A binding scheme automatic matching method based on content attribute analysis

Abstract

The application discloses a kind of based on content attribute analysis binding scheme automatic matching method, specifically related to data processing field, including by multimodal technique synchronous acquisition document text, vision, language style and field knowledge graph data, each dimension data is respectively carried out feature extraction, quantization conversion and structured processing, form the core data material of support attribute analysis;Again based on multimodal quantization model completes text theme contribution degree calculation, visual style probability analysis, language style characteristic quantization and field association strength evaluation, non-structured data is converted into the quantization index that can be fused;Finally, with attention mechanism dynamic fusion multidimensional feature, judges document core attribute by Bayes model, with preset binding scheme library is matched and filters conflict scheme, generates optimal scheme in combination with document page number, size and special printing requirement, significantly improve the efficiency and content adaptation accuracy of binding scheme matching.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, more particularly, the present application relates to a binding scheme automatic matching method based on content attribute analysis. BACKGROUND

[0002] The current document binding scheme is mainly determined by two existing technical solutions: one is the artificial decision mode, that is, the staff determines and selects the binding mode according to personal experience combined with the basic physical properties of the document such as the number of pages and size; the second is a simple rule matching mode, which realizes the preliminary screening of the binding scheme by setting a fixed rule, such as setting a fixed page number threshold, and adopting glue binding when the threshold is exceeded and saddle stitching when the threshold is below. The above two solutions are based on the physical properties of the document as the core judgment basis, and the content properties of the document such as the text theme, visual style and domain attribute are not considered, which can only meet the basic binding requirements.

[0003] However, in actual use, there are still some shortcomings, such as:

[0004] 1. Low degree of automation, relying on artificial decision, not only time-consuming and laborious, but also prone to different binding schemes for the same type of document due to differences in personnel experience, and poor consistency; based on simple rule matching, the rule flexibility is insufficient, and it cannot meet the diversified document requirements.

[0005] 2. Insufficient content adaptability, only focusing on physical properties such as page number and size, ignoring core content properties such as document text theme, visual features and domain attributes, resulting in mismatch between binding scheme and document use scenario.

[0006] 3. Single decision dimension, without integrating multi-dimensional data for comprehensive judgment, only relying on single or a small number of physical indicators to determine the scheme, which is difficult to cover the complex attribute characteristics of the document, and is prone to unreasonable scheme selection. SUMMARY

[0007] In order to overcome the above-mentioned defects of the prior art, the present application provides a binding scheme automatic matching method based on content attribute analysis, which solves the problems in the background art by the following scheme.

[0008] In order to achieve the above-mentioned purpose, the present application provides the following technical scheme: a binding scheme automatic matching method based on content attribute analysis, comprising: S1: multi-dimensional data acquisition: synchronously acquiring text, visual, language style and domain knowledge graph data of the document through multi-modal technology, based on the acquired data, performing text feature processing, obtaining visual feature vector, paragraph feature set, and constructing domain knowledge graph;

[0009] S2: Multimodal attribute analysis: Based on S1, text topic analysis, visual attribute analysis, language style analysis, and field correlation analysis are carried out respectively, and relevant modal features are generated;

[0010] S3: Fusion decision: The mutual information of each modal feature is calculated through an attention mechanism to allocate weights, and the four-dimensional features after analysis are fused into a comprehensive feature vector; based on the vector, the probability of the core attribute of the document is calculated through a Bayesian model, and the attribute result is obtained by comparing with the preset threshold;

[0011] S4: Scheme matching: The attribute result is matched with the preset binding scheme, and after filtering the conflicting schemes, the optimal binding scheme is generated.

[0012] Preferably, the multi-dimensional data collection includes: text content collection, visual element collection, language style feature collection, and domain knowledge graph data collection;

[0013] Text content collection: By recognizing the document text, combining word segmentation technology, de-duplication, extracting core words and semantic association, a set of evaluation words is obtained , is the nth evaluation word; and the number of occurrences of the nth evaluation word in the text is counted, denoted as , is the number of occurrences of the nth evaluation word; the key sentences of the document are traversed, and the number of occurrences of the evaluation word in the key sentences is counted to obtain the frequency of the evaluation word in the key sentences, denoted as , is the frequency of the nth evaluation word in the key sentences;

[0014] Visual element collection: an image segmentation model is used to batch process all pages of the document, recognize illustration and icon visual elements, count the closed contours of each element, sum to obtain the total contour number, denoted as c; count the proportion of pixels with saturation greater than 0.7 to obtain the high saturation proportion, denoted as r; calculate the ratio of the area of the visual elements to the area of the page to obtain the picture-text interaction ratio, denoted as p;

[0015] Visual feature conversion: the total contour number is logarithmically transformed, and the saturation color proportion is normalized: , and the picture-text interaction ratio is mapped: , to obtain the visual feature vector ;

[0016] Language style feature collection: the average sentence length of each paragraph and the next paragraph is calculated, and the set is denoted as , where represents the average sentence length difference value of the mth paragraph; the passive sentence proportion is counted, denoted as ; The mean value is calculated by using a 10-section sliding window, and the specific calculation is as follows: , wherein q is the paragraph number, if the total number of paragraphs is less than 10, the sliding window uses all paragraphs, h is the paragraph number in the sliding window; and the mean value set of the passive sentence proportion of all paragraphs is denoted as , wherein represents the passive sentence proportion of the mth paragraph; the difficulty level of all to-be-evaluated words is obtained and mapped into an embedding vector, and the difficulty embedding vector of each paragraph is aggregated, for each paragraph, the subset of words contained in the paragraph is extracted , and the mean value of the embedding vector of the to-be-evaluated words in the subset of words is taken: , wherein is the average word difficulty embedding vector of the mth paragraph, is the subset of to-be-evaluated words contained in the mth paragraph, is the number of words in , and is the difficulty embedding vector of the to-be-evaluated word , e is the ordinal number;

