PlusEmo2Vec at SemEval-2018 Task 1: Exploiting emotion knowledge from emoji and #hashtags

Proceedings of the 12th International Workshop on Semantic Evaluation (SemEval-2018), pages 264–272 New Orleans, Louisiana, June 5–6, 2018. ©2018 Association for Computational Linguistics PlusEmo2Vec at SemEval-2018 Task 1: Exploiting emotion knowledge from emoji and #hashtags Ji Ho Park, Peng Xu, Pascale Fung Centre for Artificial Intelligence Research (CAiRE) Hong Kong University of Science and Technology {jhpark, pxuab}@connect.ust.hk, pascale@ece.ust.hk Abstract This paper describes our system that has been submitted to SemEval-2018 Task 1: Affect in Tweets (AIT) to solve five subtasks. We fo- cus on modeling both sentence and word level representations of emotion inside texts through large distantly labeled corpora with emojis and hashtags. We transfer the emotional knowl- edge by exploiting neural network models as feature extractors and use these representa- tions for traditional machine learning mod- els such as support vector regression (SVR) and logistic regression to solve the competi- tion tasks. Our system is placed among the Top3 for all subtasks we participated. 1 Introduction Finding a good representation of texts is very chal- lenging since texts are sequences of words which are represented in a discrete space of the vocabu- lary. For this reason, many past works have inves- tigated in finding the mapping of words (Mikolov et al., 2013; Pennington et al., 2014) or sentences (Kiros et al., 2015) to continuous spaces, so that each text can be represented by a fixed-size, real- valued N-dimensional vector. This vector repre- sentation then can be applied to machine learning models to solve problems like classification and regression. A good representation should contain essential information inside each text and be a use- ful input for statistical models. Emotions in texts further deepen the complexity of modeling natural language since they not only depend on the semantics of a language but also are inherently subjective and ambiguous. Despite the difficulty, accounting for emotion is important in achieving true natural language understanding, especially in areas involving human-computer in- teractions such as dialogue systems (Fung, 2015). Humans can naturally capture and express dif- ferent emotions in texts, so machines should also be able to infer them. Many works (Tang et al., 2014; Felbo et al., 2017; Thelwall, 2017) explored modeling sentiment or emotion in texts in various directions. This work is highly related to these ef- forts. Semeval-2018 Task 1: Affect in Tweets (AIT- 2018) encourages more efforts in this area with the task of sentiment analysis, which is one of the most practical applications of modeling emo- tional text representations. We have participated in five subtasks regarding English tweets: emo- tion intensity regression, emotion intensity ordi- nal classification, valence (sentiment) regression, valence ordinal classification, and emotion classi- fication (More details on the tasks in Mohammad et al. (2018)). Although these five tasks take different for- mats, the most important objective is finding a good representation of the tweets regarding emo- tions. However, the given competition training datasets are too small to achieve our goal (Table 3). Therefore, we explore utilizing larger datasets that are distantly supervised by emojis and hash- …

Content truncated to 500 words