Social media comprehensive public opinion analysis method and related device

Abstract

The invention belongs to the field of data processing, and discloses a social media comprehensive public opinion analysis method and related device.The method comprises the steps that feature vectors of all cross-modal public opinion data are obtained through a pre-trained cross-modal large model, and public opinion events are clustered through an unsupervised clustering method according to the feature vectors; utilizing a pre-training sentiment analysis model and a pre-training topic analysis model to respectively obtain sentiment scores and topic sensitivity scores of the public opinion events in the public opinion event clusters; according to the emotion score and the topic sensitivity score of each public opinion event in each public opinion event cluster, respectively obtaining an emotion score index score and a topic sensitivity index score of each public opinion event cluster, and obtaining a comprehensive public opinion analysis score of each public opinion event cluster in combination with the exposure index score, the interactivity index score and the propagation intensity index score; and obtaining a public opinion evaluation result of each public opinion event cluster according to the comprehensive public opinion analysis score of each public opinion event cluster, thereby realizing comprehensive perception and accurate study and judgment of the public opinion situation.

Description

Technical Field

[0001] The present invention belongs to the field of data processing and relates to a method for comprehensive public opinion analysis of social media and related devices. Background Art

[0002] With the rapid development of social media in the direction of diversification and popularization, the breadth and speed of information dissemination have been significantly improved, and various public opinion events can spread rapidly to the global scope within a very short time. Such changes not only meet the diverse information needs of the public but also make social media gradually become an important platform for influencing public perception and guiding public opinion. However, the openness, real-time nature, and high interactivity of social media also bring new challenges: on the one hand, the complexity of information sources makes it easier for rumors, false information, and misleading remarks to spread on the network, triggering public misunderstandings and emotional fluctuations; on the other hand, the rapid spread of negative public opinion may have a profound impact on social stability and governance. In addition, the social media platform also includes the mixed dissemination of cross-modal information such as text and images, which further increases the difficulty of public opinion event analysis. In this context, how to timely, accurately identify, analyze, and reasonably respond to public opinion events in the environment of cross-modal information has become a key issue in the field of public opinion management in the industry.

[0003] The current public opinion analysis solutions for social media mainly include traditional public opinion analysis methods and public opinion analysis methods based on large language models. These two types of methods mainly focus on the field of public opinion analysis of text data. Traditional public opinion analysis methods use relevant technologies in machine learning or basic deep learning. First, the collected text data is initially segmented, stemmed, and stop-word filtered, etc. Then, solutions such as support vector machines, naive Bayes, or random forests in machine learning, as well as models such as RNN and LSTM, are used to capture the sentiment tendencies in long and short dependencies and complex contexts through training and optimization. Subsequently, the current public opinion situation is analyzed through methods such as sentiment aggregation, fine-grained recognition, and sequence analysis. This type of method provides a good solution for text public opinion analysis in social media and can give roughly reasonable public opinion analysis results. However, the single semantic sentiment itself is highly complex, and it is difficult to give an accurate sentiment tendency judgment through preliminary word segmentation processing, etc. At the same time, traditional machine learning and deep learning models are difficult to handle the rapid and accurate processing of a large number of samples, and there are still significant deficiencies and limitations in the real-time nature, dynamics, and accuracy of public opinion analysis, and new technical means and methods are urgently needed to overcome these difficulties and defects.

[0004] Another type of public opinion analysis method based on large language models gives full play to the text processing capabilities of large language models. After initially clustering public opinion data, it is input into the large language model to obtain a public opinion summary. Then, the summary and information such as preset sentiment categories are input into the large language model again to determine sensitive label categories and attention levels, and the public opinion situation is determined based on the numerical values of attention indicators. This type of method uses large language models as an auxiliary means to provide content such as public opinion summaries and sensitive information through large models, which to a certain extent improves the automation and intelligence levels of public opinion analysis. However, during the application of this type of method, the large language model directly jumps from text to the public opinion analysis summary, lacking the full utilization and analysis of the text features obtained by the large model in the middle. Moreover, the final judgment of sensitive information mainly relies on the quantity of public opinion events and attention indicators, lacking multi-dimensional comprehensive indicators for more reasonable and accurate public opinion identification and judgment. In particular, both of these types of methods ignore a key issue, that is, current social media not only contains text data but also a large amount of public opinion data in modalities such as images and videos. This part of cross-modal data is still an important sample source for social media public opinion analysis, while traditional models or large language models based on a single text modality cannot handle and analyze cross-modal public opinion data, so there are still some biases in terms of accuracy and credibility, and it is difficult to achieve a large improvement in the effect of public opinion analysis. Summary of the Invention

[0005] The object of the present invention is to overcome the above-mentioned shortcomings of the prior art and provide a social media comprehensive public opinion analysis method and related device.

[0006] To achieve the above object, the present invention adopts the following technical solutions:

[0007] In the first aspect of the present invention, a social media comprehensive public opinion analysis method is provided, including: obtaining cross-modal public opinion data of each public opinion event on social media; using a pre-trained cross-modal large model to obtain feature vectors of each cross-modal public opinion data, and clustering each public opinion event using an unsupervised clustering method based on the feature vectors to obtain a number of public opinion event clusters; using a pre-trained sentiment analysis model and a pre-trained topic analysis model to respectively obtain the sentiment scores and topic sensitivity scores of each public opinion event within each public opinion event cluster; obtaining the sentiment score index scores and topic sensitivity index scores of each public opinion event cluster respectively based on the sentiment scores and topic sensitivity scores of each public opinion event within each public opinion event cluster; obtaining the exposure index scores, interaction index scores, and propagation intensity index scores of each public opinion event cluster, and combining the sentiment score index scores and topic sensitivity index scores of each public opinion event cluster to obtain the comprehensive public opinion analysis scores of each public opinion event cluster; obtaining the public opinion evaluation results of each public opinion event cluster based on the comprehensive public opinion analysis scores of each public opinion event cluster.

[0008] Optionally, before obtaining the feature vectors of each cross-modal public opinion data by using the pre-trained cross-modal large model, it further includes: for the video modal data in the cross-modal public opinion data, the frame difference method is adopted to extract the key frames in the video modal data, and the extracted key frames are converted into image modal data; for the image modal data and text modal data in the cross-modal public opinion data, the similarity between the image modal data and the text modal data is obtained as the similarity of the cross-modal public opinion data, and the cross-modal public opinion data with a similarity less than the similarity threshold is removed according to the preset similarity threshold.

[0009] Optionally, obtaining the similarity between the image modal data and the text modal data includes: obtaining the similarity Sim between the image modal data and the text modal data through the following formula:

[0010]

[0011] where is the text vector representation generated by the text modal data through the pre-trained cross-modal large model, is the text description vector representation generated by the pre-trained cross-modal large model according to the image modal data.

[0012] Optionally, the pre-trained cross-modal large model includes a visual encoder, a language encoder and a fusion module; among them, the visual encoder is used to obtain the image embedding vector of the image modal data in the cross-modal public opinion data, the language encoder is used to obtain the text embedding vector of the text modal data in the cross-modal public opinion data, and the fusion module is used to fuse the image embedding vector and the text embedding vector based on the attention mechanism to obtain the feature vector of the cross-modal public opinion data; both the visual encoder and the language encoder adopt the Transformer architecture; the loss function of the pre-trained cross-modal large model adopts the contrastive learning loss function.

[0013] Optionally, clustering each public opinion event by using the unsupervised clustering method according to the feature vector includes: using the density-based spatial clustering algorithm with noise to group the feature vectors, and evaluating and optimizing the grouping results based on the silhouette coefficient to obtain several public opinion event clusters.

[0014] Optionally, both the pre-trained sentiment analysis model and the pre-trained topic analysis model adopt the Transformer architecture; the loss function of the pre-trained sentiment analysis model is the weighted cross-entropy loss function; the loss function of the pre-trained topic analysis model is the mean square error loss function; the pre-trained sentiment analysis model and the pre-trained topic analysis model adopt the low-rank adaptation fine-tuning method to update the model parameters during the pre-training process.

[0015] Optionally, obtaining the sentiment score index score and topic sensitivity index score of each public opinion event cluster according to the sentiment scores and topic sensitivity scores of each public opinion event in each public opinion event cluster includes: obtaining the sentiment score index score f of each public opinion event cluster through the following formula S :

[0016]

[0017] where NC is the number of public opinion events in the public opinion event cluster, s i is the sentiment score of the i-th public opinion event, and σ s is the sentiment fluctuation coefficient, is the average sentiment score of each public opinion event in the public opinion event cluster.

[0018] Obtaining the topic sensitivity index score f of each public opinion event cluster through the following formula H :

[0019]

[0020] where w n is the sensitivity weight of the topic category corresponding to the public opinion event cluster, and d i is the topic sensitivity score of the i-th public opinion event.

