Analysis Grammar of Japanese in the Mu-project - A Procedural Approach to Analysis Grammar

Jun-ichi Tsujii, Jun-ichi Nakamura, Makoto NagaoACL 1984

TL;DR

This paper introduces a procedural approach to Japanese analysis grammar for machine translation, emphasizing the crucial distinction between structural constraints and preference rules for disambiguation. By integrating heuristic rules and explicit control over rule application, the Mu-project's grammar aims to achieve near-deterministic analysis, producing the most probable interpretation first without extensive semantic processing.

What Problem Does This Paper Address?

Traditional Context-Free Grammars (CFGs) excel at defining structural constraints but often generate an unmanageably large number of possible interpretations for ambiguous sentences. This ambiguity is particularly problematic in languages like Japanese, which have flexible word order and numerous ways to express relationships. The paper argues that simply augmenting CFGs isn't enough; a more sophisticated approach is needed to handle the preference for certain readings over others, especially when these preferences are based on cues beyond pure syntax, such as word collocations or semantic markers. The challenge is to integrate these "heuristic" preference rules effectively into a unified grammar that can efficiently produce the most likely translation.

Methods

The core of the paper's contribution lies in its procedural approach to grammar, implemented using a custom programming language called GRADE. This approach distinguishes itself from standard CFGs in several key ways:

  • Explicit Control of Rule Applications: Heuristic rules, which guide preference, can be ordered by their "strength" (how reliably they predict the correct interpretation). This allows for a prioritized application of rules, ensuring that stronger cues are considered first.
  • Multiple Relation Representation: The grammar uses annotated trees that can simultaneously represent morphological, syntactic, semantic, and logical information. This is crucial for heuristic rules that draw on diverse types of linguistic cues.
  • Lexicon-Driven Processing: Specific heuristic rules can be tied to individual words or word collocations and stored directly in the lexicon. These rules are "strong" and invoked efficiently, preventing performance degradation.
  • Explicit Definition of Analysis Strategies: The entire analysis process is divided into distinct steps, making the grammar more readable and manageable. This strategic control is vital for preventing undesirable interpretations early on.

The paper also details the grammar's organization into several components, including post-morphological analysis, scope determination, and various levels of syntactic and semantic analysis.

Datasets

Not discussed in the paper. The paper mentions the grammar is developed for the Mu-project's Japanese-English machine translation system, which currently focuses on abstracts of scientific and technological papers. It also states the grammar has been tested on "1,000 sample sentences in real abstracts without any pre-editing."

Key Insights and Ideas

  • Heuristic Rules are Distinct from Constraint Rules: The paper strongly emphasizes that rules for selecting the most probable reading (heuristic rules) are fundamentally different from rules that define grammatical structure (constraint rules). This distinction is key to designing a more effective analysis grammar.
  • Procedural Control is Essential for Integration: Simply having heuristic rules isn't enough; their integration into a unified grammar requires explicit control over their application sequence. This allows the grammar to behave more like a human interpreter, prioritizing likely interpretations.
  • Near-Deterministic Analysis is Achievable: By leveraging a combination of strong heuristic rules, lexicon-driven processing, and strategic control, the grammar can produce the most probable interpretation as the first output, significantly reducing the need for extensive post-analysis disambiguation.
  • Lexicon-Driven Processing Enhances Efficiency: Storing strong, word-specific heuristic rules in the lexicon allows them to be invoked efficiently, avoiding the performance overhead of general-purpose rules.
  • Ambiguity Resolution Beyond Syntax: The paper highlights that many ambiguities, especially in Japanese, are resolved by cues like word collocations and semantic markers, which are not easily captured by standard CFG rules.

Key Results

The paper doesn't present quantitative results in terms of specific accuracy metrics or performance benchmarks. However, it qualitatively states that:

  • The grammar, consisting of about 900 GRADE rules, works "very well and is comprehensive enough to treat various linguistic phenomena in abstracts."
  • Experiments show that the grammar can "work almost deterministically to give the most preferable interpretation as the first output."
  • The system has been tested on "1,000 sample sentences in real abstracts without any pre-editing."

Limitations and Future Work

  • Limitations:
    • The paper acknowledges that while the grammar is comprehensive for abstracts, it might not cover all linguistic phenomena.
  • The procedural nature, while powerful, can be seen as making the grammar "complicated and hard to maintain" by critics, though the authors counter this by emphasizing GRADE's rule-based specification.
  • The paper does not provide detailed quantitative evaluation results.
  • Future Work:
    • The paper implies that the ongoing development of the Mu-project and the GRADE system will continue to refine and expand the grammar's capabilities.
  • The authors suggest that lexicon-driven processing is particularly useful for the transfer phase of machine translation systems, hinting at future applications beyond just analysis.
  • The work on integrating heuristic rules and controlling their application sequence is presented as a foundational step for more robust and efficient MT systems.

Summary updated

BibTeX
@inproceedings{tsujii1984analysis,
  author = {Jun-ichi Tsujii and Jun-ichi Nakamura and Makoto Nagao},
  title = {{Analysis Grammar of Japanese in the Mu-project - A Procedural Approach to Analysis Grammar}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {1984},
}
Figure