Research On Multiple Choice Question Answering Technology Based On Graph Neural Network

The question answering system aims to analyze users’ questions through semantic understanding and provide users with the accurate and clear answer.It is of great value in improving the understanding and cognitive ability of machines and is a research hotspot in the field of artificial intelligence.As one of the important research tasks in the question answering system,multiple choice question answering is given a question and chooses the correct answer among the candidate answers,which reflects the machine’s ability to understand the question and answer and the relationship mining ability.The research on multiple choice question answering technology is of great significance to realize a more intelligent and efficient question answering system.In practical scenarios,since users’ questions are usually short and contain very limited information,researchers usually extract question-related knowledge from large-scale knowledge graphs to construct knowledge subgraphs to provide more information.However,there are two main difficulties in combining knowledge subgraphs to complete multiple choice question answering: 1)how to effectively establish information interaction between the question and the knowledge subgraph,and fully mine the associated information between them,so as to better understand the question and provide important support for answer selection;2)how to effectively capture the relationship between nodes and edges in the knowledge subgraph,learn better knowledge subgraph representations,in order to help the model better understand the knowledge information in the subgraph.In response to the above difficulties,the main research contents of this thesis are as follows:1)To fully mine the associated information between the question and the knowledge subgraph,this thesis proposes a multiple choice question answering model based on a hybrid attention mechanism.The model consists of three parts: encoding layer,hybrid attention layer and feature fusion layer.Firstly,the encoding layer uses the pre-trained language model and graph neural network model to encode the question and the knowledge subgraph,respectively.Secondly,the hybrid attention layer incorporates node type and relation type information into the representation of the knowledge subgraph using node attention and relation attention,and focuses on the node information in the knowledge subgraph that is relevant to the question through context attention,thus enabling multidimensional and deep interaction between the question and the knowledge subgraph.Finally,the feature fusion layer further fuses the information of the question and the knowledge subgraph to guide the final answer selection.Experimental results show that the proposed model outperforms the baseline model significantly in the multiple choice question answering tasks.2)To learn a better knowledge subgraph representation,this thesis proposes a multi choice question answering model based on subgraph representation augmentation.First,the model introduces the subgraph reconstruction task as an auxiliary task to mask the edge information in the knowledge subgraph.During training,the masked edge information is predicted to better capture the relationship between nodes and edges,thereby enabling the model to learn better subgraph representations.In addition,considering that some edge information in the knowledge subgraph is crucial for understanding the question,global random masking may lead to important information loss and negative impact on model learning.Therefore,this thesis proposes a dynamic masking strategy based on relevance,which selects edges with lower relevance to the question for masking,effectively avoiding the loss of important information and ensuring the learning of the subgraph reconstruction task.Experimental results show that the proposed model effectively improves the accuracy of the multiple choice question answering task.