MCCALLUM, A. K. 1998. Distributional clustering of words for text classification.
In Proceedings of SIGIR-98, 21st ACM International Conference on Research and Development
in Information Retrieval (Melbourne, Australia,
1998), 96–103.
BELKIN, N. J. AND CROFT, W. B. 1992. Information
filtering and information retrieval: two sides
of the same coin? Commun. ACM 35, 12, 29–
38.
BIEBRICHER, P., FUHR, N., KNORZ, G., LUSTIG, G., AND
SCHWANTNER, M. 1988. The automatic indexing system AIR/PHYS. From research to application. In Proceedings of SIGIR-88, 11th ACM
International Conference on Research and Development in Information Retrieval (Grenoble,
France, 1988), 333–342. Also reprinted in Sparck
Jones and Willett [1997], pp. 513–517.
BORKO, H. AND BERNICK, M. 1963. Automatic document classification. J. Assoc. Comput. Mach. 10,
2, 151–161.
CAROPRESO, M. F., MATWIN, S., AND SEBASTIANI, F.
2001. A learner-independent evaluation of the
usefulness of statistical phrases for automated
text categorization. In Text Databases and Document Management: Theory and Practice, A. G.
Chin, ed. Idea Group Publishing, Hershey, PA,
78–102.
CAVNAR, W. B. AND TRENKLE, J. M. 1994. N-grambased text categorization. In Proceedings of
SDAIR-94, 3rd Annual Symposium on Document Analysis and Information Retrieval (Las
Vegas, NV, 1994), 161–175.
CHAKRABARTI, S., DOM, B. E., AGRAWAL, R., AND
RAGHAVAN, P. 1998a. Scalable feature selection, classification and signature generation for
organizing large text databases into hierarchical
topic taxonomies. J. Very Large Data Bases 7, 3,
163–178.
CHAKRABARTI, S., DOM, B. E., AND INDYK, P. 1998b.
Enhanced hypertext categorization using hyperlinks. In Proceedings of SIGMOD-98, ACM International Conference on Management of Data
(Seattle, WA, 1998), 307–318.
CLACK, C., FARRINGDON, J., LIDWELL, P., AND YU, T.
1997. Autonomous document classification for
business. In Proceedings of the 1st International
Conference on Autonomous Agents (Marina del
Rey, CA, 1997), 201–208.
CLEVERDON, C. 1984. Optimizing convenient online access to bibliographic databases. Inform.
Serv. Use 4, 1, 37–47. Also reprinted in Willett
[1988], pp. 32–41.
COHEN, W. W. 1995a. Learning to classify English
text with ILP methods. In Advances in Inductive


