Social sentiment analysis model representation method based on deep learning and storage medium

Abstract

The invention discloses a deep learning-based social sentiment analysis model characterization method and a storage medium, and the method comprises the steps: obtaining user uploading information and corresponding comment information under a single entry of a social platform, and carrying out the processing of the user uploading information and the comment information, obtaining first-level feature single-mode text information and second-level feature single-mode text information; analyzing the first-level single-mode text information and the second-level feature single-mode text information based on a deep learning model to obtain a first-level emotion weight matrix and a second-level emotion weight matrix; the method comprises the following steps: collecting user historical information of uploaded content under a single entry and display data of the uploaded information, analyzing the user historical information to obtain a matching coefficient of the user historical information and the entry, analyzing the display data of the uploaded information to obtain a change coefficient of the display data, and adjusting the first-level emotion weight matrix according to the second-level emotion weight matrix, the matching coefficient and the change coefficient to obtain an adjusted first-level emotion weight matrix.

Description

Technical Field

[0001] The present application relates to the technical field of social sentiment analysis, and in particular to a social sentiment analysis model characterization method and storage medium based on deep learning. Background Art

[0002] With the popularization of the Internet and mobile devices, everyone can publish content and convey opinions through social platforms. However, some false information and malicious incitement on the Internet will mislead users. Therefore, social platforms and related departments need to promptly discover, guide and manage the discussion status of social platforms. In this process, by establishing a social sentiment analysis model, it is possible to judge the emotional state of users and thus better analyze and judge the discussion status of social platforms.

[0003] The existing social sentiment analysis model establishment process is mainly established through deep learning. It collects content uploaded by users, processes it as samples, and manually annotates the emotional characteristics represented by the samples. The social sentiment analysis model is obtained through sample training. The obtained social sentiment analysis model is used to judge the content uploaded and commented by users, and then the discussion status of the social platform is obtained.

[0004] Since the existing analysis process of topic discussion status is mainly through direct social emotion judgment and accumulation of uploaded content levels, different users and different uploaded content have different influences on the topic discussion status. Therefore, there is a certain deviation in the judgment of the discussion status based on the acquired data. Therefore, how to adjust the influence of the exposure status of different users and different content on the weight and improve the accuracy of the acquired emotional feature data for judging the discussion status is the fundamental problem to be solved by the present invention. Summary of the invention

[0005] In order to improve the influence of adjusting the exposure status of different users and different contents on the weight and obtain the accuracy of emotional feature data for judging the discussion status, the present application provides a social emotion analysis model characterization method and storage medium based on deep learning.

[0006] In the first aspect, the present application provides a social sentiment analysis model characterization method based on deep learning, which adopts the following technical solutions:

[0007] A social sentiment analysis model representation method based on deep learning, including:

[0008] Obtain user uploaded information and corresponding comment information under a single entry on the social platform, process the user uploaded information and comment information, and obtain primary feature unimodal text information and secondary feature unimodal text information;

[0009] Analyze the primary single-modal text information and the secondary feature single-modal text information respectively based on a deep learning model to obtain a primary sentiment weight matrix and a secondary sentiment weight matrix;

[0010] Collect the user historical information of the uploaded content under a single entry and the display data of its uploaded information, analyze the user historical information to obtain its matching coefficient with the entry, analyze the display data of the uploaded information to obtain its change coefficient, and adjust the primary sentiment weight matrix according to the secondary sentiment weight matrix, the matching coefficient and the change coefficient to obtain an adjusted primary sentiment weight matrix;

[0011] Judge the sentiment representation of the entry according to the exposure volume of all user uploaded information under a single entry and the adjusted primary sentiment weight matrix.

