Multi-sentence level natural language generation for dialogue system

Abstract

In multi-sentence level natural language generation (NLG) system, the first task is to classify the constraint-value pairs into several groups. Following that, each group will be translate to a talk session. In this paper, we propose three classification algorithms. The first method looks for talk session group directly, the second one applies sentence frame search, and the third one utilizes dynamic programming. After classification, we propose sentence level and phrase level generation method. Although these algorithms are not sophisticate, they work well due to easy training, fast response and high quality. Dialogue system for operating in-car devices and services require fast system response time, and some PND applications are rare in memory/computational resources, so this method would be a good choice for these environments.