clustering related terms with definitions
Post on 21-Jan-2016
24 Views
Preview:
DESCRIPTION
TRANSCRIPT
LREC 2008 Marrakech 1
Clustering Related Terms with Clustering Related Terms with DefinitionsDefinitions
Scott Piao, John McNaught and Sophia Ananiadou
{scott.piao,john.mcnaught,sophia.ananiadou}@manchester.ac.uk
National Centre for Text MiningSchool of Computer ScienceThe University of Manchester
LREC 2008 Marrakech 2
Outline of talk
• Task: match related terms of ontology.• Approach: detect and cluster related terms
based on definitions.• Implementation: definition matching and term
clustering, user interface.• Evaluation on GO terms.• Conclusion.
LREC 2008 Marrakech 3
Task: matching terms for ontology enrichment
• matching similar or related terms/expressions is important task in NLP and Text Mining applications.
• Ontology term matching is also closely related to ontology enrichment.
• In the EU BOOTSTrep Project, some techniques have been tested for ontology entities matching and alignment.
• Our work focuses on testing and evaluating a text matching tool for identifying related ontology terms with their definitions.
LREC 2008 Marrakech 4
Definitions of term definitions
• Ontology terms, such as GO (Gene Ontology) terms, often contain detailed definitions:.– id: GO:0000124– name: SAGA complex– def: "A large multiprotein complex that possesses histone
acetyltransferase and is involved in regulation of transcription. The budding yeast complex includes Gcn5p, several proteins of the Spt and Ada families, and several TBP-associate proteins (TAFs); analogous complexes in other species have analogous compositions, and usually contain homologs of the yeast proteins.“
– id: GO:0005671– name: Ada2/Gcn5/Ada3 transcription activator complex– def: "A multiprotein complex that possesses histone acetyltransferase
and is involved in regulation of transcription. The budding yeast complex includes Gcn5p, two proteins of the Ada family, and two TBP-associate proteins (TAFs); analogous complexes in other species have analogous compositions, and usually contain homologs of the yeast proteins."
LREC 2008 Marrakech 5
Our approach to the issue
• The definitions can provide a fundamental information source for detecting relations between terms.
• lexicon definitions have been previously used for analyzing relations between words/terms (Castillo et al., 2003).
• We assume text matching tools can be used to detect related terms based on the definitions.
LREC 2008 Marrakech 6
A tool for clustering related texts
• Align similar sentences between texts.
• Measure the distances between texts based on the aligned sentences.
• Cluster similar texts based on a distance matrix.
LREC 2008 Marrakech 7
Metrics for pairwise text comparison
2m
llpsd ngsw
21
2
mm
lldc
ngsw
ngsw
ng
ll
lpsng
,
(δ1=0.85,δ2=0.05,δ3=0.1),psngdcpsdws 321
l
wslsd ii
)( (0 <= d <=
1).
For further details, see the paper.
LREC 2008 Marrakech 8
An effective algorithm text comparisonCited from Clough et al. (2002)
LREC 2008 Marrakech 9
Clustering texts
• Using the text comparison tool, produce distance matrix matrix elements: eij =1 – dij, (0<=eij<=1)
• Error Sum of Squares (ESS) hierarchical clustering
clustersclusterswithin
p
kkik xxESS
1
2
LREC 2008 Marrakech 10
Sample of cluster tree
{layer=9 {layer=10 {layer=11 {layer=12 GO:0009897 GO:0010339 } {layer=12 GO:0010282 } } {layer=11 {layer=12 GO:0045284 } {layer=12 GO:0045293 } } } {layer=10 {layer=11 {layer=12 GO:0017117 GO:0033202 } {layer=12 GO:0017119 } }
LREC 2008 Marrakech 11
A package for definition comparison andterm clustering
pairwise definitions
comparison
term clusterer
userinterface
check update
synonym lexicon
extended Porter’s stemmer
distancematrix
clusters
termdatabase
LREC 2008 Marrakech 12
User interface for checking and updating terms
LREC 2008 Marrakech 13
Evaluation
• The text comparison and clustering components are evaluated on a set of GO terms as test data.
• In the evaluation, we consider GO terms to be related if they:– share a parent term within three layers of ancestor trees via
IS_A relation, or– have direct parent/child relations (e.g. X is_a Y), or– have direct part-of relations (e.g. X is part of Y).
LREC 2008 Marrakech 14
Evaluation
• Test data– GO terms under the namespace of cellular_component – 2,027 found, of which 2,010 have definitions --- actual test data. – All of the 2,010 test terms are related as defined previously with
one or more other test terms.
• Our evaluation strategy is to examine:– How many clustered terms have the relations defined previously,
and – How many of the related terms can be covered by the clusters.
LREC 2008 Marrakech 15
Evaluation of bottom-layer clusters
Total_clustered_terms=1,076
depths of parentnodes considered
clustered true pairs precision(%)
coverage(%)
1 417 (834 terms) 76.09 41.49
2 489 (978 terms) 89.23 48.66
3 531 (1,062 terms) 96.90 52.84
LREC 2008 Marrakech 16
Distribution of relation types IS_A and PART_OF in the clustered terms
1 parent node 2 parent nodes 3 parent nodes
type is-a part-of is-a part-of is-a part-of
numb 122 49 128 50 128 50
percent 29.3 11.75 26.2 10.2 24.1 9.4
LREC 2008 Marrakech 17
Evaluation of the second layer clusters
depths of parent nodesconsidered
correctly clusteredterms
precision/coverage(%)
1 1,163 57.86
2 1,474 73,33
3 1,685 83,83
Total_clustered_terms=2,010
LREC 2008 Marrakech 18
Evaluation of the third layer clusters
depths of parent nodesconsidered
correctly clustered terms
precision/coverage(%)
1 1,284 63.88
2 1,642 81.69
3 1,843 91.69
Total_clustered_terms=2,010
LREC 2008 Marrakech 19
• This package can be used as an assistant tool for modifying and enriching ontology and terminology. (Brief demo of interface)
Application of this package
LREC 2008 Marrakech 20
Conclusion
• Ontology term definitions provide an important
information source for term matching.
• Text comparing and clustering tool can provide useful
tool for matching the terms.
• For a better performance, the tool needs domain
knowledge resources.
LREC 2008 Marrakech 21
Acknowledgements
This research was supported by EC BOOTStrep Project (ref. FP6-028099).
The UK National Centre for Text Mining is sponsored by the JISC/BBSRC/EPSRC.
LREC 2008 Marrakech 22
References
• BOOTStrep Project website: http://www.BOOTStrep.org.
• Castillo, Gabriel, Gerardo Sierra, John McNaught (2003). An improved Algorithm for Semantic Clustering. Proceedings of the 1st international symposium on Information and communication technologies, Dublin.
• Clough, Paul, Robert Gaizauskas, Scott Piao, Yorick Wilks (2002), METER: MEasuring TExt Reuse, In Proceedings of the ACL-2002, University of Pennsylvania, Philadelphia, USA, pp. 152-159.
• Gene Ontology http://www.geneontology.org.
• Piao, Scott and Tony McEnery (2003). A tool for text comparison. Proceedings of the Corpus Linguistics
top related