[0012] By adopting the above technical solution, analyze the user historical information to obtain its matching coefficient with the entry. The matching coefficient can reflect the overlap degree between the uploaded content and the historical uploaded content of the corresponding user. Adjust the weights of different users through the matching coefficient, thereby improving the accuracy of the judgment result; compare the display data of the uploaded information with the display data of the historical uploaded content of the corresponding user, which can judge the dissemination tendency of the current content, and then adjust the weights of different users through the change coefficient, thereby improving the accuracy of the judgment result of the feature sentiment and discussion status of the social platform.

[0013] Optionally, the user upload data includes uploaded video single-modal information, uploaded picture single-modal information, uploaded audio single-modal information and uploaded text single-modal information;

[0014] Extract video frames and perform ocr recognition to obtain primary video conversion single-modal text information; perform ocr recognition on the uploaded picture single-modal information to obtain primary picture conversion single-modal text information; perform speech recognition on the uploaded audio single-modal information to obtain primary audio conversion single-modal text information; perform data cleaning and deduplication on the primary video conversion single-modal text information, the primary picture conversion single-modal text information, the primary audio conversion single-modal text information and the uploaded text single-modal information to obtain primary single-modal text information;

[0015] The comment information includes comment picture single-modal information and comment text single-modal information; perform ocr recognition on the comment picture single-modal information to obtain secondary picture conversion single-modal text information; perform data cleaning and deduplication on the secondary picture conversion single-modal text information and the comment text single-modal information to obtain secondary single-modal text information.

[0016] Optionally, collect the historical n uploaded information of the user and its display data; obtain the historical sentiment feature matrix through formulas (1)-(2) ;

[0017] (1)

[0018] (2)

[0019] where j = 1, 2, …, n; is the weight of the i-th characteristic emotion in the j-th uploaded message in the first-level emotion weight matrix of the historically uploaded information, G is the number of types of display data, and x = 1, 2, …, G; is the total amount of the x-th type of display data, is the total amount of the x-th type of display data in the j-th uploaded content, is the adjustment coefficient of the x-th type of display data, and ;

[0020] Compare the historical emotion feature matrix with the current first-level emotion weight matrix to determine the matching coefficient.

[0021] By adopting the above technical solution, it is possible to judge the matching degree between the content currently uploaded by the user and the content historically uploaded by the user, and then adjust the first-level emotion weight matrix through the matching coefficient, so as to adaptively adjust the influence weight of different users in the professionalism and influence in a certain field in the overall social emotion analysis and discussion status judgment, thereby improving the accuracy of the discussion status judgment.

[0022] Optionally, the process of obtaining the matching coefficient further includes:

[0023] The matching coefficient is calculated by formula (3) ;

[0024] (3)

[0025] where m is the number of types of characteristic emotions set, and i = 1, 2, …, m; is the weight value of the i-th characteristic emotion in the first-level emotion weight matrix of the currently uploaded content, is the neutral word offset coefficient of the i-th characteristic emotion.

[0026] By adopting the above technical solution, the calculated matching coefficient can adjust the first-level emotion weight matrix, adjust the influence weight of different users in the overall social emotion analysis and discussion status judgment, and thus improve the accuracy of the discussion status judgment.

[0027] Optionally, the process of obtaining the change coefficient includes:

[0028] Obtain the exposure change curve in the display data;

[0029] Obtain the exposure volume change curve of the user's historical upload information, and obtain the peak curve and valley curve of the exposure volume change of the user's historical upload information based on the density clustering algorithm;

[0030] Calculate and obtain the change coefficient through formula (4) ;

[0031] (4)

[0032] Among them, t1 is the current time point, t0 is the upload time point, is the exposure volume change curve in the display data, is the peak curve of the exposure volume change of the user's historical upload information, is the valley curve of the exposure volume change of the user's historical upload information, , are tuning coefficient.

