An image-text content quality identification method and system based on artificial intelligence

Abstract

The application discloses a kind of based on artificial intelligence's graphic-text content quality identification method and system, it is related to artificial intelligence technical field, including, acquisition graphic-text content data stream, extract the structured feature of image and text, generate multi-dimensional structure vector set;Based on the association between multi-dimensional structure vector establishes graph connection, constructs heterogeneous graph structure and records the propagation path and time sequence of node content, generates dynamic graph topology unit;Dynamic graph topology unit is distributed to edge identification node, and image feature vector and text feature vector are quality score and semantic analysis, generate graphic-text node preliminary screening result;Based on graphic-text node preliminary screening result and dynamic graph topology unit, using graph neural network carries out multi-round feature aggregation, calculates content guide degree, semantic redundancy degree and propagation path offset, generates node score matrix.The application significantly improves the accuracy and intelligent level of graphic-text content quality identification.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, and in particular to a picture-text content quality identification method and system based on artificial intelligence. BACKGROUND

[0002] Under the background of high-speed development of contemporary information dissemination, multi-modal picture-text content integrating images and texts has become the mainstream information carrier in Internet platforms, social media and news pushing. In order to cope with the trend of increasing number of picture-text information, the existing technology usually adopts a static scoring mechanism, a rule base or a machine learning classification model to judge the quality of picture-text content. Generally, the image clarity, text compliance, keyword matching degree and other dimensions are taken as the basis, and the quality label is output by combining the preset index scoring rule or the trained classifier. However, the image and the text are usually treated as independent contents respectively, and the potential semantic linkage relationship between the picture and the text is ignored; at the same time, when dealing with the dynamic dissemination behaviors such as user comments, forwarding and re-creation in the evolution process of picture-text content, it is difficult to accurately track the structural changes of information quality along with the evolution of the dissemination path.

[0003] Especially in the quality evaluation scene of multi-source picture-text data, the problems such as image clarity but semantic redundancy, text coherence but disordered pictures, semantic shift of content in the dissemination process are common, and the existing methods often cannot identify the quality defects caused by the cross of picture and text, multi-round dissemination or semantic heterogeneity of upstream and downstream. The main reason is that the common model lacks the modeling ability of the structural relationship between picture and text and the dynamic changes of dissemination, and the analysis process is mostly limited to the static feature recognition of the content itself, and cannot make comprehensive judgment combined with the evolution path of information, the guiding relationship between nodes and semantic consistency and other factors. SUMMARY

[0004] In view of the above existing problems, the present application is proposed.

[0005] Therefore, the present application provides a picture-text content quality identification method based on artificial intelligence to solve the problem of lack of structured identification of picture-text content in the evolution process of dissemination path.

[0006] To solve the above technical problems, the present application provides the following technical solutions:

[0007] In a first aspect, the present application provides an artificial intelligence-based image-text content quality identification method, which comprises: collecting an image-text content data stream, extracting structured features of images and texts, and generating a multi-dimensional structure vector set; establishing a graph connection based on the association relationship between the multi-dimensional structure vectors, constructing a heterogeneous graph structure and recording the propagation path and time sequence of the node content, and generating a dynamic graph topology unit; distributing the dynamic graph topology unit to an edge identification node, performing quality scoring and semantic analysis on the image feature vector and the text feature vector, and generating an image-text node preliminary screening result; based on the image-text node preliminary screening result and the dynamic graph topology unit, performing multi-round feature aggregation using a graph neural network, calculating the content guidance degree, the semantic redundancy degree and the propagation path offset, and generating a node scoring matrix; based on the node scoring matrix and the image-text node preliminary screening result, performing normalization processing, constructing a clustering input set, and using a hierarchical clustering algorithm for grouping and quality level division, and generating an aggregated quality score set; based on the aggregated quality score set, assigning a quality level label to the image-text node, backtracking the link nodes and abnormal features of the score change, and generating an explanation score summary text.

[0008] As a preferred scheme of the artificial intelligence-based image-text content quality identification method, the generation of the multi-dimensional structure vector set comprises the following specific steps:

[0009] Collecting an image-text content data stream including images, texts, associated publishing time, author information and propagation behavior, and parsing to form an image data set and a text data set;

[0010] Performing size normalization, color space conversion and noise suppression on the image data set, and using a pre-trained feature extraction network to extract edge texture, color distribution and composition contour features to generate an image feature vector;

[0011] Performing cleaning and word segmentation on the text data set, and using a pre-trained semantic embedding model to extract semantic association, context consistency and keyword density to generate a text feature vector;

[0012] Splicing and fusing the image feature vector and the text feature vector to generate a multi-dimensional structure vector set.

[0013] As a preferred scheme of the artificial intelligence-based image-text content quality identification method, the generation of the dynamic graph topology unit comprises the following specific steps:

[0014] Performing similarity calculation and interaction feature modeling on the multi-dimensional structure vector set to generate an association relationship matrix;

[0015] Based on the association relationship matrix, a graph connection structure is constructed, and the nodes of the image feature vector, the nodes of the text feature vector and the nodes of the propagation behavior are respectively marked as types to generate an initial heterogeneous graph;

[0016] Based on the timestamps and source information in the initial heterogeneous graph, the propagation path and occurrence order are identified, a time series mapping table is constructed, and it is fused with the initial heterogeneous graph. The time attribute is assigned to the propagation edge, the content evolution direction and propagation dynamic characteristics are marked, and dynamic graph topology units are generated.

[0017] As a preferred embodiment of the AI-based image and text content quality recognition method of the present invention, the specific steps for generating the initial screening results of image and text nodes are as follows:

[0018] The heterogeneous nodes of the dynamic graph topology unit are divided according to their respective propagation regions and distributed to the corresponding edge recognition nodes, outputting a set of graph-text subgraphs;

[0019] In the edge recognition node, the image feature vector of the image-text sub-image set is extracted, the sharpness score and composition boundary score are calculated, and the visual quality score is output.

