关键词: 自然语言处理, 情感分析, 深度学习, 共情预测, 迁移学习

Abstract: Empathy prediction from texts has achieved little progress due to the lack of sufficient labeled data, while its related task, polarity classification with a large number of labeled samples, has been studied thoroughly. Therefore, we propose an empathy prediction method augmented by the polarity classification task. The proposed method erases the domain gaps and the label gaps between the empathy prediction and the polarity classification tasks, and successfully leverages the features learned from the polarity classification task to aid the empathy prediction task. First, we propose an attention-based feature fusion strategy, which assigns dynamic weights to the common features and the private features according to their contributions to empathy prediction. Second, we propose an adversarial learning strategy to disentangle the private and common features for the two tasks, which makes the common features general to different domains and the private features specific to the task-specific domain. Third, we design a hinge loss function to guide the representation learning, where the learned common and private features are general and specific to different target labels, respectively. Experimental results on benchmark databases demonstrate the effectiveness of the proposed method for empathy prediction.

Key words: Nature language processing, Sentiment analysis, Deep learning, Empathy prediction, Transfer learning