Recommendation Model And Approach Based On Emotional Compensation For Speech Environment

Recommender systems are a subclass of information filtering system that seek to predict the preference that a user would give to an item.Recommender systems promote the development of electronic business better.The speech application like online instant voice communication,internet telephony and streaming media player has developed rapidly in the multimedia and big data era.At the present stage recommender systems pay more attention to the text mining and knowledge extraction,while the recommender system for instant voice and speech emotional environment such as the speech of chat and movie lines has not appeared.However,this speech information is closely linked to user's current interests and needs.So the recommendation method based on speech content detection is essential.In contrast to textual data which be dealt with the tradition recommender systems,speech data contain more emotional information which can better reflect characteristics of individual habits and behaviors.Therefore,the speech emotional information may improve the performance of the recommendation effectively.This paper focuses on the research of recommendation model and approach based on spoken term detection for instant voice,analyzes and solves the key problems of speech detection in the recommendation environment,explores the emotional compensation based on speech signal emotion and speech content sentiment,and solves the speech recommendation problems like negative emotion backoff,cold start and inaccurate user match by using the speech emotional information.Summarily,our contributions of this work are as follows:(1)This paper has analyzed the two main problems of speech detection of emotional compensation recommendation.First,in order to improve the classification accuracy of speaker diarizaiton,a long-term feature selection based on heuristic strategy for speaker diarization is proposed.It is used to generate new long-term features with high speaker discrimination.This method has greatly improved the ratio of inter-speaker variance and intra-speaker variance of long-term features,which is compared with the measure baseline method of Fisher discriminant analysis.Furthermore,Experiments on the open corpus revealed that the improved long-term feature set increase the accuracy of the speaker diarization.Second,in order to improve the coverage of linguistic model,an improved two-stage spoken term detection method is proposed.The proposed method,which design is based on the lattice extension using the spoken words of recommendation domain knowledge,is used to accommodate the recommendation environment.(2)Speech personality traits are not only the reflection of speakers' thought patterns,but also a most efficient way to acquire the speakersf habits and behaviors.As affective computing develops and grows,speech personality assessment has attracted attention for the past few years.In order to improve the performance of speech personality assessment,a speech personality assessment model based on multilayer perceptron is proposed.This model is based on multilayer perceptron and prosodic features selection to acquire speaker's speech personality traits.All of the parameters of the proposed model are determined accurately with gradient descent algorithm.Furthermore,by combining syllables per minute of the clip feature and syllables per minute of the speaker feature,the optimum feature subset has been successfully implemented using the feature selection based on mutual information.In practical instances of Chinese speech personality assessment,Chinese speaker personality corpus base on Big Five personality is constructed.Besides,the specialization of feature selection for Chinese speech is simulated and analyzed.The results showed that the proposed model can achieve superior performance than other state-of-the-art speech personality assessment method.(3)Sentiment tendency of speech content provide critical information for recommender systems.This paper has presented two improved text sentiment classification methods on the foundation of speech-to-text conversion.First,for the lack of an effective method for adjusting to the user's key representation,a supervised sentiment classification method adapt for text representation is proposed.Though adjusting the weighting parameters of user's text representation,the sentiment classification performance is improved.The results demonstrated the effectiveness of our proposed method.Second,writing style features,such as lexical and word-based features,are inappropriate when directly applied to the sentiment classification model.A sentiment classification model using group characteristics of writing style features is proposed.This paper determined the optimum clustering number of user reviews based on writing style features distribution.According to the classification model trained on a training subset with specific writing style clustering tags,this paper determined that the model trained on the data set of a specific writing style group has an optimal effect on the classification accuracy,which is better than the model trained on the entire data set in a particular positive or negative polarity.Through the polarity characteristics of specific writing style groups,a general model in improving the performance of the existing classification approach is presented.Results of the experiments on sentiment classification using the sentiment data set demonstrated that the proposed model improves the state-of-the-art performance in terms of classification accuracy.(4)On the basis of the previous research,this paper has studied the recommendation method based on speech emotion compensation.First of all,for the current recommendation methods are to deal with the text content,this paper has presented a recommended category detection model based on instant voice content.Real-time interest weight vectors and historical interest weight vectors of user preference categories are established using the results of spoken term detection.The cosine distance of the weighted vector space description model is then used to derive the subject category of speech.Then,the recommendation model is compensated from the two aspects of speech personality and speech content emotion.On the one hand,in order to avoid recommending when the emotion of a particular topic category is negative,this paper proposed a recommended back-off strategy based on speech content emotional compensation.The backoff function combines the commendatory semantic favorability and the topic category weight.When the detection result is lower than the recommended threshold,the recommended category retreat can improve the user experience.On the other hand,a new recommendation method based on user speech personality is proposed to solve the common cold start problem of collaborative filtering.The speech personality does not need long time to collect user behavior,which solves the cold start problem of new users and the similarity problem of user preference.Experiments show that the proposed method can not only detect the recommended topic categories from instant voice,but also can compensate the recommendation performance through two kinds of speech emotion,which has important guiding significance and application prospect.