UCB System Description for the WMT 2007 Shared Task

Proceedings of the Second Workshop on Statistical Machine Translation, pages 212–215, Prague, June 2007. c⃝2007 Association for Computational Linguistics UCB System Description for the WMT 2007 Shared Task Preslav Nakov EECS, CS division University of California at Berkeley Berkeley, CA 94720 nakov@cs.berkeley.edu Marti Hearst School of Information University of California at Berkeley Berkeley, CA 94720 hearst@ischool.berkeley.edu Abstract For the WMT 2007 shared task, the UC Berkeley team employed three techniques of interest. First, we used monolingual syntac- tic paraphrases to provide syntactic variety to the source training set sentences. Sec- ond, we trained two language models: a small in-domain model and a large out-of- domain model. Finally, we made use of re- sults from prior research that shows that cog- nate pairs can improve word alignments. We contributed runs translating English to Span- ish, French, and German using various com- binations of these techniques. 1 Introduction Modern Statistical Machine Translation (SMT) sys- tems are trained on aligned sentences of bilingual corpora, typically from one domain. When tested on text from that same domain, such systems demon- strate state-of-the art performance; however, on out-of-domain text the results can get significantly worse. For example, on the WMT 2006 Shared Task evaluation, the French to English translation BLEU scores dropped from about 30 to about 20 for nearly all systems, when tested on News Commen- tary rather than Europarl (Koehn and Monz, 2006). Therefore, this year the shared task organizers have provided 1M words of bilingual News Com- mentary training data in addition to the Europarl data (about 30M words), thus challenging the par- ticipants to experiment with domain adaptation. Below we describe our domain adaptation exper- iments, trying to achieve better results on the News Commentary data. In addition to training on both data sets, we make use of monolingual syntactic paraphrases of the English side of the data. 2 Monolingual Syntactic Paraphrasing In many cases, the testing text contains “phrases” that are equivalent, but syntactically different from the phrases learned on training, and the potential for a high-quality translation is missed. We address this problem by using nearly equivalent syntactic para- phrases of the original sentences. Each paraphrased sentence is paired with the foreign translation that is associated with the original sentence in the training data. This augmented training corpus can then be used to train an SMT system. Alternatively, we can paraphrase the test sentences making them closer to the target language syntax. Given an English sentence, we parse it with the Stanford parser (Klein and Manning, 2003) and then generate paraphrases using the following syntactic transformations: 1. [NP NP1 P NP2] ⇒ [NP NP2 NP1]. inequality in income ⇒ income inequality. 2. [NP NP1 of NP2] ⇒ [NP NP2 poss NP1]. inequality of income ⇒ income’s inequality. 3. NPposs ⇒ NP. income’s inequality ⇒ income inequality. 4. NPposs ⇒ NPPPof . income’s inequality ⇒ inequality of income. 5. NPNC ⇒ NPposs. income inequality ⇒ income’s inequality. 6. NPNC ⇒ NPPP . income inequality ⇒ inequality in incomes. 212 Sharply …

Content truncated to 500 words