Logic Programming, L. De Raedt, ed. IOS Press,
Amsterdam, The Netherlands, 124–143.
COHEN, W. W. 1995b. Text categorization and relational learning. In Proceedings of ICML-95, 12th
International Conference on Machine Learning
(Lake Tahoe, CA, 1995), 124–132.
COHEN, W. W. AND HIRSH, H. 1998. Joins that generalize: text classification using WHIRL. In Proceedings of KDD-98, 4th International Conference on Knowledge Discovery and Data Mining
(New York, NY, 1998), 169–173.
COHEN, W. W. AND SINGER, Y. 1999. Contextsensitive learning methods for text categorization. ACM Trans. Inform. Syst. 17, 2, 141–
173.
COOPER, W. S. 1995. Some inconsistencies and misnomers in probabilistic information retrieval.
ACM Trans. Inform. Syst. 13, 1, 100–111.
CREECY, R. M., MASAND, B. M., SMITH, S. J., AND WALTZ,
D. L. 1992. Trading MIPS and memory for
knowledge engineering: classifying census returns on the Connection Machine. Commun.
ACM 35, 8, 48–63.
CRESTANI, F., LALMAS, M., VAN RIJSBERGEN, C. J., AND
CAMPBELL, I. 1998. “Is this document relevant? . . . probably.” A survey of probabilistic
models in information retrieval. ACM Comput.
Surv. 30, 4, 528–552.
DAGAN, I., KAROV, Y., AND ROTH, D. 1997. Mistakedriven learning in text categorization. In Proceedings of EMNLP-97, 2nd Conference on Empirical Methods in Natural Language Processing
(Providence, RI, 1997), 55–63.
DEERWESTER, S., DUMAIS, S. T., FURNAS, G. W.,
LANDAUER, T. K., AND HARSHMAN, R. 1990. Indexing by latent semantic indexing. J. Amer. Soc.
Inform. Sci. 41, 6, 391–407.
DENOYER, L., ZARAGOZA, H., AND GALLINARI, P. 2001.
HMM-based passage models for document classification and ranking. In Proceedings of ECIR01, 23rd European Colloquium on Information
Retrieval Research (Darmstadt, Germany, 2001).
˜
DÁZ ESTEBAN, A., DE BUENAGA RODR´GUEZ, M., URENA
I
I
´
LOPEZ, L. A., AND GARCÁ VEGA, M. 1998. InI
tegrating linguistic resources in an uniform
way for text classification tasks. In Proceedings of LREC-98, 1st International Conference on
Language Resources and Evaluation (Grenada,
Spain, 1998), 1197–1204.
DOMINGOS, P. AND PAZZANI, M. J. 1997. On the the
optimality of the simple Bayesian classifier under zero-one loss. Mach. Learn. 29, 2–3, 103–130.
DRUCKER, H., VAPNIK, V., AND WU, D. 1999. Automatic text categorization and its applications to
text retrieval. IEEE Trans. Neural Netw. 10, 5,
1048–1054.
DUMAIS, S. T. AND CHEN, H. 2000. Hierarchical classification of Web content. In Proceedings of
SIGIR-00, 23rd ACM International Conference
on Research and Development in Information
Retrieval (Athens, Greece, 2000), 256–263.


43

DUMAIS, S. T., PLATT, J., HECKERMAN, D., AND SAHAMI,
M. 1998. Inductive learning algorithms and
representations for text categorization. In Proceedings of CIKM-98, 7th ACM International
Conference on Information and Knowledge Management (Bethesda, MD, 1998), 148–155.
ESCUDERO, G., M` RQUEZ, L., AND RIGAU, G. 2000.
A
Boosting applied to word sense disambiguation.
In Proceedings of ECML-00, 11th European Conference on Machine Learning (Barcelona, Spain,
2000), 129–141.
FIELD, B. 1975. Towards automatic indexing: automatic assignment of controlled-language indexing and classification from free indexing. J. Document. 31, 4, 246–265.
FORSYTH, R. S. 1999. New directions in text categorization. In Causal Models and Intelligent Data
Management, A. Gammerman, ed. Springer,
Heidelberg, Germany, 151–185.
FRASCONI, P., SODA, G., AND VULLO, A. 2002. Text
categorization for multi-page documents: A
hybrid naive Bayes HMM approach. J. Intell.
Inform. Syst. 18, 2/3 (March–May), 195–217.
FUHR, N. 1985. A probabilistic model of dictionarybased automatic indexing. In Proceedings of
RIAO-85, 1st International Conference “Recherche d’Information Assistee par Ordinateur”
(Grenoble, France, 1985), 207–216.
FUHR, N. 1989. Models for retrieval with probabilistic indexing. Inform. Process. Man. 25, 1, 55–
72.
FUHR, N. AND BUCKLEY, C. 1991. A probabilistic
learning approach for document indexing. ACM
Trans. Inform. Syst. 9, 3, 223–248.
FUHR, N., HARTMANN, S., KNORZ, G., LUSTIG, G.,
SCHWANTNER, M., AND TZERAS, K. 1991.
AIR/X—a rule-based multistage indexing
system for large subject fields. In Proceedings of RIAO-91, 3rd International Conference
“Recherche d’Information Assistee par Ordinateur” (Barcelona, Spain, 1991), 606–623.
FUHR, N. AND KNORZ, G. 1984. Retrieval test
evaluation of a rule-based automated indexing (AIR/PHYS). In Proceedings of SIGIR-84,
7th ACM International Conference on Research
and Development in Information Retrieval
(Cambridge, UK, 1984), 391–408.
FUHR, N. AND PFEIFER, U. 1994. Probabilistic information retrieval as combination of abstraction inductive learning and probabilistic assumptions. ACM Trans. Inform. Syst. 12, 1,
92–115.
¨
FURNKRANZ, J. 1999. Exploiting structural information for text classification on the WWW.
In Proceedings of IDA-99, 3rd Symposium on
Intelligent Data Analysis (Amsterdam, The
Netherlands, 1999), 487–497.
GALAVOTTI, L., SEBASTIANI, F., AND SIMI, M. 2000.
Experiments on the use of feature selection and negative evidence in automated text
categorization. In Proceedings of ECDL-00,
4th European Conference on Research and

