Synthesizing Compound Words for Machine Translation

Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics, pages 1085–1094, Berlin, Germany, August 7-12, 2016. c⃝2016 Association for Computational Linguistics Synthesizing Compound Words for Machine Translation Austin Matthews and Eva Schlinger and Alon Lavie and Chris Dyer Language Technologies Institute Carnegie Mellon University Pittsburgh, PA 15213, USA {austinma,eschling,alavie,cdyer}@cs.cmu.edu Abstract Most machine translation systems con- struct translations from a closed vocabu- lary of target word forms, posing problems for translating into languages that have productive compounding processes. We present a simple and effective approach that deals with this problem in two phases. First, we build a classifier that identifies spans of the input text that can be trans- lated into a single compound word in the target language. Then, for each identi- fied span, we generate a pool of possible compounds which are added to the trans- lation model as “synthetic” phrase trans- lations. Experiments reveal that (i) we can effectively predict what spans can be compounded; (ii) our compound gener- ation model produces good compounds; and (iii) modest improvements are pos- sible in end-to-end English–German and English–Finnish translation tasks. We ad- ditionally introduce KomposEval, a new multi-reference dataset of English phrases and their translations into German com- pounds. 1 Introduction Machine translation systems make a closed- vocabulary assumption: with the exception of ba- sic rules for copying unknown word types from the input to the output, they can produce words in the target language only from a fixed, finite vo- cabulary. While this is always a naïve assumption given the long-tailed distributions that character- ize natural language, it is particularly challenging in languages such as German and Finnish that have productive compounding processes. In such languages, expressing compositions of basic concepts can require an unbounded num- ber of words. For example, English multiword phrases like market for bananas, market for pears, and market for plums are expressed in German with single compound words (respectively, as Ba- nanenmarkt, Birnenmarkt, and Pflaumenmarkt). Second, while they are individually rare, com- pound words are, on the whole, frequent in native texts (Baroni et al., 2002; Fritzinger and Fraser, 2010). Third, compounds are crucial for transla- tion quality. Not only does generating them make the output seem more natural, but they are content- rich. Since each compound has, by definition, at least two stems, they are intuitively (at least) dou- bly important for translation adequacy. Fortunately, compounding is a relatively regular process (as the above examples also illustrate), and it is amenable to modeling. In this paper we intro- duce a two-stage method (§2) to dynamically gen- erate novel compound word forms given a source language input text and incorporate these as “syn- thetic rules” in a standard phrase-based transla- tion system (Bhatia et al., 2014; Chahuneau et al., 2013; Tsvetkov et al., 2013). First, a binary classi- fier examines each source-language sentence and labels each span therein with whether that span could become a compound word when translated into the target language. Second, we transduce the identified phrase into the …

Content truncated to 500 words