A Study On Neural Response Generation Algorithm For Open-Domain Conversations

In recent years,the field of human-computer interaction has increasingly become a research hotspot,especially the open domain dialogue task.With the growth of massive data on the Internet and the development of deep learning technology,data-driven dialogue generation has received extensive attention.The current mainstream approach is to use a Sequence-to-Sequence(Seq2Seq)framework to encode the conversation context and then generate the corresponding response during decoding.However,such methods are prone to ”safe response” problem,that is,the generated responses are uninformative and meaningless.According to this problem,we analyze the shortcomings of existing work and propose corresponding solutions based on the three elements of the dialogues(speaking content,different personas,and the way of human thinking during conversation): some work attempts to improve the understanding of context semantics to improve the semantic consistency of the response,but ignores the multiple dialogue patterns(transitions between topics)in the open domain dialogue,which will make the generated response much general;Other work uses character information as an additional part of the model.Although it improves the replier-consistency to some extent,such operations cannot be migrated to other corpora;most of the work uses the attention mechanism from machine translation tasks to simulate the dialogue process based on the Seq2 Seq model.However,it is different from the way of human thinking.The dynamic attention calculation process will destroy the core thought of the response.Therefore,based on these three considerations,we propose a Neural Variational Scaling Reasoning Network(NSSRN),a Semi-Supervised Stable Variational Network(SSVN)and a dynamic static Static-Dynamically Attentive Variational Network(Sdav Net).Among them,we set up a neural dialog pattern reasoner in the NVSRN model,which uses the advantage of von Mises-Fisher(v MF)distribution characterizing the directional data to learn the transfer between topics in the dialog.Then the topic scaling mechanism is designed to predict the degree of the topic shifting,and assist the generator in producing an initiative and interesting response;we introduce an unsupervised personal feature extractor in the SSVN model to capture the language style characteristics of the replier,which can improve the replier-consistency in the response generator;to simulate the way of human thinking,we design a static-dynamic attention mechanism,where the static attention is responsible for extracting the leitmotiv of the response,and the dynamic attention part expands it to a complete response.Finally,the experimental results show that the proposed networks(NVSRN,SSVN and Sdav Net)make the diversity metrics increase by 14.56%,34.09% and 63.72% on the Cornell movie dataset.This further verifies that the improvements based on these three points are effective in the open domain dialogue generation task.