Deriving Paraphrases for Highly Inflected Languages from Comparable Documents

Proceedings of COLING 2012: Technical Papers, pages 185–200, COLING 2012, Mumbai, December 2012. Deriving Paraphrases for Highly Inflected Languages from Comparable Documents Kfir BAR Nachum DERSHOWITZ School of Computer Science, Tel Aviv University, Ramat Aviv, Israel kfirbar@post.tau.ac.il, nachumd@tau.ac.il ABSTRACT We describe an automatic paraphrase-inference procedure for a highly inflected language like Arabic. Paraphrases are derived from comparable documents, that is, distinct documents dealing with the same topic. A co-training approach is taken, with two classifiers, one designed to model the contexts surrounding occurrences of paraphrases, and the other trained to identify significant features of the words within paraphrases. In particular, we use morpho-syntactic features calculated for both classifiers, as is to be expected when working with highly inflected languages. We provide some experimental results for Arabic, and for the simpler English, which we find to be encouraging. Our immediate interest is to incorporate such paraphrases within an Arabic-to- English translation system. KEYWORDS : Paraphrases, highly inflected languages, morphologically rich languages, co- training, comparable documents, Arabic 185 1 Introduction Paraphrases are pairs of sequences of words, both in the same language, that have the same meaning in at least some contexts. Given a text, “paraphrasing” is the act of generating an alternate sequence of words that conveys the same meaning. Since the meaning of a text is determined only when its context is given, paraphrases are sometimes referred to as “dynamic translations” or “semantic equivalents”. Identifying paraphrases is an important capability for many natural language processing applications, including machine translation, as a possible workaround for the problem of limited coverage inherent in a corpus-based translation approach (Callison-Burch et al., 2006; Marton et el., 2009). Other applications of paraphrasing include automatic evaluation of summaries (Zhao et al., 2008) and question answering (Duboue and Chu- Carroll, 2006; Riezler et al., 2007). There are two main directions of work on paraphrases that one can find in this field: investigating an automatic approach for uncovering paraphrases in a given corpus and using paraphrases to improve the performance of a specific task. In this paper, we introduce a novel method for extracting Arabic paraphrases from a corpus of comparable documents as part of our work on improving Arabic-to-English machine translation. Comparable documents are ones that deal with the same topic, such as two newspaper reports of the same event. The extraction technique is based on a learning method, known as “co-training” (Blum and Mitchel, 1998) and inspired by the work of Barzilay and McKeown (2001) for finding paraphrases in a parallel monolingual English corpus. In order to validate our technique, we have applied it also to a similar English corpus. Like many other Semitic languages, Arabic is highly inflected; therefore, data sparseness becomes even more noticeable than in English and extracting paraphrases from a corpus turns out to be even more complicated. Arabic words are derived from a root and a pattern (template), combined with prefixes, suffixes and circumfixes. Using the same root with different patterns may yield words with different meanings. Words are inflected for …

Content truncated to 500 words