Extending CCG-based Syntactic Constraints in Hierarchical Phrase-Based SMT

Extending CCG-based Syntactic Constraints in Hierarchical Phrase-Based SMT Hala Almaghout CNGL, School of Computing Dublin City University Dublin, Ireland halmaghout@computing.dcu.ie Jie Jiang Applied Language Solutions Delph UK jie.jiang@appliedlanguage.com Andy Way Applied Language Solutions Delph UK andy.way@appliedlanguage.com Abstract In this paper, we describe two approaches to extending syntactic constraints in the Hierarchical Phrase-Based (HPB) Statis- tical Machine Translation (SMT) model using Combinatory Categorial Grammar (CCG). These extensions target the limita- tions of previous syntax-augmented HPB SMT systems which limit the coverage of the syntactic constraints applied. We present experiments on Arabic–English and Chinese–English translation. Our ex- periments show that using extended CCG labels helps to increase nonterminal la- bel coverage and achieve significant im- provements over the baseline for Arabic– English translation. In addition, com- bining extended CCG labels with CCG- augmented glue grammar helps to improve the performance of the Chinese–English translation over the baseline systems. 1 Introduction Hierarchical Phrase-Based (HPB) Statistical Ma- chine Translation (SMT) (Chiang, 2005) has been demonstrated to be one of the most successful SMT approaches nowadays. Its main idea is to imitate Context-Free Grammar (CFG) produc- tion rules in modelling translation rules while maintaining the strength of statistically extracted phrases. However, HPB SMT only models the hi- erarchical aspect of the language and does not use any linguistic information in rule extraction. set of approaches (Zollmann and Venugopal, 2006;A Almaghout et al., 2010) have tried to incorporate c⃝ 2012 European Association for Machine Translation. syntactic information extracted according to differ- ent grammar theories in the HPB SMT model by annotating phrases and nonterminals with syntac- tic labels. These systems face many challenges in integrating their syntax-based constraints with the syntax-free statistically extracted HPB SMT trans- lation grammar, which limits the coverage of these syntactic constraints and thus minimizes the bene- fit obtained from applying them. In this paper, we try to extend the scope of target-side syntactic constraints in syntax- augmented HPB SMT. More specifically, we try to exploit the flexibility of Combinatory Categorial Grammar (CCG) (Steedman, 2000) to increase the coverage of syntactic labels used to label phrases and nonterminals in hierarchical rules. In addition, we augment HPB glue grammar rules with CCG combinatory operators with the aim of directing the decoding process towards building a full parse tree of the translation output. We apply these con- straints in a soft manner through a feature in the log-linear model (Venugopal et al., 2009). The rest of this paper is organized as follows. Section 2 reviews related work. Section 3 gives an introduction to HPB SMT. Section 4 introduces CCG. Section 5 describes our approach. Section 6 presents our experiments. Finally, Section 7 con- cludes and provides avenues for future work. 2 Related Work Syntax Augmented Machine Translation (SAMT) (Zollmann and Venugopal, 2006) tries to improve the grammaticality of the HPB SMT translation output by attaching syntactic labels to target-side phrases and nonterminals. These labels are extracted from context-free phrase structure grammar parse trees of the target- side of the parallel corpus. The function of …

Content truncated to 500 words