TL;DR
This paper presents a supervised word sense disambiguation (WSD) system that leverages maximum entropy (ME) models. The key innovation lies in demonstrating that optimizing feature selection for individual words, rather than using a one-size-fits-all approach, significantly improves WSD performance.
What Problem Does This Paper Address?
Word Sense Disambiguation (WSD) is a fundamental challenge in Natural Language Processing (NLP). The goal is to determine the correct meaning (sense) of a word in a given context, especially when a word has multiple possible definitions. For instance, the word "bank" can refer to a financial institution or the side of a river. Accurately disambiguating word senses is crucial for downstream NLP tasks like machine translation, information retrieval, and question answering. Traditional approaches often rely on hand-crafted rules or generic machine learning models. This paper tackles the problem of how to best represent and learn linguistic knowledge for WSD, specifically investigating whether a uniform feature set is optimal or if word-specific feature optimization yields better results.
Methods
The core of the system is built upon Maximum Entropy (ME) probability models. These models are chosen for their flexibility in integrating diverse information sources and their ability to handle problem-specific knowledge through features. The WSD system is supervised, meaning it learns from a corpus where words have already been assigned their correct senses.
Key methodological aspects include:
- Feature Engineering: Linguistic information from the context surrounding an ambiguous word is encoded into feature vectors. These features can include word collocations, part-of-speech (POS) tags, keywords, and grammatical relationships.
- "Non-relaxed" vs. "Relaxed" Features: The paper distinguishes between "non-relaxed" features (which are very specific, e.g., "POS tag of the previous word is 'adjective'") and "relaxed" features (which are more general, e.g., "the POS tag of the previous word belongs to the set {adjective, noun}"). Relaxed features help reduce the number of features and improve efficiency without significant performance degradation.
- Best Feature Selection: A crucial contribution is the investigation into optimizing feature selection. Instead of applying the same set of features to all words, the authors perform an n-fold cross-validation to identify the best combination of features for each individual word. This "best-feature-selection" strategy is central to their approach.
- Classifier Combination: The paper explores different classifier configurations, including individual ME classifiers trained with optimized feature sets, and a voting system that combines the outputs of multiple classifiers.
Datasets
The primary dataset used for training and evaluation is the SENSEVAL-2 Spanish lexical sample task corpus. This corpus is semantically annotated, providing the necessary sense labels for supervised learning. The authors also mention using a parsed version of this corpus with the Conexor Functional Dependency Grammar parser for Spanish.
Key Insights and Ideas
- Word-Specific Feature Optimization is Key: The most significant finding is that a "one-size-fits-all" feature set is suboptimal for WSD. Each word benefits from a unique combination of features, and identifying these word-specific optimal sets leads to improved accuracy.
- "Relaxed" Features Offer Efficiency: The introduction and use of "relaxed" features significantly reduce the feature space and computational complexity without a substantial loss in accuracy, making the system more scalable.
- Feature Complementarity: Different types of features (e.g., word lemmas vs. collocations, POS tags vs. grammatical relations) often complement each other, suggesting that a rich set of diverse features is beneficial.
- Part-of-Speech (POS) Specific Optimization: Optimizing feature selection based on the part-of-speech of the ambiguous word (MEbfs.pos) proved to be a strong strategy, outperforming a single best feature set for all words.
Key Results
- The best-feature-selection strategy (MEbfs) generally outperformed a fixed feature set (MEfix).
- The MEbfs.pos system, which used the best feature selection for each POS tag (nouns, verbs, adjectives), achieved high accuracy, particularly for nouns and adjectives. For example, it achieved an accuracy of 0.774 for adjectives and 0.683 for nouns.
- Verbs were found to be more difficult to disambiguate, with lower accuracy scores across most systems.
- A voting system (VME) combining multiple ME systems did not significantly improve accuracy over the best individual systems.
- When compared to other systems in the SENSEVAL-2 Spanish task, the MEbfs.pos system achieved competitive results, particularly for nouns and adjectives. The VME+SM system, which incorporated a knowledge-based method, achieved the second-highest overall score.
Limitations and Future Work
- Difficulty with Verbs: The paper acknowledges that verbs are generally harder to disambiguate, suggesting that current feature sets might be unsuitable or that more information (potentially knowledge-based) is needed.
- Future Work: The authors plan to explore cooperative strategies between different WSD methods (both knowledge-based and corpus-based). They also intend to incorporate domain information (e.g., from WordNet Domains) as an additional feature source, potentially evaluating domain disambiguation as a proxy for sense disambiguation.
Summary updated