Selecting the optimal dialogue response once for all from a panoramic view

As an essential component of dialogue systems, response selection aims to pick out the optimal response among candidates to continue the dialogue. In existing studies, this task is usually regarded as a binary classification problem, where every candidate is ranked respectively for appropriateness. To improve its performance, we reformulate this task as a multiple-choice problem that allows the best selection to be made in one-shot inference. This new view inspires us to propose an architecture called Panoramic-encoder (Our work will be open-source for reproducibility and future research.) with a novel Candidates Attention Mechanism (CAM), which allows context-wise attention between responses and leads to fine-grained comparisons. Furthermore, we investigate and incorporate several techniques that have been proven effective for improving response selection. Experiments on three benchmarks show that our method pushes the state-of-the-art while achieving approximately 3X faster inference speed.