Affective State Prediction of Contextualized Concepts

Most studies on affective analysis of text focus on the sentiment or emotion expressed by a whole sentence or document. In this paper, we propose a novel approach to predict the affective states of a described event through the predictions of the corresponding subject, action and object involved in the described event. Rather than using a sentiment label or discrete emotion categories, the affective state is represented using the three dimensional evaluation-potency-activity (EPA) model. The main idea is to use automatically obtained word embedding as word representation and to use the Long Short-Term Memory (LSTM) network as the prediction model. Compared to the linear model used in the Affective Control Theory which uses manually annotated EPA lexicon, our proposed LSTM learning method using word embedding outperforms the linear model and word embedding also performs better than EPA lexicon. Most importantly, our work shows that automatically obtained word embedding outperforms manually constructed affective lexicons.