[0020] Simultaneously extract text feature vectors from the image-text sub-image set, analyze semantic coherence and context jump, and output semantic scores for text nodes;

[0021] Visual quality scores and text node semantic scores are matched and fused according to node identifiers to generate preliminary screening results for image and text nodes.

[0022] As a preferred embodiment of the AI-based image and text content quality recognition method of the present invention, the specific steps for generating the node scoring matrix are as follows:

[0023] The initial screening results of the text and image nodes are structurally aligned with the dynamic graph topology units, and a graph neural network is used to perform multiple rounds of feature aggregation and message passing to generate a content evolution guidance matrix.

[0024] Based on the image feature vectors and text feature vectors of the heterogeneous graph data input set, and the semantic similarity between them and the upstream and downstream neighboring nodes, the semantic redundancy is calculated, and the redundancy score vector is output.

[0025] By utilizing the node propagation order and historical path structure recorded in the dynamic graph topology unit, the temporal and structural offset of the propagation path between nodes is analyzed, and a propagation path offset score vector is output.

[0026] The content evolution guidance matrix, redundancy score vector, and propagation path offset score vector are normalized and fused to generate a node score matrix.

[0027] As a preferred embodiment of the AI-based image and text content quality recognition method of the present invention, the specific steps for generating the aggregated quality score set are as follows:

[0028] The node scoring matrix and the initial screening results of the image and text nodes are aligned in terms of features and fused in terms of dimensions to generate a fused feature set.

[0029] Based on the fusion feature set, the propagation path structure, timestamp sequence and propagation behavior features of each graph node in the dynamic graph topology unit are extracted, and feature-level concatenation is performed to output a multi-dimensional aggregated feature vector set;

[0030] The multidimensional aggregated feature vector set is normalized in dimension and weighted in feature processing to construct a weighted clustering input set. Then, the weighted clustering input set is grouped by similarity and classified into quality levels using a hierarchical clustering algorithm to generate an aggregated quality score set.

[0031] As a preferred embodiment of the AI-based image and text content quality recognition method of the present invention, the specific steps for generating the explanation and scoring summary text are as follows:

[0032] Based on the scoring results of the aggregated quality score set, and according to the scoring interval division standard, corresponding quality level labels are assigned to the image and text nodes, generating a set of image and text nodes with labeled levels.

[0033] Based on the set of labeled text and image nodes, a backtracking analysis is performed on the propagation path structure in the dynamic graph topology unit to extract the key link nodes that cause score changes. Combined with node features, semantic breakpoints and visual anomaly factors are identified, and a set of content anomaly markers is output.

[0034] The set of image and text nodes with labeled levels and the set of content anomaly markers are merged to generate an explanatory score summary text.

[0035] Secondly, this invention provides an AI-based image and text content quality recognition system, comprising: a feature acquisition module for acquiring image and text content data streams, extracting structured features of images and text, and generating a multi-dimensional structure vector set; a graph construction module for establishing graph connections based on the relationships between multi-dimensional structure vectors, constructing a heterogeneous graph structure, recording the propagation path and time series of node content, and generating dynamic graph topology units; an edge analysis module for distributing dynamic graph topology units to edge recognition nodes, performing quality scoring and semantic analysis on image feature vectors and text feature vectors, and generating preliminary screening results for image and text nodes; and a guided evaluation module. The system comprises four modules: a node rating matrix and a clustering identification module. The first module performs multi-round feature aggregation using a graph neural network based on the initial screening results of text and image nodes and dynamic graph topology units. It calculates content guidance degree, semantic redundancy degree, and propagation path offset to generate a node rating matrix. The second module performs normalization processing based on the node rating matrix and the initial screening results of text and image nodes, constructs a clustering input set, and uses a hierarchical clustering algorithm to group and classify quality levels, generating an aggregated quality rating set. The third module assigns quality level labels to text and image nodes based on the aggregated quality rating set, traces back the propagation path to identify link nodes and abnormal features that indicate rating changes, and generates explanatory rating summary text.

[0036] Thirdly, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program is executed by the processor, it implements any step of the AI-based image and text content quality recognition method as described in the first aspect of the present invention.

[0037] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the AI-based image and text content quality recognition method as described in the first aspect of the present invention.

[0038] The beneficial effects of this invention are as follows: By using a graph neural network to perform multi-round feature aggregation on the initial screening results of graph and text nodes and the dynamic graph topology, the content guidance degree, semantic redundancy degree, and propagation path offset scores of the nodes are extracted to generate a node scoring matrix, realizing the transformation from static quality scoring to dynamic propagation modeling. The scoring matrix integrates semantic features and propagation evolution information, possessing stronger structural relevance and interpretability, providing a high-quality feature foundation for subsequent clustering analysis and content level classification, and significantly improving the accuracy and intelligence level of graph and text content quality identification. Attached Figure Description

[0039] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0040] Fig. 1 This is a flowchart of an AI-based image and text content quality recognition method.

[0041] Fig. 2 This is a schematic diagram of an AI-based image and text content quality recognition system.

[0042] Fig. 3 A flowchart for generating a multidimensional structure vector set.

[0043] Fig. 4 The flowchart generated from the initial screening results of the image and text nodes. Detailed Implementation

[0044] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0045] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0046] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0047] Reference Figs. 1-4 This is one embodiment of the present invention, which provides an artificial intelligence-based method for identifying the quality of text and image content, including the following steps:

[0048] S1. Collect the image and text content data stream, extract the structured features of the images and text, and generate a multi-dimensional structure vector set.

