Learning to Respond with Deep Neural Networks forrnRetrieval-Based Human-Computer Conversation System

摘要

To establish an automatic conversation system between humansrnand computers is regarded as one of the most hardcore problemsrnin computer science, which involves interdisciplinary techniques inrninformation retrieval, natural language processing, artificial intelli-rngence, etc. The challenges lie in how to respond so as to maintainrna relevant and continuous conversation with humans. Along withrnthe prosperity of Web 2.0, we are now able to collect extremelyrnmassive conversational data, which are publicly available. It castsrna great opportunity to launch automatic conversation systems. Ow-rning to the diversity of Web resources, a retrieval-based conversa-rntion system will be able to find at least some responses from thernmassive repository for any user inputs. Given a human issued mes-rnsage, i.e., query, our system would provide a reply after adequaterntraining and learning of how to respond. In this paper, we proposerna retrieval-based conversation system with the deep learning-to-rnrespond schema through a deep neural network framework drivenrnby web data. The proposed model is general and unified for dif-rnferent conversation scenarios in open domain. We incorporate thernimpact of multiple data inputs, and formulate various features andrnfactors with optimization into the deep learning framework. In thernexperiments, we investigate the effectiveness of the proposed deeprnneural network structures with better combinations of all differen-rnt evidence. We demonstrate significant performance improvementrnagainst a series of standard and state-of-art baselines in terms ofrnp@1, MAP, nDCG, and MRR for conversational purposes.