Linguistically Annotated BTG for Statistical Machine Translation

Proceedings of the 22nd International Conference on Computational Linguistics (Coling 2008), pages 1009–1016 Manchester, August 2008 Linguistically Annotated BTG for Statistical Machine Translation Deyi Xiong, Min Zhang, Aiti Aw and Haizhou Li Human Language Technology Institute for Infocomm Research 21 Heng Mui Keng Terrace, Singapore 119613 {dyxiong, mzhang, aaiti}@i2r.a-star.edu.sg Abstract Bracketing Transduction Grammar (BTG) is a natural choice for effective integration of desired linguistic knowledge into sta- tistical machine translation (SMT). In this paper, we propose a Linguistically Anno- tated BTG (LABTG) for SMT. It conveys linguistic knowledge of source-side syn- tax structures to BTG hierarchical struc- tures through linguistic annotation. From the linguistically annotated data, we learn annotated BTG rules and train linguisti- cally motivated phrase translation model and reordering model. We also present an annotation algorithm that captures syntac- tic information for BTG nodes. The ex- periments show that the LABTG approach significantly outperforms a baseline BTG- based system and a state-of-the-art phrase- based system on the NIST MT-05 Chinese- to-English translation task. Moreover, we empirically demonstrate that the proposed method achieves better translation selec- tion and phrase reordering. 1 Introduction Formal grammar used in statistical machine trans- lation (SMT), such as Bracketing Transduction Grammar (BTG) proposed by (Wu, 1997) and the synchronous CFG presented by (Chiang, 2005), provides a natural platform for integrating lin- guistic knowledge into SMT because hierarchical structures produced by the formal grammar resem- ble linguistic structures.1 Chiang (2005) attempts to integrate linguistic information into his formally c⃝ 2008. Licensed under the Creative Commons Attribution-Noncommercial-Share Alike 3.0 Unported li- cense (http://creativecommons.org/licenses/by-nc-sa/3.0/). Some rights reserved. 1We inherit the definitions of formal and linguistic from (Chiang, 2005) which makes a distinction between formally syntax-based SMT and linguistically syntax-based SMT. syntax-based system by adding a constituent fea- ture. Unfortunately, the linguistic feature does not show significant improvement on the test set. In this paper, we further this effort by integrating lin- guistic knowledge into BTG. We want to augment BTG’s formal structures with linguistic structures since they are both hier- archical. In particular, our goal is to learn a more linguistically meaningful BTG from real-world bi- texts by projecting linguistic structures onto BTG formal structures. In doing so, we hope to (1) maintain the strength of phrase-based approach since phrases are still used on BTG leaf nodes; (2) obtain a tight integration of linguistic knowledge in the translation model; (3) and finally avoid induc- ing a complicated linguistic synchronous grammar with expensive computation. The challenge, of course, is that BTG hierarchical structures are not always aligned with the linguistic structures in the syntactic parse trees of source or target language. Along this line, we propose a novel approach: Linguistically Annotated BTG (LABTG) for SMT. The LABTG annotates BTG rules with linguistic elements that are learned from syntactic parse trees on the source side through an annotation algo- rithm, which is capable of labelling both syntactic and non-syntactic phrases. The linguistic elements extracted from parse trees capture both internal lexical content and external context of phrases. With these …

Content truncated to 500 words