Extractive Machine Reading Comprehension Integrating Dialogue Graph Information

Extractive machine reading comprehension is a very important technology in the field of natural language processing.Whether it is mature or not will directly affect the intelligence level of human-computer interaction.In a large number of application scenarios such as intelligent customer service and information retrieval,this technology enables machines to select answers that users need from reference documents based on people’s questions for feedback.In the past decades of development,academic research on extractive machine reading comprehension has focused on a single round of question answering,without considering the impact of historical rounds of dialogue on current issues.Although this approach has achieved a certain degree of success,it has been unable to meet people’s new requirements for artificial intelligence today.When people ask questions to machines,they usually use pronouns to replace or even omit some entities in the questions based on historical dialogues.This requires machines to have the ability to remember and connect historical dialogues when answering.In response to this demand,this paper proposes the idea of using dialogue graphs to capture and abstract the connections between entities in historical dialogues,and develops a dynamic composition tool to dynamically improve the dialogue graph during the progress of the dialogue.At the same time,in order to clarify which part of the dialogue graph structure has the most direct connection to the current problem,this paper proposes an optimized graph module based on pointer networks to select historical semantic information in the dialogue graph.Finally,this paper uses a extractive machine reading comprehension based on the attention mechanism to fuse the dialogue graph information,and extracts the final answer from the document on the basis of fusing the historical information of the dialogue graph.In summary,the main contributions of this paper are as follows:(1)Dialogue graph construction:This paper proposes a new idea of structuring the semantic information of historical dialogues,that is,to develop a high-quality dynamic composition tool that continuously extracts the semantic connections between contextual entities during the progress of the dialogue,and dynamically supplements the existing dialogue diagram.This tool uses the existing natural language processing toolkit to extract semantic triples in dialogue sentences,and on this basis,performs a series of processing such as deduplication and supplementation,and finally obtains structured semantic information that can be supplemented into dialogue graphs.Finally,in order to visualize the dialogue graph for easy testing and modification,a web project is developed in this thesis.Users can input dialog information through the browser and display the generated dialog graph.(2)Optimization of Dialogue Graph Based on Pointer Network:With the continuous advancement of the dialogue,the scale of the dialogue graph obtained by structuring the historical dialogue will become larger and larger,and the historical semantic information contained in it will also increase.If the huge dialogue graph structure is directly fused by the subsequent machine reading comprehension model,it will bring a lot of redundant information,which will directly affect the accuracy of answer extraction.On the other hand,graphs,as a special data structure,are difficult to directly integrate into natural language pre-training coding models.Based on these two reasons,this paper proposes a dialogue graph optimization module based on the pointer network,which is used to refine the structure of the dialogue graph and select the most appropriate historical semantic information in the dialogue graph before the answer is extracted.And restore this part of historical information to an optimized graph that can be directly input into the pre-trained model.(3)Extractive machine reading comprehension based on fusion of dialogue graph information based on attention mechanism:Existing extractive machine reading comprehension models cannot directly incorporate historical semantic information from the dialogue graph in this paper.Therefore,this paper proposes a new extractive machine reading comprehension model,which uses an attention mechanism to fuse optimized graph information extracted from dialogue graph optimization,while considering the previous round of historical question and answer.The experiments in this paper prove that this approach can effectively improve the final accuracy of the task.To sum up,in order to cope with the challenges brought by conversational scenes to extractive machine reading comprehension,this paper develops a dynamic composition tool to abstract and structure the semantic information of historical conversations.At the same time,this paper proposes a dialogue graph optimization module to refine the semantic information in the dialogue graph.Finally,based on the attention mechanism,this paper proposes an extractive machine reading comprehension model that can integrate dialogue graph information,and extracts answers on the basis of fully considering the optimized graph information.This paper conducts multiple comparative ablation experiments on public datasets to verify the effectiveness of the model proposed in this paper for conversational extractive machine reading comprehension tasks.