[0049] S1.1 Collect image and text content data streams, including publication time, author information, and dissemination behavior, and parse them to form image datasets and text datasets.

[0050] Specifically, the process involves collecting image content, text content, publication time data, author information data, and dissemination behavior data, and uniformly labeling them using image-text association identifiers. Image content and text content are matched and bound according to these identifiers to construct structured image-text pairs. Each image is extracted based on its format and saved as an image element in the image dataset, and the corresponding text is extracted based on its encoding format and saved as a text element in the text dataset. The publication time data, author information data, and dissemination behavior data of the structured image-text pairs are recorded as time fields, author fields, and dissemination fields, respectively, and an index mapping relationship is established with the image elements and text elements to complete the parsing of the image-text content data stream, forming image datasets and text datasets.

[0051] S1.2. The image dataset is normalized in size, converted in color space and suppressed in noise. The pre-trained feature extraction network is used to extract edge texture, color distribution and composition contour features to generate image feature vectors.

[0052] Specifically, each image in the image dataset is resized to a uniform size, for example, by scaling the images to 224×224 pixels to achieve size normalization. The normalized images are then converted from the RGB color space to the YCbCr color space to separate luminance and chrominance information. Gaussian filtering is used to suppress noise in the luminance channel, reducing random noise interference in the image. The processed images are then input into a pre-trained feature extraction network, such as a ResNet50 network, to extract edge texture features, color distribution features, and compositional contour features. The output of the pre-trained feature extraction network is then mapped through a fully connected layer to generate a fixed-dimensional image feature vector.

[0053] It should also be explained that the specific steps of the pre-trained feature extraction network are as follows: the input normalized image tensor is fed into the input layer of the network for preliminary convolution to extract low-level edge and texture features; through multiple convolutional and pooling layers, mid-to-high-level semantic features of the image are gradually extracted, and activation functions (such as ReLU) are used to enhance non-linear expressive power; the training effect is optimized and gradient vanishing is prevented through residual connections or normalization layers; the final convolutional layer of the feature extraction network outputs a multi-channel feature map, representing the multi-dimensional feature expression of the image; the multi-channel feature maps are aggregated through a global average pooling layer to obtain a fixed-length image feature vector.

[0054] S1.3 Clean and segment the text dataset, and use a pre-trained semantic embedding model to extract semantic relevance, contextual consistency and keyword density to generate text feature vectors.

[0055] Specifically, the text dataset undergoes preprocessing operations such as removing special symbols, HTML tags, and redundant spaces to obtain clean text. A word segmentation tool is then used to segment the clean text into word sequences. Next, stop word filtering is applied to the segmentation results to remove words with no actual semantic contribution, and the remaining words are classified according to part-of-speech tagging methods to form a structured text unit sequence. This structured text unit sequence is input into a pre-trained semantic embedding model to extract semantic vector representations of the words. Semantic relevance and contextual consistency features of the text are obtained by calculating the cosine similarity between word vectors. Simultaneously, keyword frequency is statistically analyzed, and keyword density is calculated using word frequency statistics. Finally, the semantic relevance, contextual consistency, and keyword density features are fused to generate a text feature vector.

[0056] It should also be explained that the specific steps of the pre-trained semantic embedding model are as follows: the input structured text unit sequence is converted into the corresponding word index sequence; the word index sequence is mapped to the pre-trained word vector space to obtain the initial embedding vector of each word; then, the multi-layer neural network structure in the pre-trained semantic embedding model is used to perform context encoding on the word vectors to capture the semantic association and contextual dependency between words; the context-sensitive embedding representation of each word is calculated through the forward propagation of the semantic embedding model; the context embedding vector of each word is aggregated according to certain rules (such as average pooling or max pooling) to generate the semantic vector representation of the entire text, which is output as the text feature vector.

[0057] S1.4. Concatenate and fuse the image feature vector and the text feature vector to generate a multi-dimensional structure vector set.

[0058] Specifically, based on image feature vectors and text feature vectors, ensuring that both are fixed-length vectors, the image feature vectors and text feature vectors are directly concatenated in the dimensional direction to form a high-dimensional vector. The concatenated high-dimensional vector is then standardized, for example, by using Z-score normalization, to ensure that each dimensional feature has a uniform numerical range. Finally, the standardized high-dimensional vectors are aggregated to form a multi-dimensional structure vector set.

[0059] S2. Based on the correlation between multidimensional structural vectors, establish graph connections, construct heterogeneous graph structures, record the propagation path and time series of node content, and generate dynamic graph topology units.

[0060] S2.1 Perform similarity calculation and interaction feature modeling on the multidimensional structure vector set to generate an association matrix.

[0061] Specifically, for each pair of image-text fusion feature vectors in the multidimensional structure vector set, the cosine similarity algorithm is used to calculate the similarity value between the image-text fusion feature vectors, and the expression is:

[0062] ;

[0063] in, Represents the image-text fusion feature vector Feature vector fusion with image and text Similarity score between them Indicates the first Image-text fusion feature vectors Indicates the first Image-text fusion feature vectors This represents the first index number in the set of multidimensional structured vectors, with a value range of 1. , For the total quantity, This represents the second index number in the multidimensional structure vector set, with a value range of 1. ,and ;

[0064] The similarity values ​​calculated for each pair of image-text fusion feature vectors are filled into a two-dimensional matrix to form a similarity matrix. Based on the similarity matrix, the adjacency matrix construction method in graph theory is used, combining the co-occurrence frequency and association strength of each pair of image-text fusion feature vectors, to calculate the edge interaction weights, expressed as:

[0065] ;

[0066] in, Represents the image-text fusion feature vector Feature vector fusion with image and text Edge weights between them Represents the fusion function. Represents the image-text fusion feature vector Feature vector fusion with image and text The frequency of common occurrence in the data on the dissemination of text and image content Represents the image-text fusion feature vector Feature vector fusion with image and text The semantic or content association strength between them is determined; finally, according to the set edge weight filtering rules, an association matrix representing the association relationship and strength between image and text fusion feature vectors is obtained.