[0017] For each paragraph, the average sentence length difference value, the passive sentence proportion, and the paragraph-level difficulty embedding vector are aggregated and combined into a paragraph feature triple: , and all paragraph features are combined into a set: , wherein is the paragraph feature triple of the mth paragraph;

[0018] Domain knowledge graph data collection: core terms related to the to-be-evaluated words are extracted from professional databases, and the extracted terms are subjected to data cleaning processing operations such as deduplication, standardization, and screening to obtain a term set , wherein represents the tth cleaned domain core term, t is the ordinal number; based on the cleaned term set D, each term is mapped into an independent node in the knowledge graph, and a node set is constructed; the mapping rule is that each node corresponds to a term one by one, and the node attributes include the basic information of the term; the logical relationship between the terms in the term set D is analyzed, the connection relationship of the nodes in the knowledge graph is defined, and an edge set is formed, s is the number of edges; wherein the relationship types include: hierarchical relationship, association relationship, causal relationship; each edge needs to specify the two nodes connected and the relationship type.

[0019] Preferably, the text topic analysis comprises:

[0020] The position entropy of each to-be-evaluated word is calculated based on ​, n is the number of words to be evaluated, and position entropy reflects the distribution dispersion of the words to be evaluated in the key sentence; three features are integrated by an exponential function: word occurrence frequency , semantic distance , and position entropy . Finally, the result is constrained in the interval [0, 1] by a normalization constant Z to construct a topic contribution function, and the specific mathematical function is as follows:

[0021]

[0022] Where t is the ordinal number, is a set of hyperparameters, Z is a normalization constant, , and n is the number of words to be evaluated;

[0023] Based on the topic contribution function, a set of topic concepts is generated , wherein represents the contribution of the word to be evaluated to the topic.

[0024] Preferably, the visual attribute analysis includes:

[0025] a visual feature vector , and a predefined K visual attribute category, a complex distribution of visual features is fitted by weighted summation of multiple Gaussian distributions corresponding to different visual styles, and a GMM model function is constructed, and the specific mathematical function is as follows:

[0026]

[0027] Where is a set of GMM model parameters, which is updated by iterative calculation through EM algorithm; is the prior probability of the kth category, and ; is the mean vector of the kth category; is the covariance matrix of the kth category; is the Gaussian distribution density function;

[0028] For the feature vector of the current document, based on the converged parameters, the posterior probability of belonging to each category is calculated by Bayes formula: , wherein is the posterior probability of the kth category, is the ordinal number;

[0029] The visual attribute probability is obtained, and the output visual attribute probability is a K-dimensional vector: , wherein represents the probability that the current document belongs to the Kth visual attribute.

[0030] Preferably, the language style analysis includes:

[0031] Predefined language style categories: defined according to business scenarios a language style category, is a positive integer;

[0032] Build a bidirectional LSTM network function: capture the forward and reverse context of the sequence through bidirectional LSTM, and comprehensively extract style-related time sequence features. The specific mathematical function is as follows:

[0033]

[0034] wherein, : set of network parameters;

[0035] : bidirectional LSTM layer weight matrix, the matrix dimension ; : bidirectional LSTM layer bias vector, dimension: ; : fully connected layer weight matrix, dimension: ; : fully connected layer bias vector, dimension: ; : softmax activation function; : output of bidirectional LSTM layer on paragraph language feature sequence, dimension: ;

[0036] For , , use normal distribution N(0, 0.01) for random initialization; for , , initialize to all-0 vector; based on the function output, get a set of language style probabilities , each element is dimensional probability vector.

[0037] Preferably, the field relevance analysis includes:

[0038] Graph structure definition: with as nodes, as edges, build an undirected graph ;

[0039] Original adjacency matrix AB construction: initialize a zero matrix AB of y*y; if node and have an edge , set , indicating that the two nodes are related; if there is no edge, set , indicating no relation.

[0040] Normalized adjacency matrix Construction: To avoid the imbalance of relationship weight caused by the difference of node degree, the degree matrix is introduced , is a y*y diagonal matrix, the diagonal elements , that is, the degree of node ; through the formula Normalization: I is a t*t unit matrix is the inverse square root of the degree matrix, which is used to balance the weight of nodes with different degrees, and i and j are ordinal numbers

[0041] Using a pre-trained word embedding model, each term in the term set D is converted into a fixed-dimensional real number vector as the initial embedding of the corresponding node; integrate the initial embeddings of all nodes to form an initial embedding matrix of y*d as the input feature matrix of GCN

[0042] Initial embedding matrix Input the trained GCN, and calculate the final layer node embedding through the formula iteration, get the final layer node embedding ; map the final node embedding to a scalar value , that is ; where Linear is a fully connected layer that converts high-dimensional embedding into a value in the interval [0, 1]; is an activation function; collect the correlation degrees of all matching terms to form a domain correlation degree set .

[0043] Preferably, the comprehensive feature vector comprises:

[0044] Based on four modal features, since the dimensions and ranges of each modal feature are different, the format needs to be unified first: for set type features, take the mean or maximum element to convert to a vector; L2 normalize all vectors to ensure that the scales of each modal feature are consistent, so that they are all mapped to the interval [0, 1];

[0045] Calculate the mutual information of any two modal features , based on mutual information to calculate the attention weight of each modal , the specific mathematical function is:

[0046]

[0047] where is the average feature of all modalities, and ; u1, u2 are ordinal numbers

[0048] The comprehensive feature vector VZ is generated by weighted sum of the normalized modality vectors with attention weights, and the formula is: Wherein is the normalized gth modality vector; and the L2 norm x of VZ is calculated.