44
Advanced Technology for Digital Libraries
(Lisbon, Portugal, 2000), 59–68.
GALE, W. A., CHURCH, K. W., AND YAROWSKY, D. 1993.
A method for disambiguating word senses in a
large corpus. Comput. Human. 26, 5, 415–439.
¨
GOVERT, N., LALMAS, M., AND FUHR, N. 1999. A
probabillistic description-oriented approach for
categorising Web documents. In Proceedings of
CIKM-99, 8th ACM International Conference
on Information and Knowledge Management
(Kansas City, MO, 1999), 475–482.
GRAY, W. A. AND HARLEY, A. J. 1971. Computerassisted indexing. Inform. Storage Retrieval 7,
4, 167–174.
GUTHRIE, L., WALKER, E., AND GUTHRIE, J. A. 1994.
Document classification by machine: theory
and practice. In Proceedings of COLING-94, 15th
International Conference on Computational Linguistics (Kyoto, Japan, 1994), 1059–1063.
HAYES, P. J., ANDERSEN, P. M., NIRENBURG, I. B.,
AND SCHMANDT, L. M. 1990. Tcs: a shell for
content-based text categorization. In Proceedings of CAIA-90, 6th IEEE Conference on Artificial Intelligence Applications (Santa Barbara,
CA, 1990), 320–326.
HEAPS, H. 1973. A theory of relevance for automatic document classification. Inform. Control 22, 3, 268–278.
HERSH, W., BUCKLEY, C., LEONE, T., AND HICKMAN, D.
1994. OHSUMED: an interactive retrieval evaluation and new large text collection for research.
In Proceedings of SIGIR-94, 17th ACM International Conference on Research and Development
in Information Retrieval (Dublin, Ireland, 1994),
192–201.
HULL, D. A. 1994. Improving text retrieval for the
routing problem using latent semantic indexing.
in Information Retrieval (Dublin, Ireland, 1994),
282–289.
¨
HULL, D. A., PEDERSEN, J. O., AND SCHUTZE, H. 1996.
Method combination for document filtering. In
Proceedings of SIGIR-96, 19th ACM International Conference on Research and Development
¨
in Information Retrieval (Zurich, Switzerland,
1996), 279–288.
ITTNER, D. J., LEWIS, D. D., AND AHN, D. D. 1995.
Text categorization of low quality images. In
Proceedings of SDAIR-95, 4th Annual Symposium on Document Analysis and Information
Retrieval (Las Vegas, NV, 1995), 301–315.
IWAYAMA, M. AND TOKUNAGA, T. 1995. Cluster-based
text categorization: a comparison of category
search strategies. In Proceedings of SIGIR-95,
18th ACM International Conference on Research
and Development in Information Retrieval
(Seattle, WA, 1995), 273–281.
IYER, R. D., LEWIS, D. D., SCHAPIRE, R. E., SINGER, Y.,
AND SINGHAL, A. 2000. Boosting for document
routing. In Proceedings of CIKM-00, 9th ACM
International Conference on Information and

