A Machine Translation System From Japanese Into English - Another Perspective of MT Systems -

M. Nagao, J. Tsujii, K. Mitamura, H. Hirakawa, M. KumeCOLING 1980

TL;DR

This paper presents a novel machine translation system from Japanese to English, focusing on computer manuals. It champions a lexicon-driven "transfer approach" where linguistic phenomena are primarily handled by lexical rules and descriptions rather than general syntactic rules, a departure from traditional methods. This approach is particularly effective for bridging the significant structural differences between Japanese and English.

What Problem Does This Paper Address?

Machine translation (MT) systems often struggle with language pairs that have vastly different grammatical structures, such as Japanese and English. Traditional MT systems, especially those translating between closely related European languages, rely heavily on general syntactic rules. However, for languages like Japanese and English, this approach becomes cumbersome due to the extensive structural transformations required. The paper addresses the challenge of effectively translating between these structurally dissimilar languages by proposing a system that leverages detailed lexical information and rules to manage these complex transfers. The goal is to create a more robust and adaptable MT system for domains like technical documentation.

Methods

The system employs a "transfer approach," breaking down the translation process into three main components: Japanese analysis, transfer, and English synthesis.

  • Lexicon-Driven Analysis: The core innovation is its heavy reliance on lexical descriptions. Instead of general syntactic rules, most linguistic phenomena are handled by lexical rules and descriptions stored in dictionaries. This is inspired by the idea that humans have a vast lexical capacity.
  • Intermediate Representation (JIS): Japanese analysis produces a "Japanese Intermediate Structure" (JIS), which is primarily a dependency structure based on case notions. However, it also incorporates "relation descriptors" for certain syntactic constructions, making it a hybrid of dependency and phrase structures.
  • Transfer Component: The transfer phase bridges the JIS and an "English Intermediate Structure" (EIS). This is also lexicon-driven, using a bilingual dictionary that contains transfer procedures for each word. These procedures are often parameterized and can be tailored for specific words or word classes.
  • English Intermediate Structure (EIS): The EIS is similar to phrase structure trees but includes attribute-value pairs to capture information beyond simple categories, aiding in the subsequent synthesis.
  • English Synthesis: The synthesis component transforms the EIS into English surface structures. This involves applying transformation rules, which are represented as tree-to-tree mappings. The system distinguishes between obligatory and optional rules, and the applicability of rules is checked not only by pattern matching but also by user-defined procedures.

Datasets

Not explicitly discussed in the paper, beyond mentioning that the system is designed for translating computer manuals. The paper focuses on the system's architecture and principles rather than specific training or evaluation datasets.

Key Insights and Ideas

  • Lexicon Over Syntax for Dissimilar Languages: The paper strongly argues that for languages with significant structural differences (like Japanese and English), a lexicon-driven approach is more effective than relying on general syntactic rules. This is because lexical rules can capture the specific, often complex, transformations required for individual words and constructions.
  • Hybrid Intermediate Representation: The JIS, a mix of dependency and phrase structures with "relation descriptors," offers a flexible way to represent Japanese linguistic phenomena, including complex syntactic constructions like relative clauses, in a way that facilitates transfer.
  • Pragmatic Approach to Case Representation: The system's case representation is pragmatic and oriented towards machine translation, with cases defined based on individual verbs and categorized as "intrinsic" (handled by lexical rules) or "extrinsic" (handled by general rules).
  • Stereotyped Expressions are Key: The system effectively utilizes stereotyped or semi-stereotyped expressions common in computer manuals, treating them as valuable linguistic units.

Key Results

The paper presents a progress report on an experimental prototype system. While specific quantitative results (e.g., BLEU scores) are not provided, the paper demonstrates the system's ability to handle complex sentences and structural transformations through examples. The authors highlight the successful implementation of analysis for various transformation types, including those caused by adverbial suffixes, post-verbial expressions, verbal complements, and relative clauses. The system's flexibility in handling exceptions and its lexicon-driven nature are presented as key strengths.

Limitations and Future Work

  • Definite/Indefinite Articles: The system currently struggles to distinguish between definite and indefinite noun phrases in English, as Japanese surface sentences lack explicit markers for this.
  • Contextual Elements: Supplying appropriate elements from context for omitted expressions in Japanese is a significant challenge. The system attempts to find suitable English words and structures without explicit contextual processing, but this is acknowledged as a limitation.
  • Limited Vocabulary: The current experimental prototype has a limited dictionary (around 80 verbs).
  • Future Work: The authors are designing a new, more advanced system that aims to improve upon these limitations. They emphasize the importance of integrating syntactic processing with semantic and pragmatic knowledge, advocating for their lexicon-based approach as a suitable framework for this integration.

Summary updated

BibTeX
@inproceedings{nagao1980machine,
  author = {M. Nagao and J. Tsujii and K. Mitamura and H. Hirakawa and M. Kume},
  title = {{A Machine Translation System From Japanese Into English - Another Perspective of MT Systems -}},
  booktitle = {International Conference on Computational Linguistics},
  year = {1980},
}
Figure