[0021] Optionally, obtaining the exposure index score, interactivity index score, and communication intensity index score of each public opinion event cluster includes: obtaining the exposure index score f of each public opinion event cluster through the following formula E :

[0022]

[0023] where T is the corresponding time period, and w t is the push weight of the platform for the public opinion event cluster, with a value range of [0, 1]; V t represents the number of views on the platform at the t-th moment, and λ is the time decay coefficient, with a value range of [0, 1].

[0024] Obtaining the interactivity index score f of each public opinion event cluster through the following formula I :

[0025]

[0026] where l, m, and h represent the number of likes, comments, and forwards of the public opinion event cluster respectively, a, b, and c represent the weights of the number of likes, comments, and forwards respectively, and p1, p2, and p3 represent the adjustment factors of the number of likes, comments, and forwards respectively.

[0027] Obtaining the communication intensity index score f of each public opinion event cluster through the following formula T :

[0028]

[0029] Among them, L represents the propagation level, N L is the number of forwarding times at the L-th level, v L is the propagation speed at the L-th level, L max is the maximum propagation level.

[0030] Optionally, the obtaining of the public opinion evaluation results of each public opinion event cluster according to the comprehensive public opinion analysis scores of each public opinion event cluster includes: obtaining the comprehensive public opinion analysis score threshold through the following formula

[0031]

[0032] Among them, μ is the upper bound center of the comprehensive public opinion analysis scores of theoretical normal public opinion events, ξ is the strict exponent, and ∈ is the variance of the Gaussian distribution of the upper bound center of the comprehensive public opinion analysis scores of theoretical normal public opinion events.

[0033] When the comprehensive public opinion analysis score of a public opinion event cluster is greater than the comprehensive public opinion analysis score threshold, the public opinion evaluation result of the public opinion event cluster is a high-risk public opinion event; when the comprehensive public opinion analysis score of a public opinion event cluster is not greater than the comprehensive public opinion analysis score threshold, the public opinion evaluation result of the public opinion event cluster is a low-risk public opinion event; when the public opinion evaluation result of a public opinion event cluster is a high-risk public opinion event, an early warning and measure prompt are generated.

[0034] In the second aspect of the present invention, a social media comprehensive public opinion analysis system is provided, including: a data acquisition module for acquiring cross-modal public opinion data of each public opinion event on social media; an event clustering module for using a pre-trained cross-modal large model to obtain feature vectors of each cross-modal public opinion data, and clustering each public opinion event using an unsupervised clustering method according to the feature vectors to obtain a number of public opinion event clusters; a data analysis module for using a pre-trained sentiment analysis model and a pre-trained topic analysis model to respectively obtain the sentiment scores and topic sensitivity scores of each public opinion event within each public opinion event cluster; an index acquisition module for respectively obtaining the sentiment score index scores and topic sensitivity index scores of each public opinion event cluster according to the sentiment scores and topic sensitivity scores of each public opinion event within each public opinion event cluster; an index integration module for obtaining the exposure index scores, interaction index scores and propagation intensity index scores of each public opinion event cluster, and combining the sentiment score index scores and topic sensitivity index scores of each public opinion event cluster to obtain the comprehensive public opinion analysis scores of each public opinion event cluster; a public opinion analysis module for obtaining the public opinion evaluation results of each public opinion event cluster according to the comprehensive public opinion analysis scores of each public opinion event cluster.

[0035] Optionally, before obtaining the feature vectors of each cross-modal public opinion data by using the pre-trained cross-modal large model, it further includes: for the video modal data in the cross-modal public opinion data, the frame difference method is adopted to extract the key frames in the video modal data, and the extracted key frames are converted into image modal data; for the image modal data and text modal data in the cross-modal public opinion data, the similarity between the image modal data and the text modal data is obtained as the similarity of the cross-modal public opinion data, and the cross-modal public opinion data with a similarity less than the similarity threshold is removed according to the preset similarity threshold.

[0036] Optionally, obtaining the similarity between the image modal data and the text modal data includes: obtaining the similarity Sim between the image modal data and the text modal data through the following formula:

[0037]

[0038] where is the text vector representation generated by the text modal data through the pre-trained cross-modal large model, is the text description vector representation generated by the pre-trained cross-modal large model according to the image modal data.

[0039] Optionally, the pre-trained cross-modal large model includes a visual encoder, a language encoder, and a fusion module; among them, the visual encoder is used to obtain the image embedding vector of the image modal data in the cross-modal public opinion data, the language encoder is used to obtain the text embedding vector of the text modal data in the cross-modal public opinion data, and the fusion module is used to fuse the image embedding vector and the text embedding vector based on the attention mechanism to obtain the feature vector of the cross-modal public opinion data; both the visual encoder and the language encoder adopt the Transformer architecture; the loss function of the pre-trained cross-modal large model adopts the contrastive learning loss function.

[0040] Optionally, clustering each public opinion event by using the unsupervised clustering method according to the feature vector includes: using the density-based spatial clustering algorithm with noise to group the feature vectors, and evaluating and optimizing the grouping results based on the silhouette coefficient to obtain several public opinion event clusters.

[0041] Optionally, both the pre-trained sentiment analysis model and the pre-trained topic analysis model adopt the Transformer architecture; the loss function of the pre-trained sentiment analysis model is the weighted cross-entropy loss function; the loss function of the pre-trained topic analysis model is the mean square error loss function; the pre-trained sentiment analysis model and the pre-trained topic analysis model adopt the low-rank adaptive fine-tuning method to update the model parameters during the pre-training process.

[0042] Optionally, obtaining the sentiment score index score and topic sensitivity index score of each public opinion event cluster based on the sentiment scores and topic sensitivity scores of each public opinion event in each public opinion event cluster includes: obtaining the sentiment score index score f of each public opinion event cluster through the following formula S :

[0043]

[0044] where NC is the number of public opinion events in the public opinion event cluster, s i is the sentiment score of the i-th public opinion event, and σ s is the sentiment fluctuation coefficient, is the average sentiment score of each public opinion event in the public opinion event cluster.

[0045] Obtaining the topic sensitivity index score f of each public opinion event cluster through the following formula H :

[0046]

[0047] where w n is the sensitivity weight of the topic category corresponding to the public opinion event cluster, and d i is the topic sensitivity score of the i-th public opinion event.

[0048] Optionally, obtaining the exposure index score, interactivity index score, and propagation intensity index score of each public opinion event cluster includes: obtaining the exposure index score f of each public opinion event cluster through the following formula E :

[0049]

[0050] where T is the corresponding time period, and w t is the push weight of the platform for the public opinion event cluster, with a value range of [0, 1]; V t represents the number of views of the platform at the t-th moment, and λ is the time decay coefficient, with a value range of [0, 1].[[]]END]]

[0051] Obtaining the interactivity index score f of each public opinion event cluster through the following formula I :

[0052]

[0053] where l, m, and h respectively represent the number of likes, comments, and forwards of the public opinion event cluster, a, b, and c respectively represent the weights of the number of likes, comments, and forwards, and p1, p2, and p3 respectively represent the adjustment factors of the number of likes, comments, and forwards.

[0054] Obtaining the propagation intensity index score f of each public opinion event cluster through the following formula T :

[0055]

[0056] Among them, L represents the propagation level, N L is the number of forwarding times at the L-th level, v L is the propagation speed at the L-th level, L max is the maximum propagation level.

[0057] Optionally, the obtaining of the public opinion evaluation results of each public opinion event cluster based on the comprehensive public opinion analysis scores of each public opinion event cluster includes: obtaining a comprehensive public opinion analysis score threshold through the following formula

[0058]

[0059] Among them, μ is the upper bound center of the comprehensive public opinion analysis scores of theoretical normal public opinion events, ξ is a strict exponent, and ∈ is the variance of the Gaussian distribution of the upper bound center of the comprehensive public opinion analysis scores of theoretical normal public opinion events.

[0060] When the comprehensive public opinion analysis score of a public opinion event cluster is greater than the comprehensive public opinion analysis score threshold, the public opinion evaluation result of the public opinion event cluster is a high-risk public opinion event; when the comprehensive public opinion analysis score of a public opinion event cluster is not greater than the comprehensive public opinion analysis score threshold, the public opinion evaluation result of the public opinion event cluster is a low-risk public opinion event; when the public opinion evaluation result of a public opinion event cluster is a high-risk public opinion event, a warning and measure prompt are generated.

[0061] In the third aspect of the present invention, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of the above-mentioned social media comprehensive public opinion analysis method are implemented.

[0062] In the fourth aspect of the present invention, a computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the above-mentioned social media comprehensive public opinion analysis method are implemented.