Sebastiani
Knowledge Management (McLean, VA, 2000),
70–77.
JOACHIMS, T. 1997. A probabilistic analysis of the
Rocchio algorithm with TFIDF for text categorization. In Proceedings of ICML-97, 14th
(Nashville, TN, 1997), 143–151.
JOACHIMS, T. 1998. Text categorization with support vector machines: learning with many relevant features. In Proceedings of ECML-98,
10th European Conference on Machine Learning
(Chemnitz, Germany, 1998), 137–142.
JOACHIMS, T. 1999. Transductive inference for text
classification using support vector machines. In
Proceedings of ICML-99, 16th International Conference on Machine Learning (Bled, Slovenia,
1999), 200–209.
JOACHIMS, T. AND SEBASTIANI, F. 2002. Guest editors’
introduction to the special issue on automated
text categorization. J. Intell. Inform. Syst. 18, 2/3
(March-May), 103–105.
JOHN, G. H., KOHAVI, R., AND PFLEGER, K. 1994. Irrelevant features and the subset selection problem. In Proceedings of ICML-94, 11th International Conference on Machine Learning (New
Brunswick, NJ, 1994), 121–129.
JUNKER, M. AND ABECKER, A. 1997. Exploiting thesaurus knowledge in rule induction for text classification. In Proceedings of RANLP-97, 2nd International Conference on Recent Advances in
Natural Language Processing (Tzigov Chark,
Bulgaria, 1997), 202–207.
JUNKER, M. AND HOCH, R. 1998. An experimental evaluation of OCR text representations for
learning document classifiers. Internat. J. Document Analysis and Recognition 1, 2, 116–122.
¨
KESSLER, B., NUNBERG, G., AND SCHUTZE, H. 1997.
Automatic detection of text genre. In Proceedings of ACL-97, 35th Annual Meeting of the Association for Computational Linguistics (Madrid,
Spain, 1997), 32–38.
KIM, Y.-H., HAHN, S.-Y., AND ZHANG, B.-T. 2000. Text
filtering by boosting naive Bayes classifiers. In
Proceedings of SIGIR-00, 23rd ACM International Conference on Research and Development
in Information Retrieval (Athens, Greece, 2000),
168–175.
KLINKENBERG, R. AND JOACHIMS, T. 2000. Detecting concept drift with support vector machines.
In Proceedings of ICML-00, 17th International
Conference on Machine Learning (Stanford, CA,
2000), 487–494.
KNIGHT, K. 1999. Mining online text. Commun.
ACM 42, 11, 58–61.
KNORZ, G. 1982. A decision theory approach to
optimal automated indexing. In Proceedings of
SIGIR-82, 5th ACM International Conference
Retrieval (Berlin, Germany, 1982), 174–193.
KOLLER, D. AND SAHAMI, M. 1997. Hierarchically
classifying documents using very few words. In


