Word-embedding based bilingual terminology alignment

Abstract

The ability to accurately align concepts between languages can provide significant benefits in many practical applications. In this paper, we extend a machine learning approach using dictionary and cognate-based features with novel cross-lingual embedding features using pretrained fastText embeddings. We use the tool VecMap to align the embeddings between Slovenian and English and then for every word calculate the top 3 closest word embeddings in the opposite language based on cosine distance. These alignments are then used as features for the machine learning algorithm. With one configuration of the input parameters, we managed to improve the overall F-score compared to previous work, while another configuration yielded improved precision (96%) at a cost of lower recall. Using embedding-based features as a replacement for dictionary-based features provides a significant benefit: while a large bilingual parallel corpus is required to generate the Giza++ word alignment lists, no such data is required for embedding-based features where the only required inputs are two unrelated monolingual corpora and a small bilingual dictionary from which the embedding alignments are calculated.