Intelligent document understanding method and system combining natural language processing and deep learning

Abstract

The invention provides an intelligent document understanding method and system combining natural language processing and deep learning, and relates to the technical field of natural language processing and information management. Fusing the semantic pre-annotation result and the original text feature of the document, inputting the fused semantic pre-annotation result and the original text feature of the document into a cross-modal semantic enhancement model to obtain a document enhanced semantic representation, performing bidirectional semantic interaction with a dynamic business knowledge network based on the document enhanced semantic representation, generating an associated interaction result, and constructing a structured semantic asset; finally, the structured semantic assets are input into an intelligent document application engine, semantic feedback data are collected to optimize model parameters, document understanding accuracy and practicability can be improved, and diversified business requirements are met.

Description

Technical Field

[0001] This invention relates to the fields of natural language processing and information management technology, and more specifically, to an intelligent document understanding method and system that combines natural language processing and deep learning. Background Technology

[0002] In today's era of digital information explosion, the number of unstructured documents is growing exponentially. These documents contain a wealth of valuable information, but how to efficiently and accurately understand and extract their semantic content has become a critical problem that urgently needs to be solved.

[0003] Traditional document understanding methods primarily rely on manual reading and annotation. These methods are not only inefficient but also susceptible to subjective influences, making it difficult to guarantee the consistency and accuracy of the understanding results. With the development of natural language processing technology, some rule-based and statistical methods have been applied to document understanding. However, these methods often only handle simple semantic relationships and have limited ability to understand complex semantic structures and deeper semantic levels.

[0004] In recent years, deep learning technology has made significant progress in the field of natural language processing. However, most existing deep learning-based document understanding methods only focus on the surface features of documents, such as vocabulary and grammar, while neglecting the deep semantic features and their connection with business knowledge. Furthermore, these methods typically lack effective feedback mechanisms, making it impossible to dynamically optimize the model based on actual application results, thus limiting the accuracy and practicality of document understanding. Summary of the Invention

[0005] In view of the aforementioned problems, and in conjunction with the first aspect of the present invention, embodiments of the present invention provide an intelligent document understanding method combining natural language processing and deep learning, the method comprising: Semantic pre-annotation is performed on the target unstructured document collection to generate semantic pre-annotation results containing document semantic boundaries, candidate semantic labels, and label confidence. The semantic pre-annotation results are fused with the original text features of the document and input into the cross-modal semantic enhancement model to perform semantic deepening encoding processing, resulting in a document-enhanced semantic representation that integrates surface text features and deep semantic features; Based on document-enhanced semantic representation and dynamic business knowledge network, bidirectional semantic interaction processing is performed to generate the association and interaction results between document semantics and business knowledge; Based on the results of related interactions, document semantic asset construction processing is performed to generate structured semantic assets containing entity association networks, event logic chains, and core viewpoint networks; Structured semantic assets are input into the intelligent document application engine, driving multiple application modules to collaboratively execute document understanding application operations. At the same time, semantic feedback data is collected during the application operation process, and the semantic feedback data is fed back into the cross-modal semantic enhancement model to complete parameter optimization.

[0006] In another aspect, embodiments of the present invention also provide an intelligent document understanding system combining natural language processing and deep learning, including a processor and a machine-readable storage medium connected to the processor. The machine-readable storage medium is used to store programs, instructions, or code, and the processor is used to run the programs, instructions, or code in the machine-readable storage medium to implement the above-described method.

[0007] Based on the above, this embodiment of the invention, through semantic pre-annotation processing, endows unstructured document sets with clear semantic boundaries, candidate semantic labels, and label confidence levels. The semantic pre-annotation results are then fused with the original text features of the documents and input into a cross-modal semantic enhancement model for semantic deepening encoding. This fully mines the surface and deep semantic features of the documents, generating highly integrated document-enhanced semantic representations, thereby capturing the semantic information of the documents more comprehensively and accurately. Based on the document-enhanced semantic representation and dynamic business knowledge network, bidirectional semantic interaction processing is performed, closely integrating document semantics with business knowledge. The generated associated interaction results better reflect the semantic value of the documents in actual business scenarios. Through document semantic asset construction processing, the associated interaction results are transformed into structured semantic assets containing entity association networks, event logic chains, and core viewpoint networks, achieving standardized and systematic storage of document semantics, facilitating subsequent querying and application. By inputting structured semantic assets into the intelligent document application engine, multiple application modules are driven to collaboratively execute document understanding application operations. Semantic feedback data is collected and fed back into the cross-modal semantic enhancement model to complete parameter optimization. The model parameters can be continuously adjusted according to the actual application effect, thereby improving the accuracy and practicality of document understanding and meeting the diverse needs of document understanding in different business scenarios. Attached Figure Description

[0008] Figure 1 This is a schematic diagram of the execution flow of the intelligent document understanding method combining natural language processing and deep learning provided in an embodiment of the present invention.

