Film and television work network propagation public opinion analysis method

Abstract

The invention relates to the technical field of network public opinion monitoring, and particularly discloses a film and television work network propagation public opinion analysis method, which comprises the steps of multi-modal data acquisition, theme tracking, sentiment analysis and public opinion situation awareness. According to the scheme, a mode of combining LDA topic modeling and hierarchical clustering is adopted, a tree hierarchical structure and a vocabulary evolution relation of a topic are constructed, potential topics of film and television works in network propagation public opinions are learned from word frequency co-occurrence, topic popularity is quantized and mapped to a timeline through dynamic popularity calculation, and therefore the topic popularity is obtained. The interpretability and operability of the result are ensured; the method comprises the following steps: quantifying themes of film and television works by synthesizing popularity fluctuation and emotion dispute, automatically screening key nodes in network propagation public opinions based on a theme tree, quantifying emotion intensity by fusing confidence, and aggregating the emotion intensity to a macroscopic theme and a time window, so as to construct a macroscopic and microscopic integrated dynamic public opinion map; and the sentiment analysis accuracy is greatly improved.

Description

Technical Field

[0001] This invention relates to the field of online public opinion monitoring technology, specifically to a method for analyzing public opinion regarding the online dissemination of film and television works. Background Technology

[0002] With the development of short video platforms, social media platforms, and forum communities, the online dissemination of film and television works exhibits highly real-time, fragmented, and multimodal characteristics. Users generate a large amount of cross-platform, multimodal public opinion data through text comments, bullet comments, images, short video derivative works, emojis, and forwarding interactions. In the online dissemination of film and television works, a large number of meaningless words or high-frequency words such as titles often obscure key information, resulting in low-quality lexical features extracted for public opinion analysis. Furthermore, typical topic models produce flat, unrelated lists of topics with a flat structure, lacking in-depth insight, and resulting in poor model analysis performance. Traditional sentiment analysis methods yield isolated conclusions that cannot be effectively correlated with macro-level topic popularity or hierarchical structure, making it difficult to form a global and dynamic understanding. Directly using general MLLM (Multimodal Large Language Model) to analyze film and television public opinion suffers from insufficient model comprehension capabilities and unreliable analysis results. Summary of the Invention

[0003] To address the above issues and overcome the shortcomings of existing technologies, this invention provides a method for analyzing public opinion regarding the online dissemination of film and television works. This method addresses the problem that in the online dissemination of general film and television works, a large number of meaningless words or high-frequency words such as titles often obscure key information, resulting in low-quality lexical features extracted for public opinion analysis. Furthermore, typical topic models produce flat, unrelated lists of topics with a flat structure, lacking depth and resulting in poor model analysis performance. This solution combines LDA (Latent Dirichlet Allocation) topic modeling with hierarchical clustering to construct a tree-like hierarchical structure of topics and lexical evolution relationships. It learns the potential topics of film and television works in online public opinion dissemination from word frequency co-occurrence and uses dynamic heat calculation... This approach quantifies topic popularity and maps it to a timeline, ensuring the interpretability and operability of the results. Traditional sentiment analysis methods often yield isolated conclusions that cannot be effectively correlated with macro-level topic popularity or hierarchical structure, making it difficult to form a holistic and dynamic understanding. Directly using general MLLM analysis of film and television public opinion suffers from insufficient model comprehension and unreliable results. This solution quantifies the themes of film and television works by integrating popularity fluctuations and sentiment controversy. Based on a topic tree, it automatically filters key nodes in online public opinion dissemination, integrates confidence metrics to quantify sentiment intensity, and aggregates sentiment intensity onto macro-level themes and time windows, constructing a dynamic public opinion map that integrates macro and micro perspectives, significantly improving the accuracy of sentiment analysis.

[0004] The technical solution adopted in this invention is as follows: This invention provides a method for analyzing public opinion on the online dissemination of film and television works, which includes the following steps:

[0005] Step S1: Multimodal data acquisition, obtaining relevant data of the target film and television works from online social platforms, constructing documents and an initial vocabulary, and forming structured data records for storage in a non-relational database;

[0006] Step S2: Topic tracking. Based on the initial vocabulary, LDA topic modeling is used to assign a topic to each vocabulary. Hierarchical clustering is used to construct a complete topic tree for the topics of the target film and television works, and the dynamic popularity of each topic is calculated.

