Design And Implementation Of Incomplete Information Game Domain Question Answering System Based On Deep Learning

With human beings enter the second decade of the 21 st century,the intelligent question answering system characterized by natural language response has become an important way of human-computer interaction.Incomplete information game,such as competitive mahjong,has a huge player base.Players need to know a lot of game knowledge in advance when they play the game.Players search on the Internet,but they can't get accurate knowledge quickly.To solve this problem,this paper studies the automatic question answering system in the field of incomplete information game,and constructs an incomplete information game question answering system with deep learning technology.It aims to meet the learning needs of players,play the role of assistant decision-making,improve the game experience of players,and provide reference for the application of question answering system based on deep learning in other similar demand scenarios.The work of this paper can be divided into the following parts(1)At present,there is no public knowledge base and data set in the field of incomplete information game.This paper obtains the corpus of incomplete information game based on scrapy crawler framework by means of competitive mahjong enthusiast forum,teaching books,Baidu Encyclopedia and other ways.In this paper,the joint knowledge extraction method based on special annotation strategy is adopted.On the basis of manually annotated data,the end-to-end joint extraction model based on Bert bilstm CRF is used to mine the triple knowledge in unstructured and semi-structured text.And through the manual rule audit,it is stored in the graph database neo4 j.At the same time,in order to make up for the limitations of structured knowledge,this paper also constructs the mahjong domain question answering pair data set,and uses the manual rules to automatically construct the positive and negative example data set for training the deep semantic matching model.(2)Design a question answering scheme of incomplete information game based on structured knowledge.The question and answer is divided into two sub tasks: entity chain index and relationship detection.In the entity chain index module,the sequence annotation model based on Bert fine-tuning is used to efficiently recognize entity words in incomplete information game field.In the relationship detection module,we use the relationship detection model based on depth residual mechanism and multi granularity attention mechanism.The model uses attention mechanism to model question pattern and relation representation at multi granularity,and uses residual mechanism to enhance the trainability of the model.In this paper,the relationship detection experiments are carried out on the public data set and the competitive mahjong knowledge base respectively.On the competitive mahjong knowledge base,the model used in this paper achieves the highest accuracy of 96.77%.(3)Design an incomplete information game question and answer scheme based on question and answer pair knowledge.The question answering task is abstracted as a semantic matching problem,and a deep semantic matching model based on Bert pre training word vector and bidirectional attention mechanism is used.In the embedded layer,the model uses the best to pre train the word vector,and uses the bidirectional pooling attention mechanism to extract the row and column direction features of the attention matrix.The experiment of answer selection and interpretation recognition on mahjong Q &a pair data set and open data set proves the feasibility and superiority of this model applied to Knowledge Q &A in incomplete information game field.(4)Integrate structured and Q &A algorithm,and realize the incomplete information game Q &a system based on deep learning scheme.The system algorithm adopts Python language and keras deep learning framework,and uses the architecture design based on C / S mode.It receives user's questions on the client,analyzes and searches the answers on the server,and displays the answers on the frontend android app.Finally,the system is implemented After manual test,the answer accuracy rate reached 78.3%.