TAJJEB at SemEval-2018 Task 2: Traditional Approaches Just Do the Job with Emoji Prediction

Proceedings of the 12th International Workshop on Semantic Evaluation (SemEval-2018), pages 470–476 New Orleans, Louisiana, June 5–6, 2018. ©2018 Association for Computational Linguistics TAJJEB at SemEval-2018 Task 2: Traditional Approaches Just Do the Job with Emoji Prediction Angelo Basile University of Groningen / The Netherlands University of Malta / Msida, Malta angelo.basile.17@um.edu.mt Kenny W. Lino University of the Basque Country / San Sebasti´an, Spain University of Malta / Msida, Malta kenny.lino.17@um.edu.mt Abstract Emojis are widely used on social media; thus understanding their meaning is important for both practical purposes (e.g. opinion mining, sentiment detection) and theoretical purposes (e.g. How do different L1 speakers use them?, Do they have specific syntax?). This paper presents a set of models that predict an emoji given a tweet as a part of the SemEval-2018 Task 2: Multilingual Emoji Prediction. We built different models and we found that the test results were very different from the vali- dation results. 1 Introduction To some extent, Twitter can be considered a huge corpus and researchers exploit it in many different ways to get a proxy for various types of annota- tions. Purver and Battersby (2012) use distant su- pervision on tweets to build an emotion detection system; Bollen et al. (2011) show that it is pos- sible to use Twitter data to predict the stock mar- ket; Mohammad et al. (2016) build a corpus from tweets for modelling stance. Social media in general are very popular for building corpora for sentiment-related tasks since the users make a wide use of emojis. Pool and Nissim (2016) show that it is possible to achieve state-of-the-art performance in sentiment classifi- cation using only automatically gathered data. The wide use of emojis on the web calls for sys- tems that can automatically process them. When performing opinion mining tasks, for instance, it can be the case that all we have is just an emoji. For example, a single emoji could be used as a re- ply to an advertisement of a certain product; thus, being able to get the meaning of that emoji is im- portant. Felbo et al. (2017) shows that given a text, it is possible to successfully automatically suggest the most appropriate emoji that can accompany that text. In this paper we describe our participation in the SemEval-2018 Task 2: Multilingual Emoji Predic- tion (Barbieri et al., 2018), a challenge that origi- nates from the work of (Barbieri et al., 2017). The task is structured as follows: given a tweet that originally contained one and only one emoji, pre- dict that emoji; the emoji is removed and given as a training label. We experimented using three different ap- proaches: first we use a shallow feature repre- sentation with two different algorithms (Na¨ıve- Bayes and Support Vector Machine); then we ex- perimented with a dense feature representation (i.e. word embeddings) and a deep neural clas- sifier; lastly, we modeled the problem as a trans- lation problem (i.e. treating English and Spanish as the source language and ‘Emoji’ as the …

Content truncated to 500 words