[0007] Step S3: Sentiment analysis. Based on the node importance and sentiment uncertainty analysis method, key nodes are selected from the topic tree. Data units are constructed for key nodes in combination with topic tags. A multimodal large language model is introduced. The data units are used to fine-tune the multimodal large language model to calculate the sentiment intensity of the topic.

[0008] Step S4: Public opinion situation perception, which integrates the dynamic popularity and emotional intensity of the topic to construct a comprehensive public opinion index for early warning of public opinion risks and generates periodic reports on the online dissemination of film and television works.

[0009] Further, in step S2, the topic tracking specifically includes the following steps:

[0010] Step S21: Vocabulary filtering. Generate a document-term matrix based on the initial vocabulary list. The element value of the document-term matrix is ​​the TF-IDF (term frequency-inverse text frequency index) value of each word: TF represents the number of times a word appears in a document, and IDF represents the distinctiveness of a word in the entire corpus. Filter out the words with the highest TF-IDF values ​​to form the final vocabulary list.

[0011] Step S22: Document-term frequency matrix construction. Based on the final vocabulary, all documents are rescanned. A term frequency vector with the same dimension as the final vocabulary is created for each document. The element value of the term frequency vector is the number of times the word appears in the document. The term frequency vectors of all documents are combined to form a document-term frequency matrix.

[0012] Step S23: LDA topic modeling, preset the total number of topics in the corpus, use Gibbs sampling inference based on the document-term frequency matrix to assign topics to each word in the corpus, obtain the topic assignment results, and calculate the estimated values ​​of document-topic distribution and topic-vocabulary distribution, wherein the document-topic distribution represents the topic composition of the document, and the topic-vocabulary distribution represents the vocabulary composition of the topic;

[0013] Step S24: Hierarchical optimization, calculate the similarity between topics, iteratively merge the two topics with the highest similarity to form a parent node, normalize the topic vector of the parent node to the average value of the topic-vocabulary distribution of the two child nodes, continue to merge until all topics converge into a root node, and finally generate a complete topic tree;

[0014] Step S25: Dynamic popularity calculation, dividing the time window, calculating the popularity of each topic in each time window. The formula used to determine the popularity is as follows: ;

[0015] In the formula, Indicate the theme, Represents a document. Indicates a time window. Indicates the topic In the time window The popularity Indicates time window The total number of documents within, Indicates time window The collection of documents within, Document Belongs to the topic The probability of.

[0016] Furthermore, in step S3, the sentiment analysis specifically includes the following steps:

[0017] Step S31: Key node filtering, the specific steps are as follows:

[0018] Step S311: Node importance calculation. Calculate the dynamic popularity change entropy of each node in the topic tree. The higher the dynamic popularity change entropy, the more drastic the fluctuation of the topic's dynamic popularity over time. The formula used is as follows: ;

[0019] In the formula, Indicates the topic The dynamic heat change entropy, Indicate topic Total heat across all time windows Indicates the total number of time windows;

[0020] Step S312: Sentiment uncertainty calculation. Using a text sentiment analysis method based on a sentiment dictionary, the sentiment uncertainty of each node in the topic tree is calculated. The greater the sentiment uncertainty, the greater the sentiment divergence under that topic. The formula used is as follows: ;

[0021] In the formula, Indicate topic Emotional uncertainty, Indicates belonging to the topic The total number of documents, Document Emotional score, Indicate topic The average sentiment of all documents;

[0022] Step S313: Define the comprehensive scoring function and calculate the comprehensive score for each node. The formula used is as follows: ;

[0023] In the formula, Indicate topic Overall score A parameter representing the balance between node importance and emotional uncertainty. Represents the normalization function;

[0024] Step S314: Locate the key node, traverse the topic tree, and select the leaf node with the highest overall score as the key node;

[0025] Step S315: Data unit construction. Extract data records of all key nodes from the non-relational database and construct data units by combining them with topic tags.

[0026] Step S32: Model analysis, introduce open source MLLM, insert a two-layer MLP (multilayer perceptron) as a trainable adapter, freeze the backbone parameters of the MLLM model, generate instruction data containing topic information based on data units, fine-tune the MLLM using instruction data, and calculate the loss function using a weighted combination of cross-entropy loss and text generation loss.

