A Comparison of Pivot Methods for Phrase-Based Statistical Machine Translation

Proceedings of NAACL HLT 2007, pages 484–491, Rochester, NY, April 2007. c⃝2007 Association for Computational Linguistics A Comparison of Pivot Methods for Phrase-based Statistical Machine Translation Masao Utiyama and Hitoshi Isahara National Institute of Information and Communications Technology 3-5 Hikari-dai, Soraku-gun, Kyoto 619-0289 Japan {mutiyama,isahara}@nict.go.jp Abstract We compare two pivot strategies for phrase-based statistical machine transla- tion (SMT), namely phrase translation and sentence translation. The phrase translation strategy means that we di- rectly construct a phrase translation ta- ble (phrase-table) of the source and tar- get language pair from two phrase-tables; one constructed from the source language and English and one constructed from En- glish and the target language. We then use that phrase-table in a phrase-based SMT system. The sentence translation strat- egy means that we first translate a source language sentence into n English sen- tences and then translate these n sentences into target language sentences separately. Then, we select the highest scoring sen- tence from these target sentences. We con- ducted controlled experiments using the Europarl corpus to evaluate the perfor- mance of these pivot strategies as com- pared to directly trained SMT systems. The phrase translation strategy signifi- cantly outperformed the sentence transla- tion strategy. Its relative performance was 0.92 to 0.97 compared to directly trained SMT systems. 1 Introduction The rapid and steady progress in corpus-based ma- chine translation (Nagao, 1981; Brown et al., 1993) has been supported by large parallel corpora such as the Arabic-English and Chinese-English paral- lel corpora distributed by the Linguistic Data Con- sortium and the Europarl corpus (Koehn, 2005), which consists of 11 European languages. How- ever, large parallel corpora do not exist for many language pairs. For example, there are no pub- licly available Arabic-Chinese large-scale parallel corpora even though there are Arabic-English and Chinese-English parallel corpora. Much work has been done to overcome the lack of parallel corpora. For example, Resnik and Smith (2003) propose mining the web to collect parallel corpora for low-density language pairs. Utiyama and Isahara (2003) extract Japanese-English parallel sentences from a noisy-parallel corpus. Munteanu and Marcu (2005) extract parallel sentences from large Chinese, Arabic, and English non-parallel newspaper corpora. Researchers can also make the best use of exist- ing (small) parallel corpora. For example, Nießen and Ney (2004) use morpho-syntactic information to take into account the interdependencies of inflected forms of the same lemma in order to reduce the amount of bilingual data necessary to sufficiently cover the vocabulary in translation. Callison-Burch et al. (2006a) use paraphrases to deal with unknown source language phrases to improve coverage and translation quality. In this paper, we focus on situations where no par- allel corpus is available (except a few hundred paral- lel sentences for tuning parameters). To tackle these extremely scarce training data situations, we pro- pose using a pivot language (English) to bridge the 484 source and target languages in translation. We first translate source language sentences or phrases into English and then translate those English sentences or phrases into the target language, as …

Content truncated to 500 words