How emotional are you? Neural Architectures for Emotion Intensity Prediction in Microblogs

Proceedings of the 27th International Conference on Computational Linguistics, pages 2914–2926 Santa Fe, New Mexico, USA, August 20-26, 2018. 2914 How emotional are you? Neural Architectures for Emotion Intensity Prediction in Microblogs Devang Kulshreshtha∗, Pranav Goel∗, and Anil Kumar Singh Indian Institute of Technology (Banaras Hindu University) Varanasi Varanasi, Uttar Pradesh, India {devang.kulshreshtha.cse14, pranav.goel.cse14, aksingh.cse}@iitbhu.ac.in Abstract Social media based micro-blogging sites like Twitter have become a common source of real-time information (impacting organizations and their strategies), and are used for expressing emotions and opinions. Automated analysis of such content therefore rises in importance. To this end, we explore the viability of using deep neural networks on the specific task of emotion intensity pre- diction in tweets. We propose a neural architecture combining convolutional and fully connected layers in a non-sequential manner - done for the first time in context of natural language based tasks. Combined with lexicon-based features and transfer learning, our model achieves state- of-the-art performance, outperforming the previous best system by 4.4% Pearson correlation on the WASSA’17 EmoInt shared task dataset. We investigate the performance of deep multi-task learning models trained for all emotions at once in a unified architecture and get encouraging re- sults. Experiments performed on evaluating correlation between emotion pairs offer interesting insights into the relationship between them. The code for our experiments is publicly available. 1 Introduction & Related Work One can use text in any language not only to express a variety of emotions but also to convey the as- sociated ‘intensity’ of the emotion. Emotion intensity is the degree of an emotion (like anger or fear) expressed by the speaker or author. For example, ‘When you lose somebody close to your heart you lose yourself as well. Crying my heart out!!!’ expresses more sadness than, say, ‘All friends are out of town. Feeling a bit lonely...’. Automatic detection of emotion intensity can be useful for natural language based systems. An e-commerce firm seeking to publicize positive reviews of its products would want to use the statements expressing high amount of joy or happiness. The intensity of anger expressed in a grievance can be used to automatically decide the priority of addressing complaints. A possible example scenario: based on the tweet ‘X box one controller jockey update is the fucking worst! #fuming’ expressing a higher intensity of anger than the tweet ‘New madden bundle code for x box one is stupid, takes forever to download lol’, may point to the need of prioritizing addressing issues with the new jockey stick before looking into the madden bundle of the same product (an Xbox One). The WASSA’17 workshop conducted the EmoInt shared task (Mohammad and Bravo-Marquez, 2017b) where given a tweet and the emotion it exhibits, systems had to predict the intensity of this given emotion as a real valued score between 0 and 1. We use the shared task setting and the dataset they provide to develop and evaluate all our approaches. The main contributions of our work are: a) A neural architecture for …

Content truncated to 500 words