[0033] By adopting the above technical solution, when the exposure volume change curve approaches and exceeds the exposure volume change valley curve, it can be reflected through the change coefficient , and then the first-level sentiment weight matrix is adjusted through the change coefficient . It can adaptively adjust its influence weight in the overall social sentiment analysis and discussion state judgment for different uploaded contents according to the content fermentation degree, thereby improving the accuracy of the discussion state judgment. Optionally, the process of obtaining the adjusted first-level sentiment weight matrix includes:

[0034] Obtain the adjusted first-level sentiment weight matrix through formula (5);

[0035] (5)

[0036] Among them, is the first interval comparison table function, is the second interval comparison table function, and F is the first-level sentiment weight matrix.

[0037] By adopting the above technical solution, the corresponding coefficient can be obtained through the input matching coefficient and the change coefficient . Obtain the adjusted first-level sentiment weight matrix. By using the adjusted first-level sentiment weight matrix for comprehensive analysis in the discussion state analysis, the accuracy of the discussion state judgment can be improved.

[0038] Optionally, the process of judging the sentiment representation of the entry includes:

[0039] Calculate and obtain the discussion coefficient of the i-th characteristic sentiment through formula (6);

[0040] (6)

[0041] Among them, Qs is the total upload volume in the preset time period before the current time point under a single entry, and k = 1, 2, …, Qs; is the adjusted first-level sentiment weight matrix of the k-th uploaded content is the weight of the i-th characteristic sentiment in is the total amount of the x-th display data in the k-th uploaded content, is the weight coefficient of the x-th display data;

[0042] Judge the sentiment representation of the entry according to the discussion coefficients of various characteristic sentiments.

[0043] By adopting the above technical solution, various data under the entry can be comprehensively adjusted according to the adjusted first-level sentiment weight matrix, and then the discussion coefficients of each sentiment characteristic can be obtained. According to the discussion coefficients of different characteristic sentiments, the social sentiment and discussion status can be judged, and then it can assist the relevant personnel of the social platform to manage the content under different entries.

[0044] In the second aspect, the present application provides a storage medium, adopting the following technical solution:

[0045] A storage medium stores a program of the method for representing a social sentiment analysis model based on deep learning and the storage medium described in any one of the above.

[0046] In summary, the present application includes at least one of the following beneficial technical effects:

[0047] 1. By analyzing the user's historical information, the present invention obtains the matching coefficient between the user and the entry. The matching coefficient can reflect the overlap degree between the uploaded content and the historical uploaded content of the corresponding user. By adjusting the weights of different users through the matching coefficient, the accuracy of the judgment result is improved; by comparing the display data of the uploaded information with the display data of the historical uploaded content of the corresponding user, the propagation tendency of the current content can be judged, and then the weights of different users can be adjusted through the change coefficient, thereby improving the accuracy of the judgment result of the social platform's characteristic sentiment and discussion status.

[0048] 2. By judging the matching degree between the user's current uploaded content and the user's historical uploaded content, and then adjusting the first-level sentiment weight matrix through the matching coefficient, the present invention can adaptively adjust the influence weight of different users in the overall social sentiment analysis and discussion status judgment according to their professionalism and influence in a certain field, thereby improving the accuracy of the judgment of the discussion status.

[0049] 3. When the exposure volume change curve is approaching and exceeding the exposure volume change trough curve, the present invention can pass the change coefficient manifested, and then through the variation coefficient adjust the primary emotion weight matrix, which can adaptively adjust its influence weight in the overall social emotion analysis and discussion status judgment according to different uploaded contents in terms of the content fermentation degree, thereby improving the accuracy of the discussion status judgment. Description of the Drawings

[0050] Figure 1 is a flowchart of the steps of a method for representing a social emotion analysis model based on deep learning. Specific Embodiments

[0051] The following describes in detail the embodiments of the present application, and the examples of the embodiments are shown in the drawings.

[0052] In the description of this specification, the descriptions referring to the terms "certain embodiments", "one embodiment", "some embodiments", "schematic embodiments", "examples", "specific examples", or "some examples" mean that the specific features, structures, materials, or characteristics described in connection with the embodiments or examples are included in at least one embodiment or example of the present application. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiments or examples. Moreover, the specific features, structures, materials, or characteristics described can be combined in a suitable manner in any one or more embodiments or examples.

