Small Changes Make Big Differences: Improving Multi-turn Response Selection in Dialogue Systems via Fine-Grained Contrastive Learning

Retrieve-based dialogue response selection aims to find a proper response\nfrom a candidate set given a multi-turn context. Pre-trained language models\n(PLMs) based methods have yielded significant improvements on this task. The\nsequence representation plays a key role in the learning of matching degree\nbetween the dialogue context and the response. However, we observe that\ndifferent context-response pairs sharing the same context always have a greater\nsimilarity in the sequence representations calculated by PLMs, which makes it\nhard to distinguish positive responses from negative ones. Motivated by this,\nwe propose a novel \\textbf{F}ine-\\textbf{G}rained \\textbf{C}ontrastive (FGC)\nlearning method for the response selection task based on PLMs. This FGC\nlearning strategy helps PLMs to generate more distinguishable matching\nrepresentations of each dialogue at fine grains, and further make better\npredictions on choosing positive responses. Empirical studies on two benchmark\ndatasets demonstrate that the proposed FGC learning method can generally and\nsignificantly improve the model performance of existing PLM-based matching\nmodels.\n