Research On Quality Estimation Of Human-machine Interaction

With the increasing demand for information consultation,there are more and more scenes of human-computer dialogue interaction.In a typical conversation,the user comes with a clear purpose,and wants to get information or services that meet certain conditions,such as ordering a meal,booking a ticket,looking for music,movies,or a certain item,etc.,and because the user's needs may be complicated,There are several rounds of presentations,and users may continually modify or refine their needs during the conversation.In addition,when the user's stated needs are not specific and clear,the machine can also help the user find satisfactory results by asking,clarifying or confirming.Therefore,it is necessary to extract the information characteristics expressed by the user from the dialogue process and establish a suitable dialogue quality evaluation model,which helps people to improve the machine language dialogue skills,which is of great significance for improving user satisfaction.In the work of this paper,we focus on the semantic features of multi-round dialogues,use a variety of deep network models to extract language features,and propose relevant models for dialogue with Chinese.The main work includes the following aspects:In this paper,we first use the statistical dialogue features to learn,including LR,SVM,decision tree and other traditional machine learning methods to train the sample data with artificial time features and to extract the interactive historical features automatically with the cyclic neural network model.By comparing the results of the proposed indicators,we find that the neural network model can effectively capture the characteristics of the dialogue sequence.Then,based on the NLP related research foundation,only from the dialogue text,no more manual feature extraction is done.Bi LSTM,1D-CNN and Multi-head Attention model are proposed to automatically learn the context characteristics of each word in each interactive text.Combining with historical dialogue information,each dialogue vector is generated,thus completing the training of the model,and achieving the traditional manual mention.The results of similar interactive features are obtained,and the applicability of NLP method in the case of actual text recognition errors is verified by adding erroneous words randomly.Finally,considering that the origin of Chinese characters is image information,which is different from the origin of English pronunciation,a training corpus and training data set are constructed for Chinese interactive scene.A method of extracting Chinese glyph features in dialogue text using CNN model is proposed.Compared with traditional methods in NLP,the introduction of Chinese characters can effectively improve the expression of dialogue semantics.