[0067] The specific steps of the edge weight filtering rule should also be explained: Statistical analysis is performed on the edge weight values ​​calculated between the image and text fusion feature vectors to extract the distribution characteristics of all edge weight values; based on the distribution pattern of the edge weight set, a filtering threshold is set, which can be determined by statistical quantiles (e.g., the 50th percentile) or standard deviation range (e.g., mean plus or minus one standard deviation); all edge weight values ​​are traversed, retaining only edge connections greater than or equal to the filtering threshold, and setting or removing edge weights below the filtering threshold, thus completing the effective screening of edge connections; finally, a sparse adjacency relation matrix satisfying significant correlation features is generated. The filtering rule is set to improve the discriminative ability of subsequent graph structure representations, enhance the connection density between highly correlated nodes, and avoid interference from weak edge noise on the graph structure.

[0068] S2.2 Construct a graph connection structure based on the association matrix, and label the nodes of image feature vectors, text feature vectors and propagation behavior with their types to generate an initial heterogeneous graph.

[0069] Specifically, the image-text fusion feature vectors in the multidimensional structure vector set are read sequentially, and the image and text content corresponding to each image-text fusion feature vector is treated as an independent node, initializing the node index in the graph. Based on the correlation matrix between the image-text fusion feature vectors, edge connections are established between the node indices, and corresponding edge weight information is added to the edge connections. According to the source of the image-text fusion feature vectors, nodes originating from image datasets are marked as image nodes, nodes originating from text datasets are marked as text nodes, and nodes originating from propagation behavior information in the image-text content data stream are marked as propagation behavior nodes. Finally, a graph connection structure composed of image nodes, text nodes, and propagation behavior nodes is completed, forming an initial heterogeneous graph.

[0070] S2.3. Based on the timestamps and source information in the initial heterogeneous graph, identify the propagation path and the order of occurrence, construct a time series mapping table, and merge it with the initial heterogeneous graph. Assign time attributes to the propagation edges, mark the content evolution direction and propagation dynamic characteristics, and generate dynamic graph topology units.

[0071] Specifically, the timestamp field recorded in the text and image content data stream is extracted from each image node, text node, and propagation behavior node. The timestamp field represents the publication time information explicitly marked in each piece of text and image content data. Based on the author identifier, platform source, or propagation behavior type fields corresponding to the image node, text node, and propagation behavior node in the text and image content data stream, the source type is determined, and a mapping relationship between nodes and source fields is established. Based on the timestamp values ​​between the source and target nodes corresponding to the edges connecting the nodes, the time difference for each edge is calculated, and all edges are sorted according to the size of the time difference, resulting in a set of propagation paths arranged in chronological order. The content evolution direction of the propagation path is determined based on the source fields of the two nodes connected by the edge; for example, "User A → Platform → User B" represents a propagation link from the author node to the platform node and then to the receiving node. The sorted set of propagation paths is bound to the corresponding timestamps and source information to form a time-series mapping table, marking the propagation direction, start and end nodes, time span, and source type of each propagation path, constructing a dynamic graph topology unit with time attributes.

[0072] S3. Distribute the dynamic graph topology units to the edge recognition nodes, perform quality scoring and semantic analysis on the image feature vectors and text feature vectors, and generate the initial screening results of the graph and text nodes.

[0073] S3.1 Divide the heterogeneous nodes of the dynamic graph topology unit according to their respective propagation regions and distribute them to the corresponding edge recognition nodes, outputting a set of graph-text subgraphs.

[0074] Specifically, based on the geographic attribute fields contained in each image node, text node, and propagation behavior node in the dynamic graph topology unit, the propagation region identifier is extracted. For example, the geographic attribute field may include country, province, city, or IP geographic location information. All image nodes, text nodes, and propagation behavior nodes are divided into several subsets according to the geographic attribute fields, ensuring that all nodes in each subset have the same propagation region identifier. Based on the partitioning results, the node set and edge set in the dynamic graph topology unit are segmented, and image nodes, text nodes, propagation behavior nodes, and edges belonging to the same propagation region are extracted to form a local topology subgraph. Then, according to the allocation table of edge recognition nodes corresponding to each propagation region, each local topology subgraph is sent to the edge recognition node of the corresponding propagation region. This completes the partitioning operation of extracting the local region topology structure from the dynamic graph topology unit, outputting a set of image-text subgraphs composed of image nodes, text nodes, and propagation behavior nodes corresponding to multiple propagation regions.

[0075] S3.2 In the edge recognition node, extract the image feature vector of the image sub-image set, calculate the sharpness score and composition boundary score, and output the visual quality score.

[0076] Specifically, in the edge recognition node, each image node in the image sub-graph set is located to obtain the corresponding image data; a pre-trained feature extraction network is used to extract image feature vectors containing edge intensity information from the image data; the edge distribution density and edge gradient changes in the image feature vectors are analyzed using edge detection operators to calculate the image sharpness score, where the image sharpness score... The expression is:

[0077] ;

[0078] in, Indicates the image sharpness score. This is indicated as the first rating item. This represents the total number of pixels in the image that participate in the sharpness scoring. The index range of a pixel is 1 to , Represents the image feature vector at the th The grayscale value at each pixel location Indicates the first The magnitude of local variation in the edge gradient value of each pixel;

[0079] Simultaneously, utilizing the composition parameters in the image feature vector, a contour integrity detection method is employed to evaluate the boundary coherence and closure of the main composition elements, yielding a composition boundary score. The expression is:

[0080] ;

[0081] in, Indicates the score for the composition boundary. This indicates the second scoring item. This represents the total length of all boundary line segments in the composition. Indicates the boundary quality of the main compositional elements. This represents the metric that identifies all boundary segments. This indicates the measurement used to identify closed boundary segments;

[0082] Image sharpness rating Scoring of composition boundaries The data is synthesized to form a visual quality score. , is represented as:

[0083] ;

[0084] in, Indicates visual quality score. This represents the final comprehensive evaluation result that reflects the visual quality of the image. The weighting coefficients representing the image sharpness score. This represents the weighting coefficient for the composition boundary score.

