Influence of Part-of-Speech and Phrasal Category Universal Tag-set in Tree-to-Tree Translation Models

International Joint Conference on Natural Language Processing, pages 907–911, Nagoya, Japan, 14-18 October 2013. Influence of Part-of-Speech and Phrasal Category Universal Tag-set in Tree-to-Tree Translation Models Francisco Oliveira, Derek F. Wong, Lidia S. Chao, Liang Tian, Liangye He Department of Computer and Information Science, University of Macau, Macao SAR, China {olifran, derekfw, lidiasc}@umac.mo, {tianliang0123, wutianshui0515}@gmail.com Abstract Tree-to-tree Statistical Machine Transla- tion models require the use of syntactic tree structures of both the source and target side in learning rules to guide the trans- lation process. In order to accomplish the task, available treebanks for different languages are used as the main resources to collect necessary information to handle the translation task. However, since each treebank has its own defined tags, a bar- rier is inherently created in highlighting alignment relationships at different syn- tactic levels for different tag-sets. More- over, these models are typically over con- strained. This paper presents a unified tag- set for all languages at Part-of-Speech and Phrasal Category level in tree-to-tree mod- els. Different experiments are conducted to study for its feasibility, efficiency, and translation quality. 1 Introduction The study of Statistical Machine Translation (SMT) (Lopez, 2008) relying on syntactic infor- mation has received wide attention in recent years. In particular, syntactic information is being inte- grated either on the source or target or both side(s) in training translation models for handling the translation task. In hierarchical models (Chiang, 2007) that consider syntactic information (Zoll- mann and Venugopal, 2006), the input sentence is analyzed and translated by synchronous con- text free grammars (SCFG) hierarchically with ex- tra linguistic information. In string-to-tree SMT models (Galley et al., 2004; Zhang et al., 2011), the output of the translation always follows a grammatical syntax of the target language. In tree- to-string SMT models (Liu et al., 2006; Wu et al., 2010), source side syntax is used to generate the translation output. Finally, by considering the syn- tax of both the source and target languages, tree- to-tree SMT models (Zhang et al., 2008; Liu et al., 2009) tend to be the best among the previous mod- els. Basically, all of these models require two extra components: (1) syntax parsers (He et al., 2012; Petrov et al., 2006) in obtaining annotated syntax trees for training the models, and (2) monolingual treebanks (a detailed list can be found in Petrov et al. (2012)) for training the parsers. Currently, many of them are publicly available through Inter- net, institutions and data consortiums. Independently from the method used, although there are many treebanks available, they typically have their own tag-set defined for different lan- guages, ranging from tens to hundreds of tags, which is hard to conduct the research in a mul- tilingual environment. As a consequence, Petrov et al. (2012) developed a universal Part-of-Speech (POS) tag-set for twenty five different languages. However, at phrasal level, disagreements between the languages remain undefined. This paper presents a study of the application of universal tag-set from POS to phrasal category level in tree-to-tree translation models. In …

Content truncated to 500 words