Proceedings of ICML-97, 14th International Conference on Machine Learning (Nashville, TN,
1997), 170–178.
KORFHAGE, R. R. 1997. Information Storage and
Retrieval. Wiley Computer Publishing, New
York, NY.
LAM, S. L. AND LEE, D. L. 1999. Feature reduction for neural network based text categorization. In Proceedings of DASFAA-99, 6th IEEE
International Conference on Database Advanced
Systems for Advanced Application (Hsinchu,
Taiwan, 1999), 195–202.
LAM, W. AND HO, C. Y. 1998. Using a generalized
instance set for automatic text categorization.
1998), 81–89.
LAM, W., LOW, K. F., AND HO, C. Y. 1997. Using a
Bayesian network induction approach for text
categorization. In Proceedings of IJCAI-97, 15th
International Joint Conference on Artificial Intelligence (Nagoya, Japan, 1997), 745–750.
LAM, W., RUIZ, M. E., AND SRINIVASAN, P. 1999. Automatic text categorization and its applications to
text retrieval. IEEE Trans. Knowl. Data Engin.
11, 6, 865–879.
LANG, K. 1995. NEWSWEEDER: learning to filter netnews. In Proceedings of ICML-95, 12th International Conference on Machine Learning (Lake
Tahoe, CA, 1995), 331–339.
LARKEY, L. S. 1998. Automatic essay grading using text categorization techniques. In Proceedings of SIGIR-98, 21st ACM International
Conference on Research and Development in
Information Retrieval (Melbourne, Australia,
1998), 90–95.
LARKEY, L. S. 1999. A patent search and classification system. In Proceedings of DL-99, 4th ACM
Conference on Digital Libraries (Berkeley, CA,
1999), 179–187.
LARKEY, L. S. AND CROFT, W. B. 1996. Combining
classifiers in text categorization. In Proceedings
of SIGIR-96, 19th ACM International Conference
¨
Retrieval (Zurich, Switzerland, 1996), 289–297.
LEWIS, D. D. 1992a. An evaluation of phrasal and
clustered representations on a text categorization task. In Proceedings of SIGIR-92, 15th ACM
International Conference on Research and Development in Information Retrieval (Copenhagen,
Denmark, 1992), 37–50.
LEWIS, D. D. 1992b. Representation and Learning in Information Retrieval. Ph. D. thesis, Department of Computer Science, University of
Massachusetts, Amherst, MA.
LEWIS, D. D. 1995a. Evaluating and optmizing autonomous text classification systems. In Proceedings of SIGIR-95, 18th ACM International
Information Retrieval (Seattle, WA, 1995), 246–
254.


45

LEWIS, D. D. 1995b. A sequential algorithm for
training text classifiers: corrigendum and additional data. SIGIR Forum 29, 2, 13–19.
LEWIS, D. D. 1995c. The TREC-4 filtering track:
description and analysis. In Proceedings
of TREC-4, 4th Text Retrieval Conference
(Gaithersburg, MD, 1995), 165–180.
LEWIS, D. D. 1998. Naive (Bayes) at forty: The
independence assumption in information retrieval. In Proceedings of ECML-98, 10th
European Conference on Machine Learning
LEWIS, D. D. AND CATLETT, J. 1994. Heterogeneous
uncertainty sampling for supervised learning. In
Proceedings of ICML-94, 11th International Conference on Machine Learning (New Brunswick,
NJ, 1994), 148–156.
LEWIS, D. D. AND GALE, W. A. 1994. A sequential
algorithm for training text classifiers. In Proceedings of SIGIR-94, 17th ACM International
Information Retrieval (Dublin, Ireland, 1994),
3–12. See also Lewis [1995b].
LEWIS, D. D. AND HAYES, P. J. 1994. Guest editorial
for the special issue on text categorization. ACM
Trans. Inform. Syst. 12, 3, 231.
LEWIS, D. D. AND RINGUETTE, M. 1994. A comparison of two learning algorithms for text categorization. In Proceedings of SDAIR-94, 3rd
Annual Symposium on Document Analysis and
Information Retrieval (Las Vegas, NV, 1994),
81–93.
LEWIS, D. D., SCHAPIRE, R. E., CALLAN, J. P., AND PAPKA,
R. 1996. Training algorithms for linear text
classifiers. In Proceedings of SIGIR-96, 19th
ACM International Conference on Research and
¨
Development in Information Retrieval (Zurich,
Switzerland, 1996), 298–306.
LI, H. AND YAMANISHI, K. 1999. Text classification
using ESC-based stochastic decision lists. In
Proceedings of CIKM-99, 8th ACM International
Conference on Information and Knowledge Management (Kansas City, MO, 1999), 122–130.
LI, Y. H. AND JAIN, A. K. 1998. Classification of text
documents. Comput. J. 41, 8, 537–546.
LIDDY, E. D., PAIK, W., AND YU, E. S. 1994. Text categorization for multiple users based on semantic features from a machine-readable dictionary.
ACM Trans. Inform. Syst. 12, 3, 278–295.
LIERE, R. AND TADEPALLI, P. 1997. Active learning
with committees for text categorization. In Proceedings of AAAI-97, 14th Conference of the
American Association for Artificial Intelligence
(Providence, RI, 1997), 591–596.
LIM, J. H. 1999. Learnable visual keywords for image classification. In Proceedings of DL-99, 4th
ACM Conference on Digital Libraries (Berkeley,
CA, 1999), 139–145.
¨
MANNING, C. AND SCHUTZE, H. 1999. Foundations of
Statistical Natural Language Processing. MIT
Press, Cambridge, MA.

