INF-UFRGS-OPINION-MINING at SemEval-2016 Task 6: Automatic Generation of a Training Corpus for Unsupervised Identification of Stance in Tweets

Proceedings of SemEval-2016, pages 378–383, San Diego, California, June 16-17, 2016. c⃝2016 Association for Computational Linguistics INF-UFRGS-OPINION-MINING at SemEval-2016 Task 6: Automatic Generation of a Training Corpus for Unsupervised Identification of Stance in Tweets Marcelo Dias and Karin Becker Instituto de Inform´atica (INF) Universidade Federal do Rio Grande do Sul (UFRGS) Porto Alegre, RS, Brazil {marcelo.dias, karin.becker}@inf.ufrgs.br Abstract This paper describe a weakly supervised so- lution for detecting stance in tweets, submit- ted to the SemEval 2016 Stance Task. Our approach is based on the premise that stance can be exposed as positive or negative opin- ions, although not necessarily about the stance target itself. Our system receives as input n- grams representing opinion targets and com- mon terms used to denote stance (e.g. hash- tags), and use these features, together with the sentiment detection solutions, to automat- ically compose a large training corpus. Then, it applies a supervised learning algorithm to develop a stance prediction model. 1 Introduction Sentiment analysis involves the automatic identifi- cation of opinions, feelings, evaluations, attitudes and emotions expressed by people in the written lan- guage. Popular lines of work in this field are opin- ion mining (Liu, 2012) and emotion mining (Mo- hammad, 2016). Stance detection is a less explored problem, addressed as part of SemEval-2016 (Inter- national Workshop on Semantic Evaluation 2016), Task 61. Stance detection is defined in this task as the automatic determination from text whether its author is in favor of the given target, against the given target, or whether neither inference is likely. The present work describes the solution devel- oped for Task 6-b, which involves the unsupervised stance detection in tweets based solely in their con- tent. The target is Donald Trump, a possible repub- lican presidential candidate, for which two sets of 1http://alt.qcri.org/semeval2016/task6/ non-annotated tweets were supplied: a domain cor- pus with 78,156 tweets and a test corpus with 707 tweets for task evaluation purpose. Another set of annotated tweets for stance detection was available as part of Task 6-a (supervised stance detection), which included 639 tweets about another possible candidate, Hillary Clinton. We used this annotated data to develop the proposed method, due to the sim- ilarity of problems. Details of the Task 6 can be found at (Mohammad et al., 2016). The identification of stance can be complex even for humans (Walker et al., 2012), and our strategy was to address it partly as an opinion mining prob- lem. Stance can be exposed as positive or negative opinions, but not necessarily about the target of the stance problem. For instance, when the opinion tar- get is a politician, the target may be his/her agenda for health or education, members of the same party, or opponents. In addition, stance detection faces challenges common to sentiment analysis in general, such as use of vocabulary and slang specific of the media, orthography errors, sarcasm, etc. We developed a weakly supervised method for de- tecting stance in tweets. Our method requires some side-related targets and key stance n-grams, which …

Content truncated to 500 words