ECNU: Leveraging Word Embeddings to Boost Performance for Paraphrase in Twitter

Proceedings of the 9th International Workshop on Semantic Evaluation (SemEval 2015), pages 34–39, Denver, Colorado, June 4-5, 2015. c⃝2015 Association for Computational Linguistics ECNU: Leveraging Word Embeddings to Boost Performance for Paraphrase in Twitter Jiang Zhao, Man Lan∗ Shanghai Key Laboratory of Multidimensional Information Processing Department of Computer Science and Technology, East China Normal University Shanghai 200241, P. R. China 51121201042@ecnu.cn; mlan@cs.ecnu.edu.cn∗ Abstract This paper describes our approaches to para- phrase recognition in Twitter organized as task 1 in Semantic Evaluation 2015. Lots of ap- proaches have been proposed to address the paraphrasing task on conventional texts ( sur- veyed in (Madnani and Dorr, 2010)). In this work we examined the effectiveness of vari- ous linguistic features proposed in tradition- al paraphrasing task on informal texts, (i.e., Twitter), for example, string based, corpus based, and syntactic features, which served as input of a classification algorithm. Besides, we also proposed novel features based on distributed word representations, which were learned using deep learning paradigms. Re- sults on test dataset show that our proposed features improve the performance by a mar- gin of 1.9% in terms of F1-score and our team ranks third among 10 teams with 38 systems. 1 Introduction Generally, a paraphrase is an alternative surface form in the same language expressing the same se- mantic content as the original form and it can appear at different levels, e.g., lexical, phrasal, sentential (Madnani and Dorr, 2010). Identifying paraphrase can improve the performance of several natural lan- guage processing (NLP) applications, such as query and pattern expansion (Metzler et al., 2007), ma- chine translation (Mirkin et al., 2009), question an- swering (Duboue and Chu-Carroll, 2006), see sur- vey (Androutsopoulos and Malakasiotis, 2010) for completion. Most of previous work of paraphrase are on formal text. Recently with the rapidly growth of microblogs and social media services, the compu- tational linguistic community is moving its attention to informal genre of text (Java et al., 2007; Ritter et al., 2010). For example, (Zanzotto et al., 2011) de- fined the problem of redundancy detection in Twitter and proposed SVM models based on bag-of-word, syntactic content features to detect paraphrase. To provide a benchmark so as to compare and de- velop different paraphrasing techniques in Twitter, the paraphrase and semantic similarity task in Se- mEval 2015 (Xu et al., 2015) requires the partici- pants to determine whether two tweets express the same meaning or not and optionally a degree score between 0 and 1, which can be regarded as a bina- ry classification problem. Paraphrasing task is very close to semantic textual similarity and textual en- tailment task (Marelli et al., 2014) since substan- tially these tasks all concentrated on modeling the underlying similarity between two sentences. The commonly-used features in these tasks can be cat- egorized into several following groups: (1) string based which measures the sequence similarities of original strings with others, e.g., n-gram Overlap, cosine similarity; (2) corpus based which measures word or sentence similarities using word distribu- tional vectors learned from large …

Content truncated to 500 words