[0027] Step S33: Sentiment intensity calculation. Output sentiment label and confidence score for each data unit, and calculate sentiment intensity based on the confidence score using the following formula: ;

[0028] In the formula, Indicates the first The emotional intensity of each data unit, This represents the confidence score. Indicates emotional tags, This represents a mapping function that converts sentiment labels into numerical values;

[0029] Step S34: Topic-level sentiment aggregation, calculate the average sentiment intensity of all data units within the time window for the topic, using the following formula: ;

[0030] In the formula, This represents the average emotional intensity of all data units. Indicates time window Belongs to the topic The total number of data units.

[0031] Furthermore, in step S4, the public opinion situation perception specifically involves: combining the dynamic popularity of the topic and the intensity of negative sentiment to calculate a comprehensive risk index, and setting a threshold to issue an early warning for public opinion risks. The formula used is as follows: ;

[0032] In the formula, This represents the overall risk index. This represents the maximum value function.

[0033] The beneficial effects achieved by the present invention using the above solution are as follows:

[0034] (1) In response to the problem that in the online dissemination of general film and television works, there are a large number of meaningless words or high-frequency words such as titles that drown out key information, resulting in low quality of lexical features extracted in public opinion analysis, and that general topic models produce flat, unrelated topic lists with flat topic structures, lack of in-depth insights, and poor model analysis effects, this solution adopts a combination of LDA topic modeling and hierarchical clustering to construct a tree-like hierarchical structure of topics and lexical evolution relationships. It learns the potential topics of film and television works in online public opinion from word frequency co-occurrence. Through dynamic heat calculation, the heat of topics is quantified and mapped to the timeline to ensure the interpretability and operability of the results.

[0035] (2) Traditional sentiment analysis methods have isolated conclusions that cannot be effectively linked to the popularity of macro topics and the hierarchical structure of topics, making it difficult to form a global and dynamic understanding. Directly using the general MLLM to analyze film and television public opinion has problems such as insufficient model understanding ability and unreliable analysis results. This solution quantifies the theme of film and television works by comprehensively considering the fluctuation of popularity and the controversy of sentiment. Based on the topic tree, it automatically selects key nodes in the online public opinion, integrates confidence quantification of sentiment intensity, and aggregates sentiment intensity to macro topics and time windows to construct a dynamic public opinion map that integrates macro and micro perspectives, which greatly improves the accuracy of sentiment analysis. Attached Figure Description

[0036] Figure 1 This is a flowchart illustrating a method for analyzing public opinion on the online dissemination of film and television works proposed in this invention.

[0037] Figure 2 This is a flowchart illustrating step S2;

[0038] Figure 3 This is a flowchart illustrating step S3.

[0039] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. Detailed Implementation

[0040] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0041] Example 1, see Figure 1 This invention provides a method for analyzing public opinion on the online dissemination of film and television works, which includes the following steps:

[0042] Step S1: Multimodal data acquisition, obtaining relevant data of the target film and television works from online social platforms, constructing documents and an initial vocabulary, and forming structured data records for storage in a non-relational database;

[0043] Step S2: Topic tracking. Based on the initial vocabulary, LDA topic modeling is used to assign a topic to each vocabulary. Hierarchical clustering is used to construct a complete topic tree for the topics of the target film and television works, and the dynamic popularity of each topic is calculated.

[0044] Step S3: Sentiment analysis. Based on the node importance and sentiment uncertainty analysis method, key nodes are selected from the topic tree. Data units are constructed for key nodes in combination with topic tags. A multimodal large language model is introduced. The data units are used to fine-tune the multimodal large language model to calculate the sentiment intensity of the topic.

[0045] Step S4: Public opinion situation perception, which integrates the dynamic popularity and emotional intensity of the topic to construct a comprehensive public opinion index for early warning of public opinion risks and generates periodic reports on the online dissemination of film and television works.

[0046] Example 2, see Figure 1 This embodiment is based on the above embodiment. In step S1, multimodal data acquisition specifically includes the following steps:

[0047] Step S11: Data source location and crawling. Using API calls and web crawlers, relevant data of the target film and television works are collected from online social platforms and tagged with metadata. The data is categorized into text data, image data, and video data based on its type. Specific data content includes:

[0048] Text data: comments, bullet comments, posts, long reviews,

[0049] Image data: stills, posters, emojis,

[0050] Video data: derivative works, edited videos, and commentary videos;

[0051] Step S12: Text data cleaning. Use regular expressions to remove URLs, extra spaces, and punctuation marks from the text data, and standardize the internet slang. Build documents, use the collection of all documents as a corpus, use the Jieba word segmentation tool to divide the document content into independent words, load the stop vocabulary list to remove meaningless words, and build an initial vocabulary list.

