Augment Dependency-to-String Translation with Fixed and Floating Structures

Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 2217–2226, Dublin, Ireland, August 23-29 2014. Augment Dependency-to-String Translation with Fixed and Floating Structures Jun Xie† Jinan Xu‡ Qun Liu†§ †Key Laboratory of Intelligent Information Processing, Institute of Computing Technology,Chinese Academy of Sciences xiejun@ict.ac.cn ‡School of Computer and Information Technology, Beijing Jiaotong University xja2010@gmail.com §School of Computing, Dublin City University qliu@computing.dcu.ie Abstract In this paper, we propose an augmented dependency-to-string model to combine the merits of both the head-dependents relations at handling long distance reordering and the fixed and floating structures at handling local reordering. For this purpose, we first compactly represent both the head-dependent relation and the fixed and floating structures into translation rules; second, in decoding we build “on-the-fly” new translation rules from the compact translation rules that can incorporate non-syntactic phrases into translations, thus alleviate the non-syntactic phrase coverage problem of dependency-to-string translation (Xie et al., 2011). Large-scale experiments on Chinese-to-English translation show that our augmented dependency-to-string model gains significant improvement of averaged +0.85 BLEU scores on three test sets over the dependency- to-string model. 1 Introduction As a representation holding both syntactic and semantic information, dependency grammar has been attracting more and more attention in statistical machine translation. Lin (2004) took paths as the el- ementary structures and proposed a path-based transfer model. Quirk et al. (2005) extended path to treelets (connected subgraphs of dependency trees) and put forward dependency treelet translation. Ding and Palmer (2005) proposed a model on the basis of dependency insertion grammar. Shen et al. (2008) employed the fixed and floating structures as elementary structures and proposed a string-to-dependency model with state-of-the-art performance. Xie et al. (2011) employs head-dependents relations as elemen- tary structures and proposed a dependency-to-string model with good long distance reordering property. A head-dependents relation (HDR) is composed of a head and all its dependents, which can be viewed as an instance of a sentence pattern or phrase pattern. However, since dependency trees are much flatter than constituency trees, the dependency-to-string model suffers more severe non-syntactic phrase coverage problem (Meng et al., 2013) than constituency- based models (Galley et al., 2004; Liu et al., 2006; Huang et al., 2006). Non-syntactic phrases are those phrases that can not be covered by whole subtrees. To address this problem, Meng et al. (2013) proposed to translate with both constituency and dependency trees, which can incorporate non-syntactic phrases covered by the constituents of the constituency trees. This model requires both constituency and depen- dency trees, thus may suffer from both constituency and dependency parse errors. Additionally, there are only few languages that have both constituency and dependency parsers, which limits its practical use. In this paper, we propose to address non-syntactic phrase coverage problem of the dependency-to- string model without resort to extra resources (Section 3). To this end, we augment the dependency-to- string model at two aspects. First, we combine the merits of both the head-dependent relations and the fixed and floating structures (Shen et al., 2008), and …

Content truncated to 500 words