Building a Hierarchically Aligned Chinese-English Parallel Treebank

Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 1511–1520, Dublin, Ireland, August 23-29 2014. Building a Hierarchically Aligned Chinese-English Parallel Treebank Dun Deng and Nianwen Xue Computer Science Department, Brandeis University 415 South Street, Waltham MA, USA ddeng@brandeis.edu, xuen@brandeis.edu Abstract We construct a hierarchically aligned Chinese-English parallel treebank by manually doing word alignments and phrase alignments simultaneously on parallel phrase-based parse trees. The main innovation of our approach is that we leave words without a translation counterpart (which are mostly language-particular function words) unaligned on the word level, and locate and align the appropriate phrases which encapsulate them. In doing so, we harmonize word-level and phrase- level alignments. We show that this type of annotation can be performed with high inter-annotator consistency and have both linguistic and engineering potentials. 1 Introduction The value of human annotated syntactic structures for Statistical Machine Translation has been clearly demonstrated in string-to-tree (Galley et al., 2004; Galley et al., 2006; Huang et al., 2006), tree-to-string (Liu et al., 2006; Liu and Gildea, 2008), and tree-to-tree (Eisner, 2003; Liu et al., 2009; Chiang, 2010) models. One recurring issue which hampers the utility of syntactic structures is the incompatibility be- tween word alignments and syntactic structures (Denero and Klein, 2007; Fossum et al., 2008; Pauls et al., 2010). The incompatibility arises because word alignments and syntactic structures are established independently of each other. In the case of tree-to-tree models, there is also the issue of incompatible par- allel tree structures resulting from divergent syntactic annotation standards that have been independently conceived based on monolingual corpora (Chiang, 2010). In this paper, we report an effort in building a Hierarchically Aligned Chinese-English Parallel Treebank (HACEPT) where we manually do word- level and phrase-level alignments simultaneously on parallel phrase-based parse trees. In this process, we attempt to establish an annotation standard that harmonizes word-level and phrase-level alignments. We also analyze a common incompatibility issue between Chinese-English parallel parse trees exposed in the annotation process, with the goal of solving the issue by semi-automatically revising the trees. In the rest of this paper, we describe how we construct the HACEPT and discuss issues arising in the construction process. In Section 2, we discuss the problems of word alignment done without considering its interaction with syntactic structures. In Section 3, we describe our annotation procedure where we perform word-level and phrase-level alignments simultaneously in a coordinated manner, and show how our approach is free of the problems discussed in Section 2. In Section 4, we report a common incom- patibility issue between parse trees and propose a solution. We also compare the issue with translation divergence (Dorr, 1994) and show that they are different in nature and occurrence frequency. In Section 5, we present the results of two experiments we have done on our annotation to show the intuitiveness of our approach and the linguistic and engineering potentials of our corpus. We then describe related work in Section 6 and conclude our paper in Section …

Content truncated to 500 words