Novel Database-Centric Framework for Incremental Information Extraction

IJSRD - International Journal for Scientific Research & Development| Vol. 1, Issue 4, 2013 | ISSN (online): 2321-0613
All rights reserved by www.ijsrd.com 827
Abstract--Information extraction (IE) has been an active
research area that seeks techniques to uncover information
from a large collection of text. IE is the task of
automatically extracting structured information
from unstructured and/or semi structured machine-
readable documents. In most of the cases this activity
concerns processing human language texts by means
of natural language processing (NLP). Recent activities in
document processing like automatic annotation and content
extraction could be seen as information extraction. Many
applications call for methods to enable automatic extraction
of structured information from unstructured natural
language text. Due to the inherent challenges of natural
language processing, most of the existing methods for
information extraction from text tend to be domain specific.
In this project a new paradigm for information extraction. In
this extraction framework, intermediate output of each text
processing component is stored so that only the improved
component has to be deployed to the entire corpus.
Extraction is then performed on both the previously
processed data from the unchanged components as well as
the updated data generated by the improved component.
Performing such kind of incremental extraction can result in
a tremendous reduction of processing time and there is a
mechanism to generate extraction queries from both labeled
and unlabeled data. Query generation is critical so that
casual users can specify their information needs without
learning the query language.
I. INTRODUCTION
Information extraction (IE) is typically realized by special-
purpose programs that perform a sequence of processing
modules, including sentence splitters, tokenizers, named
entity recognizers, shallow or deep syntactic parsers, and
finally extraction based on a collection of patterns.
However, such a framework is inflexible and expensive in
face of dynamic application needs. Consider a biology-
oriented scenario when the original information extraction
goal is to extract interact- tions among proteins from a
corpus of text. Suppose later on we are interested in
finding gene-disease associations from the same corpus.
Existing approaches would have to develop a new
extraction system specifically for this new extraction goal,
and run that extraction system on the entire corpus from
scratch, which is very expensive. Consider another
application scenario where the extraction goal remains the
same, but an improved named entity recognizer becomes
available. This would also require extraction to be
performed from scratch on the entire corpus. However, we
observe that only a portion of the corpus is affected with
newly recognized entities, as the majority of the entities are
overlaps between the original and the improved recognizers.
Such expensive re-computation should be minimized. This
is particularly true for extraction in the biomedical domain,
where a full processing of all 17 million Medline abstracts
took about more than 36K hours of CPU time using a
single-core CPU with 2-GHz and 2 GB of RAM. In this
case, the Link Grammar parser [3] contributes to a large
portion of the time spent in text processing.
In this demonstration, we propose a new paradigm of infor-
mation extraction in the form of database queries. We present
a general-purpose information extraction system, UniqIE, in
the context of biomedical extraction, which can efficiently
handle diverse extraction needs and keep the extracted
information up- to-date incrementally when new knowledge
becomes available. The insight of UniqIE is that changes in
extraction goals or deployment of improved processing
modules hardly affect all sentences in the entire collection.
Thus we differentiate two phases of processing.
1) Initial Phase: we perform a one-time parse,
entity recognition and tagging (identifying individual
entries as belonging to a class of interest) on the whole
corpus based on current knowledge. The generated
syntactic parse trees and semantic entity tagging of the
processed text is stored in a parse tree database (PTDB).
2) Extraction Phase: Extracting particular kinds of
relations can be done by issuing an appropriate query to
PTDB. As query languages such as XPath and XQuery
are not suitable for extracting linguistic patterns [2], we
design and implement a query language called PTQL for
pattern extraction which effectively achieves diverse IE
goals [6]. To ease the extraction tasks for users, our system
not only allows a user to issue PTQL queries for
extraction, but it can also automatically generate queries
for high-quality extraction based on user input keyword-
based queries and feedback.
There are several advantages of the proposed approach,
which have been demonstrated in our initial experimental
evaluation. First, using database queries instead of
writing individual special-purpose programs, information
extraction becomes generic for diverse applications and
becomes easier for the user. The user can express and
analyze an extraction pattern by issuing a database query.
When a user has a new extraction goal, the user only needs
to write another query on PTDB without developing and
running new programs.
Second, upon new extraction goals, the two-phase approach
(A) Query-specified Extraction: evaluation of PTQL queries
through filtering and translation to SQL queries.
(B) Pseudo- relevance Feedback Query Generation:
generation of PTQL queries based on the common
Novel Database-Centric Framework for Incremental Information
Extraction
K. Keerthana1
Mr. P. Sasikumar2
1
student M.E-CSE 2
M.Tech (Ph.D.) 2
Project Guide- Lecturer
1, 2
Selvam College of Technology, Namakkal
S.P.B.Patel Engineering College, Mehsana, Gujarat

