An Empirical Architecture for Verb Subcategorization Frame - a Lexicon for a Real-world Scale Japanese-English Interlingual MT

Naoyuki Nomura, Kazunori MurakiCOLING 1996

TL;DR

This paper tackles the challenge of representing verb subcategorization frames in Japanese for machine translation, a task complicated by the language's flexible word order and extensive ellipsis. The authors propose a novel architecture that combines surface case frames with deep case (thematic role) frames, creating a more manageable and empirically grounded lexicon for Japanese-English MT.

What Problem Does This Paper Address?

Machine translation systems rely heavily on understanding how verbs interact with their arguments (e.g., who did what to whom). This is captured by "verb subcategorization frames," which specify the expected syntactic and semantic roles of a verb's arguments. While well-studied for languages like English, Japanese presents unique challenges. Japanese allows for significant ellipsis (omitting case elements) and flexible word order ("scrambling"), making it difficult to define consistent subcategorization frames. Existing approaches often struggle with the combinatorial explosion of possibilities or lack the empirical grounding needed for real-world MT. This paper aims to bridge this gap by developing a robust and efficient lexicon architecture specifically for Japanese.

Methods

The core contribution is a new architecture for verb subcategorization frames that integrates surface case frames (how arguments are marked syntactically, e.g., with particles like "-ga", "-wo", "-ni") and deep case frames (their semantic roles, e.g., Agent, Patient, Goal).

Key methodological aspects include:

  • Combined Frame Representation: Instead of separate surface and deep frames, they propose a unified code that maps between them.
  • Handling Ellipsis and Scrambling: The architecture is designed to accommodate Japanese's flexible word order and arbitrary ellipses by focusing on the underlying semantic roles and their potential surface realizations.
  • Reducing Combinatorial Explosion: They achieve manageability by:
  • Identifying superficially different frames that share semantic roles or can be marked by alternative case markers.
  • Introducing the concept of "surface case frame permutations" to systematically account for variations introduced by auxiliary verbs (like passive, causative, potential, etc.). This involves defining "Surface Case Permutation Frame (SCPF) codes" that dictate how case markers change with auxiliary verb usage.
  • Ambiguity Markers: For cases where multiple deep cases could fit a single slot (e.g., Agent vs. Instrument), they use ambiguity markers (like AIRM) to defer the decision to later stages of the MT process.
  • Empirical Development: The lexicon was built through an extensive empirical study and development process involving over 30,000 Japanese verbs and adjectives.

Datasets

Not explicitly discussed in terms of training or evaluation datasets. The paper describes the development of a computational Japanese lexicon based on an empirical study of 30,000 verbs and adjectives.

Key Insights and Ideas

  • Surface + Deep is Key for Japanese: The flexibility of Japanese syntax (scrambling, ellipsis) necessitates a representation that explicitly links surface case markings to underlying semantic roles.
  • Auxiliary Verbs Drive Frame Variation: Auxiliary verbs in Japanese significantly alter surface case markings, creating a need for a systematic way to generate these variations (SCPF codes).
  • Manageability Through Abstraction: By grouping similar surface frames and using permutation rules for auxiliaries, the authors avoid an unmanageable number of distinct frames.
  • Deferring Ambiguity: For certain types of ambiguity (e.g., multiple possible deep cases for a subject), deferring the decision to later MT stages using ambiguity markers is a practical strategy.
  • Empirical Grounding is Crucial: The development process emphasizes building the lexicon from real-world linguistic data rather than purely theoretical constructs.

Key Results

  • The developed lexicon architecture successfully constrains the potential combinatorial explosion of verb subcategorization frames.
  • The number of different surface/deep case mapping types was reduced to 250 after developing the code for 30,000 verbs and adjectives.
  • The lexicon specification using the proposed codes and SCPF codes improved uniformity in quality and the speed of lexicon development.
  • The system is adopted in a real-world scale interlingua-based MT system translating between English and Japanese.

Limitations and Future Work

  • Limitations:
    • The paper acknowledges that some cases might require deeper conceptual primitives than the defined deep cases (referencing NJ&B94).
  • The decision point for some ambiguities might be delayed to the generation phase of MT, implying potential complexity there.
  • Future Work:
    • Further exploration of providing a concept dictionary with more syntactic test conditions.
  • Extensions to more than two languages beyond English and Japanese.

Summary updated

BibTeX
@inproceedings{nomura1996empirical,
  author = {Naoyuki Nomura and Kazunori Muraki},
  title = {{An Empirical Architecture for Verb Subcategorization Frame - a Lexicon for a Real-world Scale Japanese-English Interlingual MT}},
  booktitle = {International Conference on Computational Linguistics},
  year = {1996},
}
Figure