2020/06/20 by Antonio H. O. Fonseca, Fonseca, Antonio H. O., David van Dijk +1
Computer Science · #Computation and Language (cs.CL) #FOS: Computer and information sciences #Machine Learning (cs.LG) #Natural Language Processing Techniques #Text Readability and Simplification #Topic Modeling
paper · pdf · doi:10.48550/arxiv.2006.11578
openalex publication_date 2020/06/20 · openalex created_date 2022/07/26 · openalex updated_date 2026/07/28
Word translation is an integral part of language translation. In machine\ntranslation, each language is considered a domain with its own word embedding.\nThe alignment between word embeddings allows linking semantically equivalent\nwords in multilingual contexts. Moreover, it offers a way to infer\ncross-lingual meaning for words without a direct translation. Current methods\nfor word embedding alignment are either supervised, i.e. they require known\nword pairs, or learn a cross-domain transformation on fixed embeddings in an\nunsupervised way. Here we propose an end-to-end approach for word embedding\nalignment that does not require known word pairs. Our method, termed Word\nAlignment through MMD (WAM), learns embeddings that are aligned during sentence\ntranslation training using a localized Maximum Mean Discrepancy (MMD)\nconstraint between the embeddings. We show that our method not only\nout-performs unsupervised methods, but also supervised methods that train on\nknown word translations.\n