46
MARON, M. 1961. Automatic indexing: an experimental inquiry. J. Assoc. Comput. Mach. 8, 3,
404–417.
MASAND, B. 1994. Optimising confidence of text
classification by evolution of symbolic expressions. In Advances in Genetic Programming,
K. E. Kinnear, ed. MIT Press, Cambridge, MA,
Chapter 21, 459–476.
MASAND, B., LINOFF, G., AND WALTZ, D. 1992. Classifying news stories using memory-based reasoning. In Proceedings of SIGIR-92, 15th ACM
International Conference on Research and Development in Information Retrieval (Copenhagen,
Denmark, 1992), 59–65.
MCCALLUM, A. K. AND NIGAM, K. 1998. Employing EM in pool-based active learning for text
classification. In Proceedings of ICML-98, 15th
(Madison, WI, 1998), 350–358.
MCCALLUM, A. K., ROSENFELD, R., MITCHELL, T. M., AND
NG, A. Y. 1998. Improving text classification
by shrinkage in a hierarchy of classes. In Proceedings of ICML-98, 15th International Conference on Machine Learning (Madison, WI, 1998),
359–367.
MERKL, D. 1998. Text classification with selforganizing maps: Some lessons learned. Neurocomputing 21, 1/3, 61–77.
MITCHELL, T. M. 1996. Machine Learning. McGraw
Hill, New York, NY.
´
MLADENIC, D. 1998. Feature subset selection in
text learning. In Proceedings of ECML-98,
10th European Conference on Machine Learning
´
MLADENIC, D. AND GROBELNIK, M. 1998. Word sequences as features in text-learning. In Proceedings of ERK-98, the Seventh Electrotechnical and Computer Science Conference (Ljubljana,
Slovenia, 1998), 145–148.
MOULINIER, I. AND GANASCIA, J.-G. 1996. Applying
an existing machine learning algorithm to text
categorization. In Connectionist, Statistical,
and Symbolic Approaches to Learning for Natural Language Processing, S. Wermter, E. Riloff,
and G. Schaler, eds. Springer Verlag, Heidelberg,
Germany, 343–354.
˘
MOULINIER, I., RASKINIS, G., AND GANASCIA, J.-G. 1996.
Text categorization: a symbolic approach. In
Proceedings of SDAIR-96, 5th Annual Symposium on Document Analysis and Information
Retrieval (Las Vegas, NV, 1996), 87–99.
MYERS, K., KEARNS, M., SINGH, S., AND WALKER,
M. A. 2000. A boosting approach to topic
spotting on subdialogues. In Proceedings of
ICML-00, 17th International Conference on Machine Learning (Stanford, CA, 2000), 655–
662.
NG, H. T., GOH, W. B., AND LOW, K. L. 1997. Feature selection, perceptron learning, and a usability case study for text categorization. In Proceedings of SIGIR-97, 20th ACM International

