PA5-2_iconf08.doc.doc

Approaches to Semantic Markup for Natural Heritage
Literature
Hong Cui
School of Information Resources and Library Science, University of Arizona.
hongcui@email.arizona.edu

ABSTRACT structured format (that allows the encoding of biological
The theme of this paper is the application of automated semantic characters and their states/values) also support faceted browsing
markup techniques on natural heritage literature to address and cross collection (i.e. federated) retrieval.
information needs of taxonomists. Two machine learning based The specimen identification task is traditionally performed by
techniques (supervised and unsupervised machine learning) are using taxonomic keys, which are decision trees (upside-down
discussed and compared on a real world corpus. A prototype trees with the root node on top) with contrasting statements
application that supports batch and online modes of converting about a biological character in each node and an identification at
free text documents to XML format is described. the end of each branch. It is worth noting that research has
shown that full text information retrieval techniques are not
Topics effective for specimen identification tasks [8]. The creation of a
Information organization, Information management, Information good taxonomic key for a group of taxa relies on the
technology and services completeness of the descriptions and exhaustive search and
comparison of the descriptions of related organisms. Any
Keywords taxonomist can testify that neither of the two conditions is
Semantic markup, natural heritage literature, taxonomic supported well by free-text descriptions: it is easy for an author
descriptions, biodiversity information, biodiversity informatics, to leave out some characters, it is tedious to repeat shared
supervised machine learning, unsupervised machine learning, characteristics again and again in many descriptions, and on top
XML of these, it is also very time-consuming to search for missing
1. INTRODUCTION characters and compare many free-text descriptions at one time.
A structured format can alleviate these difficulties because it
Since over two hundred years ago, the work has started to
allows an information system to remind an author of the missing
systematically document various living organisms, their habitats,
elements and to auto-fill the shared characteristics. It is also
geographic distribution, classification, and identification keys.
possible for computers to shoulder some work involved in
Current projects of such include the Flora of North American [6]
comparing descriptions, thanks to the structures that are made
(FNA) and the Flora of China [7](FoC). As a result, a wealth of
explicit to computers, enabling more meaningful comparisons
natural heritage literature has been produced. However, the vast
element by element, rather than blind keyword matching. If the
majority of natural heritage literature is still in print format. That
descriptions are marked-up in very fine details, for example, to
in electronic format is mainly in html, pdf, or doc formats, for
the level of characters and character states, it is even possible to
example, volumes of FNA and FoC are both in print and html
automate many aspects of the generation of taxonomic keys.
formats. Only a small percentage of such information is stored in
a structured digital format such as relational database or XML A sustainable way to transform large sets of free text
format [2]. To construct the “global commons” of natural descriptions to structured formats should not rely on grant
heritage literature, as envisioned by the Biodiversity Heritage moneys and graduate assistants’ manual labor. Automatic
Library (an international collaboration among major natural semantic markup techniques must be explored to 1) convert
heritage institutes, http://www.bhl.si.edu/), a number of newly digitized full text documents into a structured format and
digitization projects are being planned or carried out at different 2) allow taxonomists to convert a freshly composed description
levels. Related standard organizations, such as the Biodiversity into the structured format by clicking a button.
Information Standards, are working intensively on a number of
biodiversity information content/encoding standards, including 2. RELATED RESEARCH
the Structured Descriptive Data (SDD, an XML application), to The majority of work on structuring taxonomic descriptions
meet the immediate needs of digitization projects. For a review focuses on the paragraphs with limited semantics, such as taxon
of XML schemas for biodiversity literature, see [14]. names and ranks etc. [9, 10, 13]. Far less research focused on
cue-poor yet semantic-rich sections, for example morphological
Natural heritage communities have come to the consensus that description sections, largely due to the lack of consistency in
the reformatting of unstructured biodiversity literature, description contents, as shown by Lydon et al.[11]. Lydon et al
especially taxonomic descriptions of living organisms, into a manually compared descriptions of the same five species from
machine readable structured format is necessary for the long six different floras and found large variations in terms of
term and effective management of the large body of literature. information content and presentation style: only 9% of
This reformatting benefits both of the two major usages of the information was exactly the same, over 55% of information was
information, namely organism identification and information from a single flora, and around 1% of information from different
retrieval. For the attendees of the iConference, structured floras contradicted with each other. Earlier works [16, 1, 17]
formats enable more accurate retrieval is probably a well- using syntactic parsing methods to extract information to
accepted fact that calls for no more elaboration. We should populate relational databases or to mark up plant descriptions in
nonetheless note that the organism descriptions in a good an XML format have focused on a single collection. Syntactic

