A method for predicting and analyzing the trend of the public opinion field based on quantitative calculation

Abstract

The present invention discloses a method for predicting and analyzing the trend of the public opinion field based on quantitative calculation, belonging to the technical field of public opinion analysis, including the steps of: S1, obtaining basic data; S2, based on the obtained basic data, performing stance detection and stance trend analysis based on the factor of the speaking time, and constructing a sample set for generating an opinion trend judgment rule; S3, constructing a neural network model TFNN for predicting the trend of the public opinion field based on the speaking account and the content of the speech; S4, using the sample data obtained by generating the sample set of the constructed trend judgment rule in step S2 to train the TFNN model in step S3, and obtaining the prediction result of the public opinion trend. The present invention can provide scientific and effective decision-making suggestions for the scientific governance of the network society in the new era and the judgment of the trend of public opinion on media platforms.

Description

Technical Field

[0001] The present invention relates to the technical field of public opinion analysis, and more specifically, to a method for predicting and analyzing the trend of the public opinion field based on quantitative calculation. Background Art

[0002] The prior art has the following technical problems: 1) The judgment of the trend of public opinion is not comprehensive enough; 2) How to quantitatively analyze the trend of public opinion for a specific theme. Summary of the Invention

[0003] The purpose of the present invention is to overcome the deficiencies of the prior art and provide a method for predicting and analyzing the trend of the public opinion field based on quantitative calculation, which can provide scientific and effective decision-making suggestions for the scientific governance of the network society in the new era and the judgment of the trend of public opinion on media platforms, etc.

[0004] The purpose of the present invention is achieved through the following solutions:

[0005] A method for predicting and analyzing the trend of the public opinion field based on quantitative calculation, comprising the steps of:

[0006] S1, obtaining basic data;

[0007] S2, based on the obtained basic data, performing stance detection and stance trend analysis considering timing factors, and constructing a sample set for generating judgment rules for the trend of public opinion;

[0008] S3, constructing a neural network model TFNN for predicting the trend of public opinion, the neural network model TFNN for predicting the trend of public opinion including an input layer, a representation layer and an output layer, with a first fully connected layer and a first activation layer provided in the representation layer, and a second fully connected layer and a second activation layer provided in the output layer;

[0009] S4, using the sample data obtained by constructing the sample set for generating judgment rules for the trend of public opinion in step S2 to train the neural network model TFNN for predicting the trend of public opinion in step S3 to achieve prediction of the trend of public opinion.

[0010] Further, in step S1, the main post and comment content data are collected from the social network platform according to the specified theme as the basic data.

[0011] Further, in step S2, the performing stance detection and stance trend analysis considering timing factors includes sub-steps:

[0012] S21, detecting the stance of the comment text: Analyzing the stance tendency of the comment text content towards the target theme as "support", "opposition" or "neutral" based on the comment content for the main post

[0013] content and theme, and realizing the stance detection of the comment content;

[0014] S22, Position Trend Analysis: After obtaining the position of the review text content in step S21, perform a position trend analysis on all reviews under the corresponding topic to generate three position sets: "support", "oppose", and "neutral".

[0015] S23, Use the LSTM model to predict the change trend of the position in step S22, and select the best inflection point in the position change as the public opinion influence timing, which is one of the factors for judging the trend direction.

[0016] Furthermore, in step S2, the construction of the judgment rule for the public opinion trend direction to generate a sample set includes sub-steps:

[0017] SS21, Trend Smoothing: Establish a trend graph curve of the proportion of the number of people in each position with respect to the total number of posts and perform smoothing processing. Then use the least squares method to regress a small window of data onto a polynomial, and then use the polynomial to estimate the point at the center of the window. Finally, the window moves forward by one data point, use the least squares method to regress the data window onto a polynomial, and use the polynomial to estimate the center point of the window until each point is optimally adjusted relative to its neighbors.

[0018] SS22, Inflection Point Detection of the Trend Graph Curve: Adopt a sliding window algorithm and expand it to find multiple change points. Slide two adjacent windows along the signal and calculate the difference between the first window and the second window. When these two windows contain dissimilar segments, if the calculated difference value is greater than the set value, an inflection point is detected. And in the offline setting, calculate the complete difference curve and perform a peak search process to find the inflection point index. Finally, perform inflection point detection and annotation on the smoothed trend curve.

