EPUTION at SemEval-2018 Task 2: Emoji Prediction with User Adaption

Proceedings of the 12th International Workshop on Semantic Evaluation (SemEval-2018), pages 449–453 New Orleans, Louisiana, June 5–6, 2018. ©2018 Association for Computational Linguistics EPUTION at SemEval-2018 Task 2: Emoji Prediction with User Adaption Liyuan Zhou1,2 Qiongkai Xu1,2 Hanna Suominen1,2,3,4 Tom Gedeon1 1 The Australian National University, Canberra, ACT, Australia 2 Data61, CSIRO, Canberra, ACT, Australia 3 University of Canberra, Canberra, ACT, Australia 4 University of Turku, Turku, Finland {Liyuan.Zhou, Qiongkai.Xu, Hanna.Suominen, Tom.Gedeon}@anu.edu.au Abstract This paper describes our approach, called EPUTION, for the open trial of the SemEval- 2018 Task 2, Multilingual Emoji Prediction. The task relates to using social media — more precisely, Twitter — with its aim to predict the most likely associated emoji of a tweet. Our solution for this text classification prob- lem explores the idea of transfer learning for adapting the classifier based on users’ tweet- ing history. Our experiments show that our user-adaption method improves classification results by more than 6 per cent on the macro- averaged F1. Thus, our paper provides evi- dence for the rationality of enriching the orig- inal corpus longitudinally with user behaviors and transferring the lessons learned from cor- responding users to specific instances. 1 Introduction Twitter sentiment analysis is an essential prob- lem for companies and organizations to compu- tationally measure customers’ perceptions which attracts attention from fields of both social media analytics and natural language processing (Rosen- thal et al., 2017; Felbo et al., 2017; Mac Kim et al., 2017). A Twitter message, called a tweet, is generally composed of text, emojis, links, and mentioned users, known as tweeters. An emoji is a small picture or symbol of a standardized set to represent a feeling or another concept (Dictio- nary.com, 2018), contributing to the sentiment of its sender (Barbieri et al., 2017). Consequently, techniques for emoji classification are relevant and can be used to transfer information to subsequent tasks of sentiment, emotion, and sarcasm analy- sis (Felbo et al., 2017). The SemEval-2018 Task 2 challenges its partic- ipants to perform multilingual emoji prediction in English and Spanish. The top-20 most frequent emojis of each language are annotated as tweets’ class labels. To encourage systems with better performance on less frequent emojis, the macro- averaged F1 score (Macro-F) (Suominen et al., 2008) is used as the official evaluation measure. Emoji prediction is widely formalized as a text classification problem in which the state-of-the- art systems fail to perform satisfactory (Barbieri et al., 2018). Because individual users enjoy di- verse preferences in their emoji usage, it is hard to train a generalized classifier to tackle emoji pre- diction. As demonstrated in Table 1, with two ex- amples of simple tweets with various annotations from the training set, even with exactly the same tweet texts, different tweeters have various choices of emojis, such as i) a user can select one of the emojis express the same emotion and ii) a user can have different attitudes towards the same objects or topics. Tweet Emoji @user happy birthday @ new york, new york …

Content truncated to 500 words