Towards a Fast Response Selection: Selecting the Optimal Dialogue Response Once for All

Abstract

Response selector, as an essential component of dialogue systems, aims to pick out an optimal response in a candidate pool to continue the dialogue. The current state-of-the-art methods are mainly based on an encoding paradigm called Cross-Encoder, which separately encodes each context-response pair and ranks the responses according to their fitness scores. However, such a paradigm is both inefficient and ineffective. Specifically, it has to repeatedly encode the same context for each response, which results in heavy inference cost. Also, without considering the relationship among the candidates, it is difficult to tell which one is the best candidate purely based on the fitness score of each candidate. To address this problem, we propose a new model called Panoramic-Encoder, which accepts all candidates and the context as inputs at once and allows them to interact with each other through a specially designed attention mechanism. Our method also allows us to naturally integrate some of the effective training techniques, such as the in-batch negative training. Extensive experiments across four benchmark datasets show that our new method significantly outperforms the current state-of-the-art while achieving approximately 3X speed-up at inference time.

Tasks

Reproductions