[0019] SS23, Trend Segment Selection: Based on the inflection points marked on the trend curve in step SS22, calculate the segment with the largest upward amplitude in the trend curve as a sample for extraction. Label the reviews with the same position as the trend curve as positive samples, and label the reviews with other positions as negative samples. Follow this rule to label each review, and the label is considered as a factor for the trend influence score.

[0020] SS24, Formulation of the Judgment Rule for the Trend Direction of the Sample: The trend direction judgment mechanism of the sample is composed of the sum of the basic information score of the review social account and the score of the influence of the review social account on the position trend.

[0021] Score_User = w_property * Score(property) + w_behavior * Score(behavior) (1)

[0022] In formula (1), Score_User represents the total score of the person portrait, Score(property) represents the score of the person's account attributes, w_property represents the weight ratio of the basic attributes, Score(behavior) represents the score of the behavioral attributes, and w_behavior represents the weight ratio of the behavioral attributes;

[0023] Score_Final = Score_User * W_User + Score_label* W_label (2)

[0024] In formula (2), Score_label is the score of the influence of the comment social account on the position trend, and the scoring rules are as follows: if a comment successfully influences the comment of another person with a different position, the comment gets the first set numerical score; if a comment successfully increases the number of people with the same position, the comment gets the second set numerical score; if a comment fails to increase the number of people with the same position, the comment gets the third set numerical score; if a comment reduces the number of people with the same position, the comment gets the fourth set numerical score; W_label represents the weight ratio of the trend influence score; W_User represents the weight ratio of the person's account attribute; Score_Final represents the final score of the sample trend direction;

[0025] In SS25, construct sample features, and the sample features include person account attribute features, comment content semantic features, and topic position semantic features; then calculate Score_Final through the sample features, and use the result of Score_Final as the conclusion for predicting the effective influence on the public opinion trend.

[0026] Further, in step S3, it includes sub-steps: using the person account attribute features, comment content semantic features, and topic position semantic features as the input features of the neural network model TFNN model.

[0027] Further, in step S4, the factors affecting the prediction of the public opinion trend direction include the speaking time, the speaking content, and the speaking account.

[0028] Further, in step SS21, the smoothing process uses the Savitzky-Golay filter for smoothing.

[0029] Further, in step SS25, the comment semantic content features are obtained according to the output of the sentence_bert model.

[0030] Furthermore, in step SS25, by inputting the comment content, topic content, and the position tendency of the selected comment text content on the target topic into the sentence_bert model, a semantic vector is output, which can characterize the semantics, topic, and position of the comment content, and obtain the topic position semantic features.

[0031] The beneficial effects of the present invention are:

[0032] The present invention formulates sample public opinion trend judgment rules by comprehensively considering multiple factors such as comment content, topic, and comment account attributes. At the same time, it proposes a new public opinion trend prediction neural network model, which realizes the prediction of public opinion trend and can provide scientific and effective decision-making suggestions for the scientific governance of the network society in the new era and the judgment of public opinion trend on media platforms. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0034] Figure 1 A schematic diagram of a public opinion strategy generation process in an embodiment of the present invention;

[0035] Figure 2 A graph showing the trend of the proportion of people in each position versus the total number of posts in an embodiment of the present invention;

[0036] Figure 3 A schematic diagram of a prediction of a trend of position changes on a specific topic in an embodiment of the present invention;

[0037] Figure 4 A schematic diagram of trend graph curve smoothing in an embodiment of the present invention;

[0038] Figure 5 Schematic diagram of smooth curve inflection point detection in an embodiment of the present invention;

[0039] Figure 6 A schematic diagram of trend segment selection in an embodiment of the present invention;

[0040] Figure 7 Schematic diagram of the TFNN model structure proposed in an embodiment of the present invention. DETAILED DESCRIPTION

[0041] All features disclosed in all embodiments in this specification, or steps in all methods or processes implicitly disclosed, except for mutually exclusive features and / or steps, can be combined and / or expanded or replaced in any manner.

[0042] The following will further elaborate on the technical concept, working principle, efficacy, and working process of the present invention based on the attached Figures 1 to 7 drawings, etc.

[0043] In the specific implementation process, the technical solution of the present invention is described in detail as follows:

[0044] To solve the problem of how to effectively judge the trend of public opinion on a specific topic, the technical concept of the present invention is: by analyzing the change in the position trend of the specific topic, predicting the trend inflection point, and providing the judgment of the influencing trend timing; comprehensively using features such as the portrait of social account characters, individual speech content, and topic content, quantitatively analyzing the influence degree of the above factors on the trend of public opinion, constructing a prediction model for the trend of public opinion, and ultimately being able to provide relevant strategies such as the speaking account and the posted content to achieve the effective prediction of the trend of public opinion. The overall process of the technical solution of the present invention is as Figure 1 shown. In the specific implementation process, it includes the following steps:

[0045] (1) Data acquisition

[0046] According to the specified topic, collect the main posts and comment content from the social network platform, including the basic information, behavior information, content information, etc. of the social accounts that post comments, as the basic data for strategy analysis.