[0063] Compared with the prior art, the present invention has the following beneficial effects:

[0064] The comprehensive public opinion analysis method for social media of the present invention can efficiently acquire and process multi-modal public opinion sample data in social media. By using a pre-trained cross-modal large model as the core analysis tool, it realizes feature extraction, deep semantic association, and feature fusion between different modal data, achieves in-depth mining and dynamic analysis of public opinion data, and fully and effectively ensures the accuracy and credibility of the analysis results, providing rich context comprehensive representation information support for comprehensive public opinion analysis. At the same time, due to the high complexity of public opinion data in social media, it is difficult for a single sentiment tendency or topic word sensitivity to reasonably and accurately represent the public opinion situation of cross-modal samples. A comprehensive analysis index combining five key factors is constructed, namely, the sentiment score index, topic sensitivity index, exposure index, interactivity index, and propagation intensity index of public opinion events. By comprehensively obtaining the comprehensive public opinion analysis index, it realizes high-level and multi-dimensional analysis and evaluation of cross-modal public opinion data, effectively capturing the core features and potential risks of public opinion data. This index system not only improves the representation ability of cross-modal data but also adapts to the public opinion characteristics of different events and scenarios through flexible weight adjustment, providing an efficient and accurate basis for the timely and effective response to subsequent public opinion. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 It is a flowchart of the comprehensive public opinion analysis method for social media according to an embodiment of the present invention.

[0066] Figure 2 It is a schematic diagram of the principle of obtaining feature vectors and unsupervised clustering according to an embodiment of the present invention.

[0067] Figure 3 It is a schematic diagram of the principle of sentiment analysis and topic analysis according to an embodiment of the present invention.

[0068] Figure 4 It is a schematic diagram of the principle of comprehensive public opinion analysis according to an embodiment of the present invention.

[0069] Figure 5 It is a block diagram of the structure of the comprehensive public opinion analysis system for social media according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0070] In order to enable those skilled in the art to better understand the solution of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0071] It should be noted that the terms "first", "second", etc. in the description, claims and above-mentioned drawings of the present invention are used to distinguish similar objects, and do not necessarily need to be used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged under appropriate circumstances, so that the embodiments of the present invention described here can be implemented in an order other than those illustrated or described here. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device including a series of steps or units does not necessarily have to be limited to those steps or units clearly listed, but may include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0072] The present invention will be further described in detail below with reference to the drawings:

[0073] See Figure 1 , in an embodiment of the present invention, a method for comprehensive public opinion analysis of social media is provided, specifically a method for comprehensive public opinion analysis of social media based on a cross-modal large model.

[0074] Specifically, the method for comprehensive public opinion analysis of social media of the present invention includes the following steps:

[0075] S1: Obtain cross-modal public opinion data of each public opinion event on social media.

[0076] S2: Use the pre-trained cross-modal large model to obtain the feature vectors of each cross-modal public opinion data, and perform clustering of each public opinion event according to the feature vectors using an unsupervised clustering method to obtain several public opinion event clusters.

[0077] S3: Use the pre-trained sentiment analysis model and the pre-trained topic analysis model to obtain the sentiment scores and topic sensitivity scores of each public opinion event in each public opinion event cluster respectively.

[0078] S4: Obtain the sentiment score index scores and topic sensitivity index scores of each public opinion event cluster according to the sentiment scores and topic sensitivity scores of each public opinion event in each public opinion event cluster respectively.

[0079] S5: Obtain the exposure index scores, interaction index scores and propagation intensity index scores of each public opinion event cluster, and combine the sentiment score index scores and topic sensitivity index scores of each public opinion event cluster to obtain the comprehensive public opinion analysis scores of each public opinion event cluster.

[0080] S6: Obtain the public opinion evaluation results of each public opinion event cluster according to the comprehensive public opinion analysis scores of each public opinion event cluster.

[0081] The comprehensive public opinion analysis method for social media of the present invention can efficiently acquire and process multi-modal public opinion sample data in social media. By using a pre-trained cross-modal large model as the core analysis tool, it realizes feature extraction, deep semantic association and feature fusion between different modal data, achieves in-depth mining and dynamic analysis of public opinion data, and fully and effectively ensures the accuracy and credibility of the analysis results, providing rich context comprehensive representation information support for comprehensive public opinion analysis. At the same time, based on the high complexity of public opinion data in social media, it is difficult for a single sentiment tendency or topic word sensitivity to reasonably and accurately represent the public opinion situation of cross-modal samples. A comprehensive analysis index combining five key factors is constructed, namely the sentiment score index, topic sensitivity index, exposure index, interactivity index and propagation intensity index of public opinion events. By comprehensively obtaining the comprehensive public opinion analysis index, it realizes high-level and multi-dimensional analysis and evaluation of cross-modal public opinion data, effectively captures the core features and potential risks of public opinion data. This index system not only improves the representation ability of cross-modal data, but also adapts to the public opinion characteristics of different events and scenarios through flexible weight adjustment, providing an efficient and accurate basis for the timely and effective response to subsequent public opinion.

[0082] In a possible implementation manner, before using the pre-trained cross-modal large model to obtain the feature vectors of each cross-modal public opinion data, it further includes: for the video modal data in the cross-modal public opinion data, the frame difference method is adopted to extract the key frames in the video modal data, and the extracted key frames are converted into image modal data; for the image modal data and text modal data in the cross-modal public opinion data, the similarity between the image modal data and the text modal data is obtained as the similarity of the cross-modal public opinion data, and the cross-modal public opinion data with a similarity less than the similarity threshold is removed according to the preset similarity threshold.

[0083] Explanatorily, for the current numerous social media websites on various platforms, first, the Selenium crawler library is used to crawl corresponding various public opinion data for the currently popular social media platforms. The social media includes platforms such as Douyin, Kuaishou, Weibo and Xiaohongshu, etc., and the crawled public opinion data mainly includes three modalities, corresponding to samples of text, image and video respectively. Optionally, during the data crawling process, the online automatic crawling and manual review method can be used. By obtaining different web page data one by one, the text garbled code or the samples of the corresponding image or video data that cannot be obtained are supplemented or excluded through manual review.

[0084] In addition, it should be noted that there are some samples that only contain a single modality while other modalities are missing. For this small amount of single-modal sample data, no additional modality supplementation is performed.

[0085] Next, it is necessary to perform preliminary cleaning and denoising on the crawled cross-modal public opinion data. Since it has passed the manual preliminary review, the noise in this step does not consider the cases of data garbling or invalid data, but focuses on whether the text matches the image or the text matches the video frame. Samples with unrelated text and images need to be excluded to prevent interference with the subsequent accurate public opinion analysis work. Considering that the relevant matching algorithms and the subsequent application of cross-modal large models mainly focus on the text-image modality, in this step, for video modality data in particular, the frame difference method is adopted to extract the key frames in the video. Specifically, the average frame difference intensity is calculated based on the local maximum value, and then the frame sequence with the maximum difference intensity is selected for smoothing and the middle value is taken as the representation of the key frames of the video. The extracted key frames are converted into image modality data. After this step, most of the samples are converted into the bimodal information form of "image-text" pairs, and a small number of samples exist in the unimodal text form.

[0086] In a possible implementation manner, the obtaining of the similarity between the image modality data and the text modality data includes: obtaining the similarity Sim between the image modality data and the text modality data through the following formula:

[0087]

[0088] where is the text vector representation generated by the pre-trained cross-modal large model for the text modality data, is the text description vector representation generated by the pre-trained cross-modal large model according to the image modality data.

[0089] Explanatorily, for the cross-modal public opinion data of "image-text", the method of image description and similarity calculation is used to calculate the matching degree, so as to complete the preliminary denoising.

[0090] Specifically, the image description strategy is used to generate the text description of the image. In this step, the pre-trained cross-modal large model BLIP is first used to complete the image description generation by inputting a single image modality. For the convenience of subsequent vector similarity calculation, the text description obtained in this step of image description is an abstract description text vector representation, corresponding to the output representation of the last hidden layer of the pre-trained cross-modal large model. On this basis, according to the generated image description vector and the original text, the pre-trained cross-modal large model BLIP is used again, inputting a single original text modality, obtaining the vector representation of the text in the last hidden layer of the pre-trained cross-modal large model BLIP, and then calculating the cosine similarity of the two corresponding text vectors as the similarity between the image modality data and the text modality data. The specific calculation formula is as follows:

[0091]

[0092] The similarity between the image modality data and the text modality data of each image-text pair sample is calculated, and then a specific similarity threshold is set. Generally, the value ranges from 0 to 0.6. That is, samples with a Sim value less than the similarity threshold (0.6) will be removed as noise, while the remaining samples are considered relevant and valid samples. Subsequently, for these valid samples and a small number of single-modal text samples, through preliminary manual screening, a very small number of cases of text garbled or images not displayed are excluded, and the remaining samples will be used as the pre-cleaned samples for subsequent cross-modal comprehensive public opinion analysis tasks.

