Forest-based Tree Sequence to String Translation Model

Proceedings of the 47th Annual Meeting of the ACL and the 4th IJCNLP of the AFNLP, pages 172–180, Suntec, Singapore, 2-7 August 2009. c⃝2009 ACL and AFNLP Forest-based Tree Sequence to String Translation Model Hui Zhang1, 2 Min Zhang1 Haizhou Li1 Aiti Aw1 Chew Lim Tan2 1Institute for Infocomm Research 2National University of Singapore zhangh1982@gmail.com {mzhang, hli, aaiti}@i2r.a-star.edu.sg tancl@comp.nus.edu.sg Abstract This paper proposes a forest-based tree se- quence to string translation model for syntax- based statistical machine translation, which automatically learns tree sequence to string translation rules from word-aligned source- side-parsed bilingual texts. The proposed model leverages on the strengths of both tree sequence-based and forest-based translation models. Therefore, it can not only utilize forest structure that compactly encodes exponential number of parse trees but also capture non- syntactic translation equivalences with linguis- tically structured information through tree se- quence. This makes our model potentially more robust to parse errors and structure di- vergence. Experimental results on the NIST MT-2003 Chinese-English translation task show that our method statistically significantly outperforms the four baseline systems. 1 Introduction Recently syntax-based statistical machine trans- lation (SMT) methods have achieved very prom- ising results and attracted more and more inter- ests in the SMT research community. Fundamen- tally, syntax-based SMT views translation as a structural transformation process. Therefore, structure divergence and parse errors are two of the major issues that may largely compromise the performance of syntax-based SMT (Zhang et al., 2008a; Mi et al., 2008). Many solutions have been proposed to address the above two issues. Among these advances, forest-based modeling (Mi et al., 2008; Mi and Huang, 2008) and tree sequence-based modeling (Liu et al., 2007; Zhang et al., 2008a) are two interesting modeling methods with promising results reported. Forest-based modeling aims to improve translation accuracy through digging the potential better parses from n-bests (i.e. forest) while tree sequence-based modeling aims to model non-syntactic translations with structured syntactic knowledge. In nature, the two methods would be complementary to each other since they manage to solve the negative impacts of monolingual parse errors and cross-lingual struc- ture divergence on translation results from dif- ferent viewpoints. Therefore, one natural way is to combine the strengths of the two modeling methods for better performance of syntax-based SMT. However, there are many challenges in combining the two methods into a single model from both theoretical and implementation engi- neering viewpoints. In theory, one may worry about whether the advantage of tree sequence has already been covered by forest because forest encodes implicitly a huge number of parse trees and these parse trees may generate many differ- ent phrases and structure segmentations given a source sentence. In system implementation, the exponential combinations of tree sequences with forest structures make the rule extraction and decoding tasks much more complicated than that of the two individual methods. In this paper, we propose a forest-based tree sequence to string model, which is designed to integrate the strengths of the forest-based and the tree sequence-based modeling methods. We pre- sent our …

Content truncated to 500 words