parsing methods require handcraft grammar rules and extensive typically reaches a performance plateau. When the plateau is
lexicons to guide the parsing process. Grammar rules can be reached, the system is said to be trained, as more training
collection-dependent, therefore the usefulness of a parser for examples will not bring in any performance gain. Unsupervised
different collections can be limited. A specialized XML editor learning techniques do away with training examples; they derive
for biosystematics literature has also been built [14], however, it knowledge from the documents themselves by exploring the
requires the users to come up with regular expression rules to various regularities embedded in the documents. For either
mark up documents. [18] extracted plant characters from technique, there has to be some source of knowledge that can be
descriptions of five species taken from six floras, using a hand- exploited and the learning algorithm has to be able to discover
made gazetteer as a lookup list to link extracted terms with their something that is previously unknown to it.
semantic labels. It remains to be seen to what extent the manual
knowledge engineering scales with larger data sets. On a much 3.1 The Supervised Learning Algorithm
larger scale, [4] took a supervised machine learning approach to Aforementioned MARTT system is a supervised learning system
address the issues of inter-collection variations and involving two phases: the training phase and the mark-up phase.
automatically marked up descriptions of over ten thousands In the training phase, the algorithm learns markup rules/models
species from three floras down to sentence/clause level. By from training examples, while in the mark-up phase the trained
reusing domain knowledge learned from other collections in the algorithm uses the learned rules to mark up new descriptions.
past, their system (called MARTT) achieved better performance The algorithm is illustrated in List 1. Details can be found in [3],
on a new collection, without any re-adjustment or fine tuning of where the algorithm was labeled as SCCP.
the system. The marked-up taxonomic descriptions are currently
available for public access at The Training Phase
http://research.sbs.arizona.edu/gs/cgi-bin/library . Name: Training
Input: Training Examples (TEs, XML documents with nested
3. TWO MACHINE LEARNING elements)
TECHNIQUES Output: Populated Learning Hierarchy (LH)
The machine learning approach has a number of advantages over Algorithm:
other existing approaches. This approach does not require the Initialize the empty LH with the root node “description”
markup or extraction rules to be provided to the system. Foreach TE in TEs
Composing the rules by hand is difficult even for trained Extract elements from TE level by level to save
knowledge engineers or computer scientists, because one cannot element contents in the corresponding nodes in LH.
easily tell the overall effect of a rule on a large collection—it If a node does not exist in the hierarchy of LH, create
may work very well with a small portion of descriptions, but a new sibling or child node in LH to accommodate the new
results in a net negative effect if applied on the entire collection. element at the desired level.
If the goal is to equip domain experts, in this case, taxonomists, End Foreach
with the markup tool, this requirement is simply not reasonable.
Taking advantage of ever increasing computational power, the Foreach node in LH
machine learning approach, on the contrary, strives to find CreateModel(the element contents saved at the node).
global optimum rules automatically. End Foreach
Done
A properly designed learning system also promotes portability
of the technique. It is desirable for a system to be able to process Name:CreateModel
different collections of documents of the same or similar Input: element contents of a node or element
domain, with little or no reengineering. A learning system that Output: learned model
learns and makes use of domain knowledge is more portable Algorithm:
than a system that learns presentation cues, because it is more N-Grams = Word-level unigrams, 2-grams, and 3-grams
likely for the documents to share domain knowledge than the extracted from the leading words of element contents.
presentation styles, such as font, size, color, or html tags. Since a Foreach N-Gram in N-Grams
machine learning system learns markup rules from document support = the number of instances of the element
collections, the markup rules are always up to date and reflect containing the N-Gram / the total number of instances of the
the characteristics of the current collection. In practice, this element
works better than relying on hand-crafted rules that may not fit confidence = the occurrence of the N-Gram in the
the current task well. element /the total occurrence of the N-Gram
Save the rule “N-Gram ->element (confidence,
There are two main types of inductive learning techniques: support)”
supervised and unsupervised learning techniques. Supervised End Foreach
learning techniques require training examples, from which the Done
learning system derives knowledge, and against which the
learning system verifies its learning. Training examples for the
MARTT system were marked-up descriptions manually The Mark-up Phase
prepared. The number of training examples required by a system
varies, depending on the learning algorithm and the Name: Mark-up
heterogeneity of the data collection. Initially the performance of Input: LH, to-be-marked document collection C
a learning system increases with the number of training Output: marked-up document collection C
examples. At a certain point (typically before 100% accuracy) it Algorithm:

Foreach document in C frequent features. In other words, is it possible for an
MarkupDocument(LH, document) unsupervised learning technique to achieve comparable
Marked-up document = Read the complete, marked- performance as the supervised technique? To what extent the
up document off the LH regularities in the text can replace the knowledge manifested in
End Foreach the training examples?
Done
3.2 The Unsupervised Learning Algorithm
Name: MarkupDocument We started to answer these questions by investigating a
Input: LH, to-be-marked document D bootstrapping approach [12]. This markup approach is a bottom-
Output: LH with marked up segments of D up approach, established on the literary warrant principle. It
Algorithm: exploits the syntactic cues commonly found in domain corpora.
If LH contains one leaf node and one leaf node only Such cues include: sentences typically start with a subject, the
Mark up D with the name of the leaf node subjects are typically nouns or noun phrases, periods are used to
Save the marked-up D in the node. end a sentence, while commas or semicolons are used to set off
Done clauses [5]. These cues were used in the unsupervised learning
Else algorithm that marked up to the level comparable to that of
Read D into the root node of LH MARTT. List 2 shows the main steps of the algorithm.
The root node marks up segments (sentences/clauses)
Name: Unsupervised Learning Algorithm
of D with the names of its child nodes by selecting the rules with
Input: To be marked-up documents in text format
the highest confidence and support.
Output: Marked-up documents in well-formed XML format
Save marked-up segments in the node.
Algorithm:
Pass marked segments (S) of D to their corresponding
Use the following simple heuristic rule to find a set of words
child node (N)
(Nouns) used as nouns in the documents.
MarkupDocument(N,S)
Heuristic rule: A word is used as a noun iff the
documents contain its singular and plural forms but no verb
forms.
List 1: Supervised Learning Algorithm
Use English grammar rules to distinguish plural forms from
singular forms.
Note the MarkupDocument function is a recursive function that SegmentSet = Segment documents into sentences/clauses (i.e.
starts from the root node of LH (a tree with a root on the top and segments) at the punctuation marks: .,;:
many leaf nodes at the bottom) and ends at a leaf node. At each SegmentSet1 = segments starting with a plural noun
call, the function works on a successively smaller part of the LH SegmentSet2 = segments that do not belong to SegmentSet1
until all branches of the tree is traversed and all leaf nodes are Foreach segment in SegmentSet1
reached. Extract the word after the first plural noun and save
the word in BoundaryWords
Although the training and mark-up phases of MARTT were End Foreach
automatic and required no human intervention, the training (Nouns, BoundaryWords) = Bootstrap(Nouns, BoundaryWords,
examples came at a cost. MARTT used close to 700 taxonomic SegmentSet2)
descriptions for training, yet more examples could still mean a Foreach segment in SegmentSet
better performance as a performance plateau was not reached If the segment’s subject main noun can be identified
with that many training examples. Even thought it was just a tag = the subject
small fraction of total documents processed (700 out of 17000, Else
or 4%), it was time consuming and error-prone to prepare tag = “unknown”
hundreds of descriptions according to the predefined XML End If
schema (http://publish.uwo.ca/~hcui7/research/xmlschema.xsd). If the segment ends with any punctuation mark but a
On the other hand, the predefined XML schema did not cover all period
organs described in the descriptions, because a complete schema tag = tag + “_block”
can be created only after one reads through the entire collections End If
of documents, due to the inherent diversity of living organisms Use tag to mark up the segment
and authorships (e.g. 600+ authors contributed to Flora of China, End Foreach
800+ to Flora of North America). All uncovered organs were End
marked “other-features”. “Other features” included many rarely
occurred organs, which are often the most valuable information Name: Bootstrap
for specimen identification. Treating them indistinguishably as Input: Nouns, BoundaryWords, SegmentSet2
“other features” is not a satisfactory solution. Limited schema Output: A bigger set of Nouns, a bigger set of BoundaryWords
coverage is a shortcoming shared by all existing systems Algorithm:
reviewed in this article. Flag_NewDiscovery = false;
The shortcomings of the existing automated markup systems for
biodiversity literature prompted us to ask this research question: Foreach segment in SegmentsSet2
To what extent the inherent textual cues in the literature can be Generate a subject pattern based on the up to three
exploited to produce XML-based markups that are not biased for leading words. For example, given a segment “Sporangiaster
rare.”, if “rare” is in BoundaryWords but Sporangiaster is in

