Keyword-Guided Neural Conversational Model

We study the problem of imposing conversational goals/keywords on open-domain agents, where agent is required to lead conversation a target keyword smoothly and fast. Solving this enables application agents in many real-world scenarios, e.g., recommendation psychotherapy. The dominant paradigm for tackling 1) train next-turn classifier, 2) keyword-augmented response retrieval model. However, existing approaches have two limitations: training evaluation datasets classification are directly extracted from conversations without human annotations, thus, they noisy low correlation with judgements, during transition, solely rely similarities between word embeddings move closer keyword, which may not reflect how humans converse. In paper, we assume that grounded commonsense propose keyword-guided neural model can leverage external knowledge graphs (CKG) both transition retrieval. Automatic evaluations suggest improves performance prediction addition, self-play show our produces responses smoother reaches faster than competitive baselines.