Supervised Phrase Table Triangulation with Neural Word Embeddings for Low-Resource Languages

Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 1079–1083, Lisbon, Portugal, 17-21 September 2015. c⃝2015 Association for Computational Linguistics. Supervised Phrase Table Triangulation with Neural Word Embeddings for Low-Resource Languages Tomer Levinboim and David Chiang Department of Computer Science and Engineering University of Notre Dame {levinboim.1,dchiang}@nd.edu Abstract In this paper, we develop a supervised learning technique that improves noisy phrase translation scores obtained by phrase table triangulation. In particular, we extract word translation distributions from small amounts of source-target bilin- gual data (a dictionary or a parallel corpus) with which we learn to assign better scores to translation candidates obtained by trian- gulation. Our method is able to gain im- provement in translation quality on two tasks: (1) On Malagasy-to-French transla- tion via English, we use only 1k dictionary entries to gain +0.5 Bleuover triangula- tion. (2) On Spanish-to-French via English we use only 4k sentence pairs to gain +0.7 Bleuover triangulation interpolated with a phrase table extracted from the same 4k sentence pairs. 1 Introduction Phrase-based statistical machine translation sys- tems require considerable amounts of source- target parallel data to produce good quality trans- lation. However, large amounts of parallel data are available for only a fraction of language pairs, and mostly when one of the languages is English. Phrase table triangulation (Utiyama and Isa- hara, 2007; Cohn and Lapata, 2007; Wu and Wang, 2007) is a method for generating source- target phrase tables without having access to any source-target parallel data. The intuition behind triangulation (and pivoting techniques in general) is the transitivity of translation: if a source lan- guage phrase s translates to a pivot language phrase p which in turn translates to a target lan- guage phrase t, then s should likely translate to t. Following this intuition, a triangulated source- target phrase table ˆT can be composed from a source-pivot and pivot-target phrase table (§2). However, the resulting triangulated phrase table ˆT contains many spurious phrase pairs and noisy probability estimates. Therefore, early triangula- tion work (Wu and Wang, 2007) already realis- tically assumed access to a limited source-target parallel data from which a relatively high-quality source-target phrase table T can be directly esti- mated. The two phrase tables were then combined, resulting in a higher quality phrase table that pro- poses translations for many source phrases not found in T. Wu and Wang (2007) report that in- terpolation of the two phrase tables T and ˆT leads to higher quality translations. However, the trian- gulated phrase table ˆT is obtained without using the source-target bilingual data, which suggests that the source-target data is not used as fully as it could be. In this paper, we develop a supervised learning algorithm that corrects triangulated word transla- tion probabilities by relying on word translation distributions wsup derived from the limited source- target data. In particular, we represent source and target words using word embeddings (Mikolov et al., 2013) and learn a transformation between the two embedding spaces in order to approxi- …

Content truncated to 500 words