Research On Chinese Service Instruction Parser Based On Reinforcement Learning And Feature Tree Structure

With the development of society,people have more and more demands for service robots,especially for home-service robots.It is one of the research topics to parse natural language service instructions into machine instructions by means of natural language processing,and the key of parsing is how to make agent identifying the internal logic of Chinese service instructions.Based on research status at home and abroad,and taking indoor family environment as our background,this dissertation explores the parsing of Chinese service instructions by combining reinforcement learning with feature tree theory proposed in this dissertation.Firstly,collecting instruction set based on family service to build our corpus.The corresponding instructions are selected from corpus as a training set and test set,and define conjunctions as corresponding feature words,annotate training set instructions.Then using reinforcement learning method training agent to identify instruction conjunctions in logic.We input our training set through Q-Learning method to update Q table.And after that our agent can gradually understand the logical relationship among Chinese service instructions in "understanding".When dealing with relatively complex context,our training set can be combined with reinforcement learning and context model training.By constructing different contexts,our context model and reinforcement prediction are combined in a new prediction to achieve the function of simulating the actual scene context.After completing the agent training,try to start using our agent to test the test set,with observing our test results and adjusting parameters to improve prediction accuracy.If prediction results are not consistent with the actual instruction logic after prediction,a man-machine conversation module will be opened to guide identifying these Chinese service instructions by constructing a feature tree.Finally,this dissertation verified the feasibility and effectiveness of this method through experiments,compared the experiments of this model under different conditions and other dissertations in the same direction,and analyzed the experimental results.