[0052] Step S13: Image data preprocessing, extracting keyframes from video data at fixed intervals, and scaling the extracted keyframes and image data to a uniform size;

[0053] Step S14: Data integration and storage. The cleaned text data, processed image data, video frame paths and metadata are associated to form structured data records, which are then stored in a non-relational database.

[0054] Example 3, see Figure 1 and Figure 2 This embodiment is based on the above embodiment. In step S2, topic tracking specifically includes the following steps:

[0055] Step S21: Vocabulary selection. Generate a document-term matrix based on the initial vocabulary. The element value of the document-term matrix is ​​the TF-IDF value of each word. Select the V words with the highest TF-IDF values ​​to form the final vocabulary. The formula used is as follows: ;

[0056] In the formula, The term frequency-inverse text frequency index represents the frequency of words. Representing words, Represents a document. Represents a corpus, Words In the document Normalized word frequency in Words In the corpus Inverse text frequency in;

[0057] Step S22: Document-term frequency matrix construction. Based on the final vocabulary, all documents are rescanned, and a V-dimensional term frequency vector is created for each document. The element value of the term frequency vector is the number of times the word appears in the document. The term frequency vectors of all documents are combined to form a document-term frequency matrix.

[0058] Step S23: LDA topic modeling, the specific steps are as follows:

[0059] Step S231: Parameter initialization, the total number of topics in the preset corpus is... For each word in the corpus, a topic label is randomly assigned, and the Dirichlet prior hyperparameters are initialized. and ;

[0060] Step S232: Gibbs sampling inference. Traverse each word in the corpus, and based on the document-term frequency matrix, use Gibbs sampling to reassign topics to the words according to the conditional probability distribution until the topic assignment changes stabilize, obtaining the topic assignment result. The formula used is as follows: ;

[0061] In the formula, Indicate the theme, This represents the global index of the vocabulary to be assigned in the database. This indicates a new topic for vocabulary to be assigned. Indicates the topic Assigned to the One word, This indicates the topic allocation of all words except those to be assigned. Document Except for the words to be assigned, the rest are assigned to the topics. The number of words, This indicates the index of the word to be assigned in the final vocabulary. This indicates that, excluding the words to be assigned, the entire corpus contains all words. Assigned to topic Total number of times This means that all words in the entire corpus, except for those to be assigned, are assigned to topics. Quantity, , , This represents the Dirichlet prior hyperparameter. Indicates the direct proportion sign. This represents the conditional probability of Gibbs sampling.

[0062] Step S233: Parameter estimation, calculate document-topic distribution based on topic assignment results. Thematic-vocabulary distribution The estimated value;

[0063] Step S24: Hierarchical optimization, calculate the cosine similarity between topics, iteratively merge the two topics with the highest cosine similarity to form a parent node, normalize the topic vector of the parent node to the average value of the topic-vocabulary distribution of the two child nodes, continue to merge until all topics converge into a root node, and finally generate a complete topic tree;

[0064] Step S25: Dynamic popularity calculation, dividing the time window, calculating the popularity of each topic in each time window. The formula used to determine the popularity is as follows: ;

[0065] In the formula, Indicates a time window. Indicate topic In the time window The popularity Indicates time window The total number of documents within, Indicates time window The collection of documents within, Document Belongs to the topic The probability of.

[0066] By performing the aforementioned operations, this solution addresses the problem that in the online dissemination of general film and television works, a large number of meaningless words or high-frequency words such as work titles often overwhelm key information, resulting in low-quality lexical features extracted in public opinion analysis. Furthermore, general topic models produce flat, unrelated lists of topics with a flat topic structure, lacking in-depth insight and poor model analysis performance. This solution adopts a combination of LDA topic modeling and hierarchical clustering to construct a tree-like hierarchical structure of topics and lexical evolution relationships. It learns the potential topics of film and television works in online public opinion dissemination from word frequency co-occurrence. Through dynamic heat calculation, the heat of topics is quantified and mapped to a timeline, ensuring the interpretability and operability of the results.

