TransSearch : un moteur de recherche de traductions

Julien Bourdaillet, Fabrizio Gotti, Stéphane Huet, Philippe Langlais, Guy LapalmeTALN (Demonstrations) 2010 · OpenReview

Abstract: Julien Bourdaillet, Fabrizio Gotti, Stéphane Huet, Philippe Langlais, Guy Lapalme. Actes de la 17e conférence sur le Traitement Automatique des Langues Naturelles. Démonstrations.

2010.

Introduction

In the professional translation workflow, bilingual concordancers are essential tools for identifying how terms have been previously translated in large corpora. However, traditional concordancers present a cognitive bottleneck: they return a raw list of sentence pairs (bitexts), requiring the translator to manually scan dozens or hundreds of entries to find the specific translation of a query. TransSearch addresses this "manual search" problem by evolving the concordancer into an intelligent "translation finder" that automatically identifies, aggregates, and ranks potential translations.

Methodology: Translation Spotting and Advanced Querying

TransSearch moves beyond simple string matching by integrating statistical word alignment models. The system employs translation spotting, using alignment algorithms to isolate the specific word or phrase in the target sentence that corresponds to the user's query.

To handle the nuances of natural language, the system introduces specific linguistic operators that enhance recall and precision:

  • Morphological Expansion (+): Appending a + to a word (e.g., abroger+) allows the system to search for all inflected forms of that lemma (e.g., abroge, abrogeait, abrogé), which is vital for morphologically rich languages.
  • The Ellipsis (..): This proximity operator allows for non-contiguous searches, finding instances where query terms are separated by up to five words (e.g., "raise their awareness").
Advanced query using morphological expansion (+) and ellipsis (..) operators for 'abroger+ .. loi+'.

User Interface and Aggregation

The core innovation of the TransSearch interface is the aggregation of spotted translations into a sidebar, ranked by frequency. This provides immediate statistical context, helping a translator choose the most appropriate register or common usage. For a query like "subject to," the system can identify dozens of different translations across hundreds of occurrences.

TransSearch interface showing aggregated translations for the query 'subject to'.

As shown in the interface, a search for "subject to" reveals that assujettis à is the most frequent translation (232 occurrences), followed by soumis à (103) and sous réserve (65). Users can click a specific translation in the sidebar to filter the results and view only the relevant contexts, significantly reducing manual effort.

Data and Results

The system was demonstrated using the Canadian Hansards, the official bilingual records of the Canadian Parliament. This high-volume, professional-quality corpus serves as a gold standard for English-French translation. The demonstration showcased the system's ability to handle complex, lemmatized, and non-contiguous queries (e.g., abroger+ .. loi+), which significantly expands the search recall compared to literal string matching. For the query "subject to," the system successfully identified 87 different translations across 693 occurrences, providing a synthesized view that is far more efficient than a raw list of sentences.

Limitations and Future Work

While powerful, the system's performance is tied to the quality of the underlying statistical word alignment models; errors in alignment can lead to incorrect translation suggestions. Furthermore, the quality of suggestions is dependent on the domain and quality of the underlying bitext. Future work involves refining these statistical models to improve the precision of translation spotting and expanding the tool's commercial availability through platforms like Terminotix.

Conclusion

TransSearch transforms translation retrieval from a simple retrieval task into a data mining exercise. By leveraging statistical alignment and a user-centric interface, it provides professional translators with a powerful tool to navigate large bitexts with unprecedented efficiency.

Summary updated

Abstract

Julien Bourdaillet, Fabrizio Gotti, Stéphane Huet, Philippe Langlais, Guy Lapalme. Actes de la 17e conférence sur le Traitement Automatique des Langues Naturelles. Démonstrations. 2010.
BibTeX
@inproceedings{bourdaillet2010transsearch,
  author = {Julien Bourdaillet and Fabrizio Gotti and Stéphane Huet and Philippe Langlais and Guy Lapalme},
  title = {{TransSearch : un moteur de recherche de traductions}},
  booktitle = {TALN (Demonstrations)},
  year = {2010},
  url = {https://openreview.net/forum?id=74vehuXuNpa},
}
Figure