[0093] In a possible implementation, the pre-trained cross-modal large model includes a visual encoder, a language encoder, and a fusion module. Among them, the visual encoder is used to obtain the image embedding vector of the image modality data in the cross-modal public opinion data, the language encoder is used to obtain the text embedding vector of the text modality data in the cross-modal public opinion data, and the fusion module is used to fuse the image embedding vector and the text embedding vector based on the attention mechanism to obtain the feature vector of the cross-modal public opinion data. Both the visual encoder and the language encoder adopt the Transformer architecture. The loss function of the pre-trained cross-modal large model adopts the contrastive learning loss function.

[0094] Explanatorily, after obtaining and cleaning / denoising the cross-modal public opinion data of the social media platform, this step will use the pre-trained cross-modal large model BLIP to obtain the vector representation for public opinion analysis. Since the goal is to extract the feature vectors that fuse image and text pairs and use these feature vectors for unsupervised clustering to identify similar public opinion events. This process starts by feeding the input image modality data and the corresponding text modality data into the BLIP model. The BLIP model includes a visual encoder and a language encoder, which process the image modality data and the text modality data respectively.

[0095] Among them, see Figure 2, the visual encoder is based on the Transformer architecture and can convert text sequences or slices of images into a series of embedding vectors, each of which contains the meaning of the vocabulary in a specific context. In the specific cross-modal feature extraction stage, the image modality data is sliced and then converted into a series of feature maps through the visual encoder. These feature maps then pass through the attention mechanism and are transmitted to the fusion module. The text modality data is tokenized and converted into embedding vectors, which then pass through the various attention layers, batch normalization layers, and feed-forward networks of the Transformer architecture to obtain the context-related representation of each vocabulary. The fusion module combines the features of these two modalities and adopts a contrastive learning strategy of calculating the similarity between modalities. By using the following loss, the similarity of the aligned positive sample image-text pairs is made larger, and the similarity of the non-aligned negative sample image-text pairs is made smaller. The ultimate goal is to align the feature spaces of images and texts:

[0096]

[0097] where N is the number of samples, x i is the embedding vector of the i-th sample image, is the text pair matching the positive sample of the i-th sample, x k is all samples in this batch, K is the number of samples in this batch, and τ is a preset parameter.

[0098] Optionally, based on the modality fusion, a cross-modal attention mechanism can be introduced, which allows the model to dynamically allocate attention weights between image features and text features, thereby capturing the semantic relationship between the two. Finally, the fusion module outputs a unified feature vector, which synthesizes the semantic information of images and texts. This feature vector is used as the input for the subsequent unsupervised clustering algorithm.

[0099] In a possible implementation manner, the clustering of each public opinion event using the unsupervised clustering method according to the feature vector includes: grouping the feature vector using a density-based spatial clustering algorithm with noise, and evaluating and optimizing the grouping result based on the silhouette coefficient to obtain several public opinion event clusters.

[0100] Explanatorily, see again Figure 2, after extracting the feature vectors of image and text pairs, the next step is to conduct unsupervised clustering analysis. The core of this step is to use the DBSCAN algorithm (Density-Based Spatial Clustering of Applications with Noise) to group the feature vectors in order to identify similar public opinion events. This process aims to conduct a preliminary clustering of the categories for the feature vectors of the image-text pairs, so as to give public opinion clusters with different risk levels for subsequent calculations. Specifically, using the DBSCAN algorithm does not require specifying the number of clusters in advance, but forms clusters based on the density between the high-dimensional points of the feature vectors. The DBSCAN algorithm first sets two parameters: ε (neighborhood radius) and Minpts (minimum number of points in the neighborhood). In the algorithm, each vector point is checked. If the number of points contained within its neighborhood (within a sphere with radius ε) is greater than or equal to Minpts, then this point is considered a core point. Then the algorithm expands this cluster to include all points that are directly or indirectly reachable from the core point, that is, those points whose number of points in the neighborhood also meets Minpts. This process continues until all points are assigned to a cluster or marked as noise. Therefore, in the context of actual social media public opinion analysis, public opinion events with similar feature vectors can be grouped together to form a preliminary event cluster. These clusters represent public opinion events that are close to each other in the feature space, thus revealing possible patterns and associations that are difficult to detect by the human eye. The advantage of the DBSCAN algorithm adopted in this step is that it can identify clusters of arbitrary shapes and has a certain robustness to noise and outliers, which is particularly important for public opinion analysis because public opinion data often contains a large amount of noise and irrelevant information.

[0101] After unsupervised clustering, the next step is to complete the evaluation and optimization of the clustering results. This involves calculating the evaluation metric of clustering, namely the silhouette coefficient, which measures the compactness of samples within a cluster and the separation of samples between clusters. The value range of the silhouette coefficient is from -1 to 1, and generally, the larger the value, the better the corresponding clustering effect. Based on these evaluation metrics, the hyperparameters ε and Minpts are further adjusted for optimization. This step directly affects the understanding and response of the entire model framework to public opinion events. Through the optimization of the clustering results in this step, the feature vectors of the image-text pairs corresponding to similar public opinion events will be more accurately identified and grouped into the same cluster, thus providing more powerful support for subsequent public opinion analysis tasks such as sentiment analysis, topic analysis, and fine-tuning of cross-modal large models.

[0102] In a possible implementation manner, both the pre-trained sentiment analysis model and the pre-trained topic analysis model adopt the Transformer architecture; the loss function of the pre-trained sentiment analysis model is a weighted cross-entropy loss function; the loss function of the pre-trained topic analysis model is a mean square error loss function; the pre-trained sentiment analysis model and the pre-trained topic analysis model adopt a low-rank adaptation fine-tuning method to update model parameters during the pre-training process.

[0103] Explanatory, design a sentiment analysis model and a topic analysis model, and conduct fine-tuning work on the sentiment analysis model and the topic analysis model through a small amount of labeled data to complete specific analysis tasks.

[0104] Cross-modal public opinion data extracts its semantic representation in the shared feature space through a cross-modal large model. At the same time, these feature vectors are classified through unsupervised clustering to initially divide the category distribution of public opinion events, providing data support and task guidance for subsequent sentiment analysis and topic analysis. Through the preliminary results of the above clustering, the theme categories of high-risk public opinion events can be roughly located, laying a foundation for the sentiment tendency recognition and topic extraction tasks. Therefore, for annotation, first, according to the preliminary clustering results, manually give the sentiment category labels and topic label scores of the public opinion events in each category cluster. Since a large number of public opinion events have been classified through unsupervised clustering, only the clusters need to be annotated in this step, and the manual annotation cost is greatly reduced. It should be noted that only specific labels need to be given for the sentiment category, while for the topic label, in addition to giving the specific label category, the true sensitivity score corresponding to each topic category needs to be given according to prior information. For example, the true score of a sensitive topic is relatively high, while the score of a relatively insensitive topic is relatively low.

[0105] On this basis, refer to Figure 3 and design a sentiment analysis model and a topic analysis model to conduct a detailed analysis on the sentiment changes and core themes of public opinion events respectively. The goal of the sentiment analysis model is to identify the sentiment polarity in the data, that is, the sentiment tendency of positive (+1), neutral (0), or negative (-1). In the sentiment analysis model, first, the cross-modal fusion feature vectors within each cluster are input into the sentiment decoder. The decoder consists of multiple fully connected layers and is used to map the fusion features to the sentiment space. The design of the decoder adopts the Transformer structure to capture the sentiment tendency, generate sentiment analysis scores, and is trained and optimized through cross-entropy loss to handle sentiment classification problems, thereby optimizing the accuracy of sentiment prediction.

[0106] The weighted cross-entropy loss L used in the training of the sentiment analysis model S has the following calculation formula:

[0107]

[0108] In the formula, represents the unbalanced weighting coefficient of the emotion category, and p(y m(i) |x i ) represents the probability of predicting the emotion category cluster y m(i) corresponding to the public opinion event, and m(i) represents the category cluster number corresponding to the sample. This formula will be used to optimize the accurate classification probability value of the emotion analysis branch during the training process to complete accurate classification.

[0109] Refer to Figure 3 again. In the topic analysis model, the feature vectors of cross-modal fusion are input into a multi-layer fully connected decoder to extract the core information embedding of the public opinion event. The ultimate goal is to generate the topic sensitivity score based on these feature vectors. In the specific implementation process, first, the feature vectors fused within each cluster are input into the topic decoder. This decoder is used to map the fused features to a high-dimensional space and process the high-dimensional vectors to obtain the sensitivity score of each topic. This score represents the sensitivity degree of the topic corresponding to the public opinion event, thereby further evaluating the importance and sensitivity of each topic in the current public opinion event. The design in the decoder is through the Transformer model, and the output of the previous layer is passed into one or more fully connected layers and mapped to a single sensitivity score. This sensitivity score is a real value and is mapped to the value range of [0, 1] through the sigmoid function. The higher the score, the greater the sensitivity of the topic in the current public opinion and the more likely it is to attract more attention and discussion.

