Fine-Grained Tree-to-String Translation Rule Extraction

Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 325–334, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics Fine-grained Tree-to-String Translation Rule Extraction Xianchao Wu† Takuya Matsuzaki† Jun’ichi Tsujii†‡∗ †Department of Computer Science, The University of Tokyo 7-3-1 Hongo, Bunkyo-ku, Tokyo 113-0033, Japan ‡School of Computer Science, University of Manchester ∗National Centre for Text Mining (NaCTeM) Manchester Interdisciplinary Biocentre, 131 Princess Street, Manchester M1 7DN, UK {wxc, matuzaki, tsujii}@is.s.u-tokyo.ac.jp Abstract Tree-to-string translation rules are widely used in linguistically syntax-based statis- tical machine translation systems. In this paper, we propose to use deep syntac- tic information for obtaining fine-grained translation rules. A head-driven phrase structure grammar (HPSG) parser is used to obtain the deep syntactic information, which includes a fine-grained description of the syntactic property and a semantic representation of a sentence. We extract fine-grained rules from aligned HPSG tree/forest-string pairs and use them in our tree-to-string and string-to-tree sys- tems. Extensive experiments on large- scale bidirectional Japanese-English trans- lations testified the effectiveness of our ap- proach. 1 Introduction Tree-to-string translation rules are generic and ap- plicable to numerous linguistically syntax-based Statistical Machine Translation (SMT) systems, such as string-to-tree translation (Galley et al., 2004; Galley et al., 2006; Chiang et al., 2009), tree-to-string translation (Liu et al., 2006; Huang et al., 2006), and forest-to-string translation (Mi et al., 2008; Mi and Huang, 2008). The algorithms proposed by Galley et al. (2004; 2006) are fre- quently used for extracting minimal and composed rules from aligned 1-best tree-string pairs. Deal- ing with the parse error problem and rule sparse- ness problem, Mi and Huang (2008) replaced the 1-best parse tree with a packed forest which com- pactly encodes exponentially many parses for tree- to-string rule extraction. However, current tree-to-string rules only make use of Probabilistic Context-Free Grammar tree fragments, in which part-of-speech (POS) or koroshita korosareta (active) (passive) VBN(killed) 6 (6/10,6/6) 4 (4/10,4/4) VBN(killed:active) 5 (5/6,5/6) 1 (1/6,1/4) VBN(killed:passive) 1 (1/4,1/6) 3 (3/4,3/4) Table 1: Bidirectional translation probabilities of rules, denoted in the brackets, change when voice is attached to “killed”. phrasal tags are used as the tree node labels. As will be testified by our experiments, we argue that the simple POS/phrasal tags are too coarse to re- flect the accurate translation probabilities of the translation rules. For example, as shown in Table 1, sup- pose a simple tree fragment “VBN(killed)” ap- pears 6 times with “koroshita”, which is a Japanese translation of an active form of “killed”, and 4 times with “korosareta”, which is a Japanese translation of a passive form of “killed”. Then, without larger tree fragments, we will more frequently translate “VBN(killed)” into “ko- roshita” (with a probability of 0.6). But, “VBN(killed)” is indeed separable into two fine- grained tree fragments of “VBN(killed:active)” and “VBN(killed:passive)”1. Consequently, “VBN(killed:active)” appears 5 times with “ko- roshita” and 1 time with “korosareta”; and “VBN(killed:passive)” appears 1 time with “ko- roshita” and 3 times with “korosareta”. Now, by attaching the voice information to “killed”, we are gaining a …

Content truncated to 500 words