neither BoundaryWords nor Nouns, the segment’s subject
pattern is “?B”.
End Foreach

Sort (Patterns) //patterns with fewer ? are ranked higher
Do
Foreach pattern in Patterns
If (new nouns or boundary words are discovered)
/*note: depending on the pattern. The
previous example pattern “?B” would result in a new discovery:
Sporangiaster is a noun. A pattern such as “??B” would not
result in a new discovery, because it could be a “NBB”, a
“NNB”, or a “MNB” (M for modifier)*/
Flag_NewDiscovery = true
update Nouns or BoundaryWords.
End If Figure 1: The First Screen Allows
End Foreach If the user chooses to use either of the algorithms to mark up the
While Flag_NewDiscovery collection, she will provide necessary information and instruct
the application to “Learn to Mark up”. Since the learning and
List 2: The Unsupervised Learning Algorithm markup process can take a long time depending on the size of
the collection, the interface informs the user the progress (Figure
2). If the documents in the “Job Folder” (i.e. the to-be-marked
The details of the algorithm are still being refined, but the main collection) already contain XML tags, the application will
idea is to start with a small set of Nouns and a small set of automatically assume the user’s intention is to test the
BoundaryWords (those are learned from SegmentSet1) to performance of the learning algorithm. It will compare the given
discover more Nouns and BoundaryWords, using the basic tags with those generated by the computer and compute
assumption “subjects (nouns) are followed by boundary words ”. performance scores(Figure 2). The performance scores are saved
The bootstrap function runs in iterations. In each interaction, it in a file so the user can compare the performances of different
tries to make new discoveries and replaces some “?” in the runs. Note once the learning is started, the “Learn to Mark up”
subject patterns with a “B” or “N”, until it fails to make any new button is disabled, but the user is still in control where she would
discovery in an iteration. The learning process stops there. like the application to go. Impatient user may cancel the lengthy
The unsupervised learning algorithm has been tested on a small learning progress and go either to the learning setting screen
collection of algae descriptions (120 descriptions), two modest (Figure 1) to choose a different learning technique, or to the
collections of plant descriptions (400-633 descriptions), and a main interface to use other functions.
modestly large collection of plant descriptions (2000+ When the learning and mark up process is done, the user is led
descriptions). The performances on larger collections were to the main interface, where she can verify the correctness of
better. A later section compares the unsupervised and machine-generated markup (using either the supervised or
supervised learning algorithm on the 630 FNA descriptions. unsupervised approach) by using the “Batch-Mode” functions
(Figure 3) and make any corrections by invoking the popup
4. THE APPLICATION schema menu.
We implemented a prototype of an application that supports both
of the learning techniques. The user-friendliness of the The application maintains the list of files to be examined so the
application lies not only on the employment of sound interface next time the application starts, the user can continue with the
design principles, but more importantly on the reasonable remaining of the files. The application automatically records any
division of the job between domain experts and the computer. errors the user corrects, but also allows the user to add any error
The application does not assume the user has any knowledge manually by using “Record an Error” function in “Tools”. A
engineering skills, such as crafting regular expression patterns module will be implemented in the future to pass along the
for markup, but trusts the user has the ability of identifying identified errors to the learning algorithms so the algorithms can
misplaced tags. A few screenshots are included here to illustrate revise its markup rules. In the “Composing-Mode”, the user is
the design of the prototype. The application assumes the user presented a text editor window where she can compose a
either has folders of documents (i.e. batch processing) or will description, which will be marked up in XML format after the
compose a description to mark up. The first screen the user sees user clicks on the “Mark up” button. This one-click markup
once the application is started allows the user to either run a relies on the knowledge already learned by MARTT. The user
learning algorithm on a folder of documents or to skip the can check for and correct errors in the composing mode in a
learning step and go directly to the main interface if the user similar manner as in the batch mode.
wishes to access other functions (Figure 1). Note for supervised
learning, the user needs to provide training examples. The main
interface has functions that help with the preparation of training
examples (Figure 3).