[0047] (2) Position detection and position trend analysis

[0048] Step 1, comment text position detection: Analyze the position of the main post content and the topic according to the comment content to achieve the position detection of the comment content. The position detection task uses natural language processing technology to analyze whether the current comment text content has a "support", "oppose", or "neutral" position tendency towards the target topic.

[0049] Step 2, position trend analysis: After obtaining the position of the comment text content, conduct the position trend analysis of all comments under this topic, generate three position sets of "support", "oppose", and "neutral", and the trend of the proportion of the number of people in each position changing with the total number of posts is as Figure 2 shown.

[0050] Step 3, the present invention uses the LSTM model to predict the change trend of the topic position, and selects the best inflection point in the position change as one of the factors for judging the trend direction as the influencing timing.

[0051] The characteristic of LSTM is that valve nodes are added to each layer outside the RNN structure. There are three types of valves: forget gate, input gate, and output gate. The memory function of the model is realized by these valve nodes. By adjusting the opening and closing of the valves, the influence of early sequences on the final result can be achieved. LSTM performs very well in time series prediction. According to the predicted change trend, select the inflection point position of the trend reversal as the timing.

[0052] (3)Construct trend direction judgment rules to generate a sample set for trend direction judgment rules

[0053] Step 1, trend smoothing: Smooth the trend graph curve of the proportion of the number of people in each position with respect to the total number of posts using the Savitzky-Golay filter. Use the least squares method to regress a small window of data onto a polynomial, and then use the polynomial to estimate the point at the center of the window. Finally, move the window forward by one data point and repeat the process. Continue like this until each point is optimally adjusted relative to its neighbors, as Figure 4 shown.

[0054] Step 2, inflection point detection of the trend graph curve: Adopt the sliding window algorithm (a fast approximate alternative algorithm to Optimal Detection). This method relies on a single change point detection program and extends it to find multiple change points. When the algorithm is implemented, two adjacent windows slide along the signal, and the difference between the first window and the second window is calculated. When these two windows contain dissimilar segments, the calculated difference value will be large, and an inflection point is detected. In the offline setting, calculate the complete difference curve and perform a peak search process to find the inflection point index. Label the inflection points detected on the smoothed trend curve, as Figure 5 shown.

[0055] Step 3, trend segment selection: Based on the inflection points marked on the Figure 5 trend curve, calculate the segment with the largest upward amplitude in the curve as a sample to be extracted. Comments with the same position as this trend curve are labeled as positive samples, and comments with other positions are labeled as negative samples. Follow this rule to label each comment. This label is used as a consideration for the trend influence score, as Figure 6 shown.

[0056] Step 4, formulation of sample trend direction judgment rules: The trend direction judgment mechanism of the sample consists of the sum of the basic information score of a certain comment social account and the trend influence score of this comment social account on the position trend.

[0057] Score_User = w_property * Score(property) + w_behavior * Score(behavior) (1)

[0058] In formula (1), Score_User represents the total score of the user portrait, Score(property) represents the score of the user account properties, including the number of fans, the number of posts, etc., w_property represents the weight ratio of the basic properties, Score(behavior) represents the score of the behavior properties, and w_behavior represents the weight ratio of the behavior properties.

[0059] Score_Final = Score_User * W_User + Score_label * W_label (2)

[0060] In formula (2), Score_label is the score of the influence of the current comment social account on the position trend. The scoring rules are as follows: if the current comment successfully influences the comment of a person with a different position, the comment gets 4 points; if the current comment successfully increases the number of people with the same position, the comment gets 3 points; if the current comment fails to increase the number of people with the same position, the comment gets 2 points; if the current comment reduces the number of people with the same position, the comment gets 1 point. W_label represents the weight ratio of the trend influence score. W_User represents the weight ratio of the user account properties. The sum of the scores of the two parts is used as the final score Score_Final representing the trend of the sample.