[0067] Example 4, see Figure 1 and Figure 3 This embodiment is based on the above embodiment. In step S3, the sentiment analysis specifically includes the following steps:

[0068] Step S31: Key node filtering, the specific steps are as follows:

[0069] Step S311: Node importance calculation. Calculate the dynamic popularity change entropy of each node in the topic tree. The higher the dynamic popularity change entropy, the more drastic the fluctuation of the topic's dynamic popularity over time. The formula used is as follows: ;

[0070] In the formula, Indicate topic The dynamic heat change entropy, Indicate topic Total heat across all time windows Indicates the total number of time windows;

[0071] Step S312: Sentiment uncertainty calculation. Using a text sentiment analysis method based on a sentiment dictionary, the sentiment uncertainty of each node in the topic tree is calculated. The greater the sentiment uncertainty, the greater the sentiment divergence under that topic. The formula used is as follows: ;

[0072] In the formula, Indicate topic Emotional uncertainty, Indicates belonging to the topic The total number of documents, Document Emotional score, Indicate topic The average sentiment of all documents;

[0073] Step S313: Define the comprehensive scoring function and calculate the comprehensive score for each node. The formula used is as follows: ;

[0074] In the formula, Indicate topic Overall score A parameter representing the balance between node importance and emotional uncertainty. Represents the normalization function;

[0075] Step S314: Locate key nodes, traverse the topic tree, and select the top N leaf nodes with the highest comprehensive score as key nodes;

[0076] Step S315: Data unit construction. Extract data records of all key nodes from the non-relational database and construct data units by combining them with topic tags.

[0077] Step S32: Model analysis. The Qwen2-VL model is introduced, and a two-layer MLP is inserted as a trainable adapter. Simultaneously, the backbone parameters of the Qwen2-VL model are frozen. Instruction data containing topic information is generated based on data units. The MLLM is fine-tuned using this instruction data to enable the model to efficiently learn domain-specific sentiment features. The loss function is calculated using a weighted combination of cross-entropy loss and text generation loss. The two-layer MLP structure is as follows: ; ;

[0078] In the formula, This represents the original image feature vector extracted by the visual encoder of the Qwen2-VL model. , These represent the weights and biases of the first-layer MLP, respectively. Represents the activation function of the Gaussian error linear unit. This represents the feature vector after the first layer of transformation. , These represent the weights and biases of the second-layer MLP, respectively. This represents the final visual feature vector fused with the text features;

[0079] Step S33: Sentiment intensity calculation. Output sentiment label and confidence score for each data unit, and calculate sentiment intensity based on the confidence score using the following formula: ;

[0080] In the formula, Indicates the index of the data unit. Indicates the first The emotional intensity of each data unit, This represents the confidence score. Indicates emotional tags, This represents a mapping function that converts sentiment labels into numerical values;

[0081] Step S34: Topic-level sentiment aggregation, calculate the average sentiment intensity of all data units within the time window for the topic, using the following formula: ;

[0082] In the formula, This represents the average emotional intensity of all data units. Indicates time window Belongs to the topic The total number of data units.

[0083] By performing the aforementioned operations, we can address the shortcomings of traditional sentiment analysis methods. Traditional sentiment analysis methods often yield isolated conclusions that cannot be effectively correlated with macro-level topic popularity or hierarchical structure, making it difficult to form a holistic and dynamic understanding. Directly using general MLLM to analyze film and television public opinion suffers from insufficient model comprehension and unreliable analysis results. This solution quantifies the themes of film and television works by comprehensively considering popularity fluctuations and sentiment controversies. Based on a topic tree, it automatically filters key nodes in online public opinion dissemination, integrates confidence metrics to quantify sentiment intensity, and aggregates sentiment intensity onto macro-level themes and time windows. This constructs a dynamic public opinion map that integrates macro and micro perspectives, significantly improving the accuracy of sentiment analysis.

