Forced Derivation Tree based Model Training to Statistical Machine Translation

Proceedings of the 2012 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, pages 445–454, Jeju Island, Korea, 12–14 July 2012. c⃝2012 Association for Computational Linguistics Forced Derivation Tree based Model Training to Statistical Machine Translation Nan Duan Microsoft Research Asia nanduan@microsoft.com Mu Li Microsoft Research Asia muli@microsoft.com Ming Zhou Microsoft Research Asia mingzhou@microsoft.com Abstract A forced derivation tree (FDT) of a sentence pair {f, e} denotes a derivation tree that can translate f into its accurate target translation e. In this paper, we present an approach that leverages structured knowledge contained in FDTs to train component models for statistical machine translation (SMT) systems. We first describe how to generate different FDTs for each sentence pair in training corpus, and then present how to infer the optimal FDTs based on their derivation and alignment qualities. As the first step in this line of research, we verify the effectiveness of our approach in a BTG- based phrasal system, and propose four FDT- based component models. Experiments are carried out on large scale English-to-Japanese and Chinese-to-English translation tasks, and significant improvements are reported on both translation quality and alignment quality. 1 Introduction Most of today’s SMT systems depends heavily on parallel corpora aligned at the word-level to train their different component models. However, such annotations do have their drawbacks in training. On one hand, word links predicted by automatic aligners such as GIZA++ (Och and Ney, 2004) often contain errors. This problem gets even worse on lan- guage pairs that differ substantially in word orders, such as English and Japanese/Korean/German. The descent of the word alignment quality will lead to inaccurate component models straightforwardly. On the other hand, several component models are designed to supervise the decoding procedures, which usually rely on training examples extracted from word-aligned sentence pairs, such as distortion models (Tillman, 2004; Xiong et al., 2006; Galley and Manning, 2008) and sequence models (Banchs et al., 2005; Quirk and Menezes, 2006; Vaswani et al., 2011). Ideally, training examples of models are expected to match most of the situations that could be met in decoding procedures. But actually, plain structures of word alignments are too coarse to pro- vide enough knowledge to ensure this expectation. This paper presents an FDT-based model training approach to SMT systems by leveraging structured knowledge contained in FDTs. An FDT of a sen- tence pair {f, e} denotes a derivation tree that can translate f into its accurate target translation e. The principle advantage of this work is two-fold. First, using alignments induced from the 1-best FDTs of all sentence pairs, the overall alignment quality of training corpus can be improved. Second, compar- ing to word alignments, FDTs can provide richer structured knowledge for various component models to extract training instances. Our FDT-based mod- el training approach performs via three steps: (1) generation, where an FDT space composed of dif- ferent FDTs is generated for each sentence pair in training corpus by the forced decoding technique; (2) inference, where the …

Content truncated to 500 words