[0061] Step 5: Construct sample features: It is composed of user account property features and comment content features. The comment content features are output according to the sentence_bert model. The comment content, the topic content, and the position tendency of the current comment text towards the target topic are input into the sentence_bert model to output semantic vectors, which represent the semantics, topic, and position of the comment content.

[0062] The final sample features are composed of user account property features, comment content semantic features, and topic position semantic features. The Score_Final is calculated through the sample features. Using this analysis result, it can be used as a factor to judge the effective influence on the public opinion trend and as a training sample to construct a public opinion trend prediction model.

[0063] (4) Model construction

[0064] The model requires three parts of input features: the personal account attribute features, the semantic features of the comment content, and the semantic features of the topic stance. The dimensionalities of these three parts of features are different. Currently, the DSSM two-tower model in the recommendation system can handle multi-domain information. By referring to the model architecture and constructing the neural network model structure according to the content dimensionalities of the three inputs, a neural network model for predicting the trend of public opinion, the Trendforcast Neural Network (TFNN), is proposed. After passing through two fully connected layers and an activation layer, the three parts are concatenated, and then passed through another fully connected layer and an activation layer. Its principle structure is as Figure 7 shown.

[0065] (5) Prediction of the trend of public opinion

[0066] So far, using the sample data obtained by constructing the trend judgment rule generation sample set in (3), the TFNN model can be trained. By inputting the personal account attribute features, the semantic features of the comment content, and the semantic features of the topic stance, the influencing factors of public opinion can be obtained, including the timing of speech, the content of speech, and the speech account.

[0067] Due to the numerous factors affecting network information dissemination such as the heterogeneity of the social network itself, the complexity of cross-platform dissemination, and the variability of public opinion fermentation, the technical solution of the present invention can provide scientific and effective decision-making suggestions for the scientific governance of the network society in the new era and the judgment of the public opinion trend on media platforms by clarifying the laws of network information dissemination, determining the trend of network public opinion fermentation, identifying key nodes of public opinion, and intelligently recommending public opinion strategies.

[0068] Embodiment 1: A method for predicting and analyzing the trend of the public opinion field based on quantitative calculation, comprising the steps of:

[0069] S1, obtaining basic data;

[0070] S2, based on the obtained basic data, performing stance detection and stance trend analysis considering timing factors, and constructing a sample set for generating the judgment rule of the trend of public opinion;

[0071] S3, constructing a neural network model for predicting the trend of public opinion, the TFNN. The neural network model for predicting the trend of public opinion, the TFNN, includes an input layer, a representation layer, and an output layer. There is a first fully connected layer and a first activation layer in the representation layer, and a second fully connected layer and a second activation layer in the output layer;

[0072] S4, using the sample data obtained by constructing the sample set for generating the judgment rule of the trend in step S2 to train the TFNN model in step S3 to generate a public opinion strategy.

[0073] Example 2: On the basis of Example 1, in step S1, the main posts and comment content data are collected from the social network platform according to the specified topic, and used as the basic data for judging and analyzing the public opinion trend.

[0074] Example 3: On the basis of Example 1, in step S2, the stance detection and stance trend analysis considering timing factors include sub-steps:

[0075] S21, Comment text stance detection: Analyze the stance of the current comment text content towards the target topic as "support", "opposition" or "neutral" based on the comment content for the main post content and topic, so as to achieve comment content stance detection;

[0076] S22, Stance trend analysis: After obtaining the stance of the comment text content, conduct stance trend analysis on all comments under this topic, and generate three stance sets of "support", "opposition" and "neutral";

[0077] S23, Use the LSTM model to predict the change trend of the topic stance, and select the best inflection point in the stance change as the influencing timing as one of the factors for judging the trend direction.

[0078] Example 4: On the basis of Example 1, in step S2, the construction of the sample set for generating the judgment rules for the public opinion trend includes sub-steps:

[0079] SS21, Trend smoothing: Smooth the trend graph curve of the proportion of the number of people in each stance with respect to the total number of posts. Use the least squares method to regress a small window of data onto a polynomial, then use the polynomial to estimate the point at the center of the window, and finally move the window forward by one data point and repeat this process; continue like this until each point has been optimally adjusted relative to its neighbors;

[0080] SS22, Inflection point detection of the trend graph curve: Adopt the sliding window algorithm and expand it to find multiple change points. Slide two adjacent windows along the signal and calculate the difference between the first window and the second window; when the two windows contain dissimilar segments, the calculated difference value is greater than the set value, and an inflection point is detected; and in the offline setting, calculate the complete difference curve and perform a peak search process to find the inflection point index, and perform inflection point detection and annotation on the smoothed trend curve;

