A Joint Dependency Model of Morphological and Syntactic Structure for Statistical Machine Translation

Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 2081–2087, Lisbon, Portugal, 17-21 September 2015. c⃝2015 Association for Computational Linguistics. A Joint Dependency Model of Morphological and Syntactic Structure for Statistical Machine Translation Rico Sennrich and Barry Haddow School of Informatics, University of Edinburgh rico.sennrich@ed.ac.uk, bhaddow@inf.ed.ac.uk Abstract When translating between two languages that differ in their degree of morpholog- ical synthesis, syntactic structures in one language may be realized as morphologi- cal structures in the other, and SMT mod- els need a mechanism to learn such trans- lations. Prior work has used morpheme splitting with flat representations that do not encode the hierarchical structure be- tween morphemes, but this structure is rel- evant for learning morphosyntactic con- straints and selectional preferences. We propose to model syntactic and morpho- logical structure jointly in a dependency translation model, allowing the system to generalize to the level of morphemes. We present a dependency representation of German compounds and particle verbs that results in improvements in transla- tion quality of 1.4–1.8 BLEU in the WMT English–German translation task. 1 Introduction When translating between two languages that dif- fer in their degree of morphological synthesis, syntactic structures in one language may be re- alized as morphological structures in the other. Machine Translation models that treat words as atomic units have poor learning capabilities for such translation units, and morphological segmen- tations are commonly used (Koehn and Knight, 2003). Like words in a sentence, the morphemes of a word have a hierarchical structure that is rel- evant in translation. For instance, compounds in Germanic languages are head-final, and the head is the segment that determines agreement within the noun phrase, and is relevant for selectional prefer- ences of verbs. 1. sie erheben eine Hand|gepäck|gebühr. function/postion English/German example finite (main) he walks away quickly er geht schnell weg finite (sub.) [...] because he walks away quickly [...] weil er schnell weggeht bare infinitive he can walk away quickly er kann schnell weggehen to/zu-infinitive he promises to walk away quickly er verspricht, schnell wegzugehen Table 1: Surface realizations of particle verb weggehen ’walk away’. they charge a carry-on bag fee. In example 1, agreement in case, number and gender is enforced between eine ’a’ and Gebühr ’fee’, and selectional preference between erheben ’charge’ and Gebühr ’fee’. A flat representation, as is common in phrase-based SMT, does not en- code these relationships, but a dependency repre- sentation does so through dependency links. In this paper, we investigate a dependency rep- resentation of morphologically segmented words for SMT. Our representation encodes syntactic and morphological structure jointly, allowing a single model to learn the translation of both. Specifi- cally, we work with a string-to-tree model with GHKM-style rules (Galley et al., 2006), and a relational dependency language model (Sennrich, 2015). We focus on the representation of German syntax and morphology in an English-to-German system, and two morphologically complex word classes in German that are challenging for transla- tion, compounds and particle verbs. German makes heavy use of …

Content truncated to 500 words