TL;DR
This paper introduces a novel approach to Word Sense Disambiguation (WSD) that leverages both topical and semantic associations within a document. By identifying "topic discriminative terms" and constructing a "topical-semantic association graph," the method aims to improve disambiguation accuracy, particularly for ambiguous words where context is limited.
What Problem Does This Paper Address?
Word Sense Disambiguation (WSD) is a crucial task in Natural Language Processing (NLP) that aims to determine the correct meaning of a word based on its surrounding context. Despite years of research, existing WSD algorithms often struggle with accuracy and recall, especially when the context is sparse or ambiguous. A key challenge is that the immediate context of a word might not provide enough information for fine-grained disambiguation, and relying on broader contexts can increase computational complexity. This paper tackles the problem of improving WSD performance by incorporating document-level topical information alongside semantic context.
Methods
The proposed approach has three core components:
- Topic Semantic Profile Generation: This involves mapping WordNet senses to Open Directory Project (ODP) category labels to create a "topic semantic profile" (TSP) for each term. This TSP combines the term's disambiguation context (glosses, synonyms, hypernyms) with its hierarchical topic category.
- Topic Identification via Topic Discriminative Terms (TDTs): The method identifies "topic discriminative terms" (TDTs) that are strongly associated with a document's subject matter. These TDTs are extracted using features like position, statistical frequency, semantic properties (e.g., number of senses), and topic span distribution. A "topical graph" is then constructed based on the reoccurrence topic spans of these TDTs to identify the document's overall topic.
- Disambiguation using Topical-Semantic Association Graph: A "topical-semantic association graph" is built, integrating syntactic features, semantic features (from TSPs), and topical features (from the topical graph). This graph represents the relationships between ambiguous terms and their contexts. The disambiguation process then leverages this graph to determine the most likely sense of an ambiguous word by calculating similarities between topic chains and disambiguation contexts. Special attention is given to disambiguating nouns and verbs, with distinct strategies for each.
Datasets
The paper uses the SemCor corpus for evaluation. SemCor is described as the largest freely available corpus of semantically annotated words, commonly used for evaluating WSD systems. No new benchmarks are introduced.
Key Insights and Ideas
- Leveraging Document-Level Topics: The core insight is that understanding the broader topic of a document significantly aids in disambiguating individual words, especially when local context is insufficient.
- Topic Discriminative Terms (TDTs): Identifying terms that strongly signal a document's topic is crucial. The paper proposes a method to extract these TDTs using a combination of features, highlighting the importance of topic span distribution.
- Topical-Semantic Association Graph: The integration of topical information (from TDTs and topic chains) with semantic context (from WordNet senses) into a unified graph structure provides a richer representation for disambiguation.
- Differentiated Disambiguation Strategies: The paper recognizes that nouns and verbs require different disambiguation approaches, with specific methods proposed for each.
Key Results
The proposed Topical-Semantic Association (TSA) approach, when combined with WordNet and ODP, achieved better performance compared to other state-of-the-art WSD algorithms like ExtLesk and SSI on the SemCor dataset.
- Overall Performance: TSA demonstrated superior performance in disambiguating words.
- Noun vs. Verb Disambiguation: The performance for disambiguating nouns was notably higher than for verbs, underscoring the value of topical information for noun disambiguation.
- Comparison with SSI: On the noun-only subset, TSA's F1 score was comparable to SSI and significantly better (+2.6% F1) than other algorithms.
- Topic Identification: The topic identification component, using comprehensive features, achieved an accuracy of 87.8% on the SemCor dataset. The topic span distribution feature (TS) was found to be particularly important for improving document topic identification accuracy.
Limitations and Future Work
- Topic Graph Centrality: The paper notes that a potential cause for performance degradation in topic identification is the lack of high-degree centrality vertices in the topical graph, which can make it difficult to identify the document's topic.
- Improper Topic Semantic Profile: Determining the "improper unique topic semantic profile" of candidate TDTs is also identified as a cause for errors.
- Future Work: The paper suggests that the indexing of verb senses could be incrementally improved by updating the list of objects in the future. The authors also imply that further refinement of the topical graph construction and the disambiguation process could lead to even better results.
Summary updated