[0081] SS23, Trend segment selection: Based on the inflection points marked on the trend curve in step SS22, calculate the segment with the largest upward amplitude in the curve as a sample and extract it. Comments with the same stance as this trend curve are marked as positive samples, and comments with other stances are marked as negative samples. Tag each comment according to this rule; this tag is used as a consideration for the trend influence score;

[0082] SS24. Formulation of sample trend judgment rules: The sample trend judgment mechanism is composed of the sum of the basic information score of a certain comment social account and the influence score of this comment social account on the position trend;

[0083] Score_User = w_property * Score(property) + w_behavior * Score(behavior) (1)

[0084] In formula (1), Score_User represents the total score of the person profile, Score(property) represents the score of the person's account attributes, w_property represents the weight ratio of the basic attributes, Score(behavior) represents the score of the behavioral attributes, and w_behavior represents the weight ratio of the behavioral attributes;

[0085] Score_Final = Score_User * W_User + Score_label* W_label (2)

[0086] In formula (2), Score_label is the influence score of this comment social account on the position trend; the scoring rules are as follows: If this comment successfully influences the comment of a person with another position, this comment gets the first set value score; if this comment successfully increases the number of people in this position, this speech gets the second set value score; if this comment fails to increase the number of people in this position, this comment gets the third set value score; if this comment reduces the number of people in this position, this comment gets the fourth set value score; W_label represents the weight ratio of the trend influence score; W_User represents the weight ratio of the person's account attribute. Add the scores of the two parts as the final score Score_Final representing the sample trend;

[0087] SS25. Construct sample features, and the sample features include person account attribute features, comment content semantic features, and topic position semantic features; Calculate the Score_Final through the sample features, and use the analysis result as a factor to judge the effective influence on the public opinion trend, and use it as a training sample to construct a public opinion trend prediction model.

[0088] Example 5: On the basis of Example 1, in step S3, it includes sub-steps: using the person account attribute features, comment content semantic features, and topic position semantic features as the input features of the neural network model TFNN model.

[0089] Example 6: On the basis of Example 3, in step S4, the factors affecting the prediction of the public opinion trend include the speaking time, the speaking content, and the speaking account.

[0090] Example 7: On the basis of Example 1, in step SS21, the smoothing process uses Savitzky-Golay filter smoothing.

[0091] Example 8: On the basis of Example 3, in step SS25, the semantic features of the comment content are output according to the sentence_bert model.

[0092] Example 9: On the basis of Example 1, in step S25, by inputting the comment content, the topic content, and the stance tendency of the current comment text towards the target topic into the sentence_bert model, a semantic vector is output, which characterizes the semantics, topic, and stance of the comment content, and the semantic features of the topic stance are obtained.

[0093] If the functions of the present invention are implemented in the form of software function units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art or a part of this technical solution can be embodied in the form of a software product. This computer software product is stored in a storage medium and executes all or part of the steps of the methods described in various embodiments of the present invention on a computer device (which can be a personal computer, a server, or a network device, etc.) and the corresponding software. The aforementioned storage medium includes: various media that can store program codes such as USB flash drives, mobile hard disks, or optical discs. The data for testing exists in a read-only memory (Random Access Memory, RAM), a random access memory (Random Access Memory, RAM), etc. during the program implementation.

