TL;DR
This paper introduces two novel methods to improve the coverage of example-based disambiguation systems. By incorporating "conjunctive relationships" derived from corpora and a novel "unknown-word-matching" strategy, the authors significantly boost the system's ability to handle sentences with words not found in traditional thesauri or example bases. This work is crucial for building practical, broad-coverage natural language processing systems.
What Problem Does This Paper Address?
Example-based systems, which learn from existing examples to make decisions, are powerful for tasks like machine translation and disambiguation. However, a major hurdle is "coverage" – the system's ability to handle a wide range of linguistic phenomena and vocabulary. Existing systems often struggle when encountering words or relationships not present in their pre-defined example bases or thesauri. This limitation severely restricts their practical applicability, as demonstrated by a significant failure rate in previous systems when relevant examples were missing. The authors specifically address the challenge of improving the robustness and coverage of these systems, particularly in specialized domains like computer manuals, where unique terminology is common.
Methods
The paper proposes two key enhancements to the standard "best-match" algorithm used in example-based disambiguation:
- Acquisition of Conjunctive Relationships from Corpora: Traditional systems rely heavily on thesauri (synonym and "is-a" relationships). This paper argues that "conjunctive relationships" – words that appear together in coordinated structures (e.g., "retrieve, format, and store") – are also strong indicators of similarity and interchangeability. These relationships can be automatically extracted from domain-specific corpora, allowing the example base to be adapted to specific domains more effectively than relying solely on general-purpose thesauri.
- Unknown-Word-Matching using a Context-Base: When a word is not found in the example base or thesaurus, the system proposes a new strategy. It first parses the input text and stores these parse trees in a "context-base." Then, it identifies "context dependencies" for the unknown word by examining its usage in surrounding sentences within the text. These context dependencies are then used to find potential interchangeable words in the example base. Finally, the system compares the usage of these potential substitutes in the context-base with their usage in the example-base to infer the most likely interpretation.
Datasets
The paper describes an experiment using an example-base developed for disambiguating sentences in computer manuals.
- Example-Base: Contains 57,170 binary relationships derived from 9,500 sentences, covering 8,602 distinct words.
- Thesaurus: Utilizes the New Collins Thesaurus and Webster's Seventh New Collegiate Dictionary for synonym and "is-a" relationships.
- Test Data: 105 ambiguous prepositional phrase attachment cases were randomly selected from 3,000 sentences in a computer manual. These cases were specifically chosen because the unknown noun ("repository" in the example) was not present in the example-base or thesaurus.
Key Insights and Ideas
- Beyond Synonyms: The paper highlights that similarity isn't just about synonyms or hierarchical relationships. Words that frequently appear together in coordinated structures often share functional similarity, making them interchangeable in certain contexts.
- Domain Adaptation is Key: General-domain thesauri are insufficient for specialized fields. Automatically acquiring domain-specific relationships (like conjunctive ones) is crucial for practical coverage.
- Leveraging the Target Text: The text being processed itself contains valuable information. By analyzing the context of unknown words within the text, we can infer their meaning and relationships, effectively turning the text into a temporary, domain-specific knowledge source.
- Bridging the Gap: The proposed "unknown-word-matching" algorithm acts as a bridge, connecting words not explicitly known to the system with existing knowledge by analyzing their contextual usage.
Key Results
The paper reports on a small experiment for prepositional phrase attachment disambiguation:
- Unknown-Word-Matching: Achieved a 52.4% success rate on test data where the target noun was unknown to the system.
- With Heuristic-1: When combined with a simple heuristic (preferring inner attachment), the success rate for unknown words increased to 72.4%.
- Failure Analysis: Causes of failure included an imbalance in the example base (favoring frequent verbs) and the limited number of context sentences used for analysis.
Limitations and Future Work
- Limitations:
- The current example-base has an imbalance, leading to a preference for verb attachments.
- The effectiveness of the unknown-word-matching depends on the number of context sentences analyzed, and the optimal number is still an open question.
- The paper acknowledges that the "unknown-word-matching" might not work well if there are many unknown words in an ambiguity.
- Future Work:
- Developing more robust algorithms for handling multiple unknown words.
- Scaling up the example-base and thesaurus.
- Exploring various types of knowledge beyond conjunctive relationships for more robust matching.
- Investigating automatic or semi-automatic acquisition methods for these broader knowledge types.
Summary updated