Personalized response generation for customer service agents

Abstract

Natural language generation is a critical component of dialogue system and plenty of works have proved the effectiveness and efficiency of sequence-to-sequence (seq2seq) model for generation. Seq2seq model is a kind of neural networks which usual require massive data to learn its parameters. For many small shops in customer service dialogue systems, there is not large dialogue dataset to be utilized to train this model, resulting in performance of trained model cannot meet real application requirements. In this work, we present the Tensor Encoder Generative Model (TEGM) collaborating data of many shops in customer service dialogue system, and expect to alleviate the disadvantage of data insufficiency. The generator fully trained from data can be capable of encoding personalized feature of each shop. Experimental results show that the TEGM indeed can improve performance compared to baseline.