Automatic English-to-Korean Text Translation of Telegraphic Messages in a Limited Domain

Clifford Weinstein, Dinesh Tummala, Young-Suk Lee, Stephanie SeneffCOLING 1996

TL;DR

This paper presents an interlingua-based approach to automatic English-to-Korean text translation for telegraphic military messages. The system leverages natural language understanding (TINA) and generation (GENESIS) modules, tackling domain-specific ambiguities with syntactic and semantic categories to achieve accurate translation of complex sentences.

What Problem Does This Paper Address?

The authors tackle the challenge of automatic text translation, specifically for English-to-Korean in a limited, telegraphic military domain. This is a crucial step towards enabling enhanced communication in multilingual coalition environments, where rapid and accurate information exchange is vital. Traditional machine translation systems often struggle with the ambiguity inherent in natural language, especially in domain-specific, telegraphic styles where sentence structure and vocabulary can be highly compressed. The goal is to bridge the communication gap between English and Korean speakers in military operations.

Methods

The core of the system employs an interlingua approach, using two main modules:

  • TINA (Textual Interlingua Natural Language Understanding): This module takes English input and produces a semantic frame, which serves as the interlingua representation. It combines context-free grammar, augmented transition networks, and unification concepts. Crucially, TINA incorporates both syntactic and semantic categories in its analysis grammar to resolve ambiguities. For instance, it uses semantic categories like 'ship_name' to constrain co-occurrence restrictions, reducing ambiguity.
  • GENESIS (Generation System): This module takes the semantic frame and generates the target Korean output. It's driven by a language-specific lexicon, message templates, and rewrite rules. The modular design allows for easier porting to new languages by developing these submodules.

To enhance robustness, the authors propose integrating two subsystems:

  • A rule-based part-of-speech (POS) tagger to handle unknown words and constructions by replacing them with their POS tags and bootstrapping them into the analysis grammar.
  • A word-for-word translator as a fallback mechanism to produce partial translations when the main system fails.

Datasets

The primary dataset used is the MUC-II data, consisting of naval operational report messages.

  • Training Data: 105 messages for system development (Dataset B), plus an additional 154 MUC-II-like sentences created in-house (Dataset A*).
  • Evaluation Data: 40 messages (111 sentences) for system evaluation.

The MUC-II data is described as "highly telegraphic" with many sentence fragments, complex sentences, coordination, and multiple modifiers. The authors also created a subset (Dataset A*) of MUC-II-like sentences with no unknown words to specifically evaluate grammar coverage.

Key Insights and Ideas

  • Semantic Categories for Ambiguity Resolution: The paper highlights the effectiveness of incorporating semantic categories (e.g., 'ship_name') into the analysis grammar to significantly reduce ambiguity in telegraphic messages. This is a key differentiator from purely syntactic approaches.
  • Interlingua as a Flexible Bridge: The semantic frame (interlingua) proves to be a flexible representation that can capture structural differences between English and Korean, facilitating modular language generation.
  • Robustness through Fallback Mechanisms: The acknowledgment that a perfect translation system is difficult to achieve leads to the proposed integration of a POS tagger and a word-for-word translator. This pragmatic approach aims to provide useful output even when the primary system encounters unknown words or constructions.
  • Telegraphic Style Challenges: The authors explicitly address the unique challenges posed by telegraphic language, such as sentence fragments and compressed syntax, and how their system is designed to handle them.

Key Results

  • The system can translate complex sentences (up to 38 words) and sentence fragments, as well as average-length grammatical sentences (12 words).
  • For the training data (794 sentences), the system achieved 93.6% parsing and 84.8% correct translation.
  • On a specially created dataset (A*) with no unknown words, the system parsed 50.6% of sentences, with 91% of those parsed sentences being correctly translated.
  • On the test data (40 messages, 111 sentences), the system parsed 34.8% of sentences with no unknown words, and 87.0% of those parsed sentences were correctly translated.
  • The system's efficiency is noted: approximately 1.7 seconds for a 12-word sentence and about 4 seconds for a 38-word sentence.

Limitations and Future Work

  • Unknown Words and Constructions: The paper explicitly acknowledges that the current system struggles with unknown words and constructions, which is a primary limitation. This is identified as a major cause of system failure on new MUC-II sentences (41% fail due to unknown words).
  • Integration of Robustness Mechanisms: The authors are actively working on integrating the POS tagger and the word-for-word translator to address the limitations mentioned above. The POS tagger shows promising results, achieving up to 98% accuracy after training.
  • Domain Specificity: While the system is designed for a limited domain, future work could explore expanding its coverage to more general language or other specialized domains.
  • Further Evaluation: The paper suggests ongoing work on Korean-to-English translation and the development of more sophisticated generation capabilities.

Summary updated

BibTeX
@inproceedings{weinstein1996automatic,
  author = {Clifford Weinstein and Dinesh Tummala and Young-Suk Lee and Stephanie Seneff},
  title = {{Automatic English-to-Korean Text Translation of Telegraphic Messages in a Limited Domain}},
  booktitle = {International Conference on Computational Linguistics},
  year = {1996},
}

Loading related content...

Figure