Research On Emotion Cause Detection By Incorporating Domain Knowledge

Social media has transformed from official news distribution platform to public platform for daily communication and emotion expression.The research on social media analysis,opinion mining and sentiment analysis based on natural language processing(NLP)technology is becoming more and more important.Based on the sentiment analysis research which aims to mining the sentiment state in social media texts,the research on emotion cause detection,which aims to detect the emotional trigger mechanism,has shown significant academic value and practical prospects.The existing methods on emotion cause extraction may be camped into three rulebased approach,machine learning based approach,and deep learning based approach.The rule-based approach is puzzled by the quality and coverage of manually complied rule set.The machine learning based approach highly relies on manually selected features,making it difficult to build end-to-end learning models.The deep learning based approach is puzzled by the learning process of deep neural networks is uncontrollable and its weak interpretability.Meanwhile,the performance of this approach highly relies on large highquality annotated data.To solve these problems,this study introduces hierarchical attention mechanism to further improve the text representation ability of existing deep learning based model.Furthermore,this study introduces domain knowledge through knowledge distillation to improve the controllability of the model.The existing emotion cause detection methods always ignore the relationships among clause sequences.Thus,a hierarchical attention network combining position encoding and residual connection is proposed in this study.The hierarchical bidirectional gated recurrent unit(Bi-GRU)is employed to model the contextual semantics at word level and clause level,while attention mechanism is employed to capture the latent semantic relationship between clauses and emotional expressions.Besides,the relative position between clause and emotional expression is represented by a position vector.At last,the final representation of clause is obtained through residual connection.The experimental results on EMNLP2016 Chinese emotion cause extraction dataset show that the proposed models has achieved performance improvement over the baseline models.The F1 value of this method is 0.004 higher than the state-of-the-art model.Current deep learning models have the disadvantage of highly dependence on annotated data and uncontrollability.To address these problems,this study further investigates the emotion cause detection method by incorporating domain knowledge.The method for incorporating domain knowledge are proposed.One method uses retraining language model to transfer the relational domain knowledge in large unlabeled texts to the current deep learning model.Another method utilizes the manually complied rules to guide the training of neural networks,and then transfer the logical domain knowledge to the network parameters based on the knowledge distillation.The experimental results show that both of the two methods outperform the state-of-the-art model.The F1 performance is improved 0.18 and 0.05,respectively,which achieves the known highest performance on this dataset.