Novel Database-Centric Framework for Incremental Information Extraction
(IJSRD/Vol. 1/Issue 4/2013/0005)
grammatical patterns among the top-ranked sentences
relevant to the user keyword-based queries.
Fig (1): Architecture of UniqIE
Avoids performing the initial phase again, an extremely
expensive phase that has be to performed by existing
approaches.
Third, with the use of databases, UniqIE only needs to
perform extraction incrementally on the sentences that
are affected by an improved module, and thus it is much
more efficient than running the whole extraction
programs from scratch as required by existing systems.
Suppose an improved named entity recognizer that can
discover a more extensive list of protein names becomes
available. We only need to perform a delta extraction on the
database with respect to the newly recognized protein names
using queries.
Indeed, the ability of expressing information extraction
and exploiting database optimizations for the process is also
observed in [7]. While [7] proposes to use Data logs for ex-
tracting facts from “relational tables”, we focus on
extracting meaningful “tables” from “parse trees” of text
documents. Due to the variety of extraction needs, the
existence of hierarchical data structure and the lack of a
relational schema, this involves a new set of technical
challenges as outlined in Section IV.
II. SYSTEM
Figure 1 illustrates the system architecture of our UniqIE
system. The Text Processor performs the Initial Phase for
corpus processing and stores the processed information in
the Parse Tree Database (PTDB). The extraction patterns
over parse trees can be expressed in our proposed parse tree
query language (PTQL). The PTQL query evaluator takes a
PTQL query and transforms it into keyword-based queries
and SQL queries, which are evaluated by the underlying
RDBMS and IR engine. The index builder creates an
inverted index for the corpus as part of the query evaluation
by the IR engine.
The user interface provides two input modes:
query- specified extraction mode and pseudo-relevance-
feedback extraction mode. A user can directly specify
PTQL queries for extraction in query-specified extraction
mode. The user inter face also provides the capability for
users to input a keyword- based query. When a user
keyword query is issued, relevant sentences are retrieved
using an existing IR keyword search engine. With the top-
ranked sentences, their corresponding grammatical
structures are retrieved from PTDB. The PTQL query
generator then uncovers the common grammatical patterns
by considering the parse trees of the top-ranked sentences to
automatically augment the initial keyword-based queries
and generate PTQL queries. Extracted results are
presented to the users once the queries are evaluated.
Furthermore, an explainer module is available to illustrate
the provenance of query results by showing the syntactic
structures of the sentences involved in the extracted results.
This helps the users understand, and enhance their queries
accordingly.
A. Text Parsing and Parse Tree Database (PTDB)
The Text Processor parses Medline abstracts with the Link
Grammar parser [3], and identifies entities in the sentences.
Each document is represented as a hierarchical
representation called the parse tree of a document. A parse
tree is composed of a constituent tree and a linkage. A
constituent tree is a syntactic tree of a sentence with the
nodes represented by part-of-speech tags and leafs
corresponding to words in the sentence. A linkage, on the
other hand, represents the syntactic dependencies (or links)
between pairs of words in a sentence. Each node in the
parse tree has labels and attributes capturing the document
structure (such as title, sections, sentences), part-of-speech
tags, and entity types of corresponding words. Figure 2
shows a sample parse tree of a sentence, where the solid
lines indicate parent-child relationships in the constituent
tree and the dotted lines represent the linkage. Each leaf
node in a parse tree has a value and a tag attribute. The
tag attribute indicates the entity type of a leaf node.
The Parse Tree Database is a relational database for storing
parse trees and semantic types provided by the Text
Processor.
B. Information Extraction Using PTQL Queries
To perform information extraction, we propose a query
language, PTQL, to specify linguistic patterns on parse
trees. An example of a PTQL query is shown in Figure 3. A
PTQL query consists four components delimited by colons:
(i) tree patterns, (ii) a link condition, (iii) a proximity
condition, and (iv) a return expression. A tree pattern
describes the hierarchical structure and the horizontal
order of the nodes in a linguistic extraction pattern. X Path
axis are used for expressing node relationships. In the
example for Figure 3, the tree pattern specifies that there
is a node labeled as S as the root of a sub tree that
contains three nodes represented by variables i1, v and
i2. A link condition describes the linking dependencies
between nodes.
Fig. 2:an example of a parse tree
//S{//?[tag=’GENE’](i1)=>//V[Value=’regu
lates’] (v)=>//?[tag=’GENE’](i2)}: i1 !S v
and
v !O i2 :: i1.value, v.value,

