Toward Better Chinese Word Segmentation for SMT via Bilingual Constraints

Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics, pages 1360–1369, Baltimore, Maryland, USA, June 23-25 2014. c⃝2014 Association for Computational Linguistics Toward Better Chinese Word Segmentation for SMT via Bilingual Constraints Xiaodong Zeng† Lidia S. Chao† Derek F. Wong† Isabel Trancoso‡ Liang Tian† †NLP2CT Lab / Department of Computer and Information Science, University of Macau ‡INESC-ID / Instituto Superior T´enico, Lisboa, Portugal nlp2ct.samuel@gmail.com, {lidiasc, derekfw}@umac.mo, isabel.trancoso@inesc-id.pt,tianliang0123@gmail.com Abstract This study investigates on building a better Chinese word segmentation mod- el for statistical machine translation. aims at leveraging word boundary infor-It mation, automatically learned by bilin- gual character-based alignments, to induce a preferable segmentation model. We propose dealing with the induced word boundaries as soft constraints to bias the continuous learning of a supervised CRF- s model, trained by the treebank data (la- beled), on the bilingual data (unlabeled). The induced word boundary information is encoded as a graph propagation con- straint. The constrained model induction is accomplished by using posterior reg- ularization algorithm. The experiments on a Chinese-to-English machine transla- tion task reveal that the proposed model can bring positive segmentation effects to translation quality. 1 Introduction Word segmentation is regarded as a critical pro- cedure for high-level Chinese language process- ing tasks, since Chinese scripts are written in con- tinuous characters without explicit word bound- aries (e.g., space in English). The empirical works show that word segmentation can be beneficial to Chinese-to-English statistical machine translation (SMT) (Xu et al., 2005; Chang et al., 2008; Zhao et al., 2013). In fact most current SMT models assume that parallel bilingual sentences should be segmented into sequences of tokens that are meant to be “words” (Ma and Way, 2009). The practice in state-of-the-art MT systems is that Chinese sen- tences are tokenized by a monolingual supervised word segmentation model trained on the hand- annotated treebank data, e.g., Chinese treebank (CTB) (Xue et al., 2005). These models are con- ducive to MT to some extent, since they common- ly have relatively good aggregate performance and segmentation consistency (Chang et al., 2008). But one outstanding problem is that these mod- els may leave out some crucial segmentation fea- tures for SMT, since the output words conform to the treebank segmentation standard designed for monolingually linguistic intuition, rather than spe- cific to the SMT task. In recent years, a number of works (Xu et al., 2005; Chang et al., 2008; Ma and Way, 2009; Xi et al., 2012) attempted to build segmentation models for SMT based on bilingual unsegment- ed data, instead of monolingual segmented data. They proposed to learn gainful bilingual knowl- edge as golden-standard segmentation supervi- sions for training a bilingual unsupervised mod- el. Frequently, the bilingual knowledge refers to the mappings of an individual English word to one or more consecutive Chinese characters, generat- ed via statistical character-based alignment. They leverage such mappings to either constitute a Chi- nese word dictionary for maximum-matching seg- mentation (Xu et al., 2004), or form labeled data for training a sequence labeling model …

Content truncated to 500 words