UAIC at SemEval-2019 Task 3: Extracting Much from Little

Proceedings of the 13th International Workshop on Semantic Evaluation (SemEval-2019), pages 355–359 Minneapolis, Minnesota, USA, June 6–7, 2019. ©2019 Association for Computational Linguistics 355 UAIC at SemEval-2019 Task 3: Extracting Much from Little Cristian Simionescu, Ingrid Stoleru, Diana Lucaci, Gheorghe Balan, Iulian Bute, Adrian Iftene Faculty of Computer Science, ”Alexandru Ioan Cuza” University of Iasi, Romania cristian@nexusmedia.ro, ingridstoleru@gmail.com, {diana.lucaci22, balangheorghe1997}@gmail.com, iulian.bute@gmail.com, adiftene@info.uaic.ro Abstract In this paper, we present a system description for implementing a sentiment analysis agent capable of interpreting the state of an inter- locutor engaged in short three message con- versations. We present the results and obser- vations of our work and which parts could be further improved in the future. 1 Introduction It is hard to understand emotions in textual conver- sations in the absence of voice modulations and fa- cial expressions (Gupta et al., 2017). In sentiment analysis task researchers work on different levels of sentiment analysis: document (when are con- sidered single topics documents), sentence (when single sentences are classified as positive, negative or neutral), entity or aspect (which deal with find- ing the aspects in the text and then classifying in respective aspect) (Liu, 2012). Similar to sentiment analysis at sentence level, in last years tweets from Twitter were analyzed and classified (Zhang and Liu, 2017), (Kumar and Sebastian, 2012), (Mukherjee and Bhattacharyya, 2013) and (Singh and Husain, 2014). In the be- ginning, a binary classification was used, which linked opinions or opinions only to two classes: positive or negative. In (Pak and Paroubek, 2010) the authors proposed a model for classi- fying tweets in goals, positive and negative feel- ings using a classifier based on multinomial Naive Bayes to use features such as N-grams and tags POS (part- of-Speech). In (Parikh and Movassate, 2009) the authors have implemented two models, one based on the Naive Bayes bigrams model and one using Maximum Entropy to classify tweets. (Go et al., 2009) proposed a solution for analysing feelings on Twitter using distant supervision, the training data were tweets with emoticons, which are regarded as noise data. They built several well performing models using Naive Bayes, MaxEnt, and SVM. (Barbosa and Feng, 2010) have mod- eled an automated method to classify tweets us- ing space features including retweets, hashtags, links, punctuation mark amazement in combina- tion with words and POS features polarity. (Luo et al., 2013) have brought to light the difficulties that they encounter when they want to classify tweets. Spam and a variety of languages on Twit- ter make the task of identifying opinions very dif- ficult. In SemEval 2019, in Task 3, EmoContext: Contextual Emotion Detection in Text (Chatter- jee et al., 2019), the organizers ask participants to classify users messages in one of four classes: Happy, Sad, Angry or Others. These are given in the context of another two previous messages. The textual dialogue is composed of short mes- sages that appear to be from a chat conversation. In such a context, the users express their thoughts and ideas in a compact …

Content truncated to 500 words