Experiences with an On-Line Translating Dialogue System

Seiji Miike, Koichi Hasebe, Harold L. Somers, Shin'ya AmanoACL 1988 · DOI

Abstract

In 1988, researchers at Toshiba’s Research and Development Center developed one of the first real-time, bi-directional machine translation (MT) systems for networked chat. By connecting an English-Japanese MT engine to a UNIX-style interface via satellite, users in Japan and Switzerland were able to converse across language barriers. This paper evaluates the system's performance, highlighting that the primary challenge in automated dialogue is not just translation accuracy, but managing "metadialogue"—the conversational strategies users employ to navigate system limitations and linguistic ambiguities.

Introduction

In the late 1980s, machine translation was largely confined to static document processing. This study explored the feasibility of applying MT to live, interactive dialogue, which is characterized by typos, conversational fillers, and a lack of formal structure. The researchers aimed to test their AS-TRANSAC system in a real-world scenario: a transcontinental link between Japan and Switzerland.

Conceptual diagram of the satellite-linked communication system between Switzerland and Japan.

System Architecture and Interface

The experimental setup utilized Toshiba AS3260C workstations connected via a satellite link involving KDD’s Venus-P and Swiss PTT’s Telepac networks.

The Translation Engine

The core of the system was a "Lexical Transition Network Grammar" (LTNG). Unlike standard rule-based systems of the time, LTNG separated syntactic parsing from semantic analysis to improve processing speed and handle ambiguities more effectively. The system was supported by a comprehensive dictionary containing 100,000 entries for both English and Japanese.

User Interface Design

The interface was designed to provide users with maximum context to mitigate translation errors. The screen was divided into three distinct windows:

  1. Local Input: The user’s own typed text.
  2. Translated Output: The partner’s reply after translation.
  3. Raw Workspace: The original, untranslated input from the partner.
The multi-window interface showing simultaneous English and Japanese text during a live session.

This "workspace" window was particularly valuable for experienced users, allowing them to identify errors caused by typos or unrecognized proper names that the MT engine failed to process correctly.

Experimental Evaluation

The system was evaluated during 78 live sessions at the 5th World Telecommunications Exhibition (Telecom 87) in Geneva. A total of 1,429 utterances were recorded and analyzed.

Key Findings

  • High Success Rate: Approximately 90% (1,289) of utterances were successfully translated. The remaining 10% failed primarily due to spelling errors or unknown proper names.
  • Metadialogue: A significant portion of the conversation involved "metadialogue"—users discussing the translation process itself or clarifying errors (e.g., asking for the meaning of a mistranslated word).
  • Conversational Tuning: The researchers found that standard MT often failed on conversational fillers. For instance, "I see" was initially translated literally as "I am looking at something," requiring the system to be tuned to recognize idiomatic conversational markers like the Japanese "naruhodo."
  • Latency and Engagement: Despite a 15-second delay due to satellite transmission and processing, users remained engaged. A UI status indicator showing when the partner was typing was critical for maintaining the flow of conversation.

Discussion and Conclusion

The study revealed that the greatest threat to effective MT in chat is not grammatical complexity but human error, specifically typos. A single misspelling could derail the translation logic entirely. Furthermore, the system's inability to distinguish proper names from common nouns often led to confusion.

The authors concluded that future interactive MT systems should incorporate integrated spell-checkers and syntax-checkers to alert users before transmission. They also noted the emergence of "MT literacy," where users naturally adapted their language—using explicit pronouns and avoiding slang—to improve the system's success rate. This early experiment laid the groundwork for modern real-time translation tools by identifying the crucial role of interface design and user adaptation in cross-linguistic communication.

Reference: Miike, S., Hasebe, K., Somers, H., & Amano, S. (1988). Experiences with an On-Line Translating Dialogue System. Proceedings of COLING-88.

Summary updated

BibTeX
@inproceedings{miike1988experiences,
  author = {Seiji Miike and Koichi Hasebe and Harold L. Somers and Shin'ya Amano},
  title = {{Experiences with an On-Line Translating Dialogue System}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {1988},
  doi = {10.3115/982023.982042},
}
Figure