[0110] In the specific training process, the topic analysis model adopts the mean square error loss for comprehensive optimization:

[0111]

[0112] where N is the number of public opinion event samples, s p,i is the topic sensitivity score predicted by the model, and s g,m(i) is the true value of the topic label sensitivity score corresponding to the topic cluster of the m(i) category.

[0113] The cross-entropy loss in the loss optimization is used to ensure the accuracy of emotion classification, while the mean square error loss optimizes the calculation of the topic score, enabling the model to more accurately analyze public opinion emotions and evaluate the sensitivity of topics. During the training process, the model will continuously adjust the parameters to make the sentiment tendency and topic sensitivity score as accurately as possible reflect the public opinion influence of each topic. This method can not only provide accurate identification and monitoring results for public opinion analysis but also achieve quantitative evaluation of public opinion events in practical applications, helping decision-makers identify potential public opinion risks and hot topics.

[0114] Exemplarily, the sentiment analysis model and the topic analysis model can be integrated into the cross-modal large model. Furthermore, the training of the sentiment analysis model and the topic analysis model will be carried out in the fine-tuning stage of the cross-modal large model, that is, the sentiment analysis model and the topic analysis model are fine-tuned by using a small amount of labeled data and the low-rank adaptation fine-tuning method (LoRA). In this step, the parameters of the feature extraction part of the cross-modal large model are frozen, and only the parameters of the sentiment classification decoder and the topic decoder are updated. During fine-tuning, the weighted cross-entropy loss function and the mean squared error loss are jointly optimized to ensure the accuracy of sentiment classification and the accuracy of topic extraction. Finally, the sentiment analysis model and the topic analysis model use the fine-tuned parameters to achieve a deep understanding of the sentiment and topics of public opinion events, and can provide identification and real-time monitoring support according to sentiment changes and topic evolution in practical applications.

[0115] In a possible implementation manner, obtaining the sentiment score index score and the topic sensitivity index score of each public opinion event cluster according to the sentiment scores and topic sensitivity scores of each public opinion event in each public opinion event cluster includes: obtaining the sentiment score index score f of each public opinion event cluster through the following formula S :

[0116]

[0117] where NC is the number of public opinion events in the public opinion event cluster, s i is the sentiment score of the i-th public opinion event, and σ s is the sentiment fluctuation coefficient, is the average sentiment score of each public opinion event in the public opinion event cluster.

[0118] Obtaining the topic sensitivity index score f of each public opinion event cluster through the following formula H :

[0119]

[0120] where w n is the sensitivity weight of the topic category corresponding to the public opinion event cluster, and d i is the topic sensitivity score of the i-th public opinion event.

[0121] Explanatorily, after completing the feature extraction, clustering, and model fine-tuning of sentiment and topic analysis based on the pre-trained cross-modal large model, the sentiment scores and topic sensitivity scores corresponding to each public opinion event can be obtained, and then the sentiment score index scores and topic sensitivity index scores of each public opinion event cluster can be obtained.

[0122] In this step, these two indicators will be fully utilized, and combined with three other indicators to form a comprehensive public opinion analysis comprehensive indicator CPI, to complete multi-dimensional comprehensive public opinion analysis, and set specific thresholds based on the analysis results of CPI for timely high-risk public opinion early warning, and combine the text output of the cross-modal large model as the analysis results and corresponding response measures.

[0123] Regarding the sentiment score indicator, the sentiment score indicator will use the sentiment score given by the cross-modal large model and introduce an additional sentiment fluctuation coefficient σ s , this fluctuation coefficient mainly analyzes the sentiment volatility of users and the distribution of user groups for public opinion events within each cluster. The calculation formula is as follows:

[0124]

[0125] s i The value range of is [-1, 1], and σ s As described above, it represents the volatility of the sentiment score within the cluster. The greater the volatility, the more unstable the sentiment of the public opinion event. The specific calculation formula is as follows:

[0126]

[0127] Through the calculation of the sentiment score indicator, the sentiment situation of each public opinion event can be better given, and a higher risk level can be given to public opinion events with negative sentiment.

[0128] Regarding the topic sensitivity indicator, the topic sensitivity indicator will use the topic analysis result score of the cross-modal large model. The higher the score, the stronger the sensitivity of the topic. On this basis, considering the topic categories within each cluster, higher sensitivity weights will be given to some sensitive topic categories, such as political and social events. Therefore, the topic sensitivity indicator f H The calculation formula of is as follows:

[0129]

[0130] Among them, d i The value range of is [0, 1]. Topic sensitivity is a quantitative indicator of the sensitivity of the current public opinion topic, which provides a more dynamic sensitivity assessment.

[0131] In a possible implementation manner, regarding the exposure indicator, the exposure indicator design takes into account the platform's view count, the algorithm weight of platform push, and time factors. Generally, as time goes by, and the view count decreases or the platform push decreases, the risk level of the public opinion event will correspondingly decrease. Therefore, the calculation of this indicator uses an exposure function with a time decay coefficient. The calculation formula is as follows:

[0132]

[0133] Among them, T is the corresponding time period, and w t is the push weight of the platform for the public opinion event cluster, and its value range is [0, 1]; V t represents the page view volume of the platform at the t-th moment, λ is the time decay coefficient, and its value range is [0, 1]. The corresponding time period usually corresponds to about one week. The page view volume and time period in this formula will be obtained during sample crawling, while the push weight and time decay coefficient of the platform will be used as hyperparameters to control the size of the exposure index. Through the calculation of the exposure index, the exposure effect of a certain public opinion event at different times can be better reflected.

[0134] Regarding the interactivity index, the interactivity index will comprehensively evaluate the interactivity strength effect of the public opinion event by combining the number of likes, comments, and forwards of the public opinion event on the platform. For each sub-factor, the influence of different behaviors on the public opinion event is given by introducing weights and non-linear adjustment factors. The specific calculation formula is as follows:

[0135]

[0136] Among them, l, m, and h respectively represent the number of likes, comments, and forwards of the public opinion event cluster, which respectively represent the degree of interest, discussion, and dissemination willingness of users for the content of the public opinion event. a, b, and c respectively represent the weights of the number of likes, comments, and forwards, and p1, p2, and p3 respectively represent the adjustment factors of the number of likes, comments, and forwards, which are used to control the non-linear influence of different behaviors. The calculation of this interactivity index involves the crawling of the number of likes, comments, and forwards of the public opinion event, and at the same time, different weights and adjustment factor hyperparameters are designed, and logarithmic smoothing is used to suppress extreme values, so as to give a reasonable interactivity analysis index. Generally speaking, the greater the interactivity, the corresponding risk level of the public opinion event will also increase.

[0137] Regarding the communication intensity index, the communication intensity is an extension of the sub-factor number of forwards of the interactivity index in the time dimension. Because the public opinion data on the actual platform will involve multi-level forwards, secondary dissemination, etc., it is necessary to model and analyze the communication mode of the public opinion event and give an index of communication intensity to evaluate the risk level. The calculation formula is as follows:

[0138]

[0139] Among them, L represents the communication level, and N L is the number of forwards at the L-th level, and v L is the communication speed at the L-th level, which is expressed by the number of forwards per unit time, and L maxIt is the maximum propagation level, generally limited to 3 levels as the breadth of event propagation. According to the propagation intensity index, it can greatly help to judge the risk level identification of public opinion events and give more accurate public opinion indicators in the propagation dimension.

[0140] In a possible implementation manner, referring to Figure 4 , combining the above five dimensions, the calculation formula for the comprehensive public opinion analysis index CPI of social media is obtained:

[0141] CPI = αf E + βf S + γf I + δf T + εf H

[0142] where f E , f S , f I , f T , f H are the sub-dimension behavior calculation indicators of each dimension respectively, and α, β, γ, δ, ε are weight hyperparameters respectively. By controlling these hyperparameters, the value of CPI can have a minimum lower bound. At the same time, since each sub-index either uses logarithmic smoothing or limits the value range to [0, 1], the CPI values of most public opinion events will be stable within a certain range. Finally, through this complex function form design, CPI enables the identification and monitoring of public opinion to more accurately quantify user behavior and provide effective public opinion early warning and evaluation for decision-makers through multi-dimensional comprehensive analysis.

