How to Deal With Ambiguities While Parsing: EXAM - A Semantic Processing System for Japanese Language

Hidetosi SiraiCOLING 1980

TL;DR

This paper introduces EXAM, a novel system designed to effectively and efficiently handle ambiguities in Japanese language processing. EXAM integrates parsing and semantic interpretation, tackling ambiguities at the levels of word meaning, grammar, and discourse by employing a hierarchical knowledge base and a sophisticated augmented context-free parser.

What Problem Does This Paper Address?

Natural language understanding systems (NLUS) face a significant challenge in dealing with ambiguity. A core dilemma is producing plausible interpretations without succumbing to a combinatorial explosion of possibilities. Furthermore, for languages like Japanese, where grammatical structures are flexible and elements like noun phrases are often omitted, separating parsing from semantic analysis proves inefficient. The paper addresses the need for an NLUS that can effectively disambiguate meaning, grammar, and discourse while maintaining computational efficiency, particularly for the complexities of Japanese.

Methods

EXAM employs a multi-component approach:

  • Hierarchical Knowledge Sources: This includes a frame memory (containing prototype and script frames) and a text memory (holding context, frame systems, and speaker's viewpoint).
  • Semantic Interpreter: This component is responsible for generating interpretations.
  • Augmented Context-Free Parser (MELING): This parser is fundamentally breadth-first and uses both bottom-up and top-down strategies. It translates linguistic expressions into frames.

Key to EXAM's strategy is the integration of parsing and semantic processing. The system first determines the "sense" (literal meaning) of linguistic expressions and then their "meaning" (contextual interpretation). This is achieved through a two-stage semantic interpretation process:

  1. Sense Disambiguation: This stage uses techniques like semantic markers (implemented via generalization hierarchies), "GROUP" frames for homonyms, and "case ordering" heuristics to reduce the number of plausible interpretations. It prioritizes interpretations based on "likelihood."
  2. Meaning Disambiguation: This stage focuses on referential ambiguity and cohesion ambiguity. Referential ambiguity is handled by delaying disambiguation until after the entire sentence is processed, collecting referent requirements from the complete sentence. Cohesion ambiguity is addressed by recognizing sentence relationships (e.g., explanation, cause) and inferring omitted elements.

EXAM also incorporates "anticipations" during parsing, where the system predicts coreferences or relationships based on specific linguistic cues (e.g., the particle nagara).

Datasets

Not discussed in the paper.

Key Insights and Ideas

  • Integrated Parsing and Semantics: The paper argues strongly for the necessity of integrating parsing and semantic interpretation, especially for languages like Japanese, to handle omitted elements and improve efficiency.
  • Sense vs. Meaning Distinction: Explicitly distinguishing between "sense" (literal meaning) and "meaning" (contextual interpretation) provides a structured approach to disambiguation.
  • Two-Stage Semantic Interpretation: The division into sense disambiguation (early, focusing on literal meaning) and meaning disambiguation (later, focusing on context and discourse) is a robust strategy for managing complexity.
  • "Likelihood" as a Disambiguation Metric: The concept of "likelihood" (derived from semantic markers, case ordering, and frame instantiation categories like "INSTANT") is crucial for ranking and selecting interpretations.
  • Delayed Disambiguation for Referential Ambiguity: Recognizing that referential ambiguity often requires broader context, EXAM delays its resolution until after the entire sentence is processed.
  • Anticipations for Efficiency: Proactive "anticipations" during parsing help prune the search space and guide semantic interpretation.

Key Results

The paper describes EXAM as an experimental system designed to deal with ambiguities effectively and efficiently. While specific quantitative results (e.g., accuracy percentages, speedup metrics) are not detailed, the paper emphasizes the system's ability to:

  • Avoid the combinatorial explosion of interpretations.
  • Handle ambiguities at multiple linguistic levels (word meaning, grammar, discourse).
  • Improve parsing efficiency by eliminating unnecessary partial parsed trees based on semantic interpretation.
  • Produce plausible interpretations by integrating syntactic and semantic information.

Limitations and Future Work

  • Ambiguities in Meaning: The authors acknowledge that EXAM does not completely clarify all ambiguities in meaning, particularly those not resolved by preceding context. This remains an open problem.
  • Truth Maintenance System: The paper briefly mentions the potential of a Truth Maintenance System for handling ambiguities that persist even after parsing, suggesting this as a direction for future development.
  • Efficiency of Older Versions: The older version of MELING had a parsing time proportional to n² or n³, indicating that efficiency improvements were a significant motivation for the current work. The paper implies that EXAM's approach leads to greater efficiency.

Summary updated

BibTeX
@inproceedings{sirai1980how,
  author = {Hidetosi Sirai},
  title = {{How to Deal With Ambiguities While Parsing: EXAM - A Semantic Processing System for Japanese Language}},
  booktitle = {International Conference on Computational Linguistics},
  year = {1980},
}
Figure