Avoiding Echo-Responses in a Retrieval-Based Conversation System

Retrieval-based conversation systems generally tend to rank high responses that are semantically similar or even identical to the given conversation context. While the system's goal is to find the most appropriate response, rather than just semantically similar, this tendency results in low-quality responses. This challenge can be referred to as the "echoing problem". To minimize this effect, we apply a hard negative mining approach at the training stage. The evaluation shows that the resulting model reduces echoing and achieves the best quality metrics on the benchmarks.