[0143] In a possible implementation manner, the obtaining of the public opinion evaluation results of each public opinion event cluster according to the comprehensive public opinion analysis scores of each public opinion event cluster includes: obtaining the comprehensive public opinion analysis score threshold through the following formula

[0144]

[0145] where μ is the upper bound center of the comprehensive public opinion analysis scores of theoretical normal public opinion events, ξ is the strict index, and ε is the variance of the Gaussian distribution of the upper bound center of the comprehensive public opinion analysis scores of theoretical normal public opinion events.

[0146] When the comprehensive public opinion analysis score of a public opinion event cluster is greater than the comprehensive public opinion analysis score threshold, the public opinion evaluation result of the public opinion event cluster is a high-risk public opinion event; when the comprehensive public opinion analysis score of a public opinion event cluster is not greater than the comprehensive public opinion analysis score threshold, the public opinion evaluation result of the public opinion event cluster is a low-risk public opinion event; when the public opinion evaluation result of a public opinion event cluster is a high-risk public opinion event, an early warning and measure prompt are generated.

[0147] Explanatory. After completing the design of the comprehensive public opinion analysis index, this step will set a dynamic threshold according to the value range of CPI. Public opinion events exceeding the set threshold will be considered high-risk social media public opinions and require special attention. Assume that the lower bound of the comprehensive public opinion analysis index CPI is τ, and the CPI values of most normal public opinion events are stable within a range, and the upper bound of this range follows a general Gaussian distribution. Thus, the upper bound of this range will be within [μ - ε, μ + ε], where μ is the center of the upper bound of theoretical normal public opinion events, and ∈ is the variance of this Gaussian distribution, corresponding to the certainty of the upper bound position of public opinion event risks.

[0148] Based on this, the comprehensive public opinion analysis index designed in this step will be valued according to the actual strictness of the identification and monitoring of public opinion events. If the value of the strictness index ξ is higher, it means the identification and monitoring are more strict, and its value range is [0, 1]. Therefore, according to the prior strictness index, a dynamic threshold is calculated as follows:

[0149]

[0150] According to this threshold, if the comprehensive public opinion analysis index CPI of a certain public opinion event is higher than the dynamic threshold a warning will be given, warning that this public opinion event needs to be focused on and dealt with.

[0151] In addition, based on the feature information after fusing text and images by the cross-modal large model, additional prompt words are provided in the input text modality, requiring the model to perform automated calculation and threshold judgment on CPI. The model for high-risk public opinion events exceeding the threshold needs to give specific handling methods and countermeasures. Finally, through further analysis and decision-making reasoning of multi-modal features, the handling suggestions for high-risk public opinion events are output as a set of clear operation guidelines. These guidelines include but are not limited to: generating targeted public opinion guidance words, simulating potential risk diffusion paths, identifying key audience groups and formulating response strategies. Through the full-process closed-loop processing mechanism, the system can monitor public opinion dynamics in real time and update the model threshold to ensure the accuracy of early warnings and the effectiveness of countermeasures, thereby providing intelligent support for public opinion management and enhancing the risk control ability.

[0152] The following is the device embodiment of the present invention, which can be used to execute the method embodiment of the present invention. For details not disclosed in the device embodiment, please refer to the method embodiment of the present invention.

[0153] See Figure 5, in another embodiment of the present invention, a social media comprehensive public opinion analysis system is provided, which can be used to implement the above-mentioned social media comprehensive public opinion analysis method. Specifically, the social media comprehensive public opinion analysis system includes a data acquisition module, an event clustering module, a data analysis module, an index acquisition module, an index integration module, and a public opinion analysis module.

[0154] Among them, the data acquisition module is used to obtain cross-modal public opinion data of each public opinion event on social media; the event clustering module is used to obtain the feature vectors of each cross-modal public opinion data by using a pre-trained cross-modal large model, and perform clustering of each public opinion event using an unsupervised clustering method according to the feature vectors to obtain several public opinion event clusters; the data analysis module is used to obtain the sentiment scores and topic sensitivity scores of each public opinion event in each public opinion event cluster by using a pre-trained sentiment analysis model and a pre-trained topic analysis model respectively; the index acquisition module is used to obtain the sentiment score index scores and topic sensitivity index scores of each public opinion event cluster according to the sentiment scores and topic sensitivity scores of each public opinion event in each public opinion event cluster respectively; the index integration module is used to obtain the exposure index scores, interaction index scores and communication intensity index scores of each public opinion event cluster, and combine the sentiment score index scores and topic sensitivity index scores of each public opinion event cluster to obtain the comprehensive public opinion analysis scores of each public opinion event cluster; the public opinion analysis module is used to obtain the public opinion evaluation results of each public opinion event cluster according to the comprehensive public opinion analysis scores of each public opinion event cluster.

[0155] In a possible implementation manner, before obtaining the feature vectors of each cross-modal public opinion data by using the pre-trained cross-modal large model, it further includes: for the video modal data in the cross-modal public opinion data, a frame difference method is adopted to extract the key frames in the video modal data, and the extracted key frames are converted into image modal data; for the image modal data and text modal data in the cross-modal public opinion data, the similarity between the image modal data and the text modal data is obtained as the similarity of the cross-modal public opinion data, and the cross-modal public opinion data with a similarity less than the similarity threshold is removed according to a preset similarity threshold.

[0156] In a possible implementation manner, obtaining the similarity between the image modal data and the text modal data includes: obtaining the similarity Sim between the image modal data and the text modal data through the following formula:

[0157]

[0158] Among them, is the text vector representation generated by the pre-trained cross-modal large model for the text modal data, is the text description vector representation generated by the pre-trained cross-modal large model according to the image modal data.

[0159] In a possible implementation, the pre-trained cross-modal large model includes a visual encoder, a language encoder, and a fusion module. The visual encoder is used to obtain the image embedding vector of the image modality data in the cross-modal public opinion data. The language encoder is used to obtain the text embedding vector of the text modality data in the cross-modal public opinion data. The fusion module is used to fuse the image embedding vector and the text embedding vector based on the attention mechanism to obtain the feature vector of the cross-modal public opinion data. Both the visual encoder and the language encoder adopt the Transformer architecture. The loss function of the pre-trained cross-modal large model adopts the contrastive learning loss function.

[0160] In a possible implementation, the clustering of each public opinion event using the unsupervised clustering method according to the feature vector includes: grouping the feature vector using the density-based spatial clustering of applications with noise algorithm, and evaluating and optimizing the grouping result based on the silhouette coefficient to obtain several public opinion event clusters.

[0161] In a possible implementation, both the pre-trained sentiment analysis model and the pre-trained topic analysis model adopt the Transformer architecture. The loss function of the pre-trained sentiment analysis model is the weighted cross-entropy loss function. The loss function of the pre-trained topic analysis model is the mean square error loss function. The pre-trained sentiment analysis model and the pre-trained topic analysis model adopt the low-rank adaptation fine-tuning method to update the model parameters during the pre-training process.

[0162] In a possible implementation, obtaining the sentiment score index score and the topic sensitivity index score of each public opinion event cluster according to the sentiment scores and topic sensitivity scores of each public opinion event in each public opinion event cluster includes: obtaining the sentiment score index score f of each public opinion event cluster through the following formula S :

[0163]

[0164] where NC is the number of public opinion events in the public opinion event cluster, s i is the sentiment score of the i-th public opinion event, σ s is the sentiment fluctuation coefficient, is the mean value of the sentiment scores of each public opinion event in the public opinion event cluster.

[0165] Obtaining the topic sensitivity index score f of each public opinion event cluster through the following formula H :

[0166]

[0167] where w n is the sensitivity weight of the topic category corresponding to the public opinion event cluster, d iis the topic sensitivity score of the i-th public opinion event.

[0168] In a possible implementation, the obtaining of the exposure index score, interaction index score, and propagation intensity index score of each public opinion event cluster includes: obtaining the exposure index score f of each public opinion event cluster through the following formula E :

[0169]

[0170] where T is the corresponding time period, and w t is the push weight of the platform for the public opinion event cluster, and the value range is [0, 1]; V t represents the page view volume of the platform at the t-th moment, and λ is the time decay coefficient, and the value range is [0, 1].

[0171] Obtain the interaction index score f of each public opinion event cluster through the following formula I :

[0172]

[0173] where l, m, and h respectively represent the number of likes, comments, and forwards of the public opinion event cluster, a, b, and c respectively represent the weights of the number of likes, comments, and forwards, and p1, p2, and p3 respectively represent the adjustment factors of the number of likes, comments, and forwards.

[0174] Obtain the propagation intensity index score f of each public opinion event cluster through the following formula T :

[0175]

[0176] where L represents the propagation level, N L is the number of forwards at the L-th level, v L is the propagation speed at the L-th level, and L max is the maximum propagation level.

