Research And Implementation Of Task-oriented Dialogue System

Task-oriented dialogue system can accurately understand the user intention through natural language chat,automatically construct and execute tasks,and reply the results of task execution.It can accomplish various kinds of user information tasks in a more natural way,which greatly improves user experience and task execution efficiency,so it has become one of the research hotspots in both academia and industry.However,task-oriented dialogue systems applied in different domains need a large number of historical dialog data in this domain,and most platforms often lack such data accumulation.It makes task-oriented dialogue systems face the cold start problem.At the same time,the accuracy of user intention understanding needs to be further improved.In this paper,we research on the key technologies of task-oriented dialogue,and puts forward a novel approach to task-oriented dialogue systems in cold-start environment.Our approach combines rule based method and machine learning method to classify user intention,alleviating the influence of cold start problem on the performance of machine learning model;it designs an integrated CRF model to fill slots of user intention,and builds and utilizes a domain knowledge base to solve OOV(Out-of-Vocabulary)problem for helping improve the accuracy of slot filling.Based on these technologies,we design and implement a task-oriented dialogue system engine,which supports accurate understanding and acquisition of user intention through multi-round dialogues,automatically builds task API calls,and generates natural language answers according to task execution results.The main contributions of this paper include:1)We research and put forward a user intention classification approach combining rule based method and machine learning method,which effectively alleviates the cold start problem.During rule-based intention classification,we design a rule base and propose a heuristic Chinese word similarity calculation method and a rule fuzzy matching algorithm.During machine learning intention classification,we design word2 vec and n-gram features,and propose an integrated learning model based on SVM,NaiveBayes and decision tree.Then,the weighted fusion strategy of these two methods is proposed to complement each other and improves the performance of intention classification.Experiment results show that the F1-score of our approach for user intention classification can reach 82% under cold start environment,outperforming the baselines.2)We research and put forward the slot filling method based on domain knowledge base,which effectively alleviates the OOV problem and improves the accuracy of slot filling.In this paper,a domain knowledge base is constructed to normalize entities.At the same time,an integrated CRF model is proposed to train multiple CRF models for different user intend,and weighted merge the processing results of each CRF model according to the classification probability distribution results to obtain the final slot filling results,which improves the robustness of the model and the accuracy of slot filling.Experiments show that the F1 value of slot filling in the proposed integrated CRF model reaches 92%,which is higher than that in CRF model(88%)and RNN model(83%).3)Based on proposed technologies,we develop an engine of task-oriented dialogue system for software crowdsourcing domain.The engine has been integrated with the software crowdsourcing platform and passed a series of tests.The test results show that the system can help users to complete various tasks conveniently through natural language interaction.The accuracy of user tasks completion reaches 88.2%,and the average response time of single-round conversation is 0.695 seconds,which has achieved the desired goal.