selecting appropriate elements from the popup schema menu. Or
the user can use the documents marked-up by the unsupervised
algorithm as the training examples for the supervised algorithm.
In short, the application supports the preparation of training
examples, two different learning techniques for marking up
documents, two different processing modes, the verification and
correction of marked up documents, and feedback to the
learning algorithms, all in a rather user-friendly manner. An
earlier prototype was used by two non-computer science
undergraduates and seemed to be straight-forward and easy to
use. However, being a prototype, there is still room for
improvements.

5. COMPARISONS
We ran a simple experiment to demonstrate the differences
between the two learning techniques when applied to semantic
markup of taxonomic descriptions. We used 633 descriptions
from Flora of North America, manually marked-up according to
the aforementioned schema. For the supervised learning
algorithm, the descriptions were evenly divided into two sets:
training and test sets For the unsupervised learning algorithm,
since no training was needed, all descriptions were involved in
the bootstrapping.
Neither algorithms were optimized for speed at this time, but it
took 2.3 times as long for the supervised learning algorithm to
mark up 316 descriptions (28.45 mins) as for the unsupervised
learning algorithm to mark up 633 descriptions (12.35 mins) on
a Gateway 1.20 GHz Internal Core 2 CPU laptop, with the Java
heap size set to 512 MB. The supervised learning algorithm put
96% of all words in the 316 descriptions in correct tags. The
unsupervised learning marked 8400 of the 8557 sentences
correctly (98%) and identified good tags for 3765 of the 3873
unique subjects. It marked 48 sentences (37 unique subjects)
Figure 2: Learning Progress incorrectly and couldn’t make a decision on another 109
sentences (71 unique subjects).
Comparing a description for ARISTOLOCHIACEAE
ARISTOLOCHIA serpentaria, marked up using different
techniques (Table 1), we see the effects of the learning
algorithms. The unsupervised learning produces well-formed
XML documents and the tags are generated and applied as the
description contents warrant, while supervised learning assumes
a generic XML schema no matter the description contents. The
effects of this distinction are shown in different ways: the
unsupervised learning generally produced more detailed markup,
especially in the leaf description, than the supervised
learning; the unsupervised learning explicitly marked
organs/structures hidden in “other features” element produced
by the supervised learning; although not shown in this example,
the unsupervised learning is less constrained in dealing with
uncommon organization of descriptive contents, for example,
describing multiple organs/structures (“leaves and
Figure 3: The Main Interface inflorescence”) in one sentence (There were 63 sentences with
compound subjects in the 630 FNA descriptions). On the other
As mentioned earlier, the interface also supports the preparation hand, the XML schema used in supervised learning carries
of training examples. The user simply opens a list of examples, domain knowledge, although implicitly but allowing the
highlights certain text segments, and tags the segments by translations of concepts at different levels. In this example, it
marked up herbs as a type of plant habit and life style, and in