[0049] Preferably, the calculation of the establishment probability of the document core attribute comprises:

[0050] The attribute set and the threshold set are defined as follows: The attribute set AR: the core attributes related to the binding scheme are preset, and the dimension of the set is o; the threshold set: ;

[0051] For each attribute The establishment probability thereof is calculated by the Bayes formula, represents the establishment of the attribute, represents the non-establishment, and the specific mathematical function is:

[0052]

[0053] Wherein, The prior probability of the attribute is calculated based on historical data, is the ordinal number; : the likelihood function of the fusion feature L2 norm x when the attribute is established; : the prior probability of the attribute not being established; : the likelihood function when the attribute is not established;

[0054] For each attribute , the posterior probability is compared with the threshold : if , the attribute is established, and is recorded as ; if , the attribute is not established, and is recorded as ; the judgment results of all attributes are integrated to form the attribute judgment result set .

[0055] Preferably, the scheme matching comprises:

[0056] The scheme attribute set is constructed: all attributes are preset with a binding scheme that meets the minimum requirement of the attribute, and a binding scheme set corresponding to the elements in the attribute set AR is obtained according to the binding scheme, and the dimension of the set is o;

[0057] The result set R is matched with the scheme set Matching, filtering out the scheme that conflicts with R, i.e. the operation of 0, deleting the corresponding scheme set element in the result set, and obtaining the subset of the binding scheme set that meets the result set

[0058] Generating the optimal binding scheme: the binding scheme meets all the requirements of each element in the binding scheme subset, and collects the page number, size and other special printing requirements, and generates the final binding scheme based on all the above conditions.

[0059] Technical effects and advantages of the present application:

[0060] 1. Automatic efficiency improvement: the present application automatically completes document data collection, attribute analysis, feature fusion and scheme matching through multi-modal technology, without human intervention throughout, which not only replaces the shortcomings of traditional manual decision-making, but also breaks through the rigid restrictions of simple rule matching mode; at the same time, it eliminates the scheme deviation caused by personnel experience difference, so that the binding scheme of similar documents remains highly consistent, effectively solving the problems of low automation and poor consistency of existing technology.

[0061] 2. Content attribute adaptation: unlike existing technology which only focuses on physical attributes such as document page number and size, the present application includes core content attributes such as text theme, visual style, language style and domain attribute in the analysis category, and through quantitative analysis of document content features, the binding scheme is deeply matched with the actual use scene of the document, solving the technical defects of insufficient content adaptability and disconnection between scheme and scene of existing technology.

[0062] 3. Intelligent fusion of multi-modal features: the present application uses attention mechanism to dynamically allocate the weights of multi-modal features such as text, visual, language style and domain knowledge, calculates the probability of core attribute establishment based on Bayesian model, and constructs a multi-dimensional comprehensive decision-making system; this system can fully tap the correlation value of each modal data, avoid the limitations of single physical index decision-making in existing technology which is difficult to cover complex document attributes, and significantly improve the accuracy of binding scheme selection.

[0063] 4. Flexible and expandable scheme matching, adapting to diversified needs: the present application presets a binding scheme library corresponding to document attributes, can dynamically adjust the scheme combined with document page number, size and special printing requirements, and supports optimizing model parameters through historical data; compared with the fixed rules of existing technology which are difficult to adapt to diversified documents, the present application can flexibly cope with different types and needs of document binding scenes, and has strong expandability and practicality. BRIEF DESCRIPTION OF DRAWINGS

[0064] Fig. 1 It is a schematic diagram of the overall structure of the present application.

[0065] Fig. 2 ​The figure is a schematic diagram of the whole process of the present application.

[0066] Fig. 3 The figure is a schematic diagram of the collection process of the present application. DETAILED DESCRIPTION

[0067] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of protection of the present application.

[0068] Reference Figs. 1-3 The figure is a schematic diagram of the whole process of the present application.

[0069] S1: Multi-dimensional data collection: synchronously acquiring the text, vision, language style and field knowledge graph data of the document through multi-modal technology, and based on the collected data, performing text feature processing, acquiring a vision feature vector, a paragraph feature set and constructing a field knowledge graph;

[0070] S2: Multi-modal attribute analysis: based on the above-mentioned multi-dimensional data, respectively performing text theme analysis, vision attribute analysis, language style analysis and field correlation analysis, and generating relevant modal features;

[0071] S3: Fusion decision and scheme matching: calculating mutual information of each modal feature through an attention mechanism to allocate weights, fusing the analyzed four-dimensional features into a comprehensive feature vector, and based on the vector, calculating a core attribute establishment probability of the document through a Bayesian model, and comparing the attribute result with a preset threshold to obtain an attribute result;

[0072] S4: matching the attribute result with a preset binding scheme, filtering conflict schemes, and generating an optimal binding scheme.