[0177] In a possible implementation, the obtaining of the public opinion evaluation result of each public opinion event cluster according to the comprehensive public opinion analysis score of each public opinion event cluster includes: obtaining the threshold of the comprehensive public opinion analysis score through the following formula

[0178]

[0179] where μ is the upper bound center of the comprehensive public opinion analysis score of the theoretical normal public opinion event, ξ is the strict exponent, and ∈ is the variance of the Gaussian distribution of the upper bound center of the comprehensive public opinion analysis score of the theoretical normal public opinion event.

[0180] When the comprehensive public opinion analysis score of the public opinion event cluster is greater than the comprehensive public opinion analysis score threshold, the public opinion assessment result of the public opinion event cluster is a high-risk public opinion event; when the comprehensive public opinion analysis score of the public opinion event cluster is not greater than the comprehensive public opinion analysis score threshold, the public opinion assessment result of the public opinion event cluster is a low-risk public opinion event.

[0181] When the public opinion assessment result of the public opinion event cluster is a high-risk public opinion event, warnings and measure prompts are generated.

[0182] All relevant contents of each step involved in the embodiments of the foregoing social media comprehensive public opinion analysis method can be cited in the function descriptions of the corresponding function modules of the social media comprehensive public opinion analysis system in the embodiments of the present invention, and will not be elaborated here.

[0183] The division of modules in the embodiments of the present invention is illustrative. It is only a logical function division. In actual implementation, there may be other division methods. In addition, in each embodiment of the present invention, the various function modules can be integrated in one processor, or can exist separately physically, or two or more modules can be integrated in one module. The above integrated modules can be implemented in the form of hardware or in the form of software function modules.

[0184] In another embodiment of the present invention, a computer device is provided. The computer device includes a processor and a memory. The memory is used to store a computer program. The computer program includes program instructions. The processor is used to execute the program instructions stored in the computer storage medium. The processor may be a Central Processing Unit (CPU), or may also be other general-purpose processors, digital signal processors (DSPs), application specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing core and control core of the terminal, and is suitable for implementing one or more instructions. Specifically, it is suitable for loading and executing one or more instructions in the computer storage medium to implement the corresponding method flow or corresponding function; the processor described in the embodiments of the present invention can be used for the operation of the social media comprehensive public opinion analysis method.

[0185] In another embodiment of the present invention, the present invention further provides a storage medium, specifically a computer-readable storage medium (Memory). The computer-readable storage medium is a memory device in a computer device and is used to store programs and data. It can be understood that the computer-readable storage medium here can include both the built-in storage medium in the computer device and, of course, the extended storage medium supported by the computer device. The computer-readable storage medium provides a storage space, and this storage space stores the operating system of the terminal. Moreover, one or more instructions suitable for being loaded and executed by the processor are stored in this storage space, and these instructions can be one or more computer programs (including program codes). It should be noted that the computer-readable storage medium here can be a high-speed RAM memory or a non-volatile memory, such as at least one disk memory. One or more instructions stored in the computer-readable storage medium can be loaded and executed by the processor to implement the corresponding steps of the social media comprehensive public opinion analysis method in the above embodiments.

[0186] Those skilled in the art should understand that the embodiments of the present invention can be provided as a method, a system, or a computer program product. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk memories, CD-ROMs, optical memories, etc.) containing computer-usable program codes.

[0187] The present invention is described with reference to the flowcharts and / or block diagrams of methods, apparatuses (systems), and computer program products according to embodiments of the present invention. It should be understood that each flow and / or block in the flowchart and / or block diagram, as well as the combination of flows and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing devices generate a device for implementing the functions specified in Figure 1 one flow or multiple flows and / or blocks Figure 1 one block or multiple blocks.

[0188] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer-readable memory generate a manufactured article including an instruction device, and the instruction device implements the functions in Figure 1 one flow or multiple flows and / or blocksFigure 1 The functions specified in one or more boxes.

[0189] These computer program instructions can also be loaded onto a computer or other programmable data processing device, so that a series of operation steps are executed on the computer or other programmable device to generate a computer-implemented process. Thus, the instructions executed on the computer or other programmable device provide for implementing the steps of the functions specified in Figure 1 one process or more processes and / or boxes Figure 1 the functions specified in one box or more boxes.

[0190] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to the above embodiments, those of ordinary skill in the art should understand that: modifications or equivalent replacements can still be made to the specific implementation manners of the present invention. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention shall be covered by the protection scope of the claims of the present invention.

Claims

1. A comprehensive public opinion analysis method for social media, characterized in that, Including: Obtain cross-modal public opinion data of each public opinion event on social media; Use a pre-trained cross-modal large model to obtain the feature vectors of each cross-modal public opinion data, and perform clustering on each public opinion event using an unsupervised clustering method based on the feature vectors to obtain several public opinion event clusters; Use a pre-trained sentiment analysis model and a pre-trained topic analysis model to obtain the sentiment scores and topic sensitivity scores of each public opinion event within each public opinion event cluster respectively; Based on the sentiment scores and topic sensitivity scores of each public opinion event within each public opinion event cluster, obtain the sentiment score index scores and topic sensitivity index scores of each public opinion event cluster respectively; Obtain the exposure index scores, interaction index scores, and propagation intensity index scores of each public opinion event cluster, and combine the sentiment score index scores and topic sensitivity index scores of each public opinion event cluster to obtain the comprehensive public opinion analysis scores of each public opinion event cluster; Obtain the public opinion evaluation results of each public opinion event cluster based on the comprehensive public opinion analysis scores of each public opinion event cluster.

2. The social media comprehensive public opinion analysis method according to claim 1, wherein Before using the pre-trained cross-modal large model to obtain the feature vectors of each cross-modal public opinion data, it further includes: For the video modal data in the cross-modal public opinion data, adopt an inter-frame difference method to extract the key frames in the video modal data, and convert the extracted key frames into image modal data; For the image modal data and text modal data in the cross-modal public opinion data, obtain the similarity between the image modal data and the text modal data as the similarity of the cross-modal public opinion data, and remove the cross-modal public opinion data with a similarity less than the similarity threshold according to the preset similarity threshold.

3. The social media comprehensive public opinion analysis method according to claim 2, wherein The obtaining of the similarity between the image modal data and the text modal data includes: Obtain the similarity Sim between the image modal data and the text modal data through the following formula: Among them, is the text vector representation generated by the pre-trained cross-modal large model for the text modal data, is the text description vector representation generated by the pre-trained cross-modal large model according to the image modal data.

4. The social media comprehensive public opinion analysis method according to claim 1, characterized in that The pre-trained cross-modal large model includes a visual encoder, a language encoder, and a fusion module; among them, the visual encoder is used to obtain the image embedding vector of the image modal data in the cross-modal public opinion data, the language encoder is used to obtain the text embedding vector of the text modal data in the cross-modal public opinion data, and the fusion module is used to fuse the image embedding vector and the text embedding vector based on the attention mechanism to obtain the feature vector of the cross-modal public opinion data; both the visual encoder and the language encoder adopt the Transformer architecture; the loss function of the pre-trained cross-modal large model adopts a contrastive learning loss function.

5. The social media comprehensive public opinion analysis method according to claim 1, characterized in that The clustering of each public opinion event using an unsupervised clustering method based on the feature vectors includes: Use a density-based spatial clustering algorithm with noise to group the feature vectors, and evaluate and optimize the grouping results based on the silhouette coefficient to obtain several public opinion event clusters.

6. The social media comprehensive public opinion analysis method according to claim 1, characterized in that Both the pre-trained sentiment analysis model and the pre-trained topic analysis model adopt the Transformer architecture; The loss function of the pre-trained sentiment analysis model is a weighted cross-entropy loss function; The loss function of the pre-trained topic analysis model is a mean squared error loss function; The pre-trained sentiment analysis model and the pre-trained topic analysis model adopt a low-rank adaptation fine-tuning method to update the model parameters during the pre-training process.

7. The social media comprehensive public opinion analysis method according to claim 1, wherein The sentiment score index score and topic sensitivity index score of each public opinion event cluster are obtained according to the sentiment score and topic sensitivity score of each public opinion event cluster, respectively, including: The sentiment score index score f of each public opinion event cluster is obtained by the following formula S :[[-]] Among them, NC is the number of public opinion events in the public opinion event cluster, s i is the sentiment score of the i-th public opinion event, and σ s is the sentiment fluctuation coefficient, is the average sentiment score of each public opinion event in the public opinion event cluster; The topic sensitivity index score f of each public opinion event cluster is obtained by the following formula H :[[]]END]] Among them, w n is the sensitivity weight of the topic category corresponding to the public opinion event cluster, and d i is the topic sensitivity score of the i-th public opinion event.

