Subject Recognition in Chinese Sentences for Chatbots

摘要

Subject (In this paper, subject means '主体/zhu ti' in Chinese, while we use 'grammatical subject' to denote traditional '主语/zhu yu' in Chinese.) recognition plays a significant role in the conversation with a Chatbot. The misclassification of the subject of a sentence leads to the misjudgment of the intention recognition. In this paper, we build a new dataset for subject recognition and propose several systems based on pre-trained language models. We first design annotation guidelines for human-chatbot conversational data, and hire anno-tators to build a new dataset according to the guidelines. Then, classification methods based on deep neural network are proposed. Finally, extensive experiments are conducted to testify the performance of different algorithms. The results show that our method achieves 88.5% F_1 in the task of subject recognition. We also compare our systems with three other Chatbot systems and find ours perform the best.