A Bidirectional, Transfer-Driven Machine Translation System for Spoken Dialogues

Yasuhiro Sobashima, Osamu Furuse, Susumu Akamine, Jun Kawai, Hitoshi IidaCOLING 1994

TL;DR

This paper introduces TDMT, a novel bidirectional machine translation system for spoken dialogues between Japanese and English. It leverages an example-based framework and a "transfer-driven" architecture to achieve efficient, robust, and high-speed translation, particularly for domains like conference registration.

What Problem Does This Paper Address?

Spoken language translation presents unique challenges compared to written text. Spoken utterances often deviate from strict grammatical rules, and applications like automatic telephone interpretation require not only accuracy but also speed and robustness to handle diverse and potentially imperfect input. Existing machine translation systems often struggle with these characteristics. This paper aims to bridge this gap by developing a system specifically designed for the nuances of spoken dialogue, enabling seamless two-way communication between speakers of different languages.

Methods

The core of the TDMT system is its "transfer-driven" architecture. This means the translation process is primarily guided by "transfer knowledge," which is learned from actual translation examples. This knowledge is applied to map source language expressions to target language expressions. The system utilizes an example-based framework, where the similarity between linguistic expressions is calculated using a thesaurus hierarchy. This allows for efficient and robust translation by finding the "nearest" example to guide the process.

Key components and techniques include:

  • Example-Based Framework: Utilizes a database of translation examples to learn transfer knowledge.
  • Distance Calculation: Employs a thesaurus hierarchy to measure the similarity between linguistic expressions, crucial for selecting the most appropriate translation.
  • Transfer Module: The central component that applies the learned transfer knowledge to translate sentences.
  • Cooperative Modules: Lexical processing, analysis, and generation modules work in conjunction with the transfer module to refine translation quality.
  • Bidirectional Capability: The system is designed to translate seamlessly from Japanese to English and vice-versa, sharing most processing modules for efficiency.
  • Structural Disambiguation: When multiple translation structures are possible, the system calculates total distances to select the most plausible one, consistent with empirical knowledge.
  • Sentence Generation: This module ensures grammatical correctness and naturalness in the target language, including word order, inflections, and adding/deleting words for better flow.

Datasets

The paper mentions that the system was trained using corpora from ATR, specifically focusing on conference registration.

  • Training Data: 825 Japanese sentences for Japanese-to-English translation and 607 English sentences for English-to-Japanese translation.
  • Domain: Conference registration.
  • Vocabulary Size: The system can handle a vocabulary of approximately 1500 words.

Key Insights and Ideas

  • Transfer-Driven Architecture is Key: The paper strongly emphasizes that centering the translation process around a "transfer module" that learns from examples simplifies the overall system configuration and control, leading to more efficient and robust translation.
  • Example-Based Learning for Spoken Language: The success of an example-based approach is highlighted as particularly well-suited for spoken language, where variations and deviations from standard grammar are common. The system's ability to find "nearest" examples helps it cope with this variability.
  • Bidirectionality Enhances Efficiency: By designing the system to be bidirectional and sharing processing modules, the authors achieve a more streamlined and resource-efficient translation system.
  • Importance of Contextual Processing (Future Work): The authors acknowledge that current contextual processing is limited, hinting at the significant challenge and potential of incorporating deeper contextual understanding for more natural spoken dialogue translation.

Key Results

The paper reports promising experimental results for the bidirectional TDMT system:

  • Speed: On a Lisp machine with 10MIPS performance, the system achieved an average translation time of 1.9 seconds for sentences with an average length of 9.2 words.
  • Accuracy: The system demonstrated a success rate of about 70% for open test sentences within the conference registration domain.
  • Robustness: The example-based framework and transfer-centered architecture contribute to robust processing, essential for handling spoken language.
  • Bidirectional Translation: The system successfully simulates dialogues between Japanese and English speakers, demonstrating its capability for two-way communication.

Limitations and Future Work

The authors acknowledge several limitations and propose future directions:

  • Limited Domain Coverage: The current system's translation coverage is restricted to conference registration.
  • Contextual Processing: The system currently has limited contextual processing capabilities, which is crucial for truly understanding and translating spoken dialogue utterances.
  • Future Work:
    • Contextual Processing: Introduce more advanced contextual processing to better handle spoken dialogue.
  • Expanded Language Pairs: Apply TDMT to other language pairs, such as Japanese and Korean.
  • Integration with Speech Technologies: Integrate the TDMT system with speech recognition and synthesis processors to create a fully automatic telephone interpreting system.

Summary updated

BibTeX
@inproceedings{sobashima1994bidirectional,
  author = {Yasuhiro Sobashima and Osamu Furuse and Susumu Akamine and Jun Kawai and Hitoshi Iida},
  title = {{A Bidirectional, Transfer-Driven Machine Translation System for Spoken Dialogues}},
  booktitle = {International Conference on Computational Linguistics},
  year = {1994},
}
Figure