Intelligent digital media content sentiment analysis method and system

Abstract

The application discloses a kind of intelligent digital media content sentiment analysis method and system, content includes multi-source data acquisition, multi-level feature extraction, cross-modal fusion sentiment analysis modeling and intelligent analysis.The application relates to the technical field of media content sentiment analysis, specifically refers to a kind of intelligent digital media content sentiment analysis method and system, the scheme is in text passage calculation sentiment embedding offset and dynamically updates multi-center vector, in visual passage through space aggregation convolution activation measures emotional intensity, in audio passage using spectral centroid third central moment intensity is measured;At the same time, the key word intensity index and the modification mode coefficient are combined to realize the enhancement of middle-level semantics, and the global correlation of multi-modal nodes is constructed through the hypergraph structure, the context gate dynamic updating hyperedge weight is innovatively introduced, the graph attention aggregation feature is used, the time sequence self-attention and graph convolution are combined for spatiotemporal collaborative coding, which improves the accuracy, robustness and real-time adaptive ability of sentiment recognition.

Description

Technical Field

[0001] This invention relates to the field of media content analysis technology, specifically to an intelligent digital media content sentiment analysis method and system. Background Technology

[0002] With the rapid development of social media and digital platforms, user-generated text, image, and audio content has exploded. This multimodal data contains rich emotional information, but due to the diverse modes of expression and the complex cross-modal correlation of emotions, traditional sentiment analysis solutions rely heavily on manual design and single-modal signals for single feature extraction, making it difficult to take into account multi-level emotional expressions. Traditional sentiment analysis solutions based on fixed weights and simple splicing cross-modal fusion methods cannot dynamically perceive contextual relationships, are difficult to capture temporal evolution, and are easily affected by noise. Summary of the Invention

[0003] To address the aforementioned issues and overcome the shortcomings of existing technologies, this invention provides an intelligent digital media content sentiment analysis method and system. Addressing the problems of traditional sentiment analysis schemes relying heavily on manual design and single-modal signals for single-feature extraction, and struggling to accommodate multi-layered emotional expression, this solution calculates sentiment embedding offsets and dynamically updates multi-center sentiment vectors in text analysis; employs spatial aggregation convolutional activation intensity measurement in the visual pathway; introduces the third-order central moment of the spectral centroid to measure emotional intensity in the audio pathway; and simultaneously integrates sentiment keyword intensity indices and modification pattern coefficients at the semantic level, achieving a unified fusion of low-level perception and mid-level semantic enhancement of multimodal content. Furthermore, addressing the issues of traditional sentiment analysis schemes based on fixed weights and simple concatenation for cross-modal fusion, which cannot dynamically perceive contextual relationships, struggle to capture temporal evolution, and are susceptible to noise interference, this solution constructs a hypergraph structure to globally model the relationships between multimodal nodes, introduces a context-driven gating mechanism to dynamically update hyperedge weights, uses multi-head graph attention to aggregate features with adaptive weights, and combines temporal self-attention and graph convolution for spatiotemporal co-coding, achieving accurate perception and evolution capture of emotional context.

[0004] The technical solution adopted in this invention is as follows: A method for intelligent digital media content sentiment analysis, the method comprising the following steps:

[0005] Step S1: Multi-source data acquisition;

[0006] Step S2: Multi-level feature extraction;

[0007] Step S3: Cross-modal fusion sentiment analysis modeling;

[0008] Step S4: Intelligent Analysis.

[0009] Furthermore, in step S1, the data acquisition specifically involves acquiring text data, image data, audio data, and time data from digital media.

[0010] Furthermore, in step S2, the multi-level feature extraction specifically includes the following steps:

[0011] Step S21: Extract low-level features, specifically including the following steps:

[0012] Step S211: Calculate the text sentiment embedding offset. First, extract the text embedding, then calculate and update the sentiment center, and finally calculate the Euclidean norm to obtain the sentiment deviation. This measures the distance between the text and the sentiment center vector in the high-dimensional semantic space to capture the overall sentiment deviation. It also supports domain-adaptive multi-center expansion and can dynamically adjust the sentiment center according to real-time hotspots or subcategories.

[0013] Step S212: Calculate the intensity of visual emotional expression. First, input the image to be analyzed into the pre-trained CNN to extract the convolutional tensors, then calculate the aggregate activation, accumulate it according to the spatial dimension, and finally multiply it by the learned weights and sum between layers to obtain the intensity of visual emotional expression.

