Research On Task-oriented Chinese Multi-turn Dialogue Algorithm Based On End-to-end Learning

With the popularization and development of Internet,information communication and other artificial intelligence technologies,a variety of human-computer interaction products emerge in endlessly.Human-computer interaction based on man-machine dialogue is a new way to communicate with computing devices,which is widely used in daily work and life.Task-oriented dialogue in man-machine dialogue has more practicability and research value,so the research object of our paper is task-oriented dialogue.Historically,task-oriented dialogue algorithms are mainly divided into two,one is built as a pipeline,and the other is based on end-to-end learning structure.They can basically realize the function of task-oriented dialogue,but there are still the following problems in most of the current public research work:(1)the dialogue policy management is completely dependent on the state tracker.The state tracker summarizes the observable session history into state characteristics,but most methods are based on design rules or special tags.(2)the existing end-to-end models do not add domain knowledge restrictions,and most of them are trained based on separate cyclic neural networks,so that the model can not well understand the intention of users,and some of the responses generated are sometimes difficult to understand.(3)the data in the field of dialogue is often private and difficult to obtain,but we hope that there are many and diversified training data,so the data enhancement in the field of dialogue are also worth exploring.In order to solve the above problems,our paper improves the task-oriented dialogue model from two parts: state tracking and encoding-decoding,and proposes a domain restriction Chinese task-oriented multi-round dialogue model based on end-to-end learning.It can understand the dialogue input accurately and in real time and generate reasonable feedback.The specific improvements are as follows:(1)In our paper,a multi-slot dialogue state tracking model based on NBT(Neural Belief Tracker)is proposed.By reasoning on the pre-trained word vector,the user discourse and the information before and after the conversation are combined into a distributed representation.Using the BERT(Bidirectional Encoder Representation from Transformers)model as the pre-training model of expressive learning,the pre-and post-information in all layers of its neural network can be trained together to express the results more accurately.Then multiple binary classifications are performed to associate multiple values with a single state variable to achieve the goal of identifying multiple slot key-value pairs of the same type and different types in the conversation.In addition,this model can also use the connection to the underlying application to automatically expand the slot key value.Finally,the results of this module are also used as a supplement to the field of dialogue generation,making the feedback closer to the current task.(2)Our paper proposes an encoding-decoding model which combines bi-directional LSTM and self-attention mechanism,which can not only capture local key information,but also solve the long-term dependency in dialogue,and has good parallelism.(3)Our paper also proposes a data augmentation based on neural machine translation(NMT)and bilingual dictionary.The NMT model is used to translate and back-translate the original data.However,due to the particularity of the field of dialogue,a large number of UNK will be produced in translation.So our paper uses bilingual dictionaries to replace the resulting UNK.This method can not only effectively expand the data,but also optimize the UNK problem,and obtain higher quality data.In order to verify the effectiveness of the above model,the above model is deployed on the Deep Pavlov platform and verified by experiments using the open task-oriented dialogue dataset Cam Rest676-Chinese.Then the task-oriented dialogue model is compared with the existing mainstream methods and its basic model,which shows that our model can effectively improve the performance of the task-oriented dialogue system and accurately complete the task.