[0085] S3.3 Simultaneously extract text feature vectors from the set of text and image subgraphs, analyze semantic coherence and context jumps, and output semantic scores for text nodes.

[0086] Specifically, in the edge recognition node, text nodes are sequentially identified from the graph subgraph set, and the original text data corresponding to the text node is obtained through the node identifier. Word segmentation is performed on the original text data, and word vector mapping based on word embedding is used to map each word to a fixed-dimensional vector to construct the initial text feature vector. A pre-trained semantic embedding model is used to perform contextual semantic encoding on the initial text feature vector to obtain the semantic representation of the text node. Based on the semantic representation, the semantic similarity score between adjacent sentences or paragraphs in the text node is calculated to measure the semantic coherence within the text node. Further, combined with the internal context structure of the text node, logical jump positions are identified, and jump factors are constructed based on inter-sentence connections and semantic break positions to calculate the context jump score of the text node. Finally, the semantic coherence score and the context jump score are fused to output the semantic score of the text node.

[0087] S3.4 Match and merge the visual quality score and the semantic score of the text node according to the node identifier to generate the initial screening results of the image and text nodes.

[0088] Specifically, in edge recognition nodes, unique identifiers of image-text nodes are extracted and used as the basis for matching visual quality scores with text node semantic scores. Based on the node identifiers, the visual quality scores are matched one-to-one with the corresponding text node semantic scores to ensure complete association between the visual and semantic score data of each image-text node. Normalization methods are applied to the matched visual quality scores and text node semantic scores, such as mapping the score values ​​to the range of 0 to 1 to eliminate differences in units and numerical ranges. The normalized visual quality scores and text node semantic scores are then weighted, with the weight ratio set according to application requirements, for example, a visual score weight of 0.6 and a semantic score weight of 0.4. The weighted visual quality scores and text node semantic scores are then fused to obtain a fused score, which serves as the comprehensive score index for image-text nodes. Finally, the fused score and node identifiers are combined to form the initial screening results for image-text nodes.

[0089] S4. Based on the initial screening results of the graph and text nodes and the dynamic graph topology units, a graph neural network is used to perform multi-round feature aggregation, calculate the content guidance degree, semantic redundancy degree and propagation path offset, and generate a node scoring matrix.

[0090] S4.1 Align the initial screening results of the text and image nodes with the dynamic graph topology units, and use graph neural networks to perform multiple rounds of feature aggregation and message passing to generate a content evolution guidance matrix.

[0091] Specifically, the image nodes in the initial screening results of the image-text node model are mapped to the image nodes in the dynamic graph topology unit according to their node identifiers, and the text nodes in the initial screening results of the image-text node model are mapped to the text nodes in the dynamic graph topology unit according to their node identifiers, ensuring that the node identifiers of the image nodes and text nodes in the initial screening results of the image-text node model are consistent with those in the dynamic graph topology unit. The node representations of the image nodes and text nodes in the graph neural network are initialized as fused scoring vectors corresponding to the image nodes and text nodes. Multiple iterations are performed in the graph neural network. In each iteration, message passing and feature aggregation operations are performed on the feature vectors between image nodes and their neighboring nodes, and message passing and feature aggregation operations are performed on the feature vectors between text nodes and their neighboring nodes. The current representations of the image nodes and text nodes are updated using the features of the neighboring nodes and their own features. After multiple iterations, the final representations of all image nodes and text nodes are extracted to form the content evolution guidance degree matrix.

[0092] S4.2. Based on the image feature vectors and text feature vectors of the heterogeneous graph data input set, and the semantic similarity between them and the upstream and downstream neighboring nodes, calculate the semantic redundancy degree and output the redundancy degree score vector.

[0093] Specifically, semantic similarity is calculated between the image feature vectors corresponding to image nodes in the heterogeneous graph data input set and the image feature vectors of their upstream and downstream neighboring image nodes. Similarly, semantic similarity is calculated between the text feature vectors corresponding to text nodes in the heterogeneous graph data input set and the text feature vectors of their upstream and downstream neighboring text nodes. The semantic similarity calculation uses a cosine similarity algorithm, calculating similarity scores between each image node and all neighboring image nodes, and between each text node and all neighboring text nodes. The percentage of similarity scores greater than those of neighboring nodes is then used as the redundancy score for the corresponding image node or text node. Image node redundancy score vectors and text node redundancy score vectors are constructed and then merged to form a complete redundancy score vector.

[0094] S4.3 Utilize the node propagation order and historical path structure recorded in the dynamic graph topology unit to analyze the time and structural offset of the propagation path between nodes, and output the propagation path offset score vector.

[0095] Specifically, the node identifiers and corresponding timestamp values ​​of all image nodes, text nodes, and propagation behavior nodes in the dynamic graph topology unit are extracted, and the current propagation order sequence is constructed based on the timestamp values. The historical path structure recorded in the dynamic graph topology unit is read synchronously, and the order of nodes in each propagation link in the historical path structure is compared with the current propagation order sequence to identify the changes in path position, jump direction, and path length of adjacent node pairs. The historical path change features are used as structural offset indicators, and the timestamp difference of adjacent nodes in the current propagation order is used as time offset indicators. After processing them separately, they are combined to generate a propagation path offset score vector.