[0014] Step S213: Extract the emotional intensity of the audio and use the third central moment of the spectral centroid to measure the skewness and kurtosis of the audio, reflecting the intensity of the emotion;

[0015] Step S22: Extract mid-layer features, specifically including the following steps:

[0016] Step S221: Calculate the sentiment keyword strength index, multiply the sentiment word frequency by the dictionary score and normalize it to evaluate the overall strength of the sentiment keywords in the text;

[0017] Step S222: Calculate the sentiment modification pattern coefficient, quantify the impact of the modification patterns of negation, reinforcement and irony on sentiment words, and perform quantitative modeling through rate of change and distance decay.

[0018] Furthermore, in step S3, the cross-modal fusion sentiment analysis modeling specifically includes the following steps:

[0019] Step S31: Construct hypergraph nodes and hyperedges. Treat the multimodal features of text, image, and audio as node sets respectively. Construct hyperedge sets based on semantic association and prior knowledge, and output the initial hypergraph.

[0020] Step S32: Hyperedge weight update. Introduce a context-driven gating function to dynamically update the hyperedge weights, reflecting the multimodal association strength in the current context.

[0021] Step S33: Depth graph attention fusion, aggregating multimodal node information on the hypergraph using multi-head graph attention with adaptive weights;

[0022] Step S34: Temporal structure co-encoding, combining temporal self-attention and graph convolution to perform spatiotemporal co-encoding on the fused features to capture the emotion evolution pattern;

[0023] Step S35: Establish a sentiment analysis model. The model maps the spatiotemporal fusion representation of each node to the sentiment polarity and intensity space to generate sentiment analysis results.

[0024] Furthermore, in step S4, the intelligent analysis specifically involves real-time collection of digital media content, including text data, image data, audio data, and time data of the digital media. First, multi-level feature extraction is performed on the data, and then the extracted feature data is input into the sentiment analysis model. The model performs sentiment analysis on the data and outputs the sentiment analysis results.

[0025] This invention provides an intelligent digital media content sentiment analysis system, comprising a multi-source data acquisition module, a multi-level feature extraction module, a cross-modal fusion sentiment analysis modeling module, and an intelligent analysis module;

[0026] The multi-source data acquisition module collects text data, image data, audio data, and time data from digital media, and sends the data to the multi-level feature extraction module.

[0027] The multi-level feature extraction module receives data sent by the multi-source data acquisition module, calculates the emotional embedding deviation from the text, activates the emotional expression intensity of the image based on the convolutional layer, and uses the third central moment of the centroid of the audio spectrum to measure the emotional intensity of the sound. Then, it counts the frequency of emotional keywords at the text level and normalizes them to obtain the keyword intensity index. Then, it obtains the modification mode coefficient by modeling modification methods and sends the data to the cross-modal fusion emotional analysis modeling module.

[0028] The cross-modal fusion sentiment analysis modeling module receives data sent by the multi-level feature extraction module, constructs multi-modal features of text, image and audio into hypergraph nodes and hyperedges, dynamically updates hyperedge weights using context-driven gating functions, fuses node information on the hypergraph based on multi-head graph attention, and then jointly encodes spatiotemporal features through temporal self-attention and graph convolution, finally mapping and generating the sentiment polarity and intensity results of each node, and sends the data to the intelligent analysis module.

[0029] The intelligent analysis module receives data sent by the cross-modal fusion sentiment analysis modeling module, collects digital media content in real time, first performs multi-level feature extraction on the data, and then inputs the extracted feature data into the sentiment analysis model. The model performs sentiment analysis on the data and outputs the sentiment analysis results.

[0030] The beneficial effects achieved by the present invention using the above solution are as follows:

[0031] (1) In view of the problem that traditional sentiment analysis schemes rely on manual design and single modal signals for single feature extraction and are difficult to take into account multiple levels of sentiment expression, this scheme calculates the sentiment embedding offset and dynamically updates the multi-center sentiment vector in text analysis, adopts spatial aggregation convolution activation intensity measurement in the visual path, introduces the third central moment of the spectral centroid to measure the intensity of sentiment in the audio path, and simultaneously integrates the sentiment keyword intensity index and the modification mode coefficient model at the semantic level. This scheme realizes the unified integration of the bottom-level perception and the middle-level semantic enhancement of multimodal content, which significantly improves the accuracy, robustness and domain adaptability of sentiment recognition, while supporting real-time hotspot dynamic adjustment and fine-grained processing of subcategories.