Claims

1. A method for predicting and analyzing the trend of the public opinion field based on quantitative calculation, characterized in that Including the steps: S1. Obtain the basic data; S2. Based on the obtained basic data, conduct stance detection and stance trend analysis considering timing factors, and construct a sample set for generating judgment rules for the trend of public opinion; S3. Construct a neural network model TFNN for predicting the trend of public opinion. The neural network model TFNN for predicting the trend of public opinion includes an input layer, a representation layer, and an output layer. There is a first fully connected layer and a first activation layer in the representation layer, and a second fully connected layer and a second activation layer in the output layer; in step S2, the construction of the sample set for generating judgment rules for the trend of public opinion includes sub-steps: SS21. Trend smoothing: Establish a trend graph curve of the proportion of the number of people in each stance changing with the total number of posts and perform smoothing processing. Then use the least squares method to regress a small window of data onto a polynomial, and then use the polynomial to estimate the point at the center of the window. Finally, the window moves forward by one data point, use the least squares method to regress the data window onto a polynomial, and use the polynomial to estimate the center point of the window until each point is optimally adjusted relative to its neighbors; SS22. Inflection point detection of the trend graph curve: Adopt a sliding window algorithm and expand it to find multiple change points. Slide two adjacent windows along the signal and calculate the difference between the first window and the second window; when these two windows contain dissimilar segments, if the calculated difference value is greater than the set value, then an inflection point is detected; and in the offline setting, calculate the complete difference curve and perform a peak search process to find the inflection point index, and finally perform inflection point detection and annotation on the smoothed trend curve; SS23. Selection of trend segments: Based on the inflection points marked on the trend curve in step SS22, calculate the segment with the largest upward amplitude in the trend curve as a sample and extract it. Label the comments with the same stance as the trend curve as positive samples, and label the comments with other stances as negative samples. Follow this rule to label each comment, and the label is used as a consideration for the trend influence score; SS24. Formulation of the judgment rule for the trend of the sample: The judgment mechanism for the trend of the sample is composed of the sum of the basic information score of the comment social account and the influence score of the current comment social account on the stance trend; Score_User = w_property * Score(property) + w_behavior * Score(behavior) (1) In formula (1), Score_User represents the total score of the user portrait, Score(property) represents the score of the user account attribute, w_property represents the weight ratio of the basic attribute, Score(behavior) represents the score of the behavior attribute, and w_behavior represents the weight ratio of the behavior attribute; Score_Final = Score_User * W_User + Score_label* W_label (2) In formula (2), Score_label is the influence score of the comment social account on the stance trend, and the scoring rules are as follows: if a comment successfully influences the comment of another person with a different stance, the comment gets the first set numerical score; if a comment successfully increases the number of people with the same stance, the comment gets the second set numerical score; if a comment fails to increase the number of people with the same stance, the comment gets the third set numerical score; if a comment reduces the number of people with the same stance, the comment gets the fourth set numerical score; W_label represents the weight ratio of the trend influence score; W_User represents the weight ratio of the person account attribute; Score_Final represents the final score of the sample trend direction. SS25. Construct sample features, where the sample features include person account attribute features, comment content semantic features, and topic stance semantic features; then calculate Score_Final through the sample features, and use the result of Score_Final as the conclusion for predicting the effective influence on public opinion trends. S4. Use the sample data obtained by generating a sample set using the constructed public opinion trend direction judgment rule in step S2 to train the public opinion trend direction prediction neural network model TFNN in step S3 to achieve public opinion trend direction prediction.

2. The method for predicting and analyzing the trend of the public opinion field based on quantization calculation according to claim 1, wherein In step S1, collect the main post and comment content data from the social network platform according to the specified topic as the basic data.

3. The method for predicting and analyzing the trend of the public opinion field based on quantization calculation according to claim 2, characterized in that In step S2, the stance detection and stance trend analysis considering timing factors include sub-steps: S21. Comment text stance detection: Analyze the stance of the comment text on the target topic as "support", "opposition", or "neutral" based on the comment content and the stance of the main post content and topic, to achieve comment content stance detection. S22. Stance trend analysis: After obtaining the stance of the comment text content in step S21, conduct stance trend analysis on all comments under the corresponding topic to generate three stance sets of "support", "opposition", and "neutral". S23. Use the LSTM model to predict the change trend of the stance in step S22, and select the best inflection point in the stance change as the public opinion influence timing as one of the factors for judging the trend direction.

4. The method for predicting and analyzing the trend of the public opinion field based on quantization calculation according to claim 1, wherein In step S3, it includes sub-steps: Use the person account attribute features, comment content semantic features, and topic stance semantic features as the input features of the neural network model TFNN.

5. The method for predicting and analyzing the trend of the public opinion field based on quantization calculation according to claim 1, wherein In step S4, the factors affecting the prediction of public opinion trend direction include the speaking timing, speaking content, and speaking account.

6. The method for predicting and analyzing the trend of the public opinion field based on quantization calculation according to claim 1, wherein In step SS21, the smoothing process uses the Savitzky-Golay filter for smoothing.

7. The method for predicting and analyzing the trend of the public opinion field based on quantization calculation according to claim 1, wherein In step SS25, the comment content semantic features are obtained according to the output of the sentence_bert model.

8. The method for predicting and analyzing the trend of the public opinion field based on quantization calculation according to claim 1, wherein, In step SS25, by inputting the comment content, topic content, and the stance tendency of the selected comment text on the target topic into the sentence_bert model, and outputting the semantic vector, which can represent the semantics, topic, and stance of the comment content, to obtain the topic stance semantic features.

Images