[0096] S4.4 Normalize and fuse the content evolution guidance matrix, redundancy score vector, and propagation path offset score vector to generate a node score matrix.

[0097] Specifically, the content evolution guidance matrix, redundancy score vector, and propagation path offset score vector are normalized. This involves using linear normalization to calculate the maximum and minimum values ​​of image nodes, text nodes, and propagation behavior nodes in the content evolution guidance matrix, and then standardizing each score within the matrix using the same method. The same normalization process is then applied to the scores in the redundancy score vector and propagation path offset score vector. The normalized content evolution guidance matrix, redundancy score vector, and propagation path offset score vector are then combined according to the node identifiers of the image nodes, text nodes, and propagation behavior nodes, and the fused score values ​​are calculated using known weight parameters and filled into the corresponding node positions to generate a node score matrix.

[0098] S5. Based on the node scoring matrix and the initial screening results of the image and text nodes, normalization processing is performed to construct a clustering input set. Then, hierarchical clustering algorithm is used to group and classify the quality levels to generate an aggregated quality score set.

[0099] S5.1. Align the node scoring matrix with the initial screening results of the text and image nodes, perform feature alignment and dimension fusion, and generate a fused feature set.

[0100] Specifically, for each node score in the node scoring matrix, feature alignment is performed based on the node identifier and the corresponding visual quality score and semantic score of the text node in the initial screening results of the image and text nodes. During feature alignment, the node identifier is used as an index, and row-level concatenation is performed on the score vectors of each row in the node scoring matrix with the rows of the same identifier in the initial screening results of the image and text nodes to form a unified feature representation. The concatenated feature representation is then subjected to dimensionality standardization, scaling the dimensional values ​​to a uniform range. Examples of methods include min-max normalization or Z-score normalization. Finally, the standardized feature representations of all nodes are combined into a fused feature set.

[0101] S5.2. Based on the fusion feature set, extract the propagation path structure, timestamp sequence and propagation behavior features of each graph node in the dynamic graph topology unit, and perform feature-level concatenation to output a multi-dimensional aggregated feature vector set.

[0102] Specifically, based on the fused feature set, and according to the image and text node identifiers, the propagation path structure information, timestamp sequence information, and propagation behavior feature information corresponding to the image and text nodes are extracted from the dynamic graph topology units. The propagation path structure information is an ordered list of predecessor and successor image and text nodes connected to the image and text node in the dynamic graph topology. The timestamp sequence information is a time series composed of the publication times of all image and text nodes in the ordered path list. The propagation behavior feature information includes propagation behavior indicators such as the number of likes, shares, and comments for the image and text node. Subsequently, the image and text... The propagation path structure information corresponding to a node is encoded as a path structure position encoding vector of the image and text node; the timestamp sequence information corresponding to the image and text node is encoded as a time series difference encoding vector of the image and text node; and the propagation behavior feature information corresponding to the image and text node is encoded as a propagation behavior statistics encoding vector of the image and text node. The standardized feature representations corresponding to the image and text nodes in the fusion feature set are concatenated with the path structure position encoding vector, the time series difference encoding vector, and the propagation behavior statistics encoding vector of the image and text node in terms of dimensions to generate a multi-dimensional aggregated feature vector of the image and text node. Finally, the multi-dimensional aggregated feature vectors of all image and text nodes constitute a multi-dimensional aggregated feature vector set.

[0103] S5.3. Perform dimension normalization and feature weighting on the multidimensional aggregated feature vector set to construct a weighted clustering input set. Then, use a hierarchical clustering algorithm to group the weighted clustering input set by similarity and classify its quality level to generate an aggregated quality score set.

[0104] Specifically, for each multidimensional aggregated feature vector in the multidimensional aggregated feature vector set, a normalization method is used to adjust the value of each dimension to a uniform range, for example, scaling all dimension values ​​to the range of 0 to 1 to complete dimension normalization; according to the known weight parameters, each normalized dimension feature is assigned a corresponding weight, the weighted feature value is calculated, a weighted feature vector is generated, and a weighted clustering input set is constructed; based on the weighted clustering input set, a hierarchical clustering algorithm is used to divide the multidimensional aggregated feature vector into multiple quality level groups according to feature similarity. The specific steps include calculating the distance matrix between samples and iteratively merging similar clusters until the stopping condition is met; the final output is an aggregated quality score set containing the quality level corresponding to each cluster group.

[0105] S6. Assign quality level labels to image and text nodes based on the aggregated quality score set, trace back the propagation path to identify link nodes and abnormal features of score changes, and generate explanatory score summary text.

[0106] S6.1 Based on the scoring results of the aggregated quality score set, and according to the scoring interval division standard, assign corresponding quality level labels to the image and text nodes, and generate a set of image and text nodes with labeled levels.

[0107] Specifically, based on each rating result in the aggregated quality rating set, the corresponding quality level label is determined according to the rating interval division standard. This includes setting multiple rating interval ranges, such as dividing the rating into high quality, medium quality, and low quality intervals in the example; assigning the corresponding quality level label to the image and text node according to the interval to which each rating belongs; and summarizing the set of image and text nodes assigned quality level labels to generate a set of image and text nodes labeled with levels.

[0108] S6.2 Based on the set of labeled text and image nodes, perform backtracking analysis on the propagation path structure in the dynamic graph topology unit, extract the key link nodes that cause score changes, and combine node features to identify semantic breakpoints and visual anomaly factors, and output a set of content anomaly markers.