[0032] (2) In view of the problems that traditional sentiment analysis schemes based on fixed weights and simple splicing cross-modal fusion methods cannot dynamically perceive contextual associations, are difficult to capture temporal evolution, and are easily affected by noise interference, this scheme constructs a hypergraph structure to perform global association modeling of multimodal nodes, introduces a context-driven gating mechanism to dynamically update hyperedge weights, uses multi-head graph attention to aggregate features according to adaptive weights, and combines temporal self-attention and graph convolution to perform spatiotemporal co-coding, thereby achieving accurate perception and evolution capture of sentiment context, which greatly improves the accuracy, robustness and real-time early warning capability of sentiment analysis. Attached Figure Description

[0033] Figure 1 A schematic diagram illustrating an intelligent digital media content sentiment analysis method provided by the present invention;

[0034] Figure 2 A schematic diagram of an intelligent digital media content sentiment analysis system provided by the present invention;

[0035] Figure 3 This is a schematic diagram of step S2;

[0036] Figure 4 This is a schematic diagram of step S3.

[0037] 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

[0038] 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.

[0039] In the description of this invention, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0040] Example 1, see Figure 1 This invention provides an intelligent digital media content sentiment analysis method, which includes the following steps:

[0041] Step S1: Multi-source data acquisition, collecting text data, image data, audio data, and time data from digital media;

[0042] Step S2: Multi-level feature extraction, calculate the emotional embedding deviation from the text, activate the emotional expression intensity of the image based on the convolutional layer, and use the third central moment of the centroid of the audio spectrum to measure the emotional intensity of the sound. Then, count the frequency of emotional keywords at the text level and normalize them to obtain the keyword intensity index. Finally, obtain the modification mode coefficient by modeling the modification method.

[0043] Step S3: Cross-modal fusion sentiment analysis modeling, constructing hypergraph nodes and hyperedges from multimodal features of text, image and audio, dynamically updating hyperedge weights using context-driven gating functions, fusing node information on the hypergraph based on multi-head graph attention, and then jointly encoding spatiotemporal features through temporal self-attention and graph convolution, finally mapping and generating the sentiment polarity and intensity results of each node;

[0044] Step S4: Intelligent analysis involves real-time collection of digital media content. First, multi-level feature extraction is performed on the data. Then, the extracted feature data is input into the sentiment analysis model. The model performs sentiment analysis on the data and outputs the sentiment analysis results.

[0045] Example 2, see Figure 1 and Figure 3 This embodiment is based on the above embodiment. In step S2, the multi-level feature extraction specifically includes the following steps:

[0046] Step S21: Extract low-level features, specifically including the following steps:

[0047] Step S211: Calculate the text sentiment embedding offset. First, extract the text embedding, then calculate and update the sentiment center, and finally calculate the Euclidean norm to obtain the sentiment deviation. This measures the distance between the text and the sentiment center vector in the high-dimensional semantic space to capture the overall sentiment deviation and supports domain-adaptive multi-center expansion. The sentiment center can be dynamically adjusted according to real-time hotspots or subcategories, as shown below:

[0048] ;

[0049] Where T represents the text being analyzed. This represents the sentiment embedding offset of text T. The deep embedding vector of text T is extracted by the pre-trained model BERT. Represents the emotion center vector, Let K represent the Euclidean norm, K represent the number of sentiment words, and k represent the index of the sentiment word. This represents the embedding vector of the k-th sentiment word in the sentiment dictionary;

[0050] Step S212: Calculate the intensity of visual emotional expression. First, input the image to be analyzed into the pre-trained CNN to extract convolutional tensors. Then, calculate the aggregate activation, accumulate it according to the spatial dimension, and finally multiply it by the learned weights and sum between layers to obtain the intensity of visual emotional expression, as shown below:

[0051] ;

[0052] in, Indicates the input image. The index of the convolutional layer represents the intensity of visual emotional expression in the input image I. Represents the set of convolutional layer indices. This represents the emotional sensitivity weight of the l-th layer, with a value range of [0,1]. This represents the original activation tensor of the l-th layer. Presentation layer in location The activation value after averaging or weighting along the upper channel. , and These represent the height, width, and number of channels of the activation map in layer l, respectively.

