Filtering Antonymous, Trend-Contrasting, and Polarity-Dissimilar Distributional Paraphrases for Improving Statistical Machine Translation

Proceedings of the 6th Workshop on Statistical Machine Translation, pages 237–249, Edinburgh, Scotland, UK, July 30–31, 2011. c⃝2011 Association for Computational Linguistics Filtering Antonymous, Trend-Contrasting, and Polarity-Dissimilar Distributional Paraphrases for Improving Statistical Machine Translation Yuval Marton T.J. Watson Research Center IBM yymarton@us.ibm.com Ahmed El Kholy and Nizar Habash Center for Computational Learning Systems Columbia University {akholy,habash}@ccls.columbia.edu Abstract Paraphrases are useful for statistical machine translation (SMT) and natural language pro- cessing tasks. Distributional paraphrase gen- eration is independent of parallel texts and syntactic parses, and hence is suitable also for resource-poor languages, but tends to erro- neously rank antonyms, trend-contrasting, and polarity-dissimilar candidates as good para- phrases. We present here a novel method for improving distributional paraphrasing by filtering out such candidates. We evalu- ate it in simulated low and mid-resourced SMT tasks, translating from English to two quite different languages. We show statisti- cally significant gains in English-to-Chinese translation quality, up to 1 BLEU from non- filtered paraphrase-augmented models (1.6 BLEU from baseline). We also show that yielding gains in translation to Arabic, a mor- phologically rich language, is not straightfor- ward. 1 Introduction Paraphrase recognition and generation has proven useful for various natural language processing (NLP) tasks, including statistical machine transla- tion (SMT), information retrieval, query expansion, document summarization, and natural language gen- eration. We concentrate here on phrase-level (as opposed to sentence-level) paraphrasing for SMT. Paraphrasing is useful for SMT as it increases trans- lation coverage – a persistent problem, even in large- scale systems. Two common approaches are “pivot” and distributional paraphrasing. Pivot paraphrasing translates phrases of interest to other languages and back (Callison-Burch et al., 2006; Callison-Burch, 2008). It relies on parallel texts (or translation phrase tables) in various languages, which are typ- ically scarce, and hence limit its applicability. Dis- tributional paraphrasing (Marton et al., 2009) gener- ates paraphrases using a distributional semantic dis- tance measure computed over a large monolingual corpus.1 Monolingual corpora are relatively easy and inexpensive to collect, but distributional seman- tic distance measures are known to rank antonymous and polarity-dissimilar phrasal candidates high. We therefore attempt to identify and filter out such ill- suited paraphrase candidates. A phrase pair may have a varying degree of antonymy, beyond the better-known complete op- posites (hot / cold) and contradictions (did / did not), e.g., weaker contrasts (hot / cool), contrast- ing trends (covered / reduced coverage), or senti- ment polarity (happy / sad). Information extrac- tion, opinion mining and sentiment analysis litera- ture has been grappling with identifying such pairs (Pang and Lee, 2008), e.g., in order to distinguish positive and negative reviews or comments, or to de- tect contradictions (Marneffe et al., 2008; Voorhees, 2008). We transfer some of the insights, data and techniques to the area of paraphrasing and SMT. We distributionally expand a small seed set of antonyms in an unsupervised manner, following Mohammad et al. (2008). We then present a method for fil- tering antonymous and polarity-dissimilar distribu- tional paraphrases using the expanded antonymous list and a list of …

Content truncated to 500 words