[0109] Specifically, based on the set of labeled image and text nodes, the propagation path structure in the dynamic graph topology unit is back-analyzed according to time sequence and connection relationship to identify key link nodes that appear during the scoring process. For key link nodes, the image feature vectors and text feature vectors corresponding to the nodes are extracted, the continuity of the semantic content of the nodes is analyzed, and semantic similarity calculation is used to identify semantic breakpoints. Combining the visual quality scoring index in the image feature vector, visual anomaly factors are identified. The identified semantic breakpoints and visual anomaly factor nodes are summarized to generate a set of content anomaly markers.

[0110] It should also be noted that critical link nodes refer to nodes in the propagation path of dynamic graph topology units that have a significant impact on changes in the text and image content quality score. They are typically located at important positions in the propagation path; improvements or declines in content quality directly lead to fluctuations in the scores of subsequent nodes, manifesting as significant score changes, turning points in the propagation path, or anomalous nodes. By identifying critical link nodes, the specific stages of content quality change can be located, thus providing targeted analytical basis for content quality assessment and anomaly detection.

[0111] S6.3. Merge the set of graphic and text nodes with labeled levels with the set of content anomaly markers to generate an explanatory scoring summary text.

[0112] Specifically, the set of image and text nodes labeled with quality levels and the set of content anomaly markers are matched according to the unique identifier of each node. For each matched image and text node, the corresponding quality level information and anomaly marker attributes are extracted. Based on the matching results, a joint attribute vector containing quality level and anomaly features is constructed. Through predefined text template rules, combined with the specific values ​​and features in the joint attribute vector, the corresponding explanation score summary text is generated. This process is repeated to generate a complete set of explanation score summary texts for each node, and finally, the complete explanation score summary texts are summarized and output.

[0113] It should also be noted that the predefined text template rules analyze the quality level and anomaly attribute of the image and text nodes, and divide the text templates into multiple categories according to different quality levels and anomaly types. Each template category contains a fixed description structure and variable interpolation fields, which are used to fill in specific score values ​​and anomaly information. The applicable template category is determined based on the joint attribute vector of the node, and the quality level, anomaly marker, and related values ​​in the attribute vector are sequentially filled into the interpolation fields of the template. After the text is generated, the generated text is semantically and formatted to ensure that the expression is accurate and conforms to professional description standards. The text template rules are designed based on content quality assessment standards and anomaly detection results to ensure the relevance and comprehensibility of the score explanation.

[0114] This embodiment also provides an AI-based image and text content quality recognition system, including: a feature acquisition module for acquiring image and text content data streams, extracting structured features of images and text, and generating a multi-dimensional structure vector set; a graph construction module for establishing graph connections based on the relationships between multi-dimensional structure vectors, constructing a heterogeneous graph structure, recording the propagation path and time series of node content, and generating dynamic graph topology units; an edge analysis module for distributing dynamic graph topology units to edge recognition nodes, performing quality scoring and semantic analysis on image feature vectors and text feature vectors, and generating preliminary screening results for image and text nodes; and a guided evaluation module for... Based on the initial screening results of image and text nodes and dynamic graph topology units, a graph neural network is used for multi-round feature aggregation to calculate content guidance degree, semantic redundancy degree, and propagation path offset, generating a node scoring matrix. The clustering and identification module is used to normalize the node scoring matrix and the initial screening results of image and text nodes, construct a clustering input set, and use a hierarchical clustering algorithm to group and classify quality levels, generating an aggregated quality score set. The result interpretation module is used to assign quality level labels to image and text nodes based on the aggregated quality score set, trace back the propagation path to identify link nodes and abnormal features of score changes, and generate explanatory score summary text.

[0115] This embodiment also provides a computer device applicable to the case of an AI-based image and text content quality recognition method, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the AI-based image and text content quality recognition method proposed in the above embodiment.

[0116] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0117] This embodiment also provides a storage medium storing a computer program, which, when executed by a processor, implements the AI-based image and text content quality recognition method proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0118] In summary, this invention utilizes a graph neural network to perform multi-round feature aggregation on the initial screening results of graph-text nodes and the dynamic graph topology, extracting the content guidance degree, semantic redundancy degree, and propagation path offset scores of nodes to generate a node scoring matrix. This achieves a shift from static quality scoring to dynamic propagation modeling. The scoring matrix integrates semantic features and propagation evolution information, possessing stronger structural relevance and interpretability. It provides a high-quality feature foundation for subsequent clustering analysis and content level classification, significantly improving the accuracy and intelligence level of graph-text content quality identification.

[0119] 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 preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for identifying the quality of text and image content based on artificial intelligence, characterized in that: include, Collect image and text content data streams, extract structured features from images and text, and generate a multi-dimensional structured vector set; Graph connections are established based on the relationships between multidimensional structural vectors, heterogeneous graph structures are constructed, and the propagation paths and time series of node content are recorded to generate dynamic graph topology units. Dynamic graph topology units are distributed to edge recognition nodes, and quality scoring and semantic analysis are performed on image feature vectors and text feature vectors to generate preliminary screening results for graph and text nodes. Based on the initial screening results of text and image nodes and dynamic graph topology units, a graph neural network is used for multi-round feature aggregation to calculate content guidance degree, semantic redundancy degree, and propagation path offset, generating a node scoring matrix. The specific steps are as follows. The initial screening results of the text and image nodes are structurally aligned with the dynamic graph topology units, and a graph neural network is used to perform multiple rounds of feature aggregation and message passing to generate a content evolution guidance matrix. Based on the image feature vectors and text feature vectors of the heterogeneous graph data input set, and the semantic similarity between them and the upstream and downstream neighboring nodes, the semantic redundancy is calculated, and the redundancy score vector is output. By utilizing the node propagation order and historical path structure recorded in the dynamic graph topology unit, the temporal and structural offset of the propagation path between nodes is analyzed, and a propagation path offset score vector is output. The content evolution guidance matrix, redundancy score vector, and propagation path offset score vector are normalized and fused to generate a node score matrix. Based on the node scoring matrix and the initial screening results of the image and text nodes, normalization processing is performed to construct a clustering input set, and hierarchical clustering algorithm is used to group and classify the quality levels to generate an aggregated quality score set. Based on the aggregated quality score set, quality level labels are assigned to the image and text nodes. The propagation path is traced back to identify the link nodes and abnormal features of score changes, and an explanatory score summary text is generated.