[0073] The specific analysis is as follows:

[0074] S1: Multi-dimensional data collection: synchronously acquiring the text, vision, language style and field knowledge of the document through multi-modal technology, and based on the collected data, providing structured raw materials for subsequent attribute analysis; wherein, the multi-data items are presented in a set form, and the single data items are presented in a numerical form, and the specific contents are as follows:

[0075] S101: Text content collection: comprehensively extracting and characterizing the document text information, acquiring core data related to the text theme through OCR recognition, word segmentation statistics and semantic coding, including the word itself, the appearance frequency, the key sentence distribution and the semantic correlation, providing structured input for theme analysis, and the specific steps are as follows:

[0076] Text extraction: For document scans, use OCR technology to recognize and convert images into editable text; for electronic documents, extract text through API interface; use jieba library to divide the extracted text into a set of evaluation vocabularies, denoted as , is the nth evaluation vocabulary; and count the number of occurrences of the nth evaluation vocabulary in the text, denoted as , is the number of occurrences of the nth evaluation vocabulary; traverse the key sentences of the document and count the number of occurrences of the evaluation vocabulary in the key sentences to obtain the frequency of the evaluation vocabulary in the key sentences, denoted as , is the frequency of the nth evaluation vocabulary in the key sentences.

[0077] It needs to be further explained that the key sentences of the document are determined by the following method:

[0078] 1) First, use the cross-format text extraction library to automatically parse the electronic document and output it as a string of pure text;

[0079] 2) Then, perform basic cleaning on the output pure text, replacing all carriage returns, tabs, and carriage return linefeed combination symbols with linefeeds, and removing consecutive redundant spaces while preserving single spaces;

[0080] 3) Treat each linefeed as a paragraph separator, and the text information before each paragraph separator as a complete and separate paragraph;

[0081] 4) Traverse from the first character of each separate paragraph until the first period, question mark, or exclamation mark appears, and end the operation, taking the string as the first sentence of the paragraph and the first sentence of the paragraph as the key sentence, and repeating the operation for each paragraph;

[0082] Text feature processing: use the BERT model to encode each evaluation vocabulary in the evaluation vocabulary set and the article title into semantic vectors and title vectors, and calculate the Mahalanobis distance between the semantic vectors and the title vectors to obtain a set of semantic Mahalanobis distances , is the Mahalanobis distance between the semantic vector and the title vector of the nth evaluation vocabulary.

[0083] S102: Visual element collection: Systematically collect and standardize the visual features of the document, and through image segmentation, color analysis, and proportion calculation techniques, convert the outline, color, and text-image interaction of the visual elements into quantifiable feature vectors to provide uniform dimensional input data for visual attribute clustering analysis.

[0084] Contour and color collection: U-Net image segmentation model is used to process all pages of the document in batches, identify visual elements such as illustrations and icons, count the closed contours of each element, sum up the total contour number, and record it as c; the proportion of pixels with saturation greater than 0.7 in the HSV space is calculated using OpenCV to obtain the high saturation proportion, and it is recorded as r; the ratio of the area of the visual element to the area of the page is calculated to obtain the graphic-text interaction ratio, and it is recorded as p.

[0085] Visual feature conversion: the total contour number is logarithmically transformed, and the saturation color proportion is normalized: , and the graphic-text interaction ratio is mapped: , to obtain the visual feature vector ;

[0086] S103: Language style feature collection: capture the language style features of the text, and convert the rhythm, formality and complexity of the language into structured feature sequences through sentence length analysis, passive sentence statistics and vocabulary difficulty classification, to provide calculable input for language style classification and recognition.

[0087] Sentence length and passive sentence statistics: based on the method of judging key sentences in S101, the entire string format text is rescanned, the continuous string before the period, question mark or exclamation mark is marked as a complete sentence, and the total number of sentences and the length of each sentence in each paragraph are counted, the average sentence length of each paragraph is calculated, and the difference set of the average sentence length of each paragraph and the next paragraph is obtained. Since the last paragraph has no next paragraph, the difference between the average sentence length of the last paragraph and the previous paragraph is calculated, and the difference of the average sentence length of the last paragraph is calculated. The set is recorded as , where represents the average sentence length difference of the mth paragraph;

[0088] The number of times of appearing of each sentence in each paragraph is counted, and the total number of sentences in the paragraph is divided to obtain the passive sentence proportion, recorded as: ; the mean value is calculated using a 10-paragraph sliding window, which is specifically calculated as: , where q is the paragraph number, if the total number of paragraphs is less than 10, the sliding window uses all paragraphs, h is the paragraph number in the sliding window; and the mean value set of the passive sentence proportion of all paragraphs is recorded as , where represents the passive sentence proportion of the mth paragraph;

[0089] Vocabulary difficulty processing: all the to-be-evaluated vocabularies are classified according to the International Chinese Education Chinese Level Standard, and in this method, the 1-2 levels of the standard are set as the first level of difficulty, the 3-4 levels are divided into the second level of difficulty, the 5-6 levels are set as the third level of difficulty, the 7-8 levels are set as the fourth level of difficulty, and the 9 level is set as the fifth level of difficulty, forming a vocabulary difficulty set , wherein represents the difficulty level of the nth vocabulary in the to-be-evaluated vocabulary set; the difficulty level of each vocabulary is mapped into an embedding vector through a pre-trained Word2Vec model, and a set of embedding vectors is obtained, , is the embedding vector of the nth vocabulary; and the difficulty embedding vectors of each paragraph are aggregated, and for each paragraph, a subset of vocabularies contained therein is extracted , and the average of the embedding vectors corresponding to the to-be-evaluated vocabularies in the vocabulary subset is taken: , wherein is the average vocabulary difficulty embedding vector of the mth paragraph, is the subset of to-be-evaluated vocabularies contained in the mth paragraph, is the number of vocabularies in , and is the difficulty embedding vector of the to-be-evaluated vocabulary , and e is the ordinal number;

[0090] For each paragraph, the average sentence length difference, the passive sentence proportion, and the paragraph-level difficulty embedding vector are aggregated and combined into a paragraph feature triple: , and all paragraph features are combined into a set: , wherein is the paragraph feature triple of the mth paragraph;