[0009] Figure 2 This is a schematic diagram of the hardware architecture of an intelligent document understanding system that combines natural language processing and deep learning, provided in an embodiment of the present invention. Detailed Implementation

[0010] The present invention will now be described in detail with reference to the accompanying drawings. Figure 1This is a flowchart illustrating an intelligent document understanding method combining natural language processing and deep learning, provided in one embodiment of the present invention. The following is a detailed description of this intelligent document understanding method combining natural language processing and deep learning.

[0011] Step S110: Perform semantic pre-annotation processing on the target unstructured document set to generate semantic pre-annotation results containing document semantic boundaries, candidate semantic labels, and label confidence.

[0012] This embodiment uses an enterprise financial and tax management system as the unified application scenario. The target unstructured document collection includes scanned copies of VAT invoices, photos of expense reimbursement forms, PDF bank statements, Word documents of contract terms, and other financial and tax-related documents. These documents are characterized by mixed formats, coexistence of handwritten annotations and printed text, and interweaving of numbers and text. Semantic pre-annotation can achieve preliminary structuring of financial and tax information.

[0013] Step S111: Read the original data of various documents in the target unstructured document collection, parse the text format, encoding type and content structure of the documents, convert them into standardized document text in a unified format, and generate a standardized document text collection.

[0014] The system utilizes the document access interface of the enterprise's financial and tax management system to read image data from scanned VAT invoices, expense reimbursement forms, bank receipts (PDF), and contract terms (DOCX). For image documents, an OCR engine is used for text extraction. The engine's character recognition library includes financial and tax-specific fonts and handwriting recognition modules, outputting OCR results containing text content, character coordinates, and confidence levels. For PDF documents, a text stream parsing mode is used to extract text blocks, vector graphics, and font metadata. For DOCX documents, XML-formatted text nodes and style definitions are directly read. During parsing, the system automatically detects document encoding types and uniformly converts multiple encoding formats to UTF-8 without BOM. When analyzing the content structure, the system identifies header and footer areas, table borders, and stamp positions in PDF documents, extracts multi-level heading styles and nested table structures from DOCX documents, and finally converts all documents into standardized JSON-formatted text containing text content, paragraph boundaries, font sizes, and table cell coordinates, forming a standardized text collection.

[0015] Step S112: Perform semantic boundary recognition processing on each document text in the standardized document text set. Based on semantic pause features, syntactic structure features and discourse cohesion features in natural language processing, divide the document text into multiple independent semantic segment units and generate a document semantic segment set.

[0016] For the standardized VAT invoice text, semantic boundary recognition is performed to distinguish semantic units such as invoice header, taxpayer identification number, goods name, and amount.

[0017] Step S1121: Perform word segmentation on each document text in the standardized document text set. Use word segmentation algorithms in natural language processing to divide the document text into multiple lexical units and generate a document lexical unit set.

[0018] A word segmentation algorithm based on bidirectional maximum matching is used, combined with a tax and finance domain dictionary (containing professional terms such as "value-added tax invoice", "input tax", and "tax-exclusive amount") to segment the document text. For example, the invoice text "Company Name, Taxpayer Identification Number, Goods Name: Quantity A, Unit Price B" is segmented to obtain lexical units including "Company Name", "Taxpayer Identification Number", "Goods Name", "Goods", "Quantity", "Unit A", "Unit Price", and "B Yuan", forming a set of document lexical units.

[0019] Step S1122: Extract semantic pause features from the document vocabulary unit set. The semantic pause features include the distribution of punctuation marks, the position of modal particles, and the frequency of semantic transition words, and generate a document semantic pause feature set.

[0020] The punctuation marks in the above word segmentation results are analyzed. For example, the colon is used as a key-value pair separator in "Goods Name: Commodity", the comma is used to separate different attribute fields, and the period indicates the end of a paragraph. Modal particles appear very rarely in financial and tax documents; the focus is mainly on semantic transition words, such as the position of "but" and "however" in contract clauses. For example, in "This contract amount includes taxes and fees, but does not include transportation costs," "but" indicates a shift in the scope of costs. The coordinate positions and types of punctuation marks, as well as the index sequences of transition words, are recorded as a set of document semantic pause features.