2. The method for identifying the quality of text and image content based on artificial intelligence as described in claim 1, characterized in that: The specific steps for generating the multidimensional structure vector set are as follows: Collect image and text content data streams, along with associated publication time, author information, and dissemination behavior data, and parse them to form image datasets and text datasets; The image dataset is normalized in size, converted in color space and suppressed in noise. A pre-trained feature extraction network is used to extract edge texture, color distribution and composition contour features to generate image feature vectors. The text dataset is cleaned and segmented, and a pre-trained semantic embedding model is used to extract semantic relevance, contextual consistency and keyword density to generate text feature vectors. Image feature vectors and text feature vectors are concatenated and fused to generate a multi-dimensional structure vector set.

3. The method for identifying the quality of text and image content based on artificial intelligence as described in claim 1, characterized in that: The specific steps for generating dynamic graph topology units are as follows. Similarity calculation and interaction feature modeling are performed on a multidimensional structure vector set to generate an association matrix; A graph connection structure is constructed based on the association matrix, and the nodes of image feature vectors, text feature vectors, and propagation behavior are labeled with their types to generate an initial heterogeneous graph. Based on the timestamps and source information in the initial heterogeneous graph, the propagation path and occurrence order are identified, a time series mapping table is constructed, and it is fused with the initial heterogeneous graph. The time attribute is assigned to the propagation edge, the content evolution direction and propagation dynamic characteristics are marked, and dynamic graph topology units are generated.

4. The method for identifying the quality of text and image content based on artificial intelligence as described in claim 1, characterized in that: The specific steps for generating the initial screening results of the image and text nodes are as follows. The heterogeneous nodes of the dynamic graph topology unit are divided according to their respective propagation regions and distributed to the corresponding edge recognition nodes, outputting a set of graph-text subgraphs; In the edge recognition node, the image feature vector of the image-text sub-image set is extracted, the sharpness score and composition boundary score are calculated, and the visual quality score is output. Simultaneously extract text feature vectors from the image-text sub-image set, analyze semantic coherence and context jump, and output semantic scores for text nodes; Visual quality scores and text node semantic scores are matched and fused according to node identifiers to generate preliminary screening results for image and text nodes.

5. The method for identifying the quality of text and image content based on artificial intelligence as described in claim 1, characterized in that: The specific steps for generating the aggregated quality score set are as follows: The node scoring matrix and the initial screening results of the image and text nodes are aligned in terms of features and fused in terms of dimensions to generate a fused feature set. Based on the fusion feature set, the propagation path structure, timestamp sequence and propagation behavior features of each graph node in the dynamic graph topology unit are extracted, and feature-level concatenation is performed to output a multi-dimensional aggregated feature vector set; The multidimensional aggregated feature vector set is normalized in dimension and weighted in feature processing to construct a weighted clustering input set. Then, the weighted clustering input set is grouped by similarity and classified into quality levels using a hierarchical clustering algorithm to generate an aggregated quality score set.

6. The method for identifying the quality of text and image content based on artificial intelligence as described in claim 1, characterized in that: The specific steps for generating the explanation and scoring summary text are as follows: Based on the scoring results of the aggregated quality score set, and according to the scoring interval division standard, corresponding quality level labels are assigned to the image and text nodes, generating a set of image and text nodes with labeled levels. Based on the set of labeled text and image nodes, a backtracking analysis is performed on the propagation path structure in the dynamic graph topology unit to extract the key link nodes that cause score changes. Combined with node features, semantic breakpoints and visual anomaly factors are identified, and a set of content anomaly markers is output. The set of image and text nodes with labeled levels and the set of content anomaly markers are merged to generate an explanatory score summary text.

7. An AI-based image and text content quality recognition system, based on the AI-based image and text content quality recognition method according to any one of claims 1 to 6, characterized in that: include, The feature acquisition module is used to acquire image and text content data streams, extract structured features of images and text, and generate a set of multi-dimensional structured vectors. The graph construction module is used to establish graph connections based on the relationships between multidimensional structural vectors, construct heterogeneous graph structures, record the propagation path and time series of node content, and generate dynamic graph topology units. The edge analysis module is used to distribute dynamic graph topology units to edge recognition nodes, perform quality scoring and semantic analysis on image feature vectors and text feature vectors, and generate preliminary screening results for graph and text nodes. The guidance evaluation module is used to perform multi-round feature aggregation based on the initial screening results of text and image nodes and dynamic graph topology units, and to calculate the content guidance degree, semantic redundancy degree and propagation path offset, and generate a node scoring matrix. The clustering identification module is used to normalize the node scoring matrix and the initial screening results of the image and text nodes, construct a clustering input set, and use the hierarchical clustering algorithm to group and classify the quality levels to generate an aggregated quality score set. The results interpretation module is used to assign quality level labels to image and text nodes based on the aggregated quality score set, trace back the propagation path to identify link nodes and abnormal features of score changes, and generate summary text explaining the scores.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the AI-based image and text content quality recognition method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the AI-based image and text content quality recognition method as described in any one of claims 1 to 6.

Images