[0053] Step S213: Extract the emotional intensity of the audio. Use the third central moment of the spectral centroid to measure the skewness and kurtosis of the audio, reflecting the intensity of the emotion, as shown below:

[0054] ;

[0055] in, Represents the time-domain signal of the input audio. Represents audio The emotional intensity, where N represents the total number of audio frames and n represents the index of the audio frame. Let f represent the spectral centroid of the nth frame, and let f represent the index of the frequency component. This represents the short-time Fourier transform amplitude spectrum of the nth frame. This represents the average value of the centroid of the spectrum. Indicates the third-order central moment;

[0056] Step S22: Extract mid-layer features, specifically including the following steps:

[0057] Step S221: Calculate the sentiment keyword strength index by multiplying the sentiment word frequency by the dictionary score and normalizing the result to evaluate the overall strength of the sentiment keywords in the text, as shown below:

[0058] ;

[0059] in, The sentiment keyword intensity index represents the sentiment index of text T. Indicates the length of the text. This represents the set of sentiment words appearing in text T. Words that express emotion in the text Words expressing emotions The number of times it appears in text T Words expressing emotions The emotional intensity weight, with a value range of [-1, 1];

[0060] Step S222: Calculate the sentiment modification pattern coefficients to quantify the impact of negation, reinforcement, and irony modification patterns on sentiment words. Quantitative modeling is performed using rate of change and distance decay, as shown below:

[0061] ;

[0062] in, Let T represent the sentiment modification pattern coefficient of text T, Q represent the total number of modifiers used to modify sentiment words, and q represent the index of the modifiers used to modify sentiment words. This indicates the word count distance between the modifier q and the sentiment word. This indicates the original word before the word being modified by the word q. Indicator The weight of emotional intensity This indicates the word that follows the word 'q'. Indicator The weight of emotional intensity This represents the modifier type coefficient, with a value range of [-1, 3].

[0063] By performing the above operations, this solution addresses the problem that traditional sentiment analysis methods often rely on manual design and single-modal signals, making it difficult to consider multi-layered emotional expression. It achieves a unified fusion of low-level perception and mid-level semantic enhancement of multimodal content by calculating the sentiment embedding offset and dynamically updating the multi-center sentiment vector in text analysis, employing spatial aggregation convolutional activation intensity measurement in the visual pathway, introducing the third-order central moment of the spectral centroid to measure emotional intensity in the audio pathway, and simultaneously integrating the sentiment keyword intensity index and modification pattern coefficient modeling at the semantic level. This significantly improves the accuracy, robustness, and domain adaptability of sentiment recognition, while also supporting real-time dynamic adjustment of hotspots and refined processing of subcategories.

[0064] Example 3, see Figure 1 and Figure 4 This embodiment is based on the above embodiment. In step S3, the cross-modal fusion sentiment analysis modeling specifically includes the following steps:

[0065] Step S31: Construct hypergraph nodes and hyperedges, treating the multimodal features of text, image, and audio as node sets respectively. Constructing hyperedge sets based on semantic association and prior knowledge Output the initial hypergraph ;

[0066] Step S32: Hyperedge weight update. Introduce a context-driven gating function to dynamically update the hyperedge weights, reflecting the multimodal association strength in the current context, as shown below:

[0067] ;

[0068] in, Indicates the superedge The weight at time t, This represents the sigmoid function. Indicates and The training coefficients for the hyperedge weights range from [0, 0.5]. Let the hyperedge represent the vector. Represents the global context vector;

[0069] Step S33: Depth map attention fusion. Multimodal node information is aggregated on the hypergraph using multi-head graph attention with adaptive weights, as shown below:

[0070] ;

[0071] Where o represents the node index, Represents a node The fusion feature is given by p, where p represents the index of the number of attention heads, and P represents the maximum number of attention heads. This indicates that the output vectors of the P attention heads (p=1 to P) are concatenated by channel. Represents the p-th head mapping matrix. Indicates attention weights. Let represent the input feature of the o-th node, and r represent the index of the hyperedge. This represents an exponential function with the natural constant as its base. This indicates a leakage correction linear unit. Let represent the adjustment coefficient of the p-th head, with a value range of [0,1]. This indicates that the vectors are concatenated along the feature channels. Indicates the superedge Trainable representation vectors in the original feature space Indicates the superedge Trainable representation vectors in the original feature space;

[0072] Step S34: Temporal structure co-encoding, combining temporal self-attention and graph convolution, performs spatiotemporal co-encoding on the fused features to capture the sentiment evolution pattern, as shown below:

[0073] ;

[0074] in, Represents a node The temporal structure co-coding output at time t+1 Represents graph convolution of a hypergraph. This represents a temporal self-attention module. This represents the fused feature sequence over the past D time steps;

[0075] Step S35: Establish a sentiment analysis model. The model maps the spatiotemporal fusion representation of each node to the sentiment polarity and intensity space, generating the sentiment analysis results, as shown below:

[0076] ;

[0077] in, Represents a node The emotional polarity of takes values ​​in the range [-1, 1]. Represents a node The intensity of the emotional truth, with a value range of [0,1]. Represents the hyperbolic tangent function. This represents the final temporal structure co-coding output. and These represent the emotional polarity weight and bias, respectively. and These represent the weight and bias of the emotional authenticity intensity, respectively.

