Bottom-up Transfer in Example-based Machine Translation

Bottom-up transfer in Example-based Machine Translation Vincent Vandeghinste Centrum voor Computerlingu¨ıstiek Katholieke Universiteit Leuven Belgium vincent@ccl.kuleuven.be Scott Martens Centrum voor Computerlingu¨ıstiek Katholieke Universiteit Leuven Belgium scott.martens@ccl.kuleuven.be Abstract This paper describes the transfer compo- nent of a syntax-based Example-based Ma- chine Translation system. The source sen- tence parse tree is matched in a bottom-up fashion with the source language side of a parallel example treebank, which results in a target forest which is sent to the target language generation component. The re- sults on a 500 sentences test set are com- pared with a top-down approach to trans- fer of the same system, with the bottom-up approach yielding much better results. 1 Introduction In machine translation, the use of linguistics had all but disappeared during the rise of the statisti- cal machine translation (SMT) paradigm, but as “pure” SMT is reaching its ceiling, the pendulum is swinging back towards the use of linguistics in MT, even within the SMT paradigm, as demon- strated by the Workshops on Syntax and Structure in SMT (among others Wu and Chiang, 2009). The MT engine which is used in this paper is a syntax-based example-based machine translation (EBMT) system. It is example-based as it uses a large set of trans- lation examples (a parallel corpus) as training data to base its decisions on and it is syntax-based as the data in the parallel corpus is annotated with syn- tactic parse trees, both on the source and the target side. Input sentences are syntactically analysed, and the system generates target language parse trees where all ordering information is removed. These serve as input for the target language gen- c⃝ 2010 European Association for Machine Translation. eration component, which determines the output sentences. The system can also be considered a rule-based transfer system, as it conforms to the general ar- chitecture of a rule-based system, using source lan- guage syntactic analysis, syntactic transfer rules and a dictionary (lexical transfer rules), and a tar- get language generation component. Both syntac- tic and lexical transfer rules are automatically in- duced from a parallel corpus. 2 Related Work We compare the transfer component described in this paper with the transfer component of Van- deghinste and Martens (2009). The general approach towards MT is quite sim- ilar to Data-Oriented translation (DOT) (Poutsma, 1998; Hearne, 2005), differing in the fact that we use rule-based or probabilistic context-free parsers whereas they use Data-Oriented Parsing (Bod, 1992), and the DOT approach was only tested on small corpora and a limited domain, whereas we intend a general news domain using large corpora. There are also similarities with the work of Am- bati et al. (2009). They use synchronous context- free grammars (SCFGs) (Aho and Ullman, 1969), which limit the depth of the transfer rules to 2, whereas the approach described in this paper does not set a limit to the maximum depth of a trans- fer rule, just like the synchronous tree-substitution grammars (STSGs), as described in Zhang et al. (2007). The difference between our …

Content truncated to 500 words