i2.value
Fig. 3: An example of a PTQL query
In the example for Figure3, i1! S v represents that the
node denoted by i1 has to be connected to the node
denoted by v through an S link. In other words, i1 is
the subject and v is the corresponding verb. Similarly, the
link term v! O i 2 indicates that i2 is the object and v is the
corresponding verb. A proximity condition specifies words
that are within a specified word distance in the sentence.
A return expression defines the list of elements to be
returned. In the example, i1.value, v.value,
i2.value indicate to return the bindings to the
variables i1, v, i2 (i.e. two interactors and the
interaction verb) for sentences that satisfy the query. The
parse tree in Figure 2 satisfies the query. The details of the
PTQL query language and its implementation can be
found in [6].
C. Pseudo-relevance Feedback Query Generation
To ease the learning curve in issuing PTQL queries for the
users, UniqIE allows a user to issue simple keyword-based
queries, and automatically generates PTQL queries based
on the user keyword query.
To achieve this, it first performs an initial retrieval from
the inverted index of the corpus with the user keyword
query. Among the top-k% of the retrieved sentences Sk ,
the parse trees of Sk are retrieved from PTDB to find
the common grammatical patterns among Sk . Intuitively,
a sentence that bears the common grammatical patterns
among the top-ranked sentences is likely to be relevant.
Second, for each parse tree of the relevant sentence
UniqIE extracts the sub tree that is rooted at the LCA
(lowest common ancestor) lca of the query terms. Third,
to efficiently compare and find the common patterns,
UniqIE generates m-th l e v e l string encodings for each
sub tree [5]. When m = 0, the string encodes the exact
linguistic pattern in the sub tree, and thus the retrieved
sentences have the exact pattern as the relevant sentences,
potentially with a high precision. With the increase value of
m, the string encodes a more generalized linguistic pattern,
and is likely to retrieve more sentences that lead to a higher
recall with possible compromise on precision. Fourth,
identical m-th level string encodings form clusters of
common grammatical patterns Cm . Finally, a PTQL query
is generated for each of the clusters in Cm .
D. Query Evaluation and Optimization
To evaluate PTQL queries on PTDB, the Query Translator
generates SQL queries from PTQL queries. Efficiency is a
key requirement for query evaluation. One of our opti-
mizations is that for each PTQL query, the Filter module
First generates an keyword-based query to efficiently prune
irrelevant sentences, and then the Query Translator
generates a SQL query equivalent to the PTQL query, and
performs the actual extraction only on relevant sentences.
The keyword- based query captures keywords in the PTQL
query, while the extraction query captures both the structural
patterns and keywords. The keyword-based and SQL
queries are evaluated using an IR engine and a relational
database, respectively. For efficient query processing, the
Index Builder creates an inverted index that indexes
sentences according to the words, named entities and entity
types.
III. DEMONSTRATION
Fig. 4: A screenshot for the UniqIE system showing the
query results that share the same m-th level string
encoding.
What will be shown in the demo? Our web-based demon-
stration , as shown in Figure 4, will illustrate how the
UniqIE system enables generic extraction.
1) Query-specified Extraction. In the demonstration,
the user can input a PTQL query to express an extraction
pattern or select one of the PTQL query examples. We will
show that to perform extraction, a user no longer needs to
write specific extraction programs.
2) Pseudo-relevance Feedback Query Generation.
The user can express extraction patterns in the form of
keyword-based queries. This scenario illustrates the
feasibility of generating PTQL queries from keyword-
based queries through a mechanism inspired by the
pseudo-relevance feedback approach commonly found in
IR. In addition, the user can achieve extraction results for
optimal precision or recall by adjusting the value of m.
3) Two Phase Extraction and Incremental
Evaluation. We will illustrate the efficiency of UniqIE
when new extraction goals or improved processing
components emerge. For instance, as- sume that NER1 is a
currently deployed gene name recognizer, and NER2 is an
improved version NER1 to be adopted by UniqIE. The
user can browse the sentences that are affected, i.e.
sentences with genes that are recognized by NER2 but not
NER1, and vice versa. Then the user can see that the
extraction is incrementally performed on the affected
sentences only, and thus it is very efficient.
4) Provenance of Query Results and Query
Explanation. To help users develop and test their queries,
upon click, the provenance of the query results will be
displayed, which includes the original sentence along with
its parse tree. UniqIE also illustrates the flow of every step
of the query generation and PTQL query evaluation.
IV. DISCUSSION
Significance of Our Approach
The significance of our approach lies in three aspects.