[0053] The embodiments of the present application disclose a method for representing a social emotion analysis model based on deep learning, with reference to Figure 1, including: obtaining user-uploaded information and corresponding comment information under a single entry on a social platform, processing the user-uploaded information and comment information to obtain first-level feature unimodal text information and second-level feature unimodal text information; wherein the process of processing the user-uploaded information and comment information includes an identification process and a data cleaning and deduplication process. The user-uploaded data includes uploaded video unimodal information, uploaded picture unimodal information, uploaded audio unimodal information, and uploaded text unimodal information, and the comment information includes comment picture unimodal information and comment text unimodal information. The identification process mainly targets uploaded video unimodal information, uploaded picture unimodal information, uploaded audio unimodal information, and comment audio unimodal information. For the identification process of video content, it mainly obtains first-level video conversion unimodal text information by extracting video frames and performing ocr recognition; for uploaded picture unimodal information, it obtains first-level picture conversion unimodal text information by ocr recognition; for uploaded audio unimodal information, it performs speech recognition, which can be achieved by connecting the speech-to-text interface in the prior art, to obtain first-level audio conversion unimodal text information. Then, the first-level video conversion unimodal text information, first-level picture conversion unimodal text information, first-level audio conversion unimodal text information, and uploaded text unimodal information are subjected to data cleaning and deduplication to obtain first-level unimodal text information; the first-level unimodal text information is input into a deep learning model to more quickly judge the emotional characteristics of the uploaded content and comment content.

[0054] In addition, based on the deep learning model, the first-level unimodal text information and the second-level feature unimodal text information are analyzed respectively to obtain a first-level emotional weight matrix and a second-level emotional weight matrix; wherein, the first-level emotional weight matrix reflects the emotional characteristic state reflected by the uploaded content, and the second-level emotional weight matrix reflects the emotional characteristic state reflected by the comment content corresponding to the uploaded content. The user historical information of the uploaded content under a single entry and the display data of its uploaded information are collected, and the user historical information is analyzed to obtain its matching coefficient with the entry. The matching coefficient can reflect the overlap degree between the uploaded content and the historical uploaded content of the corresponding user. Obviously, when the overlap degree is high, its influence degree is high. Therefore, the weights of different users are adjusted through the matching coefficient to improve the accuracy of the judgment result;

[0055] Meanwhile, the display data of the uploaded information is analyzed to obtain its change coefficient. By comparing the display data of the uploaded information with the display data of the historical uploaded content of the corresponding user for the uploaded content, the propagation tendency of the current content can be judged. Then, the weights of different users are adjusted through the change coefficient to improve the accuracy of the judgment result of the characteristic emotion and discussion state of the social platform;

[0056] Adjust the first-level sentiment weight matrix according to the second-level sentiment weight matrix, the matching coefficient, and the change coefficient to obtain the adjusted first-level sentiment weight matrix; therefore, in the process of judging the discussion state under a certain entry, adjust during the accumulation process through the adjusted first-level sentiment weight matrix, and judge the sentiment representation of the entry according to the exposure volume of all user-uploaded information under a single entry and the adjusted first-level sentiment weight matrix, thereby improving the accuracy of the discussion state judgment.

[0057] In one embodiment, the process of obtaining the matching coefficient includes: collecting the user's historical n uploaded messages and their display data; obtaining the historical sentiment feature matrix through formulas (1)-(2) ;

[0058] (1)

[0059] (2)

