A Method for Distinguishing Exceptional and General Examples in Example-based Transfer Systems

Hideo WatanabeCOLING 1994

TL;DR

This paper tackles the "example interference" problem in example-based machine translation systems, where specific, idiomatic translation patterns can be wrongly selected over more general ones due to similarity matching. The authors propose a novel method to distinguish between general and exceptional translation patterns within the system's existing framework, allowing for more accurate translation.

What Problem Does This Paper Address?

Example-based transfer systems (EBTS) rely on a large collection of translation examples (patterns) to perform translations. A significant challenge arises when these systems uniformly treat all examples, leading to "example interference." This occurs when a general translation pattern should be used, but a more specific, exceptional pattern is chosen instead because it appears more similar to the input. This can result in awkward or incorrect translations. For instance, mistaking "I use a car" for "I share the use of a car" because the verbs are semantically similar and the overall structure matches an exceptional pattern. The paper aims to solve this by enabling the system to differentiate between general and exceptional patterns, ensuring the correct pattern is selected based on context and specificity.

Methods

The core of the proposed solution lies in how the system calculates similarity and identifies exceptional patterns.

  • Similarity Calculation: Sim Tran, the system used, calculates similarity based on both structural similarity and the similarity of lexical forms. For general translation patterns, lexical form similarity is determined using a Japanese thesaurus (Bunrui-Goi-Hyou) which assigns numerical codes to words. The distance between two words is the difference in their codes.
  • Distinguishing Exceptional Patterns: The key innovation is a mechanism to flag exceptional patterns. This is achieved by examining the lexical forms within the translation patterns.
  • General Patterns: Lexical forms are enclosed in double quotes (e.g., "computer").
  • Exceptional Patterns: At least one lexical form is enclosed in single quotes (e.g., 'kyouyousuru'). This signals that this specific lexical form requires an exact match, rather than a similarity-based match.
  • Algorithm for Identification: A multi-step algorithm is proposed to classify translation patterns:
  • Group equivalent translation patterns.
  • Mark patterns as "general" if they are equivalent to patterns in other groups, excluding root node matches.
  • Identify "extra-exceptional" patterns if their target parts have extra elements compared to equivalent general patterns.
  • Identify "intra-exceptional" patterns if they are similar to general patterns but differ in their root target node.

Datasets

The paper mentions using a Japanese-to-English transfer dictionary that was previously used in their laboratory for experiments. No new datasets were introduced.

Key Insights and Ideas

  • Lexical Form as a Signal: Using single quotes for lexical forms in exceptional patterns is a simple yet effective way to signal the need for exact matching, thereby distinguishing them from general patterns that rely on semantic similarity.
  • Integrating Exceptionality: The method integrates the handling of exceptional patterns directly into the similarity calculation framework, avoiding the need for separate pre- or post-processing steps. This makes the system more streamlined.
  • System-Specific Idiomaticity: The definition of "exceptional" is system-centric. It identifies patterns that might cause side-effects within the current system's translation pattern base, rather than strictly adhering to linguistic notions of idiomaticity. This is a pragmatic approach for improving system performance.
  • Bootstrapping Potential: The method can help bootstrap a translation pattern base by identifying potentially general patterns, which can then be used to refine the identification of exceptional ones.

Key Results

The experiments involved applying the algorithm to a Japanese-to-English transfer dictionary. The results showed that in most categories, over 90% of translation patterns were identified as exceptional. The authors interpret this high percentage not as a flaw in the algorithm, but as an indication that the previously used dictionary primarily contained exceptional cases that were not handled by default procedures. This suggests the algorithm is effective at identifying such patterns.

Limitations and Future Work

  • Data Dependency: The algorithm's effectiveness is dependent on having a sufficiently large number of translation patterns. If not enough patterns are available, exceptional patterns might not be identified correctly.
  • Human Intuition vs. System Needs: The paper acknowledges that the method's classification of exceptional patterns might not always align with human intuition about idiomatic expressions. It prioritizes patterns that are "idiomatic for the system" (i.e., those that cause side-effects).
  • Future Work: The authors suggest that deeper semantic processing might be needed to identify translation patterns that are idiomatic in the conventional linguistic sense. They also note that the classification of patterns into general, exceptional (extra and intra), and neutral categories provides a useful framework for managing translation pattern bases.

Summary updated

BibTeX
@inproceedings{watanabe1994method,
  author = {Hideo Watanabe},
  title = {{A Method for Distinguishing Exceptional and General Examples in Example-based Transfer Systems}},
  booktitle = {International Conference on Computational Linguistics},
  year = {1994},
}

Loading related content...

Figure