[0091] S104: Domain knowledge graph data collection: build the association basis of document content and domain knowledge, extract and clean domain terms and term relationships from professional databases to form structured knowledge graph elements, and provide knowledge support for calculating the correlation degree of documents and various domains. The specific analysis is as follows:

[0092] Domain core term extraction and cleaning: core terms related to to-be-evaluated vocabularies are extracted from professional databases through crawler technology, and after data cleaning processing operations such as de-duplication, standardization, and screening on the extracted terms, a term set is obtained , wherein represents the tth cleaned domain core term, and t is the ordinal number;

[0093] Mapping core terms to knowledge graph nodes: based on the cleaned term set D, each term is mapped to an independent node in the knowledge graph to construct a node set ; mapping rule: each node is associated with a term One-to-one correspondence, node attributes contain basic information of terms;

[0094] Building relationships between nodes: analyzing the logical relationships between terms in the term set D, defining the connection relationship of the nodes in the knowledge graph, and forming the edge set , s is the number of edges; wherein the relationship type includes: hierarchical relationship, association relationship, causal relationship; each edge needs to specify the two nodes connected and the relationship type;

[0095] S2: Attribute analysis: based on the collected multi-dimensional data, the text theme, visual attribute, language style and field correlation degree of the document are analyzed in four dimensions. The process of extracting core features through single-modal quantitative model converts unstructured data into quantifiable indicators that can be fused, providing input for subsequent fusion decision-making. The specific implementation is as follows:

[0096] S201: Text theme analysis: quantifying the contribution of core words in the text to the document theme, outputting a set of theme probability.

[0097] Receive the collected data output by S101, including the word set W, the appearance frequency set A, the semantic Mahalanobis distance set , the key sentence appearance probability set ; based on Calculate the position entropy of each word to be evaluated: , n is the number of words to be evaluated, and the position entropy reflects the distribution dispersion of the words to be evaluated in the key sentence; integrate the three features: word appearance frequency , semantic distance (negative contribution, the farther the distance, the lower the weight), and position entropy (negative contribution, the more dispersed the distribution, the lower the weight), and finally normalize the result to the [0, 1] interval through the normalization constant Z to construct the theme contribution function. The specific mathematical function is as follows:

[0098]

[0099] Where t is the ordinal number, is a set of hyperparameters, Z is a normalization constant, , n is the number of words to be evaluated;

[0100] Based on the theme contribution function, generate a set of theme concepts , wherein , represents the contribution of the word to be evaluated to the theme.

[0101] S202: Visual attribute analysis: cluster visual features through Gaussian mixture model, output the posterior probability that the document belongs to different visual styles, the specific analysis is as follows:

[0102] receiving the visual feature vector output by S102 and predefining K visual attribute categories, fitting the complex distribution of visual features by weighted summation of multiple Gaussian distributions corresponding to different visual styles, constructing a GMM model function, the specific mathematical function of which is as follows:

[0103]

[0104] wherein is a set of GMM model parameters, which is updated by iterative calculation through the EM algorithm;

[0105] is the prior probability of the kth category, and reflects the frequency of occurrence of the category in historical samples;

[0106] is the mean vector of the kth category, describing the typical visual features of the category;

[0107] is the covariance matrix of the kth category, describing the dispersion degree of the features of the category, and the smaller the value, the more stable the features;

[0108] is the Gaussian distribution density function, quantifying the matching degree of the feature vector v and the typical features of the kth category, and the larger the value, the higher the matching degree;

[0109] It should be further explained that the parameter set in the GMM model function needs to be learned from historical visual feature samples through the EM algorithm, and the specific steps are as follows:

[0110] Initialize parameters: initialized as a uniform distribution ; randomly select k feature vectors of historical samples as initial values; initialized as a unit matrix with uniform feature distribution.

[0111] E step: calculate the sample membership

[0112] set historical samples, denoted as , for each historical sample, calculate the posterior probability that it belongs to the kth category:

[0113]

[0114] wherein is the ordinal number, denotes the historical sample Probability of the k-th class, and .

[0115] M-step: update parameters

[0116] Based on posterior probability Update parameters to maximize the sample likelihood function:

[0117] Prior probability: (N is the total number of samples, reflecting the proportion of categories);

[0118] Mean vector: (weighted average of samples, close to the typical characteristics of the category);

[0119] Covariance matrix: (reflecting the dispersion of features within the category);

[0120] Iteration termination: repeat E-step and M-step until the parameter change is less than the threshold 10 -6 , get the converged parameter set .

[0121] For the feature vector of the current document, based on the converged parameters, calculate its posterior probability of belonging to each category by Bayes formula: , where is the posterior probability of the k-th class;

[0122] Get the visual attribute probability, then output the visual attribute probability is a K-dimensional vector: , where represents the probability that the current document belongs to the K-th visual attribute.

