Noun Paraphrasing Based on a Variety of Contexts

PACLIC 28 !644 Noun Paraphrasing Based on a Variety of Contexts Tomoyuki Kajiwara Department of Electrical Engineering Nagaoka University of Technology Nagaoka City, Niigata, Japan kajiwara@jnlp.org Kazuhide Yamamoto Department of Electrical Engineering Nagaoka University of Technology Nagaoka City, Niigata, Japan yamamoto@jnlp.org Abstract We paraphrase nouns along the contexts of sentence input on the basis of a variety of contexts obtained from a large-scale corpus. The proposed method only uses the number of types of context, not word frequency or co- occurrence frequency features. This method is based on the notion that paraphrase candidates appear more commonly with target words in the same context. The results of our experi- ment demonstrate that the approach can pro- duce more appropriate paraphrases than ap- proaches based on co-occurrence frequency and pointwise mutual information. 1 Introduction Although extensive and various forms of text data are easily available in the present age, in order for readers to gather information effectively, they need technology that overcomes any differences in their linguistic competence. For example, technology that buries the difference in the linguistic competence of foreign language learners, children, the elderly, and disabled persons is useful (Inui and Fujita, 2004). We present our research on paraphrasing to control language at the elementary school level in order to simplify texts for children. We believe that vocab- ulary simplification for children can be realized by paraphrasing text according to Basic Vocabulary to Learn (BVL) (Kai and Matsukawa, 2002) . BVL is a collection of words selected on the basis on a lexical analysis of elementary school textbooks. It contains 5,404 words that can help children write expressively. As previous work indicated, there are lexical para- phrases that define statements from a Japanese dic- tionary (Kajiwara et al., 2013). The definition state- ments from the Japanese dictionary explain a given headword in several easy words. Therefore, lexi- cal simplification and paraphrasing that conserves a particular meaning are expected by paraphrasing the headword with the words in the definitions. How- ever, definition statements are short sentences that consist of several words. Consequently, there are few paraphrase candidates, and natural paraphras- ing is difficult even if we use certain dictionaries together. In addition, the definition statement as a whole is equivalent to the headword; there is no guarantee that any individual word extracted from the definition statement can paraphrase the head- word. We propose lexical paraphrasing based on a vari- ety of contexts obtained from a large corpus without depending on existing lexical resources from such a background. The proposed method is not dependent on language, thus it can perform lexical paraphrases using a corpus of arbitrary languages. In this paper we examine and report on Japanese nouns. 2 Related Works As paraphrase acquisition from a corpus, a study with a parallel corpus and comparable corpus has been performed. Barzilay and McKeown paraphrase text using plural English translations made from the same document (Barzilay and McKeown, 2001). In addition, Shinyama and Sekine paraphrase us- ing plural newspaper articles that report the same …

Content truncated to 500 words