Semantic textual similarity based on deep learning: Can it improve matching and retrieval for Translation Memory tools?

This study proposes an original methodology to underpin the operation of new generation Translation Memory (TM) systems where the translations to be retrieved from the TM database are matched not on the basis of Levenshtein (edit) distance but by employing innovative Natural Language Processing (NLP) and Deep Learning (DL) techniques. Three DL sentence encoders were experimented with to retrieve TM matches in English-Spanish sentence pairs from the DGT TM dataset. Each sentence encoder was compared with Okapi which uses edit distance to retrieve the best match. The automatic evaluation shows the benefit of the DL technology for TM matching and holds promise for the implementation of the TM tool itself, which is our next project.