TL;DR
This paper introduces an error-driven, unsupervised method for Word Sense Disambiguation (WSD) that leverages shallow parsing to extract functional relations between words. By treating WSD as a rule-learning problem, their system achieves competitive performance, outperforming previous unsupervised approaches and showing promise for future WSD challenges.
What Problem Does This Paper Address?
Word Sense Disambiguation (WSD) is the challenging task of identifying the correct meaning of a word in a given context, especially when a word has multiple possible senses. While supervised WSD systems have achieved good results, they require large amounts of hand-labeled data. This paper focuses on unsupervised WSD, aiming to build effective disambiguation systems without relying on extensive manual annotation. The core challenge is to automatically learn rules that can accurately map words to their correct senses based on their surrounding context and grammatical relationships. Traditional approaches often struggle with the complexity of natural language, particularly with issues like interpolated clauses and passive constructions, which this paper aims to overcome.
Methods
The paper proposes an error-driven unsupervised learning algorithm adapted from Brill's (1997) POS tagging system. The key innovation lies in how they represent and utilize functional relations extracted from shallow parsing.
Here's a breakdown of their approach:
- Shallow Parsing and Functional Relations: They use a shallow parser to extract syntactic structures and identify functional relations between words (e.g., SUBJ/VERB, VERB/OBJ). This provides a richer contextual representation than simple word co-occurrence.
- Lexical Semantics Integration: They integrate semantic information from WordNet (specifically, the top 45 synsets) by tagging the words within these functional relations.
- Rule Learning: The Brill algorithm is employed to learn WSD rules. The system learns by iteratively identifying and correcting errors in its own disambiguation process. The rules capture patterns between functional relations and semantic tags.
- Input Representation: To make functional relations compatible with the Brill learner (which typically works on bigrams), they create "Functional Semantic pairs" (FS-pairs). This involves merging functional and semantic tags into a single representation for each word in a relation.
- Scoring Methods: Two scoring methods, "original" and "goodlog," are used to select the best rules during the learning process. "Goodlog" is a probabilistic measure designed to mitigate the influence of tag frequency.
Datasets
- Penn Tree Bank (PTB): Used for extracting syntactic structures and functional relations.
- WordNet (version 1.5): Provides the semantic tags (45 top-level synsets) used to label words.
- Brown Corpus: A subset of the Brown Corpus, manually tagged with the 45 WordNet tags, was used as the test corpus for evaluation.
Key Insights and Ideas
- Functional Relations are Key: The paper strongly advocates for the importance of syntactic and functional information (beyond simple word co-occurrence) for effective WSD. They argue that understanding grammatical roles (like subject-verb or verb-object) provides crucial cues for disambiguation.
- Error-Driven Learning for WSD: Adapting Brill's error-driven approach to WSD proves effective. By learning from mistakes, the system can iteratively refine its disambiguation rules.
- Bridging Syntactic and Semantic Information: The method successfully bridges the gap between syntactic structures (from shallow parsing) and semantic labels (from WordNet), creating a richer input for the learning algorithm.
- Handling Complex Constructions: By focusing on functional relations, the system can better handle constructions that are problematic for simpler bigram models, such as interpolated clauses and passive voice.
Key Results
- The system achieved competitive performance in unsupervised WSD.
- Their "Adjusted" precision measure, which accounts for non-ambiguous words, showed a significant advantage (around 10 points) over a previous information-theory-based approach by Resnik (1997).
- The "goodlog" scoring method, which smooths tag frequency, resulted in higher precision but lower recall compared to the "original" method. This suggests a trade-off between precision and recall in unsupervised WSD, with more frequent words being more ambiguous and thus harder to disambiguate.
Limitations and Future Work
- Shallow Parsing Limitations: The use of a shallow parser means that complex or incidental structures (like parenthetical clauses or coordinate structures) are discarded, potentially limiting the system's performance.
- Named Entity Handling: The current system makes a simplifying assumption that any word without a semantic tag (and not a pronoun) is a "human" entity. This is acknowledged as a weakness, and future work includes integrating a named entity recognition module.
- Word Independence Assumption: The current FS-pair representation treats lexical heads as independent, which can hinder learning and application by not capturing the identity of the same lexical head across different relations. Future work will address this with a new Java-based learner and applier.
- Lexical Resource Dependence: The performance is tied to the granularity of the semantic lexicon (WordNet). Adjectives, in particular, have limited semantic information in WordNet, which forced their exclusion from training and testing. Future work will involve using different semantic lexicons.
- Future Directions: The authors plan to integrate their system into the SENSEVAL project, extend it to other languages (French and Italian), and improve its handling of complex linguistic phenomena.
Summary updated