[0021] Step S1123: Perform syntactic analysis on the document text, identify the syntactic dependency structure, subject-verb-object relationship and modification relationship of each sentence, and generate a set of document syntactic structure features.

[0022] The document text is processed using a dependency parsing model. For example, analyzing the sentence "Company Name purchased Unit A of Goods", the model identifies the subject-verb relationship "Company Name - Purchase", the verb-object relationship "Purchase - Goods", and the quantity modification relationship "Goods - Unit A". For tax-specific sentences such as "Total Price and Tax = Amount + Tax", the model identifies mathematical relationships and marks the computational dependencies between "Total Price and Tax" and "Amount" and "Tax". These syntactic relationships are stored as a set of document syntactic structure features in the form of triples.

[0023] Step S1124: Extract the text coherence features from the document text. The text coherence features include the type of conjunctions used, the distribution of referential relationships, and the consistency of paragraph themes, and generate a set of document text coherence features.

[0024] In the contract terms document, sequence conjunctions such as "firstly," "secondly," and "again," conditional conjunctions such as "if...then...", and referential relationships in phrases like "the aforementioned amount" and "the tax rate" are extracted, where "the aforementioned amount" refers to the previously mentioned "total contract amount C yuan." Thematic consistency is assessed by calculating the TF-IDF similarity of words within paragraphs; for example, words like "airfare," "high-speed rail ticket," and "taxi ticket" within the "transportation expenses" paragraph of an expense reimbursement form show high thematic consistency. The conjunction types, referential chains, and thematic similarity values ​​are recorded as a set of document cohesion features.

[0025] Step S1125: Construct a semantic boundary recognition model by inputting the document semantic pause feature set, document syntactic structure feature set, and document chapter connection feature set into the model, configuring the model's recognition parameters, and generating a semantic boundary recognition parameter set.

[0026] A semantic boundary recognition model based on BiLSTM-CRF is constructed. The input layer receives punctuation embedding vectors from the document semantic pause feature set, dependency embedding vectors from the document syntactic structure feature set, and topic similarity feature vectors from the document coherence feature set. The model is configured with D LSTM units, the CRF layer transition matrix dimension is the number of document categories, the dropout ratio is set to E, the learning rate uses a cosine annealing strategy, and the initial value is set to F. These parameters are encapsulated into a semantic boundary recognition parameter set.

[0027] Step S1126: Based on the semantic boundary recognition parameter set, the semantic boundary recognition model performs semantic boundary prediction processing on the document text, outputs the potential semantic boundary positions in the document text, and generates a set of predicted semantic boundary positions.

[0028] The character sequences of standardized document text are converted into character embedding vectors, which are then concatenated with semantic pause features, syntactic structure features, and discourse cohesion features before being input into a semantic boundary recognition model. The model captures contextual semantic information through an LSTM layer and calculates the conditional probability of the label sequence through a CRF layer, outputting a prediction of whether each character position is a boundary. For example, in invoice text, it predicts that the semantic boundaries are between "taxpayer identification number" and subsequent codes, and between "goods name" and "commodity," generating a set of predicted semantic boundary positions containing the coordinates of these boundary locations.

[0029] Step S1127: Based on the preset minimum semantic unit length threshold, remove the boundary positions in the predicted semantic boundary position set whose mutual spacing is less than the minimum semantic unit length threshold, and generate a valid semantic boundary position set.

[0030] Set the minimum semantic unit length threshold to G characters (for short fields in financial and tax documents such as "tax rate" and "tax amount"), and check the set of predicted semantic boundary positions. If there are only H characters between two consecutive predicted boundaries (such as a misjudged space boundary), then the latter boundary position is removed. For example, in "Amount: B yuan", if there is a boundary between "gold" and "amount", it is removed because the distance is less than the threshold, and the valid boundary between "Amount:" and "B yuan" is retained to form the set of valid semantic boundary positions.

[0031] Step S1128: Based on the set of effective semantic boundary positions, divide the document text into multiple continuous text segments, each text segment corresponding to an independent semantic expression unit, and generate an initial set of semantic segments.

[0032] Based on the coordinate information in the set of valid semantic boundary locations, the standardized document text is sliced. For example, the text of a value-added tax invoice is divided into text fragments such as "Invoice Code: XXXXXXXXXXXX", "Invoice Number: XXXXXXXX", "Invoice Date: YYYY Year MM Month DD Day", and "Buyer: Company Name". Each fragment expresses a complete financial and tax attribute, generating an initial set of semantic fragments.

