Design Of Human-Computer Interaction System Based On Semantics

As a friendly interaction method,human-computer voice interaction is one of the important research directions in the domain of artificial intelligence.Due to the complexity of voice signals and the presence of environmental noise interference factors,the research on humancomputer voice interaction is still difficult to reach the level of continuous interaction between people.The human-computer interaction system implemented by the traditional method can no longer meet the needs of people to conduct multiple rounds of deep-level interaction with the machine.In order to achieve a more friendly human-computer voice interaction method,this paper uses deep learning algorithms to design and implement a semantic-based human-computer interaction system.The main research content includes:(1)Design of overall scheme of human-computer interaction system based on semantic.Based on the research and combing of the current research status of human-computer interaction system,according to the functional requirements and performance index requirements of the semantic-based human-computer interaction system,the overall scheme of the semantic-based human-computer interaction system is designed.The system consists of four modules: speech recognition,speech synthesis,single-round interaction and multiround interaction.(2)Speech recognition and speech synthesis.This dissertation combines the deep learning algorithm with the traditional hidden Markov method to complete the speech recognition test on part of the daily conversation data in the open source Weibo data.And the average word error rate(WER)remains below 4.09%.The Pixel CNN architecture-based Wave Net is adopted to perform a speech synthesis test on the above-mentioned open source Weibo text data,where the root mean squared error(RMSE)is only 0.121.(3)Design of single-round and multi-round voice interaction module.For the single-round interaction module,this dissertation crawls the noun entries and their definitions on Wikipedia as a data set,uses the Sequence to Sequence model to implement the noun question and answer mechanism.And the average accuracy rate of the experimental test reaches 98.6%.The multi-round interaction is more complicated than single-round interaction.This dissertation uses open source Weibo data as the multi-round interaction data set,and uses the generative pre-training(GPT-2)model to train the multi-round interaction model.And the best model obtained in the experiment has accuracy,recall,and F1-Measure evaluation index values of 0.81,0.85,and 0.83,respectively.And greedy matching value reaches 0.6531.(4)Design of human-machine interface,and verification of the system function and performance.This dissertation uses the wx Python GUI library provided by Python to design the system man-machine interface.After that,this dissertation conducts experiments to verify the functional requirements and performance indicators proposed in the scheme.The final experimental results show that the designed system is within the effective working range,and the average speech recognition accuracy rate reaches 98.1%.The overall response time of the system in the single-wheel and multi-wheel interactive function modes is kept within three seconds.And the system functions and performance can meet the requirements of the program design.