A Machine Translation System From Japanese Into English Based on Conceptual Structure

Hiroshi Uchida, Kenji SugiyamaCOLING 1980

TL;DR

This paper introduces a novel machine translation system for Japanese to English that leverages "conceptual structure" as an intermediate representation. By focusing on semantic meaning rather than surface-level syntax, the system aims to overcome the limitations of traditional word-for-word or transfer-based approaches, particularly for linguistically distant languages.

What Problem Does This Paper Address?

The paper tackles the growing challenge of translating vast amounts of information in our increasingly interconnected world. Traditional machine translation (MT) systems, especially those developed in the 1960s, often relied on word-for-word translations or simple syntactic transformations. While effective for closely related languages, these methods struggle with languages that have significantly different structures, like Japanese and English. The authors argue that a deeper understanding of meaning, akin to human comprehension, is necessary for accurate translation, and that existing "transfer" approaches, which map between intermediate representations of source and target languages, still carry too much linguistic baggage.

Methods

The core innovation of this system is the use of a "conceptual structure" as an intermediate representation. This structure is built upon concepts (representing objects, actions, states, and modifiers) and the relations between them, inspired by case grammar but extended for practical use.

The translation process involves:

  1. Analysis of Japanese: The Japanese sentence is first broken down into "bunsetsu" (word groups) and their syntactic relationships are analyzed to create a "bunsetsu-table."
  2. Construction of Conceptual Structure: This table is then used to build the conceptual structure, which is designed to be language-independent, focusing purely on semantic meaning.
  3. Generation of English Phrase Structure: The conceptual structure is then translated into an English phrase structure, where semantic symbols are mapped to English phrases and syntactic roles (like subject, object) are assigned to arcs.
  4. Synthesis of English Sentence: Finally, English phrases are assembled into a grammatically correct English sentence, with necessary modifications for tense, aspect, etc.

A key aspect is the system's ability to handle discourse information by representing it within the same conceptual network framework, facilitating context awareness.

Datasets

Not discussed in the paper. The authors mention an experiment on 10 pages of text from a computer system manual, but this is described as a test case rather than a formal dataset for training or evaluation.

Key Insights and Ideas

  • Meaning over Syntax: The paper strongly advocates for prioritizing semantic meaning (conceptual structure) over surface-level syntax for robust machine translation, especially across distant language families.
  • Universality of Concepts: The authors hypothesize that human-created concepts tend to be universal, reflecting shared physical laws and experiences, making a language-independent intermediate representation feasible.
  • Conceptual Structure as a Bridge: The conceptual structure acts as a semantic bridge, allowing for a cleaner separation of analysis and synthesis steps, and reducing the complexity of handling linguistic differences.
  • Discourse Integration: Representing discourse information within the same conceptual network framework offers a powerful way to integrate context into the translation process.
  • Practicality of Human-like Translation: The approach is presented as being closer to how humans translate, suggesting it's a more practical and promising direction than purely syntactic transfer methods.

Key Results

The paper describes an experiment on approximately 230 sentences from a computer system manual, reporting that the results "so far are fairly good." Specific quantitative metrics are not provided. The qualitative findings highlight the system's ability to construct conceptual structures and generate English sentences, with the potential for handling idiomatic expressions and compound concepts.

Limitations and Future Work

  • Homonyms in Japanese: The paper acknowledges that perfect word separation in Japanese is theoretically impossible due to homonyms, though they claim near-perfect separation is achieved.
  • Level of Concepts: The authors rejected a very granular, primitive-based conceptual representation, deeming it insufficient for expressing all meanings and potentially too complex for synthesis.
  • Future Work: The authors suggest extending the system to automated abstraction, which would involve distinguishing concepts at different levels to handle context across sentences and potentially resemble paragraph structure. They also mention the possibility of applying the system to other languages in the future.

Summary updated

BibTeX
@inproceedings{uchida1980machine,
  author = {Hiroshi Uchida and Kenji Sugiyama},
  title = {{A Machine Translation System From Japanese Into English Based on Conceptual Structure}},
  booktitle = {International Conference on Computational Linguistics},
  year = {1980},
}
Figure