An English Japanese Machine Translation System of the Titles of Scientific and Engineering Papers

Makoto Nagao, Jun-ichi Tsujii, Koji Yada, Toshihiro KakimotoCOLING 1982

TL;DR

This paper presents an English-to-Japanese machine translation system specifically designed for scientific and engineering paper titles. By analyzing the unique characteristics of these titles, the system achieves high translation accuracy, particularly in specialized fields like physics and mathematics, enabling easier access to English-language research for Japanese speakers.

What Problem Does This Paper Address?

Accessing and utilizing the vast amount of scientific and engineering research published in English can be a significant barrier for non-native English speakers. While general-purpose machine translation systems exist, they often struggle with the specialized terminology, concise phrasing, and unique grammatical structures found in technical titles. This paper addresses the challenge of accurately translating these specific types of text to facilitate broader research accessibility.

Methods

The core of the system relies on a "subgrammar" approach, recognizing that universal grammatical rules are impractical for specialized domains. Instead, they analyze a large corpus of scientific titles to identify a limited set of fundamental "skeleton patterns." The translation process involves several steps:

  1. Dictionary Look-up and Idiom Finding: The system first attempts to translate words and recognized idioms.
  2. Parsing of Conjunctive Phrases: It handles conjunctions like "and" and "or," though with limitations on complex structures.
  3. Parsing of Simple Noun Phrases: A finite automaton (transition network) is used to parse noun phrases, which constitute the majority of titles.
  4. Handling Special Structures and Semantic Disambiguation: Specific word sequences and ambiguous verb-ing forms are addressed through rule-based and semantic checks. For instance, semantic elements are introduced to differentiate between a verb-ing acting as a modifier versus a present participle.
  5. Skeleton Pattern Matching and Word Order Change: Identified skeleton patterns are matched with their Japanese equivalents, with word order adjustments made only when necessary.
  6. Synthesis of Japanese Title Sentence: The translated components are assembled into a Japanese title.

Datasets

The system was developed and evaluated using titles from the INSPEC database, specifically focusing on physics and mathematics. The authors analyzed ten thousand titles to derive the fundamental sentential structures and report evaluation results on a subset of these titles.

Key Insights and Ideas

  • Domain-Specific Subgrammars are Key: The paper strongly advocates for specialized grammars tailored to specific text types (like scientific titles) rather than attempting universal linguistic coverage.
  • Titles Have Predictable Structures: Scientific and engineering titles, despite their technical content, exhibit a surprisingly limited number of fundamental syntactic structures, making them amenable to rule-based translation.
  • Noun Phrases Dominate: The observation that over 98% of titles are noun phrases simplifies the translation task significantly, allowing the system to focus on this structure.
  • Semantic Information Improves Accuracy: Incorporating semantic checks, even basic ones, is crucial for disambiguating word meanings and ensuring correct translation, especially for verb-ing forms.

Key Results

The system demonstrates high translation accuracy, achieving about 93% for the specific areas of physics and mathematics from the INSPEC database. For a broader range of scientific and engineering papers, the accuracy was around 80%. The system is efficient, with an average translation time of 0.1 seconds per title. It has been successfully used to create a Japanese version of the INSPEC database, enabling Japanese-language retrieval.

Limitations and Future Work

The authors acknowledge that the system cannot handle embedded sentential structures introduced by relative pronouns or complex conjunctive phrases. They also note that the translation of prepositions can be context-dependent, and while their system assigns unique translations, further work is needed to handle these nuances more robustly. Future work could involve expanding the dictionary and improving the handling of more complex syntactic structures to further enhance translation accuracy across a wider range of scientific domains.

Summary updated

BibTeX
@inproceedings{nagao1982english,
  author = {Makoto Nagao and Jun-ichi Tsujii and Koji Yada and Toshihiro Kakimoto},
  title = {{An English Japanese Machine Translation System of the Titles of Scientific and Engineering Papers}},
  booktitle = {International Conference on Computational Linguistics},
  year = {1982},
}
Figure