12 December 2022 in IEEE Access
Users on the internet usually have conversations on interesting facts or topics along with diverse knowledge from the web. However, most existing knowledge-grounded conversation models consider only a single document regarding the topic of a conversation. The recently proposed retrieval-augmented models generate a response based on multiple documents; however, they ignore the given topic and use only the local context of the conversation. To this end, we introduce a novel retrieval-augmented response generation model that retrieves an appropriate range of documents relevant to both the topic and local context of a conversation and uses them for generating a knowledge-grounded response. Our model first accepts both topic words extracted from the whole conversation and the tokens before the response to yield multiple representations. It then chooses representations of the first N token and ones of keywords from the conversation and document encoders and compares the two groups of representation from the conversation with those groups of the document, respectively. For training, we introduce a new data-weighting scheme to encourage the model to produce knowledge-grounded responses without ground truth knowledge. Both automatic and human evaluation results with a large-scale dataset show that our models can generate more knowledgeable, diverse, and relevant responses compared to the state-of-the-art models.
8 July 2020 in IEEE Access
Knowledge-grounded conversation models aim at generating informative responses for the given dialogue context, based on external knowledge. To generate an informative and context-coherent response, it is important to conjugate dialogue context and external knowledge in a balanced manner. However, existing studies have paid less attention to finding appropriate knowledge sentences from external knowledge sources than to generating proper sentences with correct dialogue acts. In this paper, we propose two knowledge selection strategies: 1) Reduce-Match and 2) Match-Reduce and explore several neural knowledge-grounded conversation models based on each strategy. Models based on Reduce-Match strategy first distill the whole dialogue context into a single vector with salient features preserved and then compare this context vector with the representation of knowledge sentences to predict a relevant knowledge sentence. Models based on Match-Reduce strategy first match every turn of the context with knowledge sentences to capture fine-grained interactions and aggregate them while minimizing information loss to predict the knowledge sentence. Experimental results show that conversation models using each of our knowledge selection strategies outperform the competitive baselines not only in terms of knowledge selection accuracy but also in response generation performance. Our best model based on Match-Reduce outperforms the baselines in the comparative studies with the Wizard of Wikipedia dataset. Also, our best model based on Reduce-Match outperforms them with the CMU Document Grounded Conversations dataset.