Graph Neural Networks for Multiparallel Word Alignment

Anonymous

Abstract: After a period of decrease, interest in word alignments is increasing again for their usefulness in domains such as typological research, cross-lingual annotation projection and machine translation. Generally, alignment algorithms only use bitext and do not make use of the fact that many parallel corpora are multiparallel. We propose to use graph neural networks (GNNs) and community detection algorithms to exploit the graph structure of multiparallel word alignments. Our GNN approach (i) utilizes information about the meaning, position and language of the input words, (ii) incorporates information from multiple parallel sentences, (iii) can remove edges from the initial alignments, and (iv) provides a prediction model that can generalize beyond the sentences it is trained on. We show that community detection algorithms can provide valuable information for multiparallel word alignment. We show on three word alignment datasets and on a downstream task that our method outperforms previous work.