Human Computer Interaction System Based On Preference Modeling

Human-computer interaction aims to make the machine better execute the user’s instructions.The current researches have been mainly considered on the integrity of instructions and the accuracy of the computer execution,not making a foray into emotional state and preferences of users.In this vein,this thesis mainly studies the accuracy of word segmentation after speech coversion,and the influence of users’ emotion and preferences on human-computer interaction.The main research contents and results are as follows:(1)An Attention-based Bidirectional Gated Recurrence Unit-Conditional Random Field Model(AGCRF).The attention mechanism facilitates the focus on the key input.Gated Recurrence Unit(GRU)network features fewer parameters and faster training speed than LSTM network does.And Conditional Random Field(CRF)can sequentially predict the output of tags.Based on them,the text proposes a method to have a combination of the attention mechanism and BIGRU-CRF model.As such,the AGCRF model was put into practise in the corpora of People’s Daily 2014 and MSRA.The experimental results demonstrate that this model performs well in word segmentation accuracy and speed(2)An Attention-based GRU with Convolutional Neural Network(AGCNN)Model.First and foremost,the voice signal issued by users is pre-treated,Open Smile is used to obtain the voice features,and advanced features are extracted by CNN.Moreover,BIGRU network is combined to obtain the influence of the voice signals on emotion from the forward and backward time series.In addition,the introduction of Attention helps to obtain key features and extract speech features more accurately.Furthermore,Softmax model is used to classify advanced features and obtain the current emotion of users.By comparing the accuracy and recall rate of the classification results in this model with those in other models,it can be concluded that the former one performs better than the latter ones in the effect of emotion classification.(3)Hybrid recommendation algorithm.In the tag system,in order to eliminate the problems of time migration and sparse tag data,exponential attenuation function and implicit data are introduced to construct the user preference model,obtain the user preference matrix for the project,and combine the user preference model with the item based collaborative filtering algorithm to establish the tag based user preference modeling collaborative filtering recommendation algorithm.Besides,a recommendation algorithm is established based on user emotion and music emotion attribute by taking advantage of the similarity between them.The fusion of two recommendation algorithms can avoid the shortcomings of single recommendation algorithm and improve the accuracy of recommendation results.