Clustering Concept Hierarchies from Text

1
Clustering Concept Hierarchies from Text
Philipp Cimiano, Andreas Hotho, Steffen
Staab Institute AIFB University of Karlsruhe
LREC04
2
Outline

• Motivation Task
• Formal Concept Analysis
• Attribute Extraction
• Evaluation
• Conclusion Further Work

3
Motivation

• Applications of concept hierarchies
• Information Retrieval
• ML-based IE systems
• - generalize semantically over words
• Word Sense Disambiguation (WordNet)
• Syntax-Semantic Interface
• - more concise rules

4
Knowledge Acquisition Bottleneck

• Ontology Development
• difficult to agree
• time-consuming task
• Possible solution acquisition from text
• adaptivity (domain)
• speed (vs. hand-coding)
• robust (social aspects)

5
Task

• Given a set of terms (concepts) C relevant for a
  certain domain, can we derive a concept hierarchy
  ltC i.e. a partial order between these concepts
  from text?
• Clustering approach
• Formal Concept Analysis
• (set theoretical/conceptual clustering)
• Hiearchical Agglomerative Clustering (bottom-up)
• Bi-Section-Kmeans (top-down)

6
Formal Context
book rent drive ride join
appartment X X
car X X X
motor-bike X X X X
excursion X X
trip X X
7
FCA Lattice
8
Partial Order
bookable
rentable
joinable
driveable
apartment
trip
excursion
car
rideable
bike
9
Extracting Attributes

• extract syntactic dependencies from text
• verb/object, verb/subject, verb/PP relations
• car drive_obj, crash_subj, sit_in,
• LoPar, a trainable statistical left-corner
  parser
• training
• parsing
• tgrep
• syntactic relations

10
Weighting

• Parsers output
• erroneous
• not all verb/object pairs significant
• Weighting
• Resnik
• threshold t

11
Evaluation(Mädche et al. 02)

• Compare the computed taxonomy against a
• hand-crafted standard

12
Evaluation (Part II)
13
Domains Corpora

• Tourism (118 Mio. tokens)
• http//www.all-in-all.de/english
• http//www.lonelyplanet.com
• British National Corpus (BNC)
• handcrafted tourism ontology (289 concepts)
• Finance (185 Mio. tokens)
• Reuters news from 1987
• GETESS finance ontology (1178 concepts)

14
Results

• For all domains and algorithms
• SubjObjPP gt other combinations
• PP gt Obj gt Subj
• Best result F56.98 (Tourism,FCA)
• F56.12 (Reuters,HC)
• Human Performance Mädche 02
• 4 subjects (1 ontology engineerer)
• 12 tourism ontologies (ca. 300 concepts)
• average TO of 56.35 (47-87)
• assumption of an LR of 100
• gt Human F-Measure 72.08

15
Comparison of Approaches(tourism domain)
16
Other Weighting Measures(tourism domain, FCA)
17
Conclusion

• novel approach to learning concept hierarchies by
  FCA
• reasonable results compared to top-down and
  bottom-up approaches
• using SubjObjPP syntactic dependendcies
  outperforms other combinations
• use of weighting measure has some impact
• proposed a systematic evaluation method

18
Further Work

• General Features
• Levin classes
• Top Levels
• Polysemy
• first experiments
• evaluation