Emotion detection in conversations

Abstract

Emotion detection in conversations is emerging as an important research area because it has the ability to mine opinions from conversational datasets available publicly on various social media platforms. Emotion detection in conversations is a crucial research problem for developing a human-like intelligence system. Emotion detection from conversations becomes a challenging task as understanding the context from the conversations is a pivotal task. Another key point for conversations is the dataset. Generally these datasets are imbalanced in nature and some emotions verpower others, creating a majority and minority class differences. Data imbalance becomes a challenging problem during training and testing the models. A neural network model is trained biased towards one or few majority classes. Hence, when an utterance from a minority class is passed to the model, it is likely to get misclassified as one of the majority classes. This incorrect detection of the emotion can make our model inconsistent with the dynamics of the conversation. The aim of this work is to explore two methods – Class weights and Bayesian model, via which we can introduce some balance in the classes while training the model. We want to amplify the mportance of minority classes which may hold more meaning in the conversation over other classes which are not adding any relevant information. A better trained model is likely to give better testing results and improve overall performance as well. Experiments are carried out on two standard datasets – DailyDialog and IEMOCAP. The training data for DailyDialog is purely textual and contains conversations inspired by our day-to-day ommunications. The training data for IEMOCAP consists of audio, video and textual transcripts, where speakers are involved in a scripted dyadic conversation. The results of the proposed method are ompared with other recent state-of-the-art work and shows improvement in compared to the state-of-the-art works.