Sebastiani
Information Retrieval (Philadelphia, PA, 1997),
67–73.
NIGAM, K., MCCALLUM, A. K., THRUN, S., AND MITCHELL,
T. M. 2000. Text classification from labeled
and unlabeled documents using EM. Mach.
Learn. 39, 2/3, 103–134.
OH, H.-J., MYAENG, S. H., AND LEE, M.-H. 2000. A
practical hypertext categorization method using
links and incrementally available class information. In Proceedings of SIGIR-00, 23rd ACM International Conference on Research and Development in Information Retrieval (Athens, Greece,
2000), 264–271.
PAZIENZA, M. T., ed. 1997. Information Extraction.
Lecture Notes in Computer Science, Vol. 1299.
Springer, Heidelberg, Germany.
RILOFF. E. 1995. Little words can make a big difference for text classification. In Proceedings of
Retrieval (Seattle, WA, 1995), 130–136.
RILOFF, E. AND LEHNERT, W. 1994. Information extraction as a basis for high-precision text classification. ACM Trans. Inform. Syst. 12, 3, 296–333.
ROBERTSON, S. E. AND HARDING, P. 1984. Probabilistic automatic indexing by learning from human
indexers. J. Document. 40, 4, 264–270.
ROBERTSON, S. E. AND SPARCK JONES, K. 1976. Relevance weighting of search terms. J. Amer. Soc.
Inform. Sci. 27, 3, 129–146. Also reprinted in
Willett [1988], pp. 143–160.
ROTH, D. 1998. Learning to resolve natural
language ambiguities: a unified approach. In
Proceedings of AAAI-98, 15th Conference of the
American Association for Artificial Intelligence
(Madison, WI, 1998), 806–813.
RUIZ, M. E. AND SRINIVASAN, P. 1999. Hierarchical
neural networks for text categorization. In Proceedings of SIGIR-99, 22nd ACM International
Information Retrieval (Berkeley, CA, 1999),
281–282.
SABLE, C. L. AND HATZIVASSILOGLOU, V. 2000. Textbased approaches for non-topical image categorization. Internat. J. Dig. Libr. 3, 3, 261–275.
SALTON, G. AND BUCKLEY, C. 1988. Term-weighting
approaches in automatic text retrieval. Inform.
Process. Man. 24, 5, 513–523. Also reprinted in
Sparck Jones and Willett [1997], pp. 323–328.
SALTON, G., WONG, A., AND YANG, C. 1975. A vector
space model for automatic indexing. Commun.
ACM 18, 11, 613–620. Also reprinted in Sparck
Jones and Willett [1997], pp. 273–280.
SARACEVIC, T. 1975. Relevance: a review of and
a framework for the thinking on the notion in
information science. J. Amer. Soc. Inform. Sci.
26, 6, 321–343. Also reprinted in Sparck Jones
and Willett [1997], pp. 143–165.
SCHAPIRE, R. E. AND SINGER, Y. 2000. BoosTexter:
a boosting-based system for text categorization.
Mach. Learn. 39, 2/3, 135–168.


SCHAPIRE, R. E., SINGER, Y., AND SINGHAL, A. 1998.
Boosting and Rocchio applied to text filtering.
1998), 215–223.
¨
SCHUTZE, H. 1998. Automatic word sense discrimination. Computat. Ling. 24, 1, 97–124.
¨
SCHUTZE, H., HULL, D. A., AND PEDERSEN, J. O. 1995.
A comparison of classifiers and document representations for the routing problem. In Proceedings of SIGIR-95, 18th ACM International Conference on Research and Development in Information Retrieval (Seattle, WA, 1995), 229–237.
SCOTT, S. AND MATWIN, S. 1999. Feature engineering for text classification. In Proceedings of
ICML-99, 16th International Conference on Machine Learning (Bled, Slovenia, 1999), 379–388.
SEBASTIANI, F., SPERDUTI, A., AND VALDAMBRINI, N.
2000. An improved boosting algorithm and its
application to automated text categorization. In
Conference on Information and Knowledge
Management (McLean, VA, 2000), 78–85.
SINGHAL, A., MITRA, M., AND BUCKLEY, C. 1997.
Learning routing queries in a query zone. In
Proceedings of SIGIR-97, 20th ACM International Conference on Research and Development
in Information Retrieval (Philadelphia, PA,
1997), 25–32.
SINGHAL, A., SALTON, G., MITRA, M., AND BUCKLEY,
C. 1996. Document length normalization.
Inform. Process. Man. 32, 5, 619–633.
SLONIM, N. AND TISHBY, N. 2001. The power of word
clusters for text classification. In Proceedings
of ECIR-01, 23rd European Colloquium on
Information Retrieval Research (Darmstadt,
Germany, 2001).
SPARCK JONES, K. AND WILLETT, P., eds. 1997.
Readings in Information Retrieval. Morgan
Kaufmann, San Mateo, CA.
TAIRA, H. AND HARUNO, M. 1999. Feature selection
in SVM text categorization. In Proceedings
of AAAI-99, 16th Conference of the American
Association for Artificial Intelligence (Orlando,
FL, 1999), 480–486.
TAURITZ, D. R., KOK, J. N., AND SPRINKHUIZEN-KUYPER,
I. G. 2000. Adaptive information filtering
using evolutionary computation. Inform. Sci.
122, 2–4, 121–140.
TUMER, K. AND GHOSH, J. 1996. Error correlation
and error reduction in ensemble classifiers.
Connection Sci. 8, 3-4, 385–403.
TZERAS, K. AND HARTMANN, S. 1993. Automatic
indexing based on Bayesian inference networks.
in Information Retrieval (Pittsburgh, PA, 1993),
22–34.

