Contextual Fine-to-Coarse Distillation for Coarse-grained Response Selection in Open-Domain Conversations

TL;DR

This paper introduces a novel Contextual Fine-to-Coarse distillation model for improving coarse-grained response selection in open-domain dialogue systems, leveraging knowledge distillation from fine-grained models.

Contribution

It proposes a new CFC model that distills knowledge from fine-grained to coarse-grained architectures for better response retrieval.

Findings

Significant performance improvements over baselines

Effective knowledge distillation from fine to coarse models

Validated on new Reddit and Twitter datasets

Abstract

We study the problem of coarse-grained response selection in retrieval-based dialogue systems. The problem is equally important with fine-grained response selection, but is less explored in existing literature. In this paper, we propose a Contextual Fine-to-Coarse (CFC) distilled model for coarse-grained response selection in open-domain conversations. In our CFC model, dense representations of query, candidate response and corresponding context is learned based on the multi-tower architecture, and more expressive knowledge learned from the one-tower architecture (fine-grained) is distilled into the multi-tower architecture (coarse-grained) to enhance the performance of the retriever. To evaluate the performance of our proposed model, we construct two new datasets based on the Reddit comments dump and Twitter corpus. Extensive experimental results on the two datasets show that the proposed methods achieve a significant improvement over all evaluation metrics compared with traditional baseline methods.