[0060] where j = 1, 2,..., n; is the weight of the i-th characteristic sentiment in the first-level sentiment weight matrix in the j-th uploaded message in the historical uploaded information, G is the number of types of display data, and x = 1, 2,..., G; is the total amount of the x-th type of display data, is the total amount of the x-th type of display data in the j-th uploaded content. The types of display data mainly include exposure rate, number of likes, number of comments, number of bullet screens, etc., and are set according to the content actually set on the social platform and the depth of social sentiment judgment. is the adjustment coefficient of the x-th type of display data, and ; adjustment coefficient is set according to the influence degree of different types of display data on the fermentation and spread of content in the empirical data. At the same time, since the weight is adjusted during distribution, the constraint condition needs to be satisfied. Therefore, the weight of the i-th characteristic sentiment in the historical uploaded content is obtained through the calculation process of formula (1), and the historical sentiment feature matrix is formed through formula (2); then, according to the historical sentiment feature matrix and the current first-level sentiment weight matrix are compared to determine the matching coefficient, which can judge the matching degree between the user's current uploaded content and the user's historical uploaded content. Furthermore, by adjusting the first-level sentiment weight matrix through the matching coefficient, the influence weight of different users in a certain field in the overall social sentiment analysis and discussion state judgment can be adjusted adaptively according to their professionalism and influence, thereby improving the accuracy of the discussion state judgment.

[0061] Among them, the process of obtaining the matching coefficient further includes:

[0062] The matching coefficient is calculated through formula (3). ;

[0063] (3)

[0064] Among them, m is the number of types of characteristic emotions, and i = 1, 2, …, m; is the weight value of the i-th characteristic emotion in the first-level emotion weight matrix of the currently uploaded content, is the neutral word offset coefficient of the i-th characteristic emotion, which is assigned when dividing different characteristic emotions. Among them, the neutral characteristic emotion is assigned 0, the positive characteristic emotion is assigned a positive value, and the specific value is set manually. The negative characteristic emotion is assigned a negative value, and the specific value is set manually. The neutral word offset coefficient is set according to the range of the absolute value of the assignment difference corresponding to different characteristic emotions. The range of the assignment size is selected and set by itself, but the corresponding neutral word offset coefficient is set by fitting according to empirical data. Therefore, the calculated matching coefficient , can adjust the first-level emotion weight matrix, adjust its influence weight in the overall social emotion analysis and discussion status judgment for different users, and then improve the accuracy of the discussion status judgment.

[0065] In one embodiment, the process of obtaining the change coefficient includes: obtaining the exposure amount change curve in the display data; obtaining the exposure amount change curve of the user's historical upload information, and obtaining the exposure amount change peak curve and valley curve of the user's historical upload information based on the density clustering algorithm. The specific process is to display the exposure amount change curve of the user's historical upload content in the same coordinate system, obtain the area with a higher degree of aggregation of the exposure amount change curve through the density clustering algorithm, use its upper limit as the exposure amount change peak curve, and use its lower limit as the exposure amount change valley curve. This process can be implemented by MATLAB software; then the change coefficient is calculated through formula (4). ;

[0066] (4)

[0067] Among them, t1 is the current time point, t0 is the upload time point, S(t) is the exposure amount change curve in the display data, Sp(t) is the exposure amount change peak curve of the user's historical upload information, and Sq(t) is the exposure amount change valley curve of the user's historical upload information. 、 are tuning coefficients, and the tuning coefficients 、 According to the fitting setting of empirical data, by separately calculating the difference between the peak curve of exposure amount change and the peak curve and valley curve of exposure amount change in the user's historical uploaded information, and adjusting it through the tuning coefficient, it is possible to, when the exposure amount change curve approaches and exceeds the valley curve of exposure amount change, be reflected through the change coefficient After that, through the change coefficient adjust the first-level sentiment weight matrix, and it is possible to adaptively adjust its influence weight in the overall social sentiment analysis and discussion status judgment according to the content fermentation degree of different uploaded contents, thereby improving the accuracy of the discussion status judgment.

