An Empirical Study in Source Word Deletion for Phrase-Based Statistical Machine Translation

Proceedings of the Third Workshop on Statistical Machine Translation, pages 1–8, Columbus, Ohio, USA, June 2008. c⃝2008 Association for Computational Linguistics An Empirical Study in Source Word Deletion for Phrase-based Statistical Machine Translation Chi-Ho Li, Dongdong Zhang, Mu Li, Ming Zhou Microsoft Research Asia Beijing, China chl, dozhang@microsoft.com muli, mingzhou@microsoft.com Hailei Zhang Northeastern University of China Shenyang, China hailei.zh@gmail.com Abstract The treatment of ‘spurious’ words of source language is an important problem but often ignored in the discussion on phrase-based SMT. This paper explains why it is impor- tant and why it is not a trivial problem, and proposes three models to handle spurious source words. Experiments show that any source word deletion model can improve a phrase-based system by at least 1.6 BLEU points and the most sophisticated model improves by nearly 2 BLEU points. This paper also explores the impact of training data size and training data domain/genre on source word deletion. 1 Introduction It is widely known that translation is by no means word-to-word conversion. Not only be- cause sometimes a word in some language trans- lates as more than one word in another language, also every language has some ‘spurious’ words which do not have any counterpart in other lan- guages. Consequently, an MT system should be able to identify the spurious words of the source language and not translate them, as well as to gen- erate the spurious words of the target language. This paper focuses on the first task and studies how it can be handled in phrase-based SMT. An immediate reaction to the proposal of inves- tigating source word deletion (henceforth SWD) is: Is SWD itself worth our attention? Isn’t it a trivial task that can be handled easily by existing techniques? One of the reasons why we need to pay attention to SWD is its significant improve- ment to translation performance, which will be shown by the experiments results in section 4.2. Another reason is that SWD is not a trivial task. While some researchers think that the spurious words of a language are merely function words or grammatical particles, which can be handled by some simple heuristics or statistical means, there are in fact some tricky cases of SWD which need sophisticated solution. Consider the follow- ing example in Chinese-to-English translation: in English we have the subordinate clause “accord- ing to NP”, where NP refers to some source of information. The Chinese equivalent of this clause can sometimes be “ACCORDING-TO/�� NP EXPRESS/��”; that is, in Chinese we could have a clause rather than a noun phrase following the preposition ACCORDING-TO/��. There- fore, when translating Chinese into English, the content word EXPRESS/�� should be consid- ered spurious and not to be translated. Of course, the verb EXPRESS/�� is not spurious in other contexts. It is an example that SWD is not only about a few function words, and that the solu- tion to SWD has to take context-sensitive factors into account. Moreover, the solution needed for such tricky cases seems to be …

Content truncated to 500 words