[0123] S203: Language style analysis: Through the four steps of presetting style categories, constructing a bidirectional LSTM network, training network learning rules, and reasoning the current document style, the language feature sequence collected in S103 is converted into quantified language style probability; The core logic is: use bidirectional LSTM to capture the time sequence dependence of language sequence, train the model through historical labeled data, and finally output the probability of each paragraph of the current document belonging to different language styles, providing language dimension features for cross-modal fusion, which is analyzed as follows:

[0124] Preset language style categories: defined according to business scenarios language style categories, is a positive integer;

[0125] Construct a bidirectional LSTM network function: capture the forward and reverse context of the sequence through bidirectional LSTM, and comprehensively extract style-related time sequence features, whose specific mathematical function is as follows:

[0126]

[0127] wherein, : set of network parameters, the core object of model learning;

[0128] : weight matrix of bi-LSTM layer, dimension , convert input 3D language feature vector into dimension hidden layer feature, weight value determines the contribution strength of each input feature to time series feature;

[0129] : bias vector of bi-LSTM layer (dimension: ), adjust the baseline value of hidden layer output, offset the overall offset of input features (such as the passive sentence proportion of all paragraphs is generally low, which is compensated by );

[0130] : weight matrix of full connection layer (dimension: ), map the dimension time series feature output by bi-LSTM into dimension style category score, weight value determines the influence degree of each time series feature on different styles;

[0131] : bias vector of full connection layer (dimension: ), adjust the baseline value of style category score, avoid score offset caused by uneven feature distribution;

[0132] : softmax activation function, normalize the full connection layer output category score to probability, ensure that the result can be directly used as style category probability;

[0133] : output of bi-LSTM layer to paragraph language feature sequence (dimension: ), fuse the forward and backward context information of paragraph language feature sequence, output a vector that can represent the time series feature of the paragraph.

[0134] , and , are randomly initialized with normal distribution N(0, 0.01); and , are initialized as all-0 vectors;

[0135] Based on the function output, a set of language style probabilities is obtained, each element being a dimension probability vector, quantifying the language style attributes of each paragraph in the current document.

[0136] S204: Field correlation analysis: learn the relationship characteristics of the terms in the field knowledge graph constructed in S104 through the graph neural network, quantify the correlation strength between the core terms of the document and the target field, and output a set of field correlation degrees to provide a quantitative basis for cross-modal fusion in the field dimension.

[0137] Convert the terms, nodes, and edges of S104 into graph structure data that can be processed by GCN. The key is to construct an adjacency matrix that quantifies the relationship strength between nodes to provide a foundation for subsequent feature propagation:

[0138] Graph structure definition: take as nodes, as edges, and construct an undirected graph .

[0139] Original adjacency matrix AB construction: initialize a zero matrix AB of y*y; if nodes and have an edge , set , indicating that the two nodes are related; if there is no edge, set , indicating no relationship.

[0140] Normalized adjacency matrix construction: To avoid relationship weight imbalance caused by node degree differences, introduce a degree matrix , which is a y*y diagonal matrix. The diagonal elements are the degrees of nodes ; normalize through the formula I is a t*t identity matrix; is the inverse square root of the degree matrix, used to balance the weights of nodes with different degrees, and i and j are ordinal numbers.

[0141] Use a pre-trained word embedding model to convert each term in the term set D into a fixed-dimensional real number vector as the initial embedding of the corresponding node; integrate the initial embeddings of all nodes to form an initial embedding matrix of y*d as the input feature matrix of GCN;

[0142] Input the trained GCN with the initial embedding matrix , and calculate the final layer node embedding through the formula ; map the final node embedding to a scalar value through a fully connected layer ; where Linear is a fully connected layer that converts high-dimensional embeddings to values in the [0, 1] interval; is the activation function; collect the correlation degrees of all matching terms to form a set of field correlation degrees .

[0143] S3: Fusion decision: integrate the text, visual, language, and field four-dimensional features parsed in S2 into a unified input through cross-modal feature fusion, then determine the core attributes of the document through attribute judgment, and finally match the optimal binding scheme based on the rule base. The specific steps are as follows:

[0144] S301: Multi-dimensional feature fusion: dynamically allocate the weight of each modality through attention mechanism, and fuse the four-modal features into a unified comprehensive feature vector;

[0145] Receive the four-modal features output by S2. Since the dimensions and ranges of each modality feature are different, the format needs to be unified first: for set-type features, take the mean or maximum element to convert to a vector; L2 normalize all vectors to ensure that the dimensions of each modality feature are consistent, so that they are all mapped to the [0, 1] interval;

[0146] Calculate the mutual information of any two modalities The higher the mutual information, the stronger the consistency of the two modalities, and the more reliable the contribution of this modality to attribute judgment; based on the mutual information, calculate the attention weight of each modality The specific mathematical function is:

[0147]

[0148] Where is the average feature of all modalities, and ; u1, u2 are ordinal numbers;

[0149] Weighted sum the standardized modal vectors through attention weights to generate a comprehensive feature vector VZ, the formula is: Where is the standardized gth modality vector; at the same time, calculate the L2 norm x of VZ to quantify the strength of the comprehensive feature, which is used for subsequent attribute judgment.

[0150] S302: Attribute judgment result generation: based on the strength x of the fusion feature, calculate the probability of each core attribute of the document through the Bayesian model, compare it with the preset threshold, and determine the attribute set of the document to provide decision basis for scheme matching; the specific analysis is as follows:

[0151] Define the attribute set and threshold set Attribute set AR: preset core attributes related to binding schemes (such as ), the dimension of this set is o; threshold set : set the minimum probability threshold for the establishment of each attribute (such as ), the threshold is set based on historical binding requirements and production experience;

[0152] For each attribute Calculate its probability of being true by Bayes formula, Denote attribute true, Denote attribute false, its specific mathematical function is:

[0153]

[0154] Where, Denote as ordinal, Attribute Prior probability is calculated based on historical data; : Attribute True, fusion feature L2 norm x likelihood function; : Attribute False prior probability; : Attribute False likelihood function;

[0155] For each attribute , compare the posterior probability With threshold : If , attribute true, record ; If , attribute false, record ; Integration of all attribute judgment results, form attribute judgment result set (As Academic attribute true, child attribute false, chemical attribute true);

[0156] S4: Scheme matching

[0157] Build scheme attribute set : All attributes are preset to meet the minimum requirements of the binding scheme, and according to the binding scheme, an element corresponding to the attribute set AR is obtained Binding scheme set , the set dimension is o;

[0158] Match the result set R with the scheme set Filter out the schemes that conflict with R, that is, go 0 operation, delete the corresponding scheme set elements in the result set, get the binding scheme set subset that meets the result set ;

[0159] Generate the optimal binding scheme: the binding scheme meets all the requirements of each element in the binding scheme subset, and collects the number of pages, size and other special printing requirements, combined with all the above conditions, to generate the final binding scheme.