[0078] By performing the above operations, this solution addresses the problems of traditional sentiment analysis schemes, which rely on fixed weights and simple concatenation for cross-modal fusion, such as the inability to dynamically perceive contextual relationships, difficulty in capturing temporal evolution, and susceptibility to noise interference. Instead, it constructs a hypergraph structure to globally model the relationships between multimodal nodes, introduces a context-driven gating mechanism to dynamically update hyperedge weights, uses multi-head graph attention to aggregate features with adaptive weights, and combines temporal self-attention with graph convolution for spatiotemporal co-coding. This achieves accurate perception and evolution capture of sentiment context, significantly improving the accuracy, robustness, and real-time early warning capabilities of sentiment analysis.

[0079] Example 4, see Figure 1 This embodiment is based on the above embodiment. In step S4, the intelligent analysis specifically involves real-time collection of digital media content, including text data, image data, audio data, and time data of digital media. First, multi-level feature extraction is performed on the data. Then, the extracted feature data is input into the sentiment analysis model. The model performs sentiment analysis on the data and outputs the sentiment analysis results.

[0080] Example 5, see Figure 1 and Figure 2 Based on the above embodiments, this embodiment provides an intelligent digital media content sentiment analysis system, which includes a multi-source data acquisition module, a multi-level feature extraction module, a cross-modal fusion sentiment analysis modeling module, and an intelligent analysis module.

[0081] The multi-source data acquisition module collects text data, image data, audio data, and time data from digital media, and sends the data to the multi-level feature extraction module.

[0082] The multi-level feature extraction module receives data sent by the multi-source data acquisition module, calculates the emotional embedding deviation from the text, activates the emotional expression intensity of the image based on the convolutional layer, and uses the third central moment of the centroid of the audio spectrum to measure the emotional intensity of the sound. Then, it counts the frequency of emotional keywords at the text level and normalizes them to obtain the keyword intensity index. Then, it obtains the modification mode coefficient by modeling modification methods and sends the data to the cross-modal fusion emotional analysis modeling module.

[0083] The cross-modal fusion sentiment analysis modeling module receives data sent by the multi-level feature extraction module, constructs multi-modal features of text, image and audio into hypergraph nodes and hyperedges, dynamically updates hyperedge weights using context-driven gating functions, fuses node information on the hypergraph based on multi-head graph attention, and then jointly encodes spatiotemporal features through temporal self-attention and graph convolution, finally mapping and generating the sentiment polarity and intensity results of each node, and sends the data to the intelligent analysis module.

[0084] The intelligent analysis module receives data sent by the cross-modal fusion sentiment analysis modeling module, collects digital media content in real time, first performs multi-level feature extraction on the data, and then inputs the extracted feature data into the sentiment analysis model. The model performs sentiment analysis on the data and outputs the sentiment analysis results.

[0085] 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.

[0086] 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.