[0068] In addition, the process of obtaining the adjusted first-level sentiment weight matrix includes: obtaining the adjusted first-level sentiment weight matrix through formula (5) ;

[0069] (5)

[0070] Among them, is the first interval comparison table function, is the second interval comparison table function, F is the first-level sentiment weight matrix, and the first interval comparison table function and the second interval comparison table function are respectively set according to the influence degree of different matching coefficients Mat and different change coefficients on the discussion status. Therefore, through the first interval comparison table function and the second interval comparison table function, it is possible to obtain the corresponding coefficients by inputting the matching coefficient Mat and the change coefficient , and then obtain the adjusted first-level sentiment weight matrix through formula (5) . By using the adjusted first-level sentiment weight matrix in the discussion status analysis for comprehensive analysis, the accuracy of the discussion status judgment can be improved.

[0071] In addition, the process of judging the sentiment representation of the entry includes:

[0072] Calculating the discussion coefficient of the i-th characteristic sentiment through formula (6) ;

[0073] (6)

[0074] Among them, Qs is the total upload amount in the preset time period before the current time point under a single entry, k = 1, 2,..., Qs; is the weight of the i-th characteristic sentiment in the adjusted first-level sentiment weight matrix of the k-th uploaded content, is the total amount of the x-th display data in the k-th uploaded content, is the weight coefficient of the x-th display data, Set according to the influence degree of different types of display data in empirical data on the fermentation and dissemination of content; by comprehensively considering various data under the entry through formula (6) and adjusting according to the adjusted first-level sentiment weight matrix, the discussion coefficient of each sentiment feature can be obtained. Judge the social sentiment and discussion status according to the discussion coefficients of different feature sentiments, so as to assist the relevant personnel of the social platform to manage the content under different entries.

[0075] Among them, the process of judging the social sentiment and discussion status according to the discussion coefficients of different feature sentiments mainly depends on the distribution state of the discussion coefficients corresponding to different feature sentiments. It can be directly shown to the social platform managers for judgment, or by comparing the discussion coefficients of each feature sentiment with the corresponding set threshold to judge whether there is a risk. The corresponding set threshold is set according to empirical data. When the discussion coefficients of some negative feature sentiments exceed the corresponding set threshold, a warning can be issued.

[0076] The embodiment of the present application also discloses a storage medium storing a program of the method for representing a social sentiment analysis model based on deep learning described in any one of the above. The program stored in the storage medium analyzes the user's historical information to obtain the matching coefficient between the user and the entry. The matching coefficient can reflect the coincidence degree between the uploaded content and the user's historical uploaded content. Adjust the weights of different users through the matching coefficient to improve the accuracy of the judgment result; compare the display data of the uploaded information with the display data of the uploaded content and the user's historical uploaded content to judge the dissemination tendency of the current content, and then adjust the weights of different users through the change coefficient to improve the accuracy of the judgment result of the social platform's feature sentiment and discussion status.

[0077] Although the embodiments of the present application have been shown and described above, it can be understood that the above embodiments are exemplary and should not be construed as limiting the present application. Those of ordinary skill in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of the present application.

[0078] Although the embodiments of the present application have been shown and described above, it can be understood that the above embodiments are exemplary and should not be construed as limiting the present application. Those of ordinary skill in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of the present application.

Claims

1. A social sentiment analysis model representation method based on deep learning, characterized in that: include: Obtain user uploaded information and corresponding comment information under a single entry on the social platform, process the user uploaded information and comment information, and obtain primary feature unimodal text information and secondary feature unimodal text information; Based on the deep learning model, the first-level unimodal text information and the second-level feature unimodal text information are analyzed respectively to obtain the first-level sentiment weight matrix and the second-level sentiment weight matrix; Collect the user history information of the uploaded content under a single entry and the display data of the uploaded information, analyze the user history information, obtain the matching coefficient between it and the entry, analyze the display data of the uploaded information, obtain its variation coefficient, adjust the primary emotion weight matrix according to the secondary emotion weight matrix, matching coefficient and variation coefficient, and obtain the adjusted primary emotion weight matrix; The emotional representation of a single term is judged based on the exposure of all user-uploaded information under the single term and the adjusted first-level emotional weight matrix.

