Emotion Detection and Classification in a Multigenre Corpus with Joint Multi-Task Deep Learning

Proceedings of the 27th International Conference on Computational Linguistics, pages 2905–2913 Santa Fe, New Mexico, USA, August 20-26, 2018. 2905 Emotion Detection and Classification in a Multigenre Corpus with Joint Multi-Task Deep Learning Shabnam Tafreshi George Washington University Department of Computer Science GWU NLP Lab shabnamt@gwu.edu Mona Diab George Washington University Department of Computer Science GWU NLP Lab mtdiab@gwu.edu Abstract Detection and classification of emotion categories expressed by a sentence is a challenging task, due to subjectivity of emotion. To date, most of the models are trained and evaluated on single genre and when used to predict emotion in different genre, their performance drops by a large margin. To address the issue of robustness, we model the problem within a joint multi-task learning framework. We train this model with a multigenre emotion corpus to predict emotions across various genres. Each genre is represented as a separate task, we use soft parameter shared layers across the various tasks. our experimental results show that this model improves the results across the various genres, compared to a single genre training in the same neural net architecture. Figure 1: Joint Multi-Task Emotion neural net architecture. In this multi-task framework, Task-E and Task-BHM (explained in section 3) demonstrate different genres. 1 Introduction Sentence-level emotion detection is garnering a lot of attention recently. To date, most systems are trained and evaluated on a single genre resource. However, the problem is robustness, when such models are applied to new genres, the performance expectedly drops significantly. In this paper, we propose a model to address the genre robustness issue. This work is licenced under a Creative Commons Attribution 4.0 International Licence. Licence details: http:// creativecommons.org/licenses/by/4.0/ 2906 Dataset joy trust anti surprise sad fear anger disgust BLG 689 43 260 150 312 132 192 255 HLN 106 6 56 31 83 68 28 55 MOV 4875 26 119 255 258 63 20 5145 EmoNet 5670 75 435 436 653 263 240 5455 BHMT 11340 150 870 872 1306 526 480 10910 Table 1: Data statistics illustrating the distributions of the various emotion tags from PL8 across the different genres. We frame the problem as a joint multi task learning problem, where each of the genres is represented as a separate task that shares information with other tasks (genres). Our model, the Joint Multi-Task Emotion (JMTE), illustrated in figure 1, trains on emotions in four genres of data: Blog Posts, News Headlines, Movie Reviews, and Tweets. Experimental results show that using JMTE achieves better results, over a single model trained on a single of these different genres. We also study the impact of specifically adding Twitter data that is distantly supervised to the training models. The rest of this paper is organized as follows: We describe our multigenre corpus in section 2; We present our devised JMTE model in section 3; Experimental conditions and results are presented in section 4; Discussion is presented in section 5; We review related studies in section 6; Conclusions are described in section 7. 2 …

Content truncated to 500 words