A Prototype English-Japanese Machine Translation System for Translating IBM Computer Manuals

Taijiro TsutsumiCOLING 1986

TL;DR

This paper introduces a novel two-pass transfer approach for English-Japanese machine translation, specifically tailored for IBM computer manuals. By separating English transformation from English-to-Japanese conversion, the system achieves improved maintainability and modularity, paving the way for higher-quality, context-specific translations.

What Problem Does This Paper Address?

Translating between languages with vastly different grammatical structures, like English and Japanese, presents significant challenges for machine translation (MT). Unlike Indo-European languages, English and Japanese have distinct word orders, sentence styles, and often, a single English word can have multiple Japanese equivalents. This ambiguity makes direct, rule-based translation difficult and error-prone. Existing MT systems often struggle with this complexity, leading to translations that are either low-quality or require extensive manual post-editing. The authors aim to tackle this by developing an MT system specifically for IBM computer manuals, a domain where clarity and consistency are paramount, and where a practical English-Japanese terminology dictionary already exists.

Methods

The core innovation of this system is its two-pass transfer approach. Instead of a single, complex transfer phase, the process is divided into two distinct stages:

  1. English Transformation: The system first analyzes an English sentence into one or more parse trees. Then, the "English transformation component" rewrites these English parse trees into an intermediate "Japanese-like style" English tree. This is achieved using a set of English transformation rules and a transformation dictionary. The goal here is to adjust the English syntax to be closer to Japanese sentence structure, handling phenomena like "It-that" or "too-to" constructions. This stage is designed to be modular, allowing for easier maintenance and the ability to map multiple English expressions to a single Japanese expression by modifying transformation rules.
  2. English-Japanese Conversion: The transformed English tree is then converted into a Japanese tree by the "English-Japanese conversion component." This stage uses conversion rules and a conversion dictionary to determine appropriate Japanese equivalents, syntax, and additional linguistic data (tense, aspect, modality, voice). A key aspect here is the use of semantic markers specific to computer manuals (e.g., Logical Container, Work/Action, Document) to disambiguate word meanings and select the correct Japanese translations. The conversion dictionary is structured with sub-dictionaries for verbs, nouns, and prepositions, and entries are conditioned on English syntax and semantic markers.

Finally, a Japanese Generation component produces the final Japanese sentences from the Japanese tree.

Datasets

The paper mentions that the transformation rules and conversion rules are being constructed and tested using the IBM manual "VM/SP General Information" (60P). No other specific datasets are mentioned for training or evaluation.

Key Insights and Ideas

  • Modularity through Two-Pass Transfer: Separating English transformation from E-J conversion significantly improves the maintainability of the transfer component. Changes to one part don't necessarily require extensive modifications to the other.
  • Domain-Specific Semantic Markers: The introduction of 24 semantic markers tailored for computer manuals is crucial for disambiguation and achieving higher translation quality in this specific domain.
  • Handling English-Japanese Structural Differences: The English transformation phase is a clever way to bridge the syntactic gap between English and Japanese before the final conversion, making the subsequent steps more manageable.
  • Fail-Soft Analysis: The English parser is designed to be "fail-soft," meaning it can still produce partial translations even if a complete sentence analysis fails, which is vital for practical MT systems.
  • Plausibility Ranking: When multiple parse trees are generated, metric values are used to rank their plausibility, ensuring the most likely interpretation is carried forward.

Key Results

The paper describes a prototype system that is implemented in LISP and is undergoing testing. While specific quantitative results (e.g., BLEU scores, accuracy percentages) are not provided, the authors claim that the system aims for:

  • High-quality translation within the domain of IBM computer manuals.
  • An easily maintained transfer component.
  • A smaller English-Japanese terminology dictionary.

The paper highlights the successful application of rules to transform sentences and noun phrases, demonstrating the system's ability to handle structural shifts.

Limitations and Future Work

  • Limited Domain: The system is currently focused on IBM computer manuals. Its performance on other types of text is not discussed.
  • Quantitative Evaluation: The paper lacks detailed quantitative evaluation metrics, making it difficult to assess the system's performance objectively against other MT systems.
  • Future Work: The authors state their intention to continue enhancing the MT system to translate many kinds of IBM computer manuals into high-quality Japanese. This implies ongoing development and expansion of the rule sets and dictionaries.

Summary updated

BibTeX
@inproceedings{tsutsumi1986prototype,
  author = {Taijiro Tsutsumi},
  title = {{A Prototype English-Japanese Machine Translation System for Translating IBM Computer Manuals}},
  booktitle = {International Conference on Computational Linguistics},
  year = {1986},
}
Figure