Supervised Markup Unsupervised Markup
<description> <description>
<plant-habit-and-life-style> <herb>Herbs , erect to decumbent , to 0 . 6 m .</herb>
<phls-general>Herbs, erect to decumbent, to 0.6 m.</phls- <young_stem>Young stem ridged , glabrous to hispid
general> .</young_stem>
</plant-habit-and-life-style>
<leaf_block>
<leaf>Leaves :</leaf>
<petiole>petiole 0 . 5 - 3 . 5 cm .</petiole>
</leaf_block>
<blade_block><leaf_blade>Leaf blade lanceolate to ovate , 5 - 15 *
1 - 5 cm , base truncate to cordate , sinus depth 0 - 1 . 5 cm , apex
acute to acuminate ;</leaf_blade>
<surface>surfaces abaxially glabrous or hispid ;</surface>
<venation>venation pinnate .</venation>
</blade_block>
<stems>
<stem-general>Young stem ridged, glabrous to hispid.</stem-
general>
</stems>
<leaves>
<leaf-general>Leaves: petiole 0.5-3.5 cm.</leaf-general>
<leaf-blade>Leaf blade lanceolate to ovate, 5-15 Ã? 1-5 cm,
base truncate to cordate, sinus depth 0-1.5 cm, apex acute to
acuminate; surfaces abaxially glabrous or hispid; venation
pinnate.</leaf-blade> <inflorescence_block>
</leaves> <inflorescence>Inflorescences from base of stem , an additional
flower in axil of stem leaf , racemes ;</inflorescence>
<peduncle>peduncle bracteolate , to 1 . 5 cm ;</peduncle>
<bracteole>bracteoles lanceolate , to 3 mm .</bracteole>
</inflorescence_block>
<flowers>
<inflorescence-general>Inflorescences from base of stem, an <flower_block>
additional flower in axil of stem leaf, racemes;</inflorescence- <flower>Flowers :</flower>
general> <calyx>calyx brown_purple , bent ;</calyx>
<peduncle>peduncle bracteolate, to 1.5 cm;</peduncle> <utricle>utricle pendent , pear_shaped to ovoid , 0 . 5 - 5 cm
<bracteole>bracteoles lanceolate, to 3 mm.</bracteole> ;</utricle>
<syrinx>syrinx present , ringlike , 1 mm , oblique ;</syrinx>
<tube>tube bent , cylindric , 1 cm ;</tube>
<annulu>annulus smooth ;</annulu>
<limb>limb purplish brown , 3 - lobed , lobes 0 . 5 * 0 . 5 cm ,
glabrous ;</limb>
<gynostemium>gynostemium 3 - lobed , globose to crown_shaped ,
1 . 5 mm ;</gynostemium>
<anther>anthers 6 ;</anther>
<ovary>ovary 3 - locular , to 1 . 5 cm .</ovary>
</flower_block>
<flower-general>Flowers: calyx brown-purple, bent;</flower-
general>
<other-features>utricle pendent, pear-shaped to ovoid, 0.5-5
cm; syrinx present, ringlike, 1 mm, oblique; tube bent, cylindric,
1 cm; annulus smooth; limb purplish brown, 3-lobed, lobes 0.5
Ã? 0.5 cm, glabrous; gynostemium 3-lobed, globose to crown-
shaped, 1.5 mm; </other-features>
<anther>anthers 6;</anther>
<ovary>ovary 3-locular, to 1.5 cm.</ovary>
</flowers>

