Abstract
Fabrizio Gotti, Philippe Langlais, Claude Coulombe. Actes de la 13ème conférence sur le Traitement Automatique des Langues Naturelles. Posters.
2006.
Introduction
Translation Memories (TMs) are indispensable tools for professional translators, facilitating the reuse of previous translations. However, traditional TMs typically operate at the sentence level, offering limited assistance when exact or near-exact matches are unavailable. Sub-sentential TMs address this by decomposing sentences into smaller fragments. While this increases retrieval potential, it introduces the challenge of "information overload" or "flooding," where a large database provides numerous irrelevant suggestions. This paper proposes a method to integrate context through automatic domain detection to filter these suggestions, improving both quality and efficiency.
Methodology
The authors introduce a "Third Generation Translation Memory" (MT3G) comprising three core processes:
- Sub-sentential Retrieval: The system segments source text into fragments, excluding functional words. It utilizes word alignment data (generated via GIZA++) to perform "translation spotting," identifying the specific target-language counterparts for each source fragment.
- Contextual Filtering via Domain Detection: To mitigate noise, the system employs an Information Retrieval (IR) approach. The text-to-be-translated acts as a query against various sub-corpora. Using the Lucene search engine, a $tf \cdot idf$ similarity score is calculated, and the search is restricted to the most relevant sub-index (e.g., searching only the medical sub-index for a health-related text).
- Automated Evaluation: A simulator acts as an "artificial user," attempting to reconstruct reference translations by selecting the best suggestions from the TM. This allows for quantitative benchmarking using the $f2$ metric, which balances recall and precision.
Datasets
The study utilized nine diverse bilingual French-English corpora to simulate a multi-domain environment. These range from religious texts and technical manuals to legal and parliamentary proceedings.
Testing was conducted on 256 randomly selected sentences from each corpus, ensuring no overlap with the training data.
Results and Discussion
The integration of domain-based filtering yielded significant performance gains across multiple metrics.
Domain Detection Accuracy
The $tf \cdot idf$ approach proved highly effective at identifying the correct domain for a given text. The system consistently assigned the highest similarity scores to the "home" sub-index for each test corpus, validating the use of domain as a proxy for context.
Efficiency and Quality
- Processing Speed: The average execution time per sentence was reduced from 4.36 seconds to 1.70 seconds, representing a 61% speedup. In specialized cases like the BIBLE corpus, the speedup reached 97%.
- Suggestion Quality: The $f2$ metric improved by an average of 4.65%, indicating that filtering by domain successfully reduced noise and improved the relevance of the fragments presented to the user.
Conclusion
Identifying the domain of a text serves as a powerful proxy for context in sub-sentential translation memories. By restricting searches to relevant sub-indices, the proposed system significantly reduces processing time and improves the precision of translation suggestions. Future work will explore finer-grained context, such as paragraph-level detection, and more complex relevance scoring models that combine fragment frequency with domain similarity.
Summary updated