A Client/Server Architecture for Word Sense Disambiguation

Caroline BrunCOLING 2000

TL;DR

This paper presents a robust client/server system for English word sense disambiguation (WSD) that leverages a bilingual dictionary as a semantically tagged corpus to automatically extract disambiguation rules. The system's novelty lies in its rule selection strategy and its client/server architecture, enabling efficient processing of large corpora without requiring manually annotated training data.

What Problem Does This Paper Address?

Word Sense Disambiguation (WSD) is a fundamental challenge in Natural Language Processing (NLP) that aims to identify the correct meaning of a word in a given context. Many traditional WSD approaches rely on large, manually annotated corpora, which are expensive and time-consuming to create. This paper addresses the problem of building an effective WSD system that can operate on "all words" without this data acquisition bottleneck, by proposing a method that automatically generates disambiguation rules from existing lexical resources.

Methods

The core of the system is a client/server architecture built on XeLDa, a linguistic framework for developing NLP applications. The system uses the Oxford-Hachette bilingual dictionary (OHFD) as a "semantically tagged corpus." This dictionary, originally designed for human use and stored in SGML format, contains rich meta-semantic information within its entries.

  1. Rule Extraction: The system parses dictionary entries and their associated examples (collocates, idiomatic examples, etc.) using a shallow parser. Functional dependencies extracted from these examples are used to build semantic disambiguation rules. These rules can be at the "word level" (matching specific lexical contexts) or "class level" (generalizing using semantic classes from an ontology like WordNet).
  2. Rule Application: When disambiguating a new text, the system first extracts dependencies using the shallow parser. It then attempts to match these dependencies against the extracted rules. A "preference strategy" is employed to select the best rule, prioritizing rules derived from more specific linguistic information (e.g., collocates) over more general ones (e.g., usage examples).
  3. Rule Competition and Cooperation: The system handles cases where multiple rules might apply. It uses a distance metric based on semantic class overlap to resolve competition between rules. For cooperation, it emphasizes sense numbers that are matched by multiple rules, effectively leveraging semantic redundancy.
  4. Client/Server Architecture: The system is implemented as a client/server application. The server handles the computationally intensive tasks of dictionary lookup and rule application, while the client interacts with the user or other applications. This architecture provides robustness, modularity, and fast execution.

Datasets

  • Oxford-Hachette French Dictionary (OHFD): Used as the primary resource for extracting semantic disambiguation rules. It's a bilingual English/French dictionary stored in SGML format.
  • SENSEVAL Competition Test Set: A set of 34 polysemous English words and approximately 8500 sentences used for evaluation.
  • HECTOR Corpus: Used in conjunction with SENSEVAL for evaluation.
  • "All Polysemous Words" Dataset: A smaller dataset of about 400 sentences extracted from the Times newspaper, used for evaluation.

The paper highlights that the OHFD is used as a "semantically tagged corpus," meaning it's treated as if it were annotated for word senses, enabling unsupervised rule acquisition.

Key Insights and Ideas

  • Dictionaries as Tagged Corpora: The most significant insight is the effective repurposing of a bilingual dictionary as a rich source for automatically generating WSD rules, bypassing the need for manually annotated training data.
  • SGML Meta-data for Rule Quality: The SGML tags within the dictionary provide fine-grained linguistic information that can be leveraged to build more precise and reliable disambiguation rules.
  • Preference Strategy for Rule Selection: The paper demonstrates that a carefully designed preference strategy, prioritizing rules based on their linguistic origin (collocates, idiomatic examples, etc.), leads to higher precision.
  • Client/Server for Scalability: The client/server architecture is crucial for handling large-scale WSD tasks efficiently, making the system practical for real-world applications.
  • Semantic Redundancy for Robustness: The approach to rule competition and cooperation, emphasizing consensus among multiple rules, enhances the system's robustness.

Key Results

The evaluation on the SENSEVAL test set showed:

  • High Precision: The system achieved a global precision of 79.5%, which is significantly higher than the 46% reported for GINGER II in SENSEVAL. This improvement is attributed to the use of more fine-grained dictionary information.
  • Equivalent Recall: The global recall was 37.4%, which was comparable to GINGER II (37.6%), suggesting that the dictionary content is similar in coverage.
  • Preference Strategy Validation: The evaluation confirmed that the preference strategy for rule selection works as intended, with collocate rules being more precise than example rules.
  • Similar Performance on Newspaper Data: An evaluation on a smaller dataset from the Times newspaper yielded similar precision and recall figures, reinforcing the system's generalizability.

Limitations and Future Work

  • Dictionary Dependency: The system's performance is inherently tied to the quality and richness of the OHFD dictionary.
  • Limited Ontology for Adjectives/Adverbs: WordNet's ontology is noted as being relatively poor for adjectives and adverbs, which could limit the generalization capabilities of class-level rules for these parts of speech.
  • Future Work:
    • Learning More Rules: Exploring correspondences between functional dependencies to learn additional rules (e.g., deriving SUBJPASS rules from DOBJ rules).
  • Improving Semantic Information: Integrating more fine-grained semantic information for adverbs and adjectives.
  • Cross-lingual Adaptation: The modular architecture allows for quick adaptation to other languages, with initial work already started on a French semantic disambiguator by integrating a French semantic ontology extracted from the dictionary.
  • General Semantic Networks: Utilizing the technique to extract general semantic networks from dictionaries or encyclopedias.

Summary updated

BibTeX
@inproceedings{brun2000client,
  author = {Caroline Brun},
  title = {{A Client/Server Architecture for Word Sense Disambiguation}},
  booktitle = {International Conference on Computational Linguistics},
  year = {2000},
}
Figure