YNU-HPCC at SemEval-2018 Task 3: Ensemble Neural Network Models for Irony Detection on Twitter

Proceedings of the 12th International Workshop on Semantic Evaluation (SemEval-2018), pages 622–627 New Orleans, Louisiana, June 5–6, 2018. ©2018 Association for Computational Linguistics YNU-HPCC at SemEval-2018 Task 3: Ensemble Neural Network Models for Irony Detection on Twitter Bo Peng, Jin Wang and Xuejie Zhang School of Information Science and Engineering Yunnan University Kunming, P.R. China Contact:xjzhang@ynu.edu.cn Abstract This paper describes our proposed to partici- pate in the first year of the irony detection in English tweets competition. Previous works have demonstrated that long short-term mem- ory models have achieved remarkable perfor- mance in natural language processing; more- over, combining multiple classifications from various individual classifiers is generally more powerful than a single classification. In order to obtain more precise irony detection classi- fication, our system trains several individual neural network classifiers and combines their results according to the ensemble-learning al- gorithm. 1 Introduction In most sentiment analysis tasks, recognition of the precise emotional polarity of a sentence form- s the basis for further work. However, much of the corpus we used for analysis and training con- tains numerous sarcasm and irony features that will have a negative impact on the results of our analysis and training. For example, although the tweets provided by Twitter constitute a valuable and widely applicable corpus for many natural lan- guage processing tasks, Twitter users express their feelings and opinions on social networks with fre- quent irony (Amir et al., 2016). Therefore, such tweets may contain converse sentiments informa- tion compared their literal meaning. For exam- ple, @someuser Yeah keeping cricket clean, that’s what he wants #Sarcasm: ignoring the hash tag, this tweet would be positive, which would miss lead an analysis system that uses these types of tweets as input. Thus, it makes sense to discriminate whether a text is ironic, particularly for social network texts such as tweets. Further applications including tweet sentiment analysis, will benefit from auto- matic irony detection. The SemEval-2018 Twitter competition promotes research in this area, and is divided into two subtasks that involve binary and four-class classification. Subtask A is a two-class (or binary) classifica- tion task whereby the system must predict whether or not a tweet is ironic. The subtask B is a multi- class classification task where the system has to predict one out of four labels describing i) ver- bal irony realized through a polarity contrast, ii) verbal irony without such a polarity contrast (i.e., other verbal irony), iii) descriptions of situation- al irony, and iv) non-irony (Cynthia Van Hee and Hoste, 2018). For a more detailed description, please see Carman et al. (2017). In recent years, deep learning techniques have significantly outperformed traditional methods in several natural language processing (NLP) tasks (Cliche, 2017). In such task, several deep learn- ing architecture-based methods have achieved out- standing performance in irony and sarcasm detec- tion in social media. Silvio (Amir et al., 2016) presented a novel convolutional network-based method for learning user embeddings from their previous posts and used the user embeddings with lexical signals to …

Content truncated to 500 words