8. The social media comprehensive public opinion analysis method according to claim 1, characterized in that The steps of obtaining the exposure index score, interactivity index score, and communication intensity index score of each public opinion event cluster include: The exposure index score f of each public opinion event cluster is obtained by the following formula E : Among them, T is the corresponding time period, and w t is the push weight of the platform for the public opinion event cluster, and the value range is [0, 1]; V t represents the page view volume of the platform at the t-th moment, and λ is the time decay coefficient, and the value range is [0, 1]; The interactivity index score f of each public opinion event cluster is obtained by the following formula I : Where l, m, and h represent the number of likes, comments, and reposts of the public opinion event cluster, respectively; a, b, and c represent the weights of the number of likes, comments, and reposts, respectively; and p1, p2, and p3 represent the adjustment factors of the number of likes, comments, and reposts, respectively. The propagation intensity index score f of each public opinion event cluster is obtained by the following formula T :[[]] Among them, L represents the propagation level, and N L is the number of forwarding times at the L-th level, and v L is the propagation speed at the L-th level, and L max is the maximum propagation level.

9. The social media comprehensive public opinion analysis method according to claim 1, wherein The public opinion evaluation results of each public opinion event cluster obtained based on the comprehensive public opinion analysis scores of each public opinion event cluster include: Obtain the comprehensive public opinion analysis score threshold through the following formula Among them, μ is the upper bound center of the comprehensive public opinion analysis score of theoretically normal public opinion events, ξ is the strictness index, and ∈ is the variance of the Gaussian distribution of the upper bound center of the comprehensive public opinion analysis score of theoretically normal public opinion events; When the comprehensive public opinion analysis score of a public opinion event cluster is greater than the comprehensive public opinion analysis score threshold, the public opinion assessment result of the public opinion event cluster is a high-risk public opinion event; when the comprehensive public opinion analysis score of a public opinion event cluster is not greater than the comprehensive public opinion analysis score threshold, the public opinion assessment result of the public opinion event cluster is a low-risk public opinion event; When the public opinion assessment result of a public opinion event cluster is a high-risk public opinion event, an early warning and action prompts are generated.

10. A comprehensive social media public opinion analysis system, characterized in that, include: The data acquisition module is used to obtain cross-modal public opinion data of various public opinion events on social media; An event clustering module is used to obtain feature vectors of each cross-modal public opinion data using a pre-trained cross-modal large model, and cluster each public opinion event using an unsupervised clustering method based on the feature vectors to obtain a number of public opinion event clusters; The data analysis module is used to use the pre-trained sentiment analysis model and the pre-trained topic analysis model to obtain the sentiment score and topic sensitivity score of each public opinion event in each public opinion event cluster; An index acquisition module is used to obtain the sentiment score index score and topic sensitivity index score of each public opinion event cluster based on the sentiment score and topic sensitivity score of each public opinion event in each public opinion event cluster; The index synthesis module is used to obtain the exposure index score, interactivity index score and communication intensity index score of each public opinion event cluster, and combine the sentiment score index score and topic sensitivity index score of each public opinion event cluster to obtain the comprehensive public opinion analysis score of each public opinion event cluster; The public opinion analysis module is used to obtain the public opinion evaluation results of each public opinion event cluster based on the comprehensive public opinion analysis score of each public opinion event cluster.

11. The social media comprehensive public opinion analysis system according to claim 10, characterized in that, Before using the pre-trained cross-modal large model to obtain the feature vectors of each cross-modal public opinion data, the method further includes: For the video modality data in the cross-modal public opinion data, the inter-frame difference method is used to extract the key frames in the video modality data, and the extracted key frames are converted into image modality data; For the image modal data and text modal data in the cross-modal public opinion data, the similarity between the image modal data and the text modal data is obtained as the similarity of the cross-modal public opinion data, and the cross-modal public opinion data with a similarity less than the similarity threshold is eliminated according to a preset similarity threshold.

12. The social media comprehensive public opinion analysis system according to claim 11, wherein The obtaining of the similarity between the image modality data and the text modality data includes: Obtaining the similarity Sim between the image modality data and the text modality data through the following formula: Among them, is the text vector representation generated by the pre-trained cross-modal large model for text modality data, is the text description vector representation generated by the pre-trained cross-modal large model according to the image modality data.

13. The social media comprehensive public opinion analysis system according to claim 10, characterized in that The pre-trained cross-modal large model includes a visual encoder, a language encoder, and a fusion module; among them, the visual encoder is used to obtain the image embedding vector of the image modality data in the cross-modal public opinion data, the language encoder is used to obtain the text embedding vector of the text modality data in the cross-modal public opinion data, and the fusion module is used to fuse the image embedding vector and the text embedding vector based on the attention mechanism to obtain the feature vector of the cross-modal public opinion data; both the visual encoder and the language encoder adopt the Transformer architecture; the loss function of the pre-trained cross-modal large model adopts the contrastive learning loss function.

14. The social media comprehensive public opinion analysis system according to claim 10, characterized in that, The clustering of each public opinion event according to the feature vector using the unsupervised clustering method includes: Using the density-based spatial clustering algorithm with noise to group the feature vector, and evaluating and optimizing the grouping result based on the silhouette coefficient to obtain several public opinion event clusters.

15. The social media comprehensive public opinion analysis system according to claim 1, characterized in that, Both the pre-trained sentiment analysis model and the pre-trained topic analysis model adopt the Transformer architecture; The loss function of the pre-trained sentiment analysis model is the weighted cross-entropy loss function; The loss function of the pre-trained topic analysis model is the mean square error loss function; The pre-trained sentiment analysis model and the pre-trained topic analysis model adopt the low-rank adaptation fine-tuning method to update the model parameters during the pre-training process.

16. The social media comprehensive public opinion analysis system according to claim 1, wherein The obtaining of the sentiment score index score and the topic sensitivity index score of each public opinion event cluster according to the sentiment scores and topic sensitivity scores of each public opinion event in each public opinion event cluster includes: The sentiment score index score f of each public opinion event cluster is obtained by the following formula S :[[]]END]] Among them, NC is the number of public opinion events in the public opinion event cluster, s i is the sentiment score of the i-th public opinion event, and σ s is the sentiment fluctuation coefficient, is the average sentiment score of each public opinion event in the public opinion event cluster; The topic sensitivity index score f of each public opinion event cluster is obtained by the following formula H :[[-END]] Among them, w n is the sensitivity weight of the topic category corresponding to the public opinion event cluster, and d i is the topic sensitivity score of the i-th public opinion event.

17. The social media comprehensive public opinion analysis system according to claim 1, characterized in that The obtaining of the exposure index score, interaction index score, and propagation intensity index score of each public opinion event cluster includes: The exposure index score f of each public opinion event cluster is obtained by the following formula E :[[]]END]] Among them, T is the corresponding time period, and w t is the push weight of the platform for the public opinion event cluster, and its value range is [0, 1]; V t represents the page view volume of the platform at the t-th moment, and λ is the time decay coefficient, and its value range is [0, 1]; The interactivity index score f of each public opinion event cluster is obtained by the following formula I :[[]]END]] Among them, l, m, and h respectively represent the number of likes, comments, and forwards of the public opinion event cluster, a, b, and c respectively represent the weights of the number of likes, comments, and forwards, and p1, p2, and p3 respectively represent the adjustment factors of the number of likes, comments, and forwards; The propagation intensity index score f of each public opinion event cluster is obtained by the following formula T :[[-END]] Among them, L represents the propagation level, and N L is the number of forwarding times at the L-th level, and v L is the propagation speed at the L-th level, and L max is the maximum propagation level.

18. The social media comprehensive public opinion analysis system according to claim 1, characterized in that, The obtaining of the public opinion evaluation result of each public opinion event cluster according to the comprehensive public opinion analysis score of each public opinion event cluster includes: Obtain the comprehensive public opinion analysis score threshold through the following formula Among them, μ is the upper bound center of the comprehensive public opinion analysis score of the theoretical normal public opinion event, ξ is the strict exponent, and ∈ is the variance of the Gaussian distribution of the upper bound center of the comprehensive public opinion analysis score of the theoretical normal public opinion event; When the comprehensive public opinion analysis score of the public opinion event cluster is greater than the comprehensive public opinion analysis score threshold, the public opinion evaluation result of the public opinion event cluster is a high-risk public opinion event; when the comprehensive public opinion analysis score of the public opinion event cluster is not greater than the comprehensive public opinion analysis score threshold, the public opinion evaluation result of the public opinion event cluster is a low-risk public opinion event; When the public opinion evaluation result of the public opinion event cluster is a high-risk public opinion event, generate a warning and measure prompt.

19. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the social media comprehensive public opinion analysis method according to any one of claims 1 to 9.

20. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the social media comprehensive public opinion analysis method according to any one of claims 1 to 9.

Images