CDTDS: Predicting Paraphrases in Twitter via Support Vector Regression

Proceedings of the 9th International Workshop on Semantic Evaluation (SemEval 2015), pages 75–79, Denver, Colorado, June 4-5, 2015. c⃝2015 Association for Computational Linguistics CDTDS: Predicting Paraphrases in Twitter via Support Vector Regression Rafael Michael Karampatsis Department of Informatics University of Edinburgh 10 Crichton Street, EH8 9AB Edinburgh, United Kingdom mpatsis13@gmail.com Abstract In this paper we describe a system that rec- ognizes paraphrases in Twitter for tweets that refer to the same topic. The system partici- pated in Task1 of SEMEVAL-2015 and uses a support vector regression machine to predict the degree of similarity. The similarity is then thresholded to create a binary prediction. The model and experimental results are discussed along with future work that could improve the method. 1 Introduction Recently, Twitter has gained significant popularity among the social network services. Lots of users of- ten use Twitter to express feelings or opinions about a variety of subjects. Analysing this kind of content can lead to useful information for fields such as per- sonalized marketing or social profiling. However, such a task is not trivial, because the language used on Twitter is often informal, presenting new chal- lenges to text analysis. Task1 of SEMEVAL-2015 (Xu et al., 2015) fo- cuses on recognition of paraphrases and semantic similarity in Twitter i.e., recognizing if two tweets are alternative linguistic expressions of the same, or similar, meaning (Bhagat and Hovy, 2013). The task is based on a crowdsourced corpus of 18000 pairs of paraphrases and non-paraphrased sentences from Twitter (Xu et al., 2014) and each pair consists of two tweets from the same topic. A label is provided with each pair, which is the number of yes votes from 5 crowdsourced annotators when asked if the second tweet is a paraphrase of the first one. Paraphrase Example: Roberto Mancini gets the boot from Man City Roberto Mancini has been sacked by Manchester City with the Blues saying Non-Paraphrase Example: WORLD OF JENKS IS ON AT 11 World of Jenks is my favorite show on tv Figure 1: Examples of both a paraphrase and a non- paraphrase pair of the data. The method utilizes a support vector regression machine (SVR). The regression model tries to pre- dict the degree of semantic similarity between two tweets, by assuming that it can be represented by the probability that random human annotators would an- notate the pair as a paraphrase. The predicted value is transformed into a binary decision via a threshold. Section 2 describes the data provided by the or- ganizers. Sections 4 and 5 present the system and its performance respectively. Finally, Section 6 pro- vides ideas for future work and Section 7 concludes. 2 Data The objective of this task is to predict whether two sentences from Twitter sharing the same topic, im- ply the same or very similar meaning and optionally a degree score between 0 and 1. In Figure 1, a para- phrase and a non-paraphrase example taken from the task website are illustrated. The organizers released a training (Train) and …

Content truncated to 500 words