User Intention Classification in an Entities Missed In-vehicle Dialog System

摘要

In the human computer dialog system in vehicle environment, some dialogue entities are usually left out by human after several dialog turns. This causes troubles to classify user's intention in a period of chat history. A usual solution for this problem is adding context information to expand the current question. This method causes a trend to generate multiple entities in the expanded question and decreases the classification accuracy of users' intention. In this paper, an RNN based entity recognition model is built to recognize entities in the current problem. If the topic related entities are recognized, the intention and property are classified respectively using LDA and word2vec models; otherwise entities in context information are added to complete the question before intention classification. Experiments show that the proposed method has about 9.4% improvement in precision and 2.3% improvement in recall compared with the traditional context expansion method.