Research On Multi-Turn Dialogue Generation Method Based On Improved Context Modeling

Recent years have witnessed a spurt of progress in Internet of Everything(IOE)and Human-Machine Interaction(HMI).Multi-turn Dialogue system,as an extensivelyapplied interaction medium,has penetrated into human life,including but not limited to customer service dialogue,AI speaker and Auto Cockpit Technology.Also,owing to its overwhemling strengths in upgrading user experience when accessing information and assistance on the interaction between speech instructions,relevant platforms and technology have been introduced into a multitude of fields such as military,law,medicine,government and enterprise.Multi-turn Dialogue system is of great research and application value.Multi-turn Dialogue Generation is categorized as a generative dialogue focusing on continuous dialogues and intricate semantic interaction,in which meaningful,diverse replies are available to users fluently on the basis of the interactive record produced by the users and AI within a specific period of time.Nevertheless,previous researchers only attach emphasis to the progressive modeling of semantic sequence,lacking detailed exploration in scope and depth regarding complicated semantics and correlative modeling on long-sequence semantics such as the neglect on the dialogue topics’ impetus on the dialogue context,multi-turn dialogue’s long-distance dependence on the text which is far from now,and so forth.Consequently,centring on Multi-turn Dialogue Response Generation,taking intensive research and analysis on dialogue system theory as the foundation,drawing the organization together with conclusion on the related technology concerning text generation,the thesis aims to putting forward two optimum proposal around the existing problems that the modeling on dialogue context semantics confronts now:(1)The Topic Fine-Grained Information Interaction method for Multi-turn Dialogue GenerationAiming at the difficulty of fine-grained modeling for complicated semantics in multi-turn dialogues,the paper,taking the topic awareness for reference,proposes the topic fine-grained information interaction method for multi-turn dialogue generation.The method endeavoring to figure out the complicated semantics information of dialogue content through fine-grained semantic attention on different topics.Specifically speaking,the paper firstly utilizes the hierarchical semantic emcoding module to obtain the sentence semantic representation of the dialogue-level,and then the attention mechanism and topic keywords are used for fine-grained information interaction one by one to obtain the finegrained sentence semantic representation for implementing the dialogue generation.The fine-grained semantic representation can be comapred to the scalpel-cut disassemblyrecombination according to topic,faciliating the complicated semantics of the whole dialogue to be simplified.Validated by two benchmark datasets,the notable result of the corresponding model verifies the authenticity and validity of the in-depth information interaction between topic semantics and dialogue content.(2)The Dual-Channel Collaborative Semantic Modeling method for Multi-turn Dialogue GenerationAiming at the difficulty of difficult association modeling for long-distance semantics in multi-turn dialogues,the paper,combining the advantages of the graph neural network,proposes dual-channel collaborative semantic modeling method for multi-turn dialogue generation.The method endeavors to obtain obtain information correlation and semantic inference across a larger span by fusing the semantic advantages of sequential channel and graphical channel.That is,for one thing,a cognitive graph on dailogues is constructed,with the graph nodes as the integration of topic semantics and sentence cluster semantics,and the edges as the presentation by the overlap of topics in sentence clusters,following which we resort to the dual-gated graph neural network for deep learning so as to attain the semantic representation of the dialogue context on the graphical structure,for another,we acquire the sequential semantic representation on the dialogue context by embedding a hierarchical attention mechanism in the preserved sequential channels.As the final step,we coordinate the information contribution made by two semantic representations to realize the prediction.On the basis of benchmark model,the model we establish outperforms stands out in mitigate/meliorating the matter of long-distance dependence on semantics.In brief,with a watchful eye to the complicated semantics and the graphical-structure inference on long-sequence semantics respectively,the two proposals are dedicated to modelling the context,semantics of multi-turn dialogues fleetly and directly so that corresponding and valuable response can be generated.Our schemes can broaden the research ideas to a great extent when studying the Human-Machine Interaction Systems in continuous dialogues and complicated circumstance.