[0160] Secondly: the embodiment of the present application discloses only the structure related to the embodiment of the present application, other structures can refer to the general design, and in the case of no conflict, the same embodiment and different embodiments of the present application can be combined with each other;

[0161] Finally: the above only for the preferred embodiment of the present application, and does not limit the present application, any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included in the protection scope of the present application.

Claims

1. A method for automatically matching a binding scheme based on content attribute analysis, characterized by, Comprise: S1: multi-dimensional data acquisition: through multi-modal technology, the text, visual, language style and domain knowledge graph data of the document are synchronously acquired, based on the acquired data, text feature processing is carried out, visual feature vector is acquired, paragraph feature set is acquired, and domain knowledge graph is constructed; S2: multi-modal attribute analysis: based on S1, text theme analysis, visual attribute analysis, language style analysis and domain correlation analysis are carried out respectively, and relevant modal features are generated; S3: fusion decision: the mutual information of each modal feature is calculated through the attention mechanism to allocate weight, and the four-dimensional features after analysis are fused into a comprehensive feature vector; based on The comprehensive feature vector calculates the probability of the establishment of the core attribute of the document through the Bayesian model, and compares it with the preset threshold to obtain the attribute result; The calculation of the probability of the establishment of the core attribute of the document comprises: Definition attribute set With attribute threshold set : attribute set and attribute threshold set dimension are o; for each attribute Calculate its probability of being true by Bayes formula, Indicates that the attribute is true, Indicates that it is not true, and its specific mathematical function is: wherein, : attribute the prior probability that the hypothesis is true, calculated based on historical data, is an ordinal number; : attribute the likelihood function of the fused feature L2 norm x when the hypothesis is true; : attribute the prior probability that the hypothesis is not true; : attribute the likelihood function when the hypothesis is not true; For each attribute , compare the posterior probability with a threshold : if , the attribute is true, record ; if , the attribute is false, record ; integrate the results of all attributes to form the attribute judgment result set ; S4: scheme matching: the attribute result is matched with the preset binding scheme, after filtering the conflict scheme, the optimal binding scheme is generated; The scheme matching comprises: The construction scheme attribute set : all attributes are preset with a binding scheme satisfying the minimum requirement of the attribute, and a binding scheme set corresponding to the elements in the attribute set AR is obtained according to the binding scheme , the dimension of the binding scheme set is o; The result set R is matched with the scheme set The schemes in conflict with R are filtered out, i.e. the go 0 operation, the element in the result set is 0, the corresponding scheme set element is deleted, and the binding scheme set subset satisfying the result set is obtained ; Generating the optimal binding scheme: the binding scheme meets all the requirements of each element in the binding scheme subset, and combines the page number and size data to generate the final binding scheme.

2. The method of claim 1, wherein: The multi-dimensional data acquisition comprises: text content acquisition, visual element acquisition, language style feature acquisition, and domain knowledge graph data acquisition; Text content collection: by identifying document text, combining word segmentation technology, deduplication, extracting core words and semantic association, obtaining a set of to-be-evaluated words , is the nth to-be-evaluated word; and the number of occurrences of each to-be-evaluated word in the text is counted, and the set of all to-be-evaluated word text occurrence numbers is denoted as , is the nth to-be-evaluated word; and the number of occurrences of each to-be-evaluated word in the text is counted, and the set of all to-be-evaluated word text occurrence numbers is denoted as , is the nth to-be-evaluated word in the key sentence; Visual element acquisition: an image segmentation model is used to batch process all pages of the document, identify the visual elements such as illustrations and icons, count the closed contours of each element, sum to obtain the total contour number, which is denoted as c; the proportion of pixels with saturation greater than 0.7 is counted to obtain the high saturation proportion, which is denoted as r; the ratio of the area of the visual element to the area of the page is calculated to obtain the graphic-text interaction ratio, which is denoted as p; Visual feature transformation: Logarithmic transformation is performed on the total contour number, and the saturated color proportion is normalized: , and the text interaction ratio is remapped: , to obtain the visual feature vector ; Language style feature collection: calculate the average sentence length of each paragraph and the next paragraph, and denote the set as wherein represents the average sentence length difference value of the mth paragraph; calculate the proportion of passive sentences, denoted as ; calculate the mean value using a 10-paragraph sliding window, which is calculated as follows: wherein q is the paragraph number, if the total number of paragraphs is less than 10, the sliding window uses all paragraphs, h is the paragraph number in the sliding window; and the mean value set of all paragraphs is denoted as wherein represents the proportion of passive sentences of the mth paragraph; obtain the difficulty level of all the words to be evaluated and map it to an embedding vector, aggregate the difficulty embedding vector of each paragraph, and extract the words to be evaluated in each paragraph, denoted as the subset of words to be evaluated , take the average of the embedding vectors of the words to be evaluated in the subset of words: wherein is the average word difficulty embedding vector of the mth paragraph, is the subset of words to be evaluated in the mth paragraph, is the number of words in , is the difficulty embedding vector of the word to be evaluated , and e is the ordinal number; For each paragraph, the average sentence length difference, the passive sentence proportion, and the paragraph-level difficulty embedding vector are combined into a paragraph feature triplet: And all the paragraph features are combined into a set: Where is the paragraph feature triplet of the mth paragraph. Domain knowledge graph data collection: extracting core terms related to the to-be-evaluated vocabulary from professional databases, and performing data cleaning processing operations such as deduplication, standardization, and screening on the extracted terms to obtain a term set , wherein represents the t-th cleaned domain core term, and t is a serial number; based on the cleaned term set D, each term is mapped to an independent node in the knowledge graph to construct a node set ; the mapping rule is that each node corresponds to a term one-to-one, and the node attributes include the basic information of the term; the logical relationship between the terms in the term set D is analyzed, the connection relationship of the nodes in the knowledge graph is defined, and an edge set is formed, and ks is the number of edges; wherein the relationship types include: hierarchical relationship, association relationship, and causal relationship; each edge needs to specify the two nodes connected and the relationship type.

