An Integrated Architecture for Example-Based Machine Translation

Alexander Franz, Keiko Horiguchi, Lei Duan, Doris Ecker, Eugene Koontz, Kazami UchidaCOLING 2000

TL;DR

This paper introduces HARMONY, a hybrid machine translation architecture that combines example-based translation with rule-based linguistic analysis for improved flexibility and accuracy. Their English-to-Japanese prototype achieves good translation quality, handling idiomatic expressions while maintaining grammatical correctness, and operates with reasonable computational resources.

What Problem Does This Paper Address?

Traditional machine translation systems often fall into two camps: example-based systems that excel at idiomatic translations but struggle with grammatical coverage, and rule-based systems that offer broad coverage but can produce brittle, unnatural output. The challenge is to bridge this gap, creating a system that leverages the strengths of both approaches to produce high-quality, natural-sounding translations across a wider range of linguistic phenomena. This paper specifically addresses the need for a system that can handle the nuances of colloquial language, particularly in domains like travel, where users often simplify their input.

Methods

The HARMONY architecture is built upon a robust NLP infrastructure, featuring a custom Grammar Programming Language (GPL). GPL allows for the direct expression of linguistic algorithms for parsing, transfer, and generation, drawing inspiration from formalisms like Lexical-Functional Grammar.

Key methodological components include:

  • GPL Compiler: Compiles GPL grammars into C code, generating efficient action, expansion, and test functions for feature structure manipulation.
  • Feature Structure Library: Provides efficient representation, testing, and memory management for feature structures, enabling fast garbage collection.
  • Hybrid Transfer: Combines rule-based transfer grammar with example-based matching. The transfer grammar constrains example searches and handles linguistic aspects beyond the example database, while the example database (containing 14,000 pairs) provides idiomatic translations.
  • Recursive Transfer: The transfer process operates recursively, invoking example matching for various parts of the input.
  • Information Content for Similarity: Lexical similarity is calculated using a scheme based on the information content of thesaurus nodes, rather than just direct distance.
  • Travel Domain Focus: The prototype is implemented for the "travel domain," characterized by colloquial language and simplified user input, reflecting real-world communication patterns.

Datasets

  • Development Corpus: 7,000 expressions for training and refinement.
  • Test Set: 5,000 unseen expressions for evaluation.
  • Evaluation Sample: A random 500-expression sample from the test set was used for manual evaluation.
  • Example Database: 14,000 example pairs (words to sentences) used for example-based transfer.

The paper introduces a corpus specifically designed for the "travel domain," aiming to capture a wide variety of linguistic phenomena, including simplified input patterns observed in human-machine communication.

Key Insights and Ideas

  • Hybridization is Key: The core insight is that combining example-based flexibility with rule-based linguistic rigor is crucial for achieving high-quality machine translation that is both idiomatic and grammatically sound.
  • Colloquial Language Challenges: The "travel domain" highlights that real-world language is often simplified by users, requiring MT systems to be robust to shorter utterances and less formal structures.
  • Efficient Feature Structure Management: The GPL infrastructure's focus on efficient feature structure representation and memory management is critical for the performance of complex NLP tasks.
  • Information Content for Semantics: Using information content from a thesaurus for lexical similarity offers a more nuanced approach than simple word distance.

Key Results

  • Translation Quality: 84% of translations conveyed the meaning in an acceptable manner.
  • Flawless: 60%
    • Stylistic Problem: 9%
    • Minor Problem: 14%
    • Acceptable with OOV: 1%
    • Major Problem: 9%
    • Wrong Translation: 5%
    • Translation Failure: 3%
  • Computational Resources:
    • Average translation speed: 0.44 seconds on a Pentium III 500 MHz.
  • Memory requirements are detailed in Table 2, with read-only memory for code/data at 6MB and dictionary/examples at 23MB, and read/write memory for feature structures at 14MB.

Limitations and Future Work

  • Limitations: The paper acknowledges that while the system has good coverage, there are still instances of "Translation Failure" (3%) and "Wrong Translation" (5%), indicating areas where the rule-based component's coverage is insufficient. The current system is limited to English-to-Japanese.
  • Future Work:
    • Extending the input vocabulary size.
  • Developing mechanisms for closer integration with speech recognition and speech synthesis for speech-to-speech translation.
  • Working on Japanese-to-English translation direction.

Summary updated

BibTeX
@inproceedings{franz2000integrated,
  author = {Alexander Franz and Keiko Horiguchi and Lei Duan and Doris Ecker and Eugene Koontz and Kazami Uchida},
  title = {{An Integrated Architecture for Example-Based Machine Translation}},
  booktitle = {International Conference on Computational Linguistics},
  year = {2000},
}
Figure