[0084] Example 5, see Figure 1 This embodiment is based on the above embodiment. In step S4, the public opinion situation perception specifically includes the following steps:

[0085] Step S41: Construct a comprehensive public opinion index, combining the dynamic popularity of the topic and the intensity of negative sentiment to calculate a comprehensive risk index, and set a threshold to issue early warnings for public opinion risks. The formula used is as follows: ;

[0086] In the formula, This represents the overall risk index. Represents the maximum value function;

[0087] Step S42: Generate a popularity trend chart, constructing a popularity trend curve of the topic over time;

[0088] Step S43: Multimodal display, generating periodic reports on online public opinion regarding film and television works, including: topic tree structure, topic popularity trend evolution, topic popularity trend analysis based on MLLM inference, and specific action suggestions.

[0089] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0090] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

[0091] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A method for analyzing public opinion on the online dissemination of film and television works, characterized in that: The method includes the following steps: Step S1: Multimodal data acquisition, obtaining relevant data of the target film and television works from online social platforms, constructing documents and an initial vocabulary, and forming structured data records for storage in a non-relational database; Step S2: Topic tracking. Based on the initial vocabulary, LDA topic modeling is used to assign a topic to each vocabulary. Hierarchical clustering is used to construct a complete topic tree for the topics of the target film and television works, and the dynamic popularity of each topic is calculated. Step S3: Sentiment analysis. Based on the node importance and sentiment uncertainty analysis method, key nodes are selected from the topic tree. Data units are constructed for key nodes in combination with topic tags. A multimodal large language model is introduced. The data units are used to fine-tune the multimodal large language model to calculate the sentiment intensity of the topic. Step S4: Public opinion situation perception, which integrates the dynamic popularity and emotional intensity of the topic to construct a comprehensive public opinion index for early warning of public opinion risks and generates periodic reports on the online dissemination of film and television works.

2. The method for analyzing public opinion on the online dissemination of film and television works according to claim 1, characterized in that: In step S2, the topic tracking includes the following steps: Step S21: Vocabulary filtering. Generate a document-term matrix based on the initial vocabulary list. The element value of the document-term matrix is ​​the TF-IDF value of each word. Filter out the words with the highest TF-IDF values ​​to form the final vocabulary list. Step S22: Document-term frequency matrix construction. Based on the final vocabulary, all documents are rescanned. A term frequency vector with the same dimension as the final vocabulary is created for each document. The element value of the term frequency vector is the number of times the word appears in the document. The term frequency vectors of all documents are combined to form a document-term frequency matrix. Step S23: LDA topic modeling, preset the total number of topics in the corpus, use Gibbs sampling inference based on the document-term frequency matrix to assign topics to each word in the corpus, obtain the topic assignment results, and calculate the estimated values ​​of document-topic distribution and topic-word distribution; Step S24: Hierarchical optimization, calculate the similarity between topics, iteratively merge the two topics with the highest similarity to form a parent node, normalize the topic vector of the parent node to the average value of the topic-vocabulary distribution of the two child nodes, continue to merge until all topics converge into a root node, and finally generate a complete topic tree; Step S25: Dynamic popularity calculation, dividing the time window, calculating the popularity of each topic in each time window. The formula used to determine the popularity is as follows: ； In the formula, Indicate the theme, Represents a document. Indicates a time window. Indicate topic In the time window The popularity Indicates time window The total number of documents within, Indicates time window The collection of documents within, Document Belongs to the topic The probability of.

3. The method for analyzing public opinion on the online dissemination of film and television works according to claim 1, characterized in that: In step S3, the sentiment analysis specifically includes the following steps: Step S31: Key node selection. Based on dynamic popularity and topic tree structure, calculate the dynamic popularity change entropy and sentiment uncertainty of all topic nodes, define a comprehensive scoring function, traverse the topic tree, select the leaf node with the highest comprehensive score as the key node, and construct data units by combining topic tags. Step S32: Model analysis, introduce open source MLLM, insert a two-layer MLP as a trainable adapter, freeze the backbone parameters of the MLLM model, generate instruction data containing topic information based on data units, fine-tune the MLLM using instruction data, and calculate the loss function using a weighted combination of cross-entropy loss and text generation loss. Step S33: Calculate sentiment intensity. Output sentiment label and confidence score for each data unit, and calculate sentiment intensity based on confidence score. Step S34: Topic-level sentiment aggregation, calculate the average sentiment intensity of all data units within the time window for the topic, using the following formula: ； In the formula, This represents the average emotional intensity of all data units. Indicates time window Belongs to the topic The total number of data units, Indicates the index of the data unit. Indicates the first The emotional intensity of each data unit.

Images