RIJSBERGEN, C. J. 1977. A theoretical basis for
the use of co-occurrence data in information
retrieval. J. Document. 33, 2, 106–119.
VAN RIJSBERGEN, C. J. 1979. Information Retrieval,
2nd ed. Butterworths, London, UK. Available at
http://www.dcs.gla.ac.uk/Keith.
WEIGEND, A. S., WIENER, E. D., AND PEDERSEN, J. O.
1999. Exploiting hierarchy in text catagorization. Inform. Retr. 1, 3, 193–216.
´
WEISS, S. M., APTE, C., DAMERAU, F. J., JOHNSON, D.
E., OLES, F. J., GOETZ, T., AND HAMPP, T. 1999.
Maximizing text-mining performance. IEEE
Intell. Syst. 14, 4, 63–69.
WIENER, E. D., PEDERSEN, J. O., AND WEIGEND, A. S.
1995. A neural network approach to topic spotting. In Proceedings of SDAIR-95, 4th Annual
Symposium on Document Analysis and Information Retrieval (Las Vegas, NV, 1995), 317–332.
WILLETT, P., ed. 1988. Document Retrieval Systems. Taylor Graham, London, UK.
WONG, J. W., KAN, W.-K., AND YOUNG, G. H. 1996.
ACTION: automatic classification for full-text
documents. SIGIR Forum 30, 1, 26–41.
YANG, Y. 1994. Expert network: effective and
efficient learning from human decisions in text
categorisation and retrieval. In Proceedings of
Retrieval (Dublin, Ireland, 1994), 13–22.
YANG, Y. 1995. Noise reduction in a statistical approach to text categorization. In Proceedings of
Retrieval (Seattle, WA, 1995), 256–263.
YANG, Y. 1999. An evaluation of statistical approaches to text categorization. Inform. Retr. 1,
1–2, 69–90.
YANG, Y. AND CHUTE, C. G. 1994. An example-based
mapping method for text categorization and retrieval. ACM Trans. Inform. Syst. 12, 3, 252–277.
YANG, Y. AND LIU, X. 1999. A re-examination of
text categorization methods. In Proceedings of
SIGIR-99, 22nd ACM International Conference
Retrieval (Berkeley, CA, 1999), 42–49.
YANG, Y. AND PEDERSEN, J. O. 1997. A comparative
study on feature selection in text categorization.
In Proceedings of ICML-97, 14th International
Conference on Machine Learning (Nashville,
TN, 1997), 412–420.
YANG, Y., SLATTERY, S., AND GHANI, R. 2002. A study
of approaches to hypertext categorization. J. Intell. Inform. Syst. 18, 2/3 (March-May), 219–241.
YU, K. L. AND LAM, W. 1998. A new on-line learning algorithm for adaptive text filtering. In
Conference on Information and Knowledge
Management (Bethesda, MD, 1998), 156–160.
VAN

Received December 1999; revised February 2001; accepted July 2001


47

Machine learning in automated text categorization

More Related Content

What's hot (15)

Similar to Machine learning in automated text categorization (20)

More from unyil96 (20)

Recently uploaded (20)

Machine learning in automated text categorization