📄 Improvements To Monolingual English Word Sense Disambiguation

Summary

The paper addresses all-words word sense disambiguation (WSD) in English: every content word in running text has to be mapped to its WordNet sense, and no sense-tagged training data is used by the disambiguator itself. The authors build on the unsupervised graph method In-Degree from Sinha and Mihalcea (SM07). In that method, the candidate senses of nearby words become graph nodes, edges are weighted by WordNet similarity, and each word takes the sense whose incoming weights add up to the most. The authors make three changes:

• They score verb–verb pairs with JCN instead of LCH.
• They enrich the Lesk gloss overlap with glosses of related synsets ("Expand-Lesk").
• They add SemCor usage examples to WordNet glosses.

Experiments cover Senseval 2, Senseval 3 and SemEval 2007 under WordNet 1.7.1, 2.1 and 3.0. The combined system reaches 62.7% F-measure on Senseval 2 (WN 1.7.1), the best unsupervised result reported at the time. This shows that enriching the lexical resource can matter as much as changing the algorithm.

Key points

• Baseline graph method: In-Degree uses JCN for noun pairs and Lesk for adjectives, adverbs and cross-POS pairs. Scores are normalized to [0,1]: Lesk values above 240 become 1, and JCN's observed range of 0.04–0.2 is rescaled. The normalization values were calibrated on Senseval 3.
• Expand-Lesk relations: hypernym, hyponym, similar attributes, verb group, pertainym, holonym and meronym. Only the context word's sense is expanded. Expanding the target sense as well hurt accuracy because some relations are noisy.
• F-measure, Senseval 2 (WN 1.7.1):
  • random baseline: 39.9
  • reimplemented SM07: 59.7
  • JCN for verbs: 60.2
  • plus Expand-Lesk: 60.9
  • plus SemCor: 62.04
  • both enrichments: 62.7
• Other data sets: On Senseval 3 (WN 1.7.1), SemCor alone gave 59.7. Verb F-measure on Senseval 2 rose from 33.01 to 38.66. Noun scores slightly dropped with SemCor (68.7 → 68.3).
• Error analysis: Remaining errors include metonymy (e.g. "English" meaning the people rather than the language) and unrecognized multiword expressions. Coverage on Senseval 2 was 98.36%, mostly because of POS mismatches.

Interesting ideas

• I1. Choosing a sense by summing similarity-weighted votes from neighbouring words' candidate senses is a simple, fully inspectable global decision, with no hidden classifier.
• I2. Pick a different similarity measure per POS pair (taxonomy-based for nouns and verbs, gloss overlap across POS), and normalize every measure to a common [0,1] scale before combining.
• I3. Expand only the context side with related synsets, not the candidate being judged, so that noisy relations cannot distort the hypothesis under test.
• I4. Enlarge sparse dictionary glosses with real hand-annotated usage examples, re-mapped between resource versions through an index file.
• I5. The leftover errors (metonymy, multiwords) are cases where syntactic and role information would help, for example knowing that a subject acts with volition.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

• I1 (world): A WordNet sense picker in the compiled knowledge layer could store each edge's contribution, so the explanation of a chosen sense lists exactly which context senses supported it.
• I2 (world): When ontology lookups mix WordNet hierarchy distance with gloss evidence, per-category measures with explicit cutoffs keep the scores comparable and auditable.
• I3 (frames): When checking whether an argument fits a VerbNet role restriction, the argument's evidence could be widened through WordNet relations while the role definition stays fixed.
• I4 (learn): Silver-annotated sentences that several witness parsers agree on could be attached as examples to lexicon entries, giving later disambiguation more text to compare against.
• I5 (frames): Sense choice could be checked against the frame. An agent slot that requires volition would favour the "people" reading of "English" over the "language" reading.