[0033] Step S1129: Merge the semantically incomplete segments in the initial semantic segment set with adjacent segments to generate a complete semantic segment set.

[0034] The initial set of semantic fragments is examined. If a fragment only contains "total price including tax" but lacks the specific amount, it is merged with the subsequent amount fragments to form "total price including tax: D yuan". For cases where "tax amount" is separated from the specific value due to line breaks, the same merging operation is performed to ensure that each fragment contains complete key-value pairs or semantic units, forming a complete set of semantic fragments.

[0035] Step S11210: Add a unique identifier and position index to each semantic fragment in the complete semantic fragment set to generate a document semantic fragment set.

[0036] Each segment in the complete semantic segment set is assigned a UUID as a unique identifier, recording its positional information in the original document, such as page number, top-left corner coordinates, and bottom-right corner coordinates. For example, the identifier for the segment "Goods Name: Commodity" is "uuid-I", and its position index is "Invoice text area, page number J, coordinates (x1, y1, x2, y2)". The segment text, unique identifier, and position index are stored together to generate a document semantic segment set.

[0037] Step S113: Extract the core word sequence and syntactic dependency relations in each semantic segment unit. Based on the preset semantic tag system, match the corresponding candidate semantic tags for each semantic segment unit. The candidate semantic tags include entity tags, relation tags, event tags and opinion tags, and generate a set of candidate semantic segment tags.

[0038] The core vocabulary sequence "total price and tax, D yuan" is extracted from the semantic fragment "total price and tax: D yuan", with the syntactic dependency relationship being "total price and tax - value - D yuan". A pre-defined semantic tagging system for the financial and tax domain is established. Entity tags include "[amount]", "[date]", "[tax rate]", "[taxpayer identification number]", etc.; relation tags include "[equal to]", "[includes]", "[attribution]", etc.; event tags include "[reimbursement]", "[deduction]", "[declaration]", etc.; and opinion tags include "[compliance]", "[abnormal]", "[pending review]", etc. Candidate entity tags "[amount]" and relation tags "[equal to]" are matched for this fragment, generating a candidate set of semantic fragment tags.

[0039] Step S114: Construct a semantic tag matching degree calculation model, input the core word sequence, syntactic dependency relation and candidate semantic tags of the semantic segment unit, perform tag matching degree analysis, and generate the matching degree score corresponding to each candidate semantic tag.

[0040] A BERT-based semantic matching model is constructed, converting the core word sequence "total price and tax, D yuan" into word vectors and calculating their similarity with the embedding vectors of the candidate label "[amount]". Simultaneously, the matching score is weighted by incorporating the syntactic dependency relation "numerical". If the syntactic relation matches the relation type defined by the label (e.g., the "amount" label expects a numerical dependency relation), the matching score is increased. The model outputs the matching score for each candidate label; for example, "[amount]" scores K, and "[date]" scores L.

[0041] Step S115: Normalize the matching score to generate the tag confidence score corresponding to each candidate semantic tag, forming a set of semantic fragment tag confidence scores.

[0042] The Min-Max normalization method is used to map the matching scores to the [0,1] interval. For example, the K score for "[Amount]" is normalized to M, and the matching score N for the "[Equal to]" label is normalized to O. The normalized confidence values ​​are then mapped one-to-one with the labels to form a set of semantic fragment label confidence scores.

[0043] Step S116: Based on the label confidence, sort the candidate semantic labels of each semantic segment unit in descending order to generate a sorted semantic label list. Extract the candidate semantic label with the highest label confidence from the sorted semantic label list as the main label, and the remaining candidate semantic labels as auxiliary labels to generate a set of main and auxiliary labels for the semantic segment.

[0044] Candidate labels for the "Total Price and Tax" segment are sorted by confidence level: [Amount] (M), [Equal to] (O), and [Financial Indicator] (P). "[Amount]" is selected as the primary label, and "[Equal to]" and "[Financial Indicator]" are selected as secondary labels. The same operation is performed on all semantic segments to generate a set of primary and secondary labels for each semantic segment.

[0045] Step S117: Associate and integrate the boundary location information, primary and secondary label information, and label confidence information of each semantic segment unit to generate the semantic pre-annotation intermediate result of a single document.

