Learning Hierarchical Translation Spans

Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 183–188, October 25-29, 2014, Doha, Qatar. c⃝2014 Association for Computational Linguistics Learning Hierarchical Translation Spans Jingyi Zhang1,2, Masao Utiyama3, Eiichro Sumita3, Hai Zhao1,2 1Center for Brain-Like Computing and Machine Intelligence, Department of Computer Science and Engineering, Shanghai Jiao Tong Unviersity, Shanghai, 200240, China 2Key Laboratory of Shanghai Education Commission for Intelligent Interaction and Cognitive Engineering, Shanghai Jiao Tong Unviersity, Shanghai, 200240, China 3National Institute of Information and Communications Technology 3-5Hikaridai, Keihanna Science City, Kyoto, 619-0289, Japan zhangjingyizz@gmail.com, mutiyama/eiichiro.sumita@nict.go.jp, zhaohai@cs.sjtu.edu.cn Abstract We propose a simple and effective ap- proach to learn translation spans for the hierarchical phrase-based translation model. Our model evaluates if a source span should be covered by translation rules during decoding, which is integrated into the translation system as soft con- straints. Compared to syntactic con- straints, our model is directly acquired from an aligned parallel corpus and does not require parsers. Rich source side contextual features and advanced machine learning methods were utilized for this learning task. The proposed approach was evaluated on NTCIR-9 Chinese-English and Japanese-English translation tasks and showed significant improvement over the baseline system. 1 Introduction The hierarchical phrase-based (HPB) translation model (Chiang, 2005) has been widely adopted in statistical machine translation (SMT) tasks. The HPB translation rules based on the synchronous context free grammar (SCFG) are simple and pow- erful. One drawback of the HPB model is the appli- cations of translation rules to the input sentence are highly ambiguous. For example, a rule whose English side is “X1 by X2” can be applied to any word sequence that has “by” in them. In Figure 1, this rule can be applied to the whole sentence as well as to “experiment by tomorrow”. In order to tackle rule application ambiguities, a few previous works used syntax trees. Chi- ang (2005) utilized a syntactic feature in the HPB I will fnish this experiment by tomorrow 我 会 在 明天 之前 完成 这个 实验 Figure 1: A translation example. model, which represents if the source span cov- ered by a translation rule is a syntactic constituent. However, the experimental results showed this feature gave no significant improvement. Instead of using the undifferentiated constituency feature, (Marton and Resnik, 2008) defined different soft syntactic features for different constituent types and obtained substantial performance improve- ment. Later, (Mylonakis and Sima’an, 2011) in- troduced joint probability synchronous grammars to integrate flexible linguistic information. (Liu et al., 2011) proposed the soft syntactic constraint model based on discriminative classifiers for each constituent type and integrated all of them into the translation model. (Cui et al., 2010) focused on hierarchical rule selection using many features in- cluding syntax constituents. These works have demonstrated the benefits of using syntactic features in the HPB model. How- ever, high quality syntax parsers are not always easily obtained for many languages. Without this problem, word alignment constraints can also be used to guide the application of the rules. Suppose that we want to …

Content truncated to 500 words