Mapping Lexical Entries in a Verbs Database to WordNet Senses

Rebecca Green, Lisa Pearl, Bonnie J. Dorr, Philip ResnikACL 2001 · DOI

TL;DR

This paper presents an automated approach to mapping a large database of English verbs to specific WordNet senses by leveraging syntactic features and semantic similarity. Unlike traditional word sense disambiguation (WSD) that relies on sentence context, this method uses "context-free" data like thematic grids and verb classes to achieve 72% precision, significantly outperforming frequency-based baselines and approaching the limits of human agreement.

What Problem Does This Paper Address?

The researchers tackle a specialized version of Word Sense Disambiguation (WSD). Usually, WSD involves looking at a word in a sentence (e.g., "The prices dropped") and using the surrounding words to guess its meaning. However, the authors are working with a lexical database—a list of 9,611 verb entries—where there is no surrounding text.

The goal is to link these entries to WordNet, a massive database of semantic relationships. Verbs are notoriously slippery; for instance, the verb "drop" might translate to bajar (descend), disminuir (decrease), or dejar caer (let fall) in Spanish depending on its syntactic context and intended sense.

Mapping example for the verb "drop" across different Levin+ classes and WordNet senses.

By mapping the database to WordNet senses, the authors provide the "semantic glue" needed for high-quality machine translation and information retrieval.

Methods

The authors treat the problem as an ensemble classification task. Since they lack sentential context, they use the "internal" features of the verb database:

  • Levin+ Classes: Verbs are grouped into 491 classes based on their syntactic behavior (e.g., "Roll Verbs").
  • Thematic Grids ($\theta$-grids): These describe the roles a verb requires, such as [agent], [theme], or [goal].
  • WordNet Frames: Syntactic patterns associated with synsets (e.g., "Somebody ____s something").
  • Semantic Similarity: A measure (based on Resnik, 1999) that calculates how related different verbs within the same Levin class are based on the WordNet hierarchy.
  • Prior Probabilities: Frequency data pulled from SEMCOR (a sense-tagged version of the Brown Corpus).

The core of the paper explores various Voting Schemes:

  1. Majority Voting: Different probability measures "vote" for a sense. The authors tested simple voters (individual measures) and aggregate voters (products or weighted sums of measures).
  2. Threshold Voting: A measure only casts a vote if its confidence exceeds a specific, mathematically optimized threshold.

Key Insights and Ideas

  • Syntactic-Semantic Correspondence: The paper reinforces the idea that how a verb is used syntactically (its "frames" and "grids") is a powerful proxy for what it means. You can disambiguate a verb quite well without ever seeing it in a sentence, just by knowing its allowed syntactic structures.
  • The "Same-Synset Assumption": A key heuristic used was that if two verbs are in the same syntactic class and share a WordNet synset, they likely map to that same synset in the database.
  • The Difficulty of the Task: Human coders only agreed on 51% of sense assignments (a Kappa of 0.47). This highlights that WordNet senses are often so fine-grained that even experts struggle to distinguish them, setting a realistic "ceiling" for how well an algorithm can perform.

Key Results

The authors compared their automated mapping against a baseline and human performance. The automated system significantly improved recall over the baseline while maintaining high precision. The best-performing ensemble method, MajAggr, achieved 72% precision.

Performance metrics (Recall and Precision) for different voting schemes with and without Semantic Similarity (SS).
Method Precision Recall
Lower Bound (Most Frequent Sense) 62% 38%
Best Automated (MajAggr) 72% 58%
Upper Bound (Human Judgment) 87% 75%

While the system didn't reach human levels, it proved that automated mapping is a viable way to enrich lexical resources without requiring manual sentence tagging.

Limitations and Future Work

  • Fine-Grained Senses: WordNet's distinctions are often too subtle for this type of database mapping, leading to lower agreement and recall.
  • Data Sparsity: Some measures (like Grid and Levin+ probabilities) suffered from "zero-value" issues where a specific combination hadn't been seen in the training data.
  • Future Directions: The authors suggest that relaxing the voting criteria—allowing the system to assign more than two senses to an entry—might improve recall and better reflect the inherent ambiguity of verbs.

Summary updated

BibTeX
@inproceedings{green2001mapping,
  author = {Rebecca Green and Lisa Pearl and Bonnie J. Dorr and Philip Resnik},
  title = {{Mapping Lexical Entries in a Verbs Database to WordNet Senses}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2001},
  doi = {10.3115/1073012.1073044},
}
Figure