Towards a Wikipedia-extracted alpine corpus

Abstract

This paper describes a method for extracting parallel sentences from comparable texts. We present the main challenges in creating a German-French corpus for the Alpine domain. We demonstrate that it is difficult to use the Wikipedia categorization for the extraction of domain-specific articles from Wikipedia, therefore we introduce an alternative information retrieval approach. Sentence alignment algorithms were used to identify semantically equivalent sentences across the Wikipedia articles. Using this approach, we create a corpus of sentence-aligned Alpine texts, which is evaluated both manually and automatically. Results show that even a small collection of extracted texts (approximately 10000 sentence pairs) can partially improve the performance of a state-of-the-art statistical machine translation system. Thus, the approach is worth pursuing on a larger scale, as well as for other language pairs and domains.