2. The social sentiment analysis model characterization method based on deep learning according to claim 1 is characterized in that: The user uploaded data includes uploaded video unimodal information, uploaded picture unimodal information, uploaded audio unimodal information and uploaded text unimodal information; Extract video frames and perform OCR recognition to obtain first-level video conversion single-modal text information; perform OCR recognition on uploaded image single-modal information to obtain first-level image conversion single-modal text information; perform voice recognition on uploaded audio single-modal information to obtain first-level audio conversion single-modal text information; perform data cleaning and deduplication on first-level video conversion single-modal text information, first-level image conversion single-modal text information, first-level audio conversion single-modal text information and uploaded text single-modal information to obtain first-level single-modal text information; The comment information includes comment picture unimodal information and comment text unimodal information; the comment picture unimodal information is subjected to OCR recognition to obtain secondary picture conversion unimodal text information; the secondary picture conversion unimodal text information and the comment text unimodal information are subjected to data cleaning and deduplication to obtain secondary unimodal text information.

3. The social sentiment analysis model characterization method based on deep learning according to claim 1 is characterized in that: Collect n pieces of historical user uploaded information and their display data; obtain the historical sentiment feature matrix through formulas (1)-(2) ; (1) (2) Where j = 1, 2, …, n; is the weight of the i-th characteristic emotion in the j-th uploaded information in the first-level emotion weight matrix of the historical uploaded information, G is the number of types of displayed data, x=1, 2, …, G; is the total amount of the x-th display data, is the total amount of display data of the xth type of the jth uploaded content, is the adjustment factor for the x-th display data, and ; According to the historical sentiment feature matrix Compare with the current first-level sentiment weight matrix to determine the matching coefficient.

4. The social sentiment analysis model characterization method based on deep learning according to claim 3 is characterized in that: The process of obtaining the matching coefficient also includes: The matching coefficient is calculated by formula (3): ; (3) Where m is the number of types of characteristic emotion settings, i=1, 2, …, m; is the weight value of the i-th characteristic emotion in the first-level emotion weight matrix of the current uploaded content, is the neutral word offset coefficient of the i-th characteristic sentiment.

5. The social sentiment analysis model characterization method based on deep learning according to claim 4 is characterized in that: The process of obtaining the coefficient of variation includes: Obtain the exposure change curve in the display data; Obtain the exposure change curve of the user's historical uploaded information, and obtain the peak curve and valley curve of the exposure change of the user's historical uploaded information based on the density clustering algorithm; The coefficient of variation is calculated by formula (4): ; (4) Among them, t1 is the current time point, t0 is the upload time point, To show the exposure change curve in the data, It is the peak curve of exposure change of the user's historical uploaded information. It is the valley curve of the exposure change of the user's historical uploaded information. , is the tuning coefficient.

6. The social sentiment analysis model characterization method based on deep learning according to claim 5 is characterized in that: The process of obtaining the adjusted first-level sentiment weight matrix includes: The adjusted first-level sentiment weight matrix is ​​obtained through formula (5); (5) in, is the first interval lookup table function, is the second interval comparison table function, and F is the first-level sentiment weight matrix.

7. The social sentiment analysis model characterization method based on deep learning according to claim 6 is characterized in that: The process of judging the sentiment representation of a term includes: The discussion coefficient of the i-th characteristic emotion is calculated by formula (6); (6) Where Qs is the total upload amount of a single entry in a preset period before the current time point, k=1, 2, ..., Qs; is the adjusted first-level sentiment weight matrix of the kth uploaded content The weight of the i-th feature sentiment in , is the total amount of the x-th type of display data in the k-th uploaded content, is the weight coefficient of the x-th display data; The sentiment representation of the entry is judged based on the discussion coefficients of various characteristic sentiments.

8. A social sentiment analysis model storage medium based on deep learning, characterized in that: A program storing a social sentiment analysis model characterization method based on deep learning as described in any one of claims 1 to 7.

Images