Ambiguity Resolution for Machine Translation of Telegraphic Messages

Young-Suk Lee, Clifford J. Weinstein, Stephanie Seneff, Dinesh TummalaACL 1997 · DOI

Abstract

Telegraphic messages, such as military situational reports (SITREPs), present significant challenges for machine translation due to rampant omissions and structural ambiguity. This paper details a "Mixed Grammar" approach developed at MIT Lincoln Laboratory that strategically combines domain-specific semantic rules with general syntactic structures. By utilizing selective lexicalization and integrating a part-of-speech tagger, the system achieves higher parsing coverage and lower error rates than purely semantic or syntactic models in English-to-Korean translation tasks.

Introduction

Telegraphic messages are characterized by extreme brevity, often omitting functional words like articles and prepositions. For example, a phrase like "TU-95 destroyed 220 nm" is highly ambiguous: it could mean the aircraft was destroyed at that distance, or that it destroyed a target at that distance. In machine translation (MT), particularly between structurally divergent languages like English and Korean, these ambiguities frequently lead to "misparses" that result in nonsensical output.

Traditional natural language processing (NLP) approaches typically fall into two categories: purely semantic grammars, which are accurate but brittle (low coverage), and purely syntactic grammars, which have high coverage but are easily confused by telegraphic omissions (high ambiguity). This research proposes a middle ground.

System Architecture

The translation pipeline utilizes two primary frameworks: TINA for language understanding and GENESIS for language generation. TINA transforms input English text into a semantic frame by navigating a parse tree, while GENESIS uses these frames to generate the target Korean text.

The system architecture for English-to-Korean machine translation, showing the flow from TINA to GENESIS.

The Mixed Grammar Approach

The core innovation is a "Mixed Grammar" that finds an optimal balance between semantic precision and syntactic flexibility. Key strategies include:

  • Selective Lexicalization: High-stakes words like verbs and prepositions are lexicalized to define their specific complement requirements and resolve attachment problems (e.g., deciding what a phrase like "with 12 rounds" modifies).
  • Domain-Specific Phrases: The grammar includes specialized rules for common telegraphic omissions, such as time expressions (e.g., "1300 z") and distances (e.g., "220 nm").
  • POS Tagger Integration: A rule-based Brill Part-of-Speech tagger acts as a preprocessor, allowing the system to handle unknown technical terms by assigning them a general category (like "noun") rather than failing the parse.
A parse tree illustrating how domain-specific rules handle telegraphic time expressions like "1410 z" within a larger fragment.

Experimental Results

The researchers evaluated the Mixed Grammar against the MUC-II corpus of naval operational reports. The results demonstrate that the mixed approach significantly outperforms both purely semantic and purely syntactic models.

Grammar Type Parsing Coverage (TEST') Misparse Rate
Semantic (Lexicalized) 43.1% 14.6%
Syntactic (POS-only) 76.5% 28.0%
Mixed (Proposed) 82.0% 10.0%

The Mixed Grammar achieved the highest coverage (82.0%) while simultaneously maintaining the lowest error rate (10.0%), proving that a hybrid approach is most effective for the telegraphic domain.

Conclusion

By anchoring parses with lexicalized verbs and prepositions while allowing for general syntactic flexibility elsewhere, the MIT Lincoln Laboratory team successfully mitigated the inherent ambiguities of telegraphic speech. While the system requires some manual domain adaptation for new fields, its robustness and accuracy represent a significant step forward for military and technical machine translation applications.

Summary updated

BibTeX
@inproceedings{lee1997ambiguity,
  author = {Young-Suk Lee and Clifford J. Weinstein and Stephanie Seneff and Dinesh Tummala},
  title = {{Ambiguity Resolution for Machine Translation of Telegraphic Messages}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {1997},
  doi = {10.3115/976909.979633},
}
Figure