TL;DR
This paper introduces an adaptive approach to word sense disambiguation (WSD) that leverages conceptual knowledge from dictionaries and thesauri. By iteratively refining its understanding of word senses based on the context of a given text, the system achieves improved disambiguation accuracy compared to static methods, especially for high-frequency words with contrasting senses.
What Problem Does This Paper Address?
Word Sense Disambiguation (WSD) is a fundamental challenge in Natural Language Processing (NLP), aiming to identify the correct meaning of a word in a given context. For unrestricted text, this is particularly difficult due to the vastness and diversity of language. Traditional WSD systems often rely on static, lexicalized models trained on large corpora. However, these models suffer from data sparseness, a lack of abstraction, and inflexibility, making them struggle with diverse contexts and difficult to port to new domains. The authors aim to overcome these limitations by proposing a more robust and adaptable WSD system.
Methods
The core of the proposed approach is an adaptive conceptual WSD framework. It begins with knowledge acquisition from Machine-Readable Dictionaries (MRDs). Specifically, they use a "TopSense" algorithm to extract contextual representations (CR) for each word sense. This CR is built by linking MRD sense definitions to thesaurus categories, essentially creating a weighted vector of concepts.
The process then involves:
- Initial Disambiguation: The system uses the initial CRs to perform a first pass at disambiguating words in the input text.
- Adaptation Step: Crucially, the system then adapts its knowledge base. It gleans knowledge from the partially disambiguated text and uses this to refine the CRs, making them more relevant to the specific text at hand.
- Final Disambiguation: The adjusted knowledge base is then applied to the text again to finalize the disambiguation results.
This adaptive strategy allows the system to learn from the text it's processing, improving its accuracy iteratively.
Datasets
- Training Materials: Definitions and example sentences from the Longman Dictionary of Contemporary English (LDOCE) were used. The Longman Lexicon of Contemporary English (LLOCE) was also utilized for topical information.
- Testing Materials: Passages from the Brown Corpus and Wall Street Journal (WSJ) articles were used for testing.
Key Insights and Ideas
- Adaptation is Key: The most significant insight is that an adaptive approach, which refines its knowledge based on the target text, significantly outperforms static models. This is especially true for words with contrasting senses.
- Conceptual Knowledge Matters: Moving beyond purely lexical information and incorporating conceptual representations derived from dictionaries and thesauri provides a crucial level of abstraction.
- MRDs as a Rich Source: Machine-readable dictionaries, when combined with thesauri, offer a rich, unbiased source of knowledge that can be leveraged for WSD without extensive manual annotation.
- Trade-offs in Sense Definitions: The quality of sense definitions in MRDs directly impacts the effectiveness of the contextual representation. Vague definitions can lead to the inclusion of irrelevant topics.
Key Results
The experimental results show that the adaptive algorithm achieved:
- 71% accuracy on the Brown corpus.
- 77% accuracy on the Wall Street Journal sample.
- These results represent an improvement over static approaches, particularly for high-frequency words with topically contrasting senses. For instance, the adaptation step helped resolve instances of "bank" to its MONEY sense by leveraging co-occurring words like "hold" and "paper" in the partially disambiguated text.
Limitations and Future Work
- Low-Frequency and Non-Topical Senses: The paper acknowledges that the adaptive approach is less effective for low-frequency senses or non-topical senses (like the ROW sense of "bank" in a passage about a shoreline). These are difficult to disambiguate due to data sparseness, a problem that affects most WSD approaches.
- Quality of MRD Definitions: The effectiveness of the approach is limited by the quality and vagueness of the sense definitions in the MRD.
- Future Directions: The authors suggest that exploiting constraints like "one sense per discourse" and bootstrapping strategies could further boost coverage while maintaining precision. Local contextual representations might also be more effective for certain types of senses.
Summary updated