English–Welsh Cross-Lingual Embeddings

Cross-lingual embeddings are vector space representations where word translations tend to be co-located. These enable learning transfer across languages, thus bridging the gap between data-rich languages such as English and others. In this paper, we present evaluate a suite of cross-lingual for English–Welsh language pair. To train bilingual embeddings, Welsh corpus approximately 145 M words was combined with an Wikipedia corpus. We used dictionary frame problem mappings supervised machine task, is first learned independently on monolingual corpus, after which linear alignment strategy applied map common space. Two approaches were learn including word2vec fastText. Three cross-language strategies explored, cosine similarity, inverted softmax cross-domain similarity local scaling (CSLS). evaluated different combinations these using two tasks, induction, sentiment analysis. The best results achieved fastText CSLS metric. also demonstrated that by few automatically translated training documents, performance text classifier can increase 20 percent points.