<fruits> <capsule_block>
<fruit-general>Capsule globose, 0.8-2 Ã? 1-2 cm, dehiscence <capsule>Capsule globose , 0 . 8 - 2 * 1 - 2 cm , dehiscence basipetal
basipetal;</fruit-general> ;</capsule>
<valve>valves 6;</valve> <valve>valves 6 ;</valve>
<septum>septa absent.</septum> <septum>septa absent .</septum>
</fruits> </capsule_block>
<seeds> <seed>Seeds rounded , ovate , 0 . 5 * 0 . 4 cm .</seed>
<seed-general>Seeds rounded, ovate, 0.5 Ã? 0.4 cm.</seed-
general>
</seeds>
<chromosomes>2n = 28.</chromosomes> <2n>2n = 28 .</2n>
</description> </description>

effect translated <herb> to its broader concept <plant-habit-and- 7. REFERENCES
life-style>. Similarly it took <capsule> as a type of <fruit>, and
<2n> <chromosome> counts. [1]. Abascal, R. & Sánchez. (1999) X-tract: Structure Extraction
from Botanical Textual Descriptions. Proceeding of the
There were also 50 instances of organ/structure descriptions String Processing & Information Retrieval Symposium and
marked as “other features” by the supervised algorithm, including International Workshop on Groupware, SPIRE/CRIWG,
those about staminodial ring, utricle, 1st leaves, glandular hairs, pp2-7.
intravaginal squamules, pigment cells, stomates, ring meristem, [2]. Blum, S. D. (2000). An Overview of Biodiversity Informatics.
ostiole, air chambers, leaf primordia, terminal leaflet, abaxial Accessed 1/7/2008.
ridges, stranded plants, hastula, cells, sporangiasters, staminate http://www.calacademy.org/research/informatics/sblum/pub/b
scales, and cystoliths, totally 19 organs/structures. Unsupervised iodiv_informatics.html
learning, on the other hand, discovered good tags for 14 of these
structures and marked 40 of these instances with correct tags. [3]. Cui, H. (2005a). MARTT: Using knowledge based approach
to automatically mark up plant taxonomic descriptions with
6. DISCUSSIONS XML. Proceedings of the Annual Meeting of American
What have we learned from the experiment on the randomly Association of Information and Technology. Oct 28-Nov 2.
selected 633 FNA descriptions? First we learned that the 2005 Charlotte, North Carolina, USA.
documents from the real world will contain information that is not [4]. Cui, H., & Heidorn, P.B. (2007). The reusability of induced
covered by a predefined XML schema. In the 633 descriptions (a knowledge for the automatic semantic markup of taxonomic
tiny portion of FNA, new volumes of which are still being descriptions. Journal of the American Society for Information
published), 19 organs/structures were not found in the XML Science and Technology. 58(1). 133-149.
schema. In addition, there are also 63 descriptive sentences with [5]. FNA Editorial Committee (2006). Flora of North America
compound subjects, which cannot be tagged as any single North of Mexico Guide for Contributors. Accessed July 10,
organ/structure. We believe a semantic schema, let it be an XML 2007 from http://www.fna.org/FNA/Guide/guide_2006.pdf.
schema or an ontology, is necessary for any semantic markup task [6]. FNA Flora of North America Editorial Committee. (Eds.).
to ensure consistent interpretation. The top-down approach to the (1993-). Flora of North America North of Mexico. Accessed
creation of such a schema, however, does not seem to be a viable July 10, 2007 from http://www.fna.org/
approach. In order to discover all organs/structures described in a
collection, we believe a bottom-up approach based on the literature [7]. FOC Flora of China Editorial Committee. (Eds.). (1994-).
warrant principle such as the unsupervised learning method Flora of China. Beijing/St. Louis: Science Press/Missouri
described in this paper is more feasible. To say the least, the Botanical Garden Press. Accessed July 10, 2007 from
unsupervised learning method can be used to generate element http://flora.huh.harvard.edu/china/
