A Multilingual Dataset for Evaluating Parallel Sentence Extraction from Comparable Corpora
Résumé
Comparable corpora can be seen as a reservoir for parallel sentences and phrases to overcome limitations in variety and quantity encountered in existing parallel corpora. This has motivated the design of methods to extract parallel sentences from comparable corporad. Despite this interest and work, no shared dataset has been made available for this task until the 2017 BUCC Shared Task. We present the challenges faced to build such a dataset and the solutions adopted to design and create the 2017 BUCC Shared Task dataset, emphasizing issues we had to cope with to include Chinese as one of the languages. The resulting corpus contains a total of about 3.5 million distinct sentences in English, French, German, Russian, and Chinese, mostly from Wikipedia. We illustrate the use of this dataset in the shared task and summarize the main results obtained by its participants. We finally outline remaining issues.