A. Novel Database-Centric Framework for Information
Ex- traction.
Information extraction is traditionally realized by writing
special-purpose programs for each specific extraction goal.
In this demonstration, we will illustrate a new extraction
framework, where extraction is formulated as queries on a
database that stores the parsed data. The benefits of such a
framework for information extraction include: (i)
incremental evaluation is achieved in the presence of new
extraction goals and deployment of improved processing
components; (ii) database query optimization is leveraged
for efficiency.
B. Proven Success of Information Extraction in Biomedical
Domain with Promises to General Domains
The underlying framework of the UniqIE system has been
tested on information extraction of biomedical literature
[5], and performed among the top in the BioNLP’09 shared
task on event extraction [4]. Our two-phase extraction
framework and query generation are not specific to only the
biomedical domain, but can be adapted to information
extraction in general domain.
C. Performing Diverse Extraction Goals without
Training Data
Typical IE systems, such as Snowball [1], adopt the
supervised learning approach that takes annotated data in
generating extraction patterns. However, training data is
scarce and it is known to be expensive to assemble. This
can limit the opportunity for a trained IE system to perform
another extraction goal. Our automated query generation
approach forms PTQL extraction queries by exploiting the
linguistic features of the top-ranked relevant results.
Without the use of training data, our approach is readily
available to extract different kinds of extraction goals. Such
approach serves diverse information needs among different
users.
1) Database Challenges. Our general framework
for two-phase information extraction opens up a lot of new
opportunities and challenges for data management research.
D. Languages for Information Extraction
The parse tree database is complex, and extraction patterns
involve traversals of paths in constituent trees, as well as
links and link types between node pairs. Without user-
defined functions, existing query languages fail to specify
required extraction patterns due to missing axes (XPath,
XQuery) or unable to traverse linkages as a first class citizen
(XPath, XQuery, LPath [2]). The design of query languages
for information extraction on parsed documents demands
investigation.
1) Optimization Challenges for Query Optimization on
Large- scale Data. UniqIE handles 1.5 terabytes of parsed
text data. Thus efficiency and scalability are essential
elements of the system. During prototyping, we found that
directly evaluating SQL queries translated from PTQL
queries was very slow due to the complexity of the
extraction patterns. In UniqIE, we significantly improved the
efficiency by leveraging keyword- based queries for pruning.
However, further query optimization is essential to handle
cases when only a small number of sentences can be filtered
by keyword-based queries.
2) Automated Query Generation. Query generation is
critical so that casual users can specify their information
needs without learning a query language, although our
current attempts of automated query generation already show
promises, many further technical challenges need to be
addressed. For instance, how to strike the balance of
precision and recall when generating PTQL queries that may
generalize the linguistic tree patterns in relevant sentences?
How to estimate the “quality” of the generated PTQL queries
before the execution?
UniqIE presents our attempts in providing a versatile
approach for information extraction. The elegance of our
approach is that unlike typical extraction frameworks, intro-
ducing new knowledge in our framework does not require
the reprocessing of all modules. Simple SQL insert
statements can be issued to store the new entities in PTDB.
We believe that studying fundamental database management
issues on information extraction – a well-known important
problem – opens up a lot of new opportunities and
challenges.
REFERENCES
[1] E. Agichtein and L. Gravano Snowball: extracting
relations from large plain-text collections. In ACM Digital
libraries, 2000.
[2] S. Bird, Y. Chen, et. al. Designing and Evaluating
an XPathDialect for Linguistic Queries. In ICDE, 2006.
[3] D. Grinberg, et. al.. A Robust Parsing Algorithm For
LINK Grammars. CMU-CS-TR-95-125, Pittsburgh, PA,
1995.
[4] J. Hakenberg, et. al. Molecular event extraction
from Link Grammar parse trees. In Proc. of BioNLP’09,
2009.
[5] L. Tari, et. al.. Querying parse tree database of
Medline text to synthesize user-specific biomolecular
networks. In PSB’09, 2009.
[6] P. H. Tu, et. al.. Generalized text extraction from
molecular biology text using parse tree database
querying. TR-08-004, Arizona State University, 2008.
[7] S. Warren, et. al. Declarative IE using datalog with
embedded extraction predicates. In VLDB ’07, 2007.

Novel Database-Centric Framework for Incremental Information Extraction

More Related Content

What's hot (16)

Viewers also liked (13)

Similar to Novel Database-Centric Framework for Incremental Information Extraction (20)

More from ijsrd.com (20)

Recently uploaded (20)

Novel Database-Centric Framework for Incremental Information Extraction