A Dependency-Constrained Hierarchical Model with Moses

Proceedings of the Eighth Workshop on Statistical Machine Translation, pages 464–471, Sofia, Bulgaria, August 8-9, 2013 c⃝2013 Association for Computational Linguistics A Dependency-Constrained Hierarchical Model with Moses Yvette Graham†‡ †Department of Computing and Information Systems, The University of Melbourne ‡Centre for Next Generation Localisation, Dublin City University ygraham@unimelb.edu.au Abstract This paper presents a dependency- constrained hierarchical machine transla- tion model that uses Moses open-source toolkit for rule extraction and decoding. Experiments are carried out for the German-English language pair in both di- rections for projective and non-projective dependencies. We examine effects on SCFG size and automatic evaluation results when constraints are applied with respect to projective or non-projective dependency structures and on the source or target language side. 1 Introduction A fundamental element of natural language syntax is the dependency structure encoding the binary asymmetric head-dependent relations captured in dependency grammar theory. A main criteria for determining the dependency structure of a given sentence is the following: The linear position of dependent, D, is specified with reference to its head, H (K¨ubler et al., 2009). This runs in parallel with that which hierarchical machine translation SCFG rules encode: The linear position of a trans- lated phrase, Xi, is specified with reference to the lexicalised words in the rule. Figure 1 shows de- She lives � � �in � a white house � � . Ella vive � � �en � una casa � � blanca . Figure 1: Projective Dependency Structures pendency structures for She lives in a white house and its Spanish translation, with example SCFG (1) X →< white house , casa blanca > (2) X →< white , blanca > (3) X →< house , casa > (4) X →< X0 house , casa X0 > (5) X →< white X0 , X0 blanca > Figure 2: Initial rules (1), (2) and (3), with hierar- chical rules (4) and (5) rules shown in Figure 2. Given the existence of initial rules (1), (2) and (3), hierarchical rules (4) and (5) can be created. Rule (4) specifies the lin- ear position of the translation of the English phrase that precedes house with reference to lexicalised casa. For hierarchical machine translation models (Chiang, 2005), there is no requirement for a syn- tactic relationship to exist between the lexicalised words of a rule and the words replaced by non- terminals, the only requirement being that substi- tuted words form an SMT phrase (Koehn et al., 2003). The dependency structure of either the source or target (or indeed both) can, however, be used to constrain rule extraction as to only allow hierarchical rules in which the linear position of dependents are specified with reference to the po- sition of their lexicalised heads. For example, in the case of the hierarchical rules in Figure 2, rule (4) satisfies such a constraint according to both the source and target language dependency structures (since white is the dependent of house and blanca is the dependent of casa, and it is both white and blanca …

Content truncated to 500 words