3. The method of claim 1, wherein the method further comprises: determining a content attribute of the content; and determining the binding scheme based on the content attribute. The text theme analysis comprises: Based on The position entropy of each evaluated vocabulary is calculated: , n is the number of evaluated vocabularies, and the position entropy reflects the distribution dispersion of the evaluated vocabulary in the key sentence; three features are integrated by an exponential function: the number of vocabulary occurrences , semantic distance , and position entropy . Finally, the result is constrained in the [0, 1] interval by a normalization constant Z to construct a theme contribution function, and its specific mathematical function is as follows: wherein t is an ordinal number, is a set of hyperparameters, and Z is a normalization constant, n is the number of words to be evaluated. generating a topic concept set based on a topic contribution function wherein denotes a vocabulary to be evaluated contribution to a topic.

4. The method of claim 1, wherein: The visual attribute analysis comprises: visual feature vector And predefine K visual attribute categories, by multiple corresponding different visual style of Gaussian distribution weighted sum, fitting the complex distribution of visual features, construct GMM model function, its specific mathematical function as follows: wherein is a set of GMM model parameters, updated by iterative computation through EM algorithm; is the prior probability of the kth class, and ; is the mean vector of the kth class; is the covariance matrix of the kth class; is the Gaussian distribution density function; For the feature vector of the current document, based on the converged parameters, the posterior probability of its belonging to each class is calculated by Bayes formula: where is the posterior probability of the kth class, is the ordinal number; obtaining the visual attribute probability, the output visual attribute probability is a K-dimensional vector: wherein represents a probability that the current document belongs to the Kth visual attribute.

5. The method of claim 1, wherein: The language style analysis comprises: Predefined language style categories: defined according to business scenarios a language style category, is a positive integer; Constructing a bidirectional LSTM network function: through bidirectional LSTM, the forward and reverse contexts of the sequence are captured at the same time, and the timing features related to the style are comprehensively extracted, and the specific mathematical function is as follows: wherein : set of network parameters; : Bi-LSTM layer weight matrix, dimension ; : Bi-LSTM layer bias vector, dimension: ; : Fully connected layer weight matrix, dimension: ; : Fully connected layer bias vector, dimension: ; : Softmax activation function; : Bi-LSTM layer output for a sequence of paragraph linguistic features, dimension: ; right , The system is initialized using a normal distribution N(0, 0.01); , Initialize it as a vector of all zeros; obtain the language style probability set based on the output of this function. Each element is A probability vector of dimension.

6. The method of claim 1, wherein: The domain correlation analysis comprises: Graph structure definition: construct an undirected graph with as nodes, as edges. ; Construction of the original adjacency matrix AB: Initialize the zero matrix AB of y*y; if the node and There are edges Then let If there is no edge, then the two nodes are related; if there is no edge, then... , indicating irrelevant; Normalized adjacency matrix Construction: To avoid the imbalance of relationship weight caused by the difference of node degree, the degree matrix is introduced , is a y*y diagonal matrix, the diagonal elements , that is, the degree of node ; through the formula Normalization: I is a t*t unit matrix; is the inverse square root of the degree matrix, which is used to balance the weight of nodes with different degrees, and i and j are ordinal numbers; Each term in the term set D is converted into a fixed-dimension real number vector as the initial embedding of the corresponding node by using a pre-trained word embedding model; the initial embeddings of all nodes are integrated to form an initial embedding matrix of y*d as the input feature matrix of the GCN Initial embedding matrix Input the trained GCN, and calculate the node embedding of the final layer by formula Iterative calculation, get the final layer node embedding ; map the final node embedding to scalar value through the fully connected layer , i.e. ; where Linear is a fully connected layer that converts high-dimensional embedding into a value in the interval [0, 1]; and is an activation function; collect the correlation degrees of all matching terms to form a domain correlation degree set .​ 7. The method of claim 1, wherein: The comprehensive feature vector comprises: Based on the four modal features, since the dimensions and ranges of each modal feature are different, the format needs to be unified: for set type features, the mean value or the maximum element is converted into a vector; all vectors are L2 normalized to ensure that the modal feature scales are consistent, so that they are all mapped to the [0, 1] interval; computing mutual information of any two modalities computing attention weights for each modality based on mutual information with a specific mathematical function: wherein is the average feature for all modalities, and ; ui, u2are ordinal numbers; The comprehensive feature vector VZ is generated by weighted summation of the normalized modality vectors with attention weights, and the formula is as follows: wherein is the gth normalized modality vector; and the L2 norm x of VZ is calculated simultaneously.

Images