Parallel Data Extraction using Word Embeddings

Abstract

Building a robust MT system requires a sufficiently large parallel corpus to be available as training data. In this paper, we propose to automatically extract parallel sentences from comparable corpora without using any MT system or even any parallel corpus at all. Instead, we use crosslingual information retrieval (CLIR), average word embeddings, text similarity and a bilingual dictionary, thus saving a significant amount of time and effort as no MT system is involved in this process. We conduct experiments on two different kinds of data: (i) formal texts from news domain, and (ii) user-generated content (UGC) from hotel reviews. The automatically extracted sentence pairs are then added to the already available parallel training data and the extended translation models are built from the concatenated data sets. Finally, we compare the performance of our new extended models against the baseline models built from the available data. The experimental evaluation reveals that our proposed approach is capable of improving the translation outputs for both the formal texts and UGC.