OPAL at SemEval-2016 Task 4: the Challenge of Porting a Sentiment Analysis System to the “Real” World

Proceedings of SemEval-2016, pages 262–265, San Diego, California, June 16-17, 2016. c⃝2016 Association for Computational Linguistics OPAL at SemEval-2016 Task 4: the Challenge of Porting a Sentiment Analysis System to the "Real" World Alexandra Balahur European Commission Joint Research Centre Institute for the Protection and Security of the Citizen Via E. Fermi 2749, 21027 Ispra (VA), Italy alexandra.balahur@jrc.ec.europa.eu Abstract Sentiment analysis has become a well- established task in Natural Language Pro- cessing. As such, a high variety of methods have been proposed to tackle it, for different types of texts, text levels, languages, domains and formality levels. Although state-of-the- art systems have obtained promising results, a big challenge that still remains is to port the systems to the “real world” – i.e. to implement systems that are running around the clock, dealing with information of heterogeneous na- ture, from different domains, written in differ- ent styles and diverse in formality levels. The present paper describes our efforts to imple- ment such a system, using a variety of strate- gies to homogenize the input and comparing various approaches to tackle the task. Specifi- cally, we are tackling the task using two dif- ferent approaches: a) one that is unsupervised, based on dictionaries of sentiment-bearing words and heuristics to compute final polarity of the text considered; b) the second, super- vised, trained on previously annotated data from different domains. For both approaches, the data is first normalized and the slang is re- placed with its expanded version. 1 Introduction Sentiment analysis is the task in Natural Language Processing (NLP) that deals with classifying opin- ions according to the polarity of the sentiment they express. Due to the large quantities of user- generated online contents available on different Internet sites (forums, social networks, blogs, re- view sites, microblogs, etc.) and the possible value they can have for different domains (Marketing, e- Rulemaking, Political Science, etc.), Sentiment Analysis has become a very popular task in the field in the past decade. As such, a high variety of methods have been proposed to tackle it, for different types of texts, text levels, languages, domains and formality lev- els. Although state of the art systems have ob- tained promising results (most of all in priory defined datasets – domains, languages, styles, for- mality levels), a big challenge that still remains is to port the systems to the “real world” – i.e. sys- tems that are running around the clock, dealing with information of heterogeneous nature, from different domains, written in different styles and diverse in formality levels. The present paper describes our efforts to build such a system. The challenge is not straightforward to tackle, as this entails building a system that is: a) on the one hand, robust enough to obtain similar levels of performance across domains, languages, text types and formality levels; b) on the other hand, flexible enough to treat all these types of texts. Further on, this entails that such a system must have components to treat the input to make it …

Content truncated to 500 words