Combining External Sentiment Knowledge for Emotion Cause Detection

Emotion cause detection (ECD) that aims to extract the trigger event of a certain emotion explicitly expressed in text has become a hot topic in natural language processing. However, the performance of existing models all suffers from inadequate sentiment information fusion and the limited size of corpora. In this paper, we propose a novel model to combine external sentiment knowledge for ECD task, namely ExSenti-ECD, to try to solve these problems. First, in order to fully fuse sentiment information, we utilize a sentiment-specific embedding method to encode external sentiment knowledge contained in emotional text into word vectors. Meanwhile a new sentiment polarity corpus is merged from multiple corpora. Then, a pre-training method is adopted to mitigate the impact of the limitation of annotated data for ECD task instead of simply expanding samples. Furthermore, we apply attention mechanism to take emotional context into consideration based on the observation that the context around emotion keywords can provide emotion cause clues. Experimental results show that our model greatly outperforms the state-of-the-art baseline models.