Conversational Scaffolding: An Analogy-based Approach to Response Prioritization in Open-domain Dialogs

We present Conversational Scaffolding, a response-prioritization technique that capitalizes on the structural properties of existing linguistic embedding spaces. Vector offset operations within the embedding space are used to identify an ‘ideal’ response for each set of inputs. Candidate utterances are scored based on their cosine distance from this ideal response, and the top-scoring candidate is selected as conversational output. We apply our method in an open-domain dialog setting and show that the most effective analogy-based strategy outperforms both an Approximate Nearest-Neighbor approach and a naive nearest neighbor baseline. We also demonstrate the method’s ability to retrieve relevant dialog responses from a repository containing 19,665 random sentences. As an additional contribution we present the Chit-Chat dataset, a high-quality conversational dataset containing 483,112 lines of friendly, respectful chat exchanges between university students.