candidates and to identify difficult cases (“unknown”) for the [8]. Heidorn, P.B. & Cui, H. (2000). The Interaction of Result Set
human experts to base their work on. Display Dimensionality and Cognitive Factors in Information
Second we saw that the strength and weakness of the two Retrieval Systems. Proceedings of the Annual Meeting of the
techniques are quite evident. The unsupervised technique is more American Society for Information Science, ASIS 2002, pp.
flexible, more efficient, and capable of uncover new/rare 258-270.
organs/structures. But it has the shortcoming of unable to associate [9]. Kirkup, D., Malcolm, P., Christian, G., & Paton, A. (2005).
concepts with their broader concepts. For example, it marks Towards a digital African Flora. Taxon, 54(2). 457-466.
different fruit types explicitly yet unable to tell that they share one [10].Koning, D., Sarkar, I.N., & Moritz, T (2005). TaxonGrad:
broader concept “fruit”. We often assume that the semantic Extracting Taxonomic Names from Text. Biodiversity
relationships among domain concepts have long been codified in a Informatics. 2. 79-82.
thesaurus or something similar. While such assumptions are often [11].Lydon, S.J., Wood, M. M., Huxley, R., & Sutton, D.(2003).
unfounded, even if they are, we still face the issue of discrepancies Data Patterns in Multiple Botanical Descriptions: implications
among different thesauri: for example, if you trace the hypernymy for automatic processing of legacy data. Systematics and
relation of the word nut in WordNet, you will find nut is a seed. Biodiversity 1(2). 151-157.
However, in the Oxford Plant Characters thesaurus, nut’s parent
term is fruit. In WordNet, a capsule is a seed vessel, but in Oxford [12].Riloff, E. & Jones, R. (1999). Learning Dictionaries for
it is a fruit. Fortunately, the unsupervised learning algorithm Information Extraction by Multi-Level Bootstrapping.
reduced the workload for human experts to make situated Proceedings of the 16th National Conference on Artificial
associations among concepts from reading X sentences to examine Intelligence. pp. 474-479
Y distinct tags. In case of the 633 FNA descriptions, X=8557, [13].Sautter, G., Agosti, D., & Böhm, K. (2006). A Combining
Y=280. Approach to Find All Taxon Names(FAT). Biodiversity
Informatics. 3, 46-58.
This identifies an additional component to be implemented as a
part of the MARTT Interface application. In addition to support [14].Sautter, G., Agosti, D., & Böhm, K. (2007). Semi-Automated
supervised and unsupervised learning, a module is needed to allow XML Markup of Biosystematics Legacy Literature with the
a human expert to establish semantic relationships among existing GoldenGATE Editor, Proceedings of PSB 2007, Wailea, HI,
or introduced concepts. This is what we will do next. USA. Accessed July 10, 2007 from
http://psb.stanford.edu/psb-
online/proceedings/psb07/sautter.pdf

[15].Sautter, G., Böhm, K., & Agosti, D. (2007). A Quantitative [18].Wood, M., Lydon, S., Tablan, V., Maynard, D. &
Comparison of XML Schemas for Taxonomic Publications. Cunningham, H. (2004). Populating a database from parallel
Biodiversity Informatics. 4, 1-13. texts using ontology-based information extraction. In
[16].Taylor, A.(1995). Extracting Knowledge from Biological Meziane, F. and M_etais, E., (editors) Proceedings of Natural
Descriptions. Proceedings of 2nd International Conference on Language Processing and Information Systems, 9th
Building and Sharing Very Large-Scale Knowledge Bases. pp. International Conference on Applications of Natural
114-119. Languages to Information Systems. pp.254-264.
[17].Vanel, J.-M. (2004). Worldwide Botanical Knowledge Base.
Accessed July 5, 2007 from http://wwbota.free.fr/.

PA5-2_iconf08.doc.doc

More Related Content

What's hot (11)

Similar to PA5-2_iconf08.doc.doc (20)

More from butest (20)

PA5-2_iconf08.doc.doc