Paraphrase Detection Based on Identical Phrase and Similar Word Matching

Paraphrase Detection Based on Identical Phrase and Similar Word Matching Hoang-Quoc Nguyen-Son 1, Yusuke Miyao 2, and Isao Echizen 2 1University of Science, VNU-HCM, Hochiminh, Vietnam nshquoc@fit.hcmus.edu.vn 2National Institute of Informatics, Tokyo, Japan {yusuke,iechizen}@nii.ac.jp Abstract Paraphrase detection has numerous important applications in natural language processing (such as clustering, summarizing, and detect- ing plagiarism). One approach to detecting paraphrases is to use predicate argument tu- ples. Although this approach achieves high paraphrase recall, its accuracy is generally low. Other approaches focus on matching similar words, but word meaning is often contextual (e.g., ‘get along with,’ ‘look for- ward to’). An effective approach to detect- ing plagiarism would take into account the fact that plagiarists frequently cut and paste whole phrases and/or replace several words with similar words. This generally results in the paraphrased text containing identical phrases and similar words. Moreover, pla- giarists usually insert and/or remove various minor words (prepositions, conjunctions, etc.) to both improve the naturalness and disguise the paraphrasing. We have developed a sim- ilarity matching (SimMat) metric for de- tecting paraphrases that is based on match- ing identical phrases and similar words and quantifying the minor words. The metric achieved the highest paraphrase detection ac- curacy (77.6%) when it was combined with eight standard machine translation metrics. This accuracy is better than the 77.4% rate achieved with the state-of-the-art approach for paraphrase detection. 1 Introduction Paraphrase detection is used to determine whether two texts (phrases, sentences, paragraphs, docu- ments, etc.) of arbitrary lengths have the same meaning. Such detection is widely used to remove the tremendous amount of duplicate information on the Internet. It is also used to handle the over- lap of semantic components in texts. Such compo- nents are used in various natural language applica- tions such as word sense discrimination, summariza- tion, automatic thesaurus extraction, question-and- answer generation, machine translation, and plagia- rist or analogical relation identification. Some researchers in the field of paraphrase de- tection have used vector-based similarity to iden- tify the differences between two sentences (Mihal- cea et al., 2006; Blacoe and Lapata, 2012). The two sentences are represented by two vectors based on the frequency of their words in text corpora. The vectors are compared to estimate sentence similar- ity. Plagiarists attempt to thwart this comparison by modifying the copied sentence by inserting or re- moving a few minor words, replacing words with similar words that have different usage frequencies, etc. Such modification reduces the effectiveness of vector-based similarity analysis. Other researchers have analyzed the difference in meaning between two sentences on the basis of their syntactic parsing trees (Socher et al., 2011; Qiu et al., 2006; Das and Smith, 2009). The structure of the trees is a major factor used various sophisticated algorithms such as recursive autoencoders (Socher et al., 2011), heuristic similarity (Qiu et al., 2006), and probabilistic inference (Das and Smith, 2009). However, these algorithms are affected by manipu- lation (deleting, inserting, reordering, etc.) of the words in the sentences. Such manipulations can sig- nificantly change the …

Content truncated to 500 words