Discriminative Induction of Sub-Tree Alignment using Limited Labeled Data

Proceedings of the 23rd International Conference on Computational Linguistics (Coling 2010), pages 1047–1055, Beijing, August 2010 Discriminative Induction of Sub-Tree Alignment using Limited Labeled Data Jun Sun1,2 Min Zhang1 Chew Lim Tan2 1Institute for Infocomm Research 2School of Computing, National University of Singapore sunjun@comp.nus.edu.sg mzhang@i2r.a-star.edu.sg tancl@comp.nus.edu.sg Abstract We employ Maximum Entropy model to con- duct sub-tree alignment between bilingual phrasal structure trees. Various lexical and structural knowledge is explored to measure the syntactic similarity across Chinese-English bi- lingual tree pairs. In the experiment, we evalu- ate the sub-tree alignment using both gold standard tree bank and the automatically parsed corpus with manually annotated sub-tree align- ment. Compared with a heuristic similarity based method, the proposed method significant- ly improves the performance with only limited sub-tree aligned data. To examine its effective- ness for multilingual applications, we further at- tempt different approaches to apply the sub-tree alignment in both phrase and syntax based SMT systems. We then compare the performance with that of the widely used word alignment. Experimental results on benchmark data show that sub-tree alignment benefits both systems by relaxing the constraint of the word alignment. 1 Introduction Recent research in Statistical Machine Translation (SMT) tends to incorporate more linguistically grammatical information into the translation mod- el known as linguistically motivated syntax-based models. To develop such models, the phrasal structure parse tree is usually adopted as the repre- sentation of bilingual sentence pairs either on the source side (Huang et al., 2006; Liu et al., 2006) or on the target side (Galley et al., 2006; Marcu et al., 2006), or even on both sides (Graehl and Knight, 2004; Zhang et al., 2007). Most of the above models either construct a pipeline to trans- form from/to tree structure, or synchronously gen- erate two trees in parallel (i.e., synchronous pars- ing). Both cases require syntactically rich transla- tional equivalences to handle non-local reordering. However, most current works obtain the syntactic translational equivalences by initially conducting alignment on the word level. To employ word alignment as a hard constraint for rule extraction has difficulty in capturing such non-local phenom- ena and will fully propagate the word alignment error to the later stage of rule extraction. Alternatively, some initial attempts have been made to directly conduct syntactic structure alignment. As mentioned in Tinsley et al. (2007), the early work usually constructs the structure alignment by hand, which is time-consuming. Re- cent research tries to automatically align the bilin- gual syntactic sub-trees. However, most of these works suffer from the following problems. Firstly, the alignment is conducted based on heuristic rules, which may lose extensibility and generality in spite of accommodating some common cases (Groves et al., 2004). Secondly, various similarity computation methods are used based merely on lexical translation probabilities (Tinsley et al., 2007; Imamura, 2001) regardless of structural fea- tures. We believe the structure information is an important issue to capture the non-local structural divergence of languages by modeling beyond the plain text. To address the above issues, we present a statis- …

Content truncated to 500 words