[0087] 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 intelligent digital media content sentiment analysis, the method comprising: The method comprises the following steps: Step S1: multi-source data collection, collecting text data, image data, audio data and time data of digital media; Step S2: multi-level feature extraction, calculating sentiment embedding deviation from text, activating sentiment expression intensity based on convolution layer, and measuring the sentiment intensity of sound by using the third order central moment of audio spectrum centroid, then normalizing the frequency of sentiment keywords to obtain the keyword intensity index, and then obtaining the modification mode coefficient by modeling modification; Step S21: extract bottom-level features, specifically including the following steps: Step S211: calculate the text sentiment embedding deviation, first extract the text embedding, then calculate and update the sentiment center, and finally calculate the Euclidean norm to obtain the sentiment deviation, measure the distance between the text in the high-dimensional semantic space and the sentiment center vector to capture the overall sentiment deviation, and support the multi-center expansion of the field adaptation, which can dynamically adjust the sentiment center according to real-time hotspots or fine classification; Step S212: calculate the visual sentiment expression intensity, first input the analysis image into the pre-trained CNN to extract the convolution tensor, then calculate the aggregated activation, accumulate according to the spatial dimension, and finally multiply the learning weight and sum between layers to obtain the visual sentiment expression intensity; Step S213: extract audio sentiment intensity, use the third order central moment of the spectrum centroid to measure the skewness and kurtosis of the audio to reflect the sentiment intensity; Step S22: extract middle-level features, specifically including the following steps: Step S221: calculate the sentiment keyword intensity index, multiply the sentiment keyword frequency by the dictionary score and normalize it to evaluate the overall strength of the text sentiment keywords; Step S222: calculate the sentiment modification mode coefficient, quantify the influence of negative, reinforcement and irony modification modes on sentiment words, and model quantization through change rate and distance decay; Step S3: cross-modal fusion sentiment analysis modeling, construct multi-modal features of text, image and audio as hypergraph nodes and hyperedges, dynamically update hyperedge weights using context-driven gating functions, fuse node information on the hypergraph based on multi-head graph attention, then encode spatio-temporal features through temporal self-attention and graph convolution, and finally map to generate sentiment polarity and intensity results for each node; Step S4: intelligent analysis, real-time collection of digital media content, multi-level feature extraction of the data, then input the extracted feature data into the sentiment analysis model, the model performs sentiment analysis on the data, and outputs the sentiment analysis result. 2.The intelligent digital media content sentiment analysis method of claim 1, wherein: In step S3, the cross-modal fusion sentiment analysis modeling specifically includes the following steps: Step S31: construct hypergraph nodes and hyperedges, regard multi-modal features of text, image and audio as node sets respectively, construct hyperedge sets based on semantic association and prior knowledge, and output the initial hypergraph; Step S32: hyperedge weight update, introduce context-driven gating function to dynamically update hyperedge weights to reflect the multi-modal association strength in the current context; Step S33: deep graph attention fusion, use multi-head graph attention to aggregate multi-modal node information on the hypergraph with adaptive weights; Step S34: Temporal structure collaborative coding, joint temporal self-attention and graph convolution, spatiotemporal collaborative coding of the fusion feature, capturing the emotional evolution pattern; Step S35: Establishing an emotion analysis model, the model maps the spatiotemporal fusion representation of each node to the emotion polarity and intensity space to generate an emotion analysis result. 3.The intelligent digital media content sentiment analysis method of claim 1, wherein: In step S4, the intelligent analysis, specifically real-time collection of digital media content, including text data, image data, audio data and time data of digital media, first multi-level feature extraction is performed on the data, and then the extracted feature data is input into an emotion analysis model, the model performs emotion analysis on the data, and outputs an emotion analysis result.

4. An intelligent digital media content sentiment analysis system for implementing an intelligent digital media content sentiment analysis method according to any one of claims 1-3, characterized in that: It comprises a multi-source data acquisition module, a multi-level feature extraction module, a cross-modal fusion emotion analysis modeling module and an intelligent analysis module.

5. The intelligent digital media content sentiment analysis system of claim 4, wherein: The multi-source data acquisition module acquires text data, image data, audio data and time data of digital media, and sends the data to the multi-level feature extraction module; The multi-level feature extraction module receives the data sent by the multi-source data acquisition module, calculates the sentiment embedding deviation from the text, activates the sentiment expression intensity of the image based on the convolution layer, and measures the sentiment intensity of the sound by using the third order central moment of the audio spectrum centroid, then normalizes the keyword intensity index obtained by statistically analyzing the frequency of the emotional keywords in the text, and obtains the modification mode coefficient by modeling modification method, and sends the data to the cross-modal fusion emotion analysis modeling module; The cross-modal fusion emotion analysis modeling module receives the data sent by the multi-level feature extraction module, constructs the multi-modal features of text, image and audio into hypergraph nodes and hyperedges, dynamically updates the hyperedge weight by using the context-driven gating function, fuses node information on the hypergraph based on multi-head graph attention, then encodes the spatiotemporal features by joint temporal self-attention and graph convolution, finally maps to generate the emotion polarity and intensity result of each node, and sends the data to the intelligent analysis module; The intelligent analysis module receives the data sent by the cross-modal fusion emotion analysis modeling module, collects digital media content in real time, first multi-level feature extraction is performed on the data, then the extracted feature data is input into an emotion analysis model, the model performs emotion analysis on the data, and outputs an emotion analysis result.

Images