CCG Contextual labels in Hierarchical Phrase-Based SMT

CCG Contextual Labels in Hierarchical Phrase-Based SMT Hala Almaghout, Jie Jiang, Andy Way CNGL, School of Computing Dublin City University Dublin, Ireland {halmaghout, jjiang, away}@computing.dcu.ie Abstract In this paper, we present a method to em- ploy target-side syntactic contextual infor- mation in a Hierarchical Phrase-Based sys- tem. Our method uses Combinatory Cate- gorial Grammar (CCG) to annotate train- ing data with labels that represent the left and right syntactic context of target-side phrases. These labels are then used to as- sign labels to nonterminals in hierarchical rules. CCG-based contextual labels help to produce more grammatical translations by forcing phrases which replace nonter- minals during translations to comply with the contextual constraints imposed by the labels. We present experiments which ex- amine the performance of CCG contextual labels on Chinese–English and Arabic– English translation in the news and speech expressions domains using different data sizes and CCG-labeling settings. Our ex- periments show that our CCG contextual labels-based system achieved a 2.42% rel- ative BLEU improvement over a Phrase- Based baseline on Arabic–English transla- tion and a 1% relative BLEU improvement over a Hierarchical Phrase-Based system baseline on Chinese–English translation. 1 Introduction After the successful introduction of Hierarchical Phrase-Based (HPB) MT system (Chiang, 2005), many approaches tried to provide the HPB model with syntactic information extracted from the tar- get side of the training corpus in order to improve translation quality. Methods such as (Zollmann c⃝ 2011 European Association for Machine Translation. and Venugopal, 2006; Almaghout et al., 2010) augment nonterminals in hierarchical rules with syntactic labels extracted from target-side parse trees of the training corpus. These syntactic la- bels act as syntactic constraints on phrases replac- ing nonterminals during decoding. However, there are some problems which affect the performance of syntax-augmented systems. One problem is the strong syntactic constraints imposed by syntax- augmented rules which restrict the search space of translation and prevent the system in many cases from finding good translations. Another problem is the sparse syntactic labels used in such systems, which cause the generation of low-probability, less reliable rules. This weakens the ability of the sys- tem to generalize. As a solution to these prob- lems comes approaches which try to soften syn- tactic constraints imposed by labeled synchronous rules (Venugopal et al., 2009; Chiang, 2010) have been advanced. In this paper, we propose a method to label nonterminals in hierarchical rules with target-side syntactic contextual labels extracted using Com- binatory Categorial Grammar (CCG) (Steedman, 2000). For each target-side phrase in the training corpus, our method uses CCG supertags assigned to its words to extract the left and right syntactic context, represented in the left and right contextual CCG categories of this phrase. Then we annotate the target side of the training corpus by assigning to each phrase a syntactic label that results from combining its left and right contextual CCG cat- egories. Finally, hierarchical rules are extracted from the annotated training corpus with nonter- minals bearing the CCG-based contextual labels, which represent rich syntactic information while imposing …

Content truncated to 500 words