Unsupervised Emotion Detection from Text

Abstract

Artificial intelligence is not only the ability of a machine to think or interact with end user smartly but also to act humanly and rationally. Emotion detection from text plays a key role in human- computer interaction which is a main field in affecting computing. Lately emotion detection from text has attracted the attention of many researchers. That is due to the great revolution of emotional data available on social and web applications of computers and much more in mobile devices. Although some past approaches focused on addressing emotion from text, but still there is less effort on completely unsupervised direction emotion detection from text. This research proposes a completely unsupervised model for textual emotion detection using hybrid technique of lexicon and word embedding concept. The proposed model represents sentences and their meanings in terms of word vectors. To enhance the overall accuracy, emotion ratios were assigned to short sentences and word lexicon. The proposed approach has been validated using the International Survey on Emotion Detection Antecedents and Reactions (ISEAR) and twitter datasets. The evaluation results show that the proposed approach successfully classifies ISEAR sentences based on hybrid technique of lexicon and word embedding with and overall accuracy of 81% which is pretty good result comparing to other unsupervised techniques.