Transcript Finding High-frequent Synonyms of A Domain

Finding High-frequent Synonyms of a Domainspecific Verb in English Sub-language of
MEDLINE Abstracts Using WordNet
Chun Xiao and Dietmar Rösner
Institut für Wissensund Sprachverarbeitung (IWS),
Faculty of Computer Science,
University of Magdeburg,
39016 Magdeburg, Germany
Introduction — MEDLINE Abstract
• MEDLINE®:
– Domain: clinical medicine, biomedicine, biological and
physical sciences;
– Source: articles from over 4,600 journals published
throughout the world;
– Coverage: abstracts are included for about 52% of the
articles.
• PubMed®, an application of UMLS (unified medical
language system), provides links within MEDLINE® to the
full text of 15 clinical medical journals .
– Available at: http://www.ncbi.nlm.nih.gov/PubMed/
Available Resources in the Experiment
• The test corpus consists of 800 MEDLINE
abstracts extracted from the GENIA Corpus
V3.0p and V3.01.
- Available at: http://www-tsujii.is.s.u-tokyo.ac.jp/GENIA/
• WordNet 1.7.1
Extraction of a Specific Relation
•
Inhibitory relation
–
•
Example: Secreted from activated T cells and
macrophages, bone marrow-derived MIP-1
alpha/GOS19 inhibits primitive hematopoietic stem
cells and appears to be involved in the homeostatic
control of stem cell proliferation.
Semantic annotations in the GENIA corpus:
 protein_molecule
 cell_type
High-frequent Verbs in the Test Corpus
5
Synonym Sets (Synsets) of Verb inhibit
• Synset in WordNet
Sense 1
suppress, stamp down, inhibit, subdue, conquer, curb
=> control, hold in, hold, contain, check, curb,
moderate
Sense 2
inhibit
=> restrict, restrain, trammel, limit, bound, confine,
throttle
• Synset in test corpus of MEDLINE abstracts
Inhibit, block, prevent, etc.
Problem
• Occurrences of verbs in the two synsets in the test
corpus of MEDLINE abstracts
– WN-synonyms: suppress (69), limit (16), restrict (5)
– non WN-synonyms: block (124), reduce (119), prevent(53)
• How can WordNet synsets and information from the
corpus be combined to create domain-specific verb
synsets?
7
Three Definitions
• Language unit — a text segment (a sentence,
several sentences, or a paragraph, etc.) that
expresses one semantic topic.
• Core word — the verb, whose synset in the
test corpus is to be found out. E.g., in this test
inhibit is the core word.
• Keyword — the word, whose corresponding
verb base form is the core word. E.g., in this test
inhibitor, inhibiting, and so on are keywords.
Example
We performed an analysis of the mechanisms by which two
PKC inhibitors, Calphostin C and Staurosporine, prevent
the FN-induced IL-1beta response. Both inhibitors blocked
the secretion of IL-1beta protein into the media of
peripheral blood mononuclear cells exposed to FN.
•
•
•
•
•
•

Language unit: two sentences
Core word: inhibit
Keyword: inhibitor (2 times)
Local context: searching window size >=3
Verbs around the first keyword: perform, prevent, block, expose
Verbs around the second keyword: prevent, perform, block, expose
In the following test, the language unit is selected to be the whole
abstract.
Idea Description
• Assumption:
The synonyms of a verb co-occur much more frequently
together with the keywords of the verb than together
with other words in the language unit.
• Method:
Thus the verb chunks around the keywords are
collected, from which the synonyms of the core word
will be selected and filtered, using WordNet synset
information.
- One resource:
WordNet synset information
- The other resource:
Local context information in the test corpus
Distribution of Keywords of inhibit in the Test Corpus
Verbs around the Keywords in the Test Corpus
Method Description I
• Expansion of WordNet Synsets (Si)
– S1 : the verb collection of synonyms of all synonyms of
the core word;
– S2 : the verb collection of synonyms of all verbs in S1;
– …
• Expansion of Stoplist (STOPk)
– STOP0: manually select 15 stop-verbs from the highfrequent verbs in the test corpus (e.g., suggest, indicate,
including the high-frequent antonyms of the core word);
– STOP1: the verb collection of synonyms of all verbs in
STOP0;
– …
Method Description II
• Verb list from the corpus (Vj)
Verbs around the keywords in a local context of
searching window size of j are collected.
• Synonym candidate list (Sg)
If a verb is in Vj and also in Si, but not in
STOPk, then add it to Sg.
Evaluation
• Golden standard list (SG)
– A manually created synonym list, which is extracted
from the test corpus.
– Consist of 10 verbs with the most frequent occurrences,
in which 3 verbs come directly from the WordNet
synset of “inhibit”, the rest 7 verbs come from its
hypernym set or the expanded list of its synonyms.
• Recall & Precision
15
Result
 60% recall of SG <=> 93.05% occurrences in the test corpus
Conclusions and Future Work
• Conclusions
–
–
–
–
English sublanguage of MEDLINE abstract;
The core word and its keywords were high-frequent;
Multiword verb structures were not considered yet;
Balance between recall and precision: expansion of Si and
STOPk should be limited.
• Future works
–
–
–
–
Consideration of other WordNet information besides synsets;
Automatic creation of stoplists;
Extraction of multiword verb structures;
Utilization of syntactic information.
Thanks!
Looking forward to your questions!
20
Possible Errors
• Errors of POS tags between
Adjectives <=> Past participles
• Errors of manual works when selecting stop-verbs
21
Question or Hope
Can WordNet provide the possibility for accessing
multiword expressions?
22