[0046] The boundary coordinates (x1, y1, x2, y2), main label "[Amount]", auxiliary label list ("[Equal to]", "[Financial Indicator]"), and corresponding confidence scores (M, O, P) of the "Total Price and Tax" segment are integrated into a JSON object containing fields such as "segment_id", "bounding_box", "main_label", "auxiliary_labels", and "confidence_scores". This integration operation is performed on all semantic segments in the document to generate intermediate semantic pre-annotation results for a single document.

[0047] Step S118: Perform format unification processing on the intermediate results of semantic pre-annotation for all documents, standardize the description format of semantic boundaries, the naming convention of tags, and the representation of confidence, and generate a set of document semantic pre-annotations.

[0048] The unified boundary location format is "page number, top-left x, top-left y, bottom-right x, bottom-right y", and label naming uses the format "[Chinese label name]", with confidence scores retained to two decimal places. For example, "confidence:Q" is standardized to "confidence:R", and "label:money" is unified to "label:[amount]". Format validation and conversion are performed on the intermediate pre-annotation results of all documents to generate a set of pre-annotated semantic documents.

[0049] Step S119: Associate and bind the document semantic pre-annotation set with the corresponding standardized document text set to generate a semantic pre-annotation result containing document semantic boundaries, candidate semantic labels and label confidence.

[0050] By associating the set of semantically pre-annotated documents with a set of standardized document text using unique document identifiers, an index mapping from standardized text to semantic annotations is established. The final semantically pre-annotated result includes document-level metadata (filename, upload time), text content, and an array of semantic fragments (each fragment includes boundaries, labels, and confidence scores), stored in JSONLines format, supporting subsequent module calls in the enterprise's financial and tax management system.

[0051] Step S120: The semantic pre-annotation results are fused with the original text features of the document, and the input is given to the cross-modal semantic enhancement model to perform semantic deepening encoding processing, so as to obtain the document enhanced semantic representation that integrates the surface features and deep semantic features of the text.

[0052] In the enterprise financial and tax management system, the semantic pre-annotation results of VAT invoices (including labels such as [amount], [tax rate] and boundary information) are fused with the original OCR text features (font size, character color, spatial layout). Through cross-modal modeling, the association between the surface visual features and deep semantics of financial and tax information is learned. For example, "text in the red stamp area" usually corresponds to the entity "invoice special stamp", and its semantic importance is higher than that of ordinary text.

[0053] Step S121: Extract the main label features, auxiliary label features and label confidence features of each semantic segment unit from the semantic pre-annotation results, convert them into semantic label feature vectors in vector format, and generate a set of semantic label feature vectors.

[0054] Extract the one-hot encoding vector of the main label "[Amount]" (dimension is the total number of categories in the label system), the multi-hot encoding vectors of the auxiliary labels "[Equal to]" and "[Financial Indicator]", and the numerical vector of the confidence score M. Concatenate these three vectors into a semantic label feature vector of fixed length. For example, the main label vector (1,0,...0), the auxiliary label vector (0,1,1,...0), and the confidence score vector (M) are concatenated to form a vector of dimension (S+T+1) (S is the number of main labels, and T is the number of auxiliary labels). The vectors of all semantic segments constitute the semantic label feature vector set.

[0055] Step S122: Perform text surface feature extraction processing on each document text in the standardized document text set, extracting lexical features, syntactic features and discourse structure features of the text. Lexical features include part-of-speech distribution, word length distribution and word frequency. Syntactic features include syntactic dependency structure and sentence type. Discourse structure features include paragraph connection mode and sentence association type, generating a text surface feature set.

[0056] Lexical features were extracted from the standardized invoice text: statistical part-of-speech distribution (noun percentage U, verb percentage V, numeral percentage W), word length distribution (average word length X characters, longest word Y characters), and word frequency ("invoice" appears Z times, "amount" appears AA times). Syntactic features included dependency triples and sentence type (declarative sentences AB%, imperative sentences AC%). Discourse structure features were also extracted: paragraphs were separated by semicolons (AD%), and sentence relationships were predominantly parallel (AE%). These features were then quantified to generate a set of surface-level text features.

[0057] Step S123: Convert the text surface feature set into a vector format text surface feature vector to generate a text surface feature vector set.

[0058] The part-of-speech distribution is converted into a probability vector (e.g., noun probability U, verb probability V, etc.), the word length distribution is converted into a histogram vector, and the word frequency is converted into a TF-IDF vector. Syntactic dependency structures are converted into structure vectors using a graph neural network, and sentence types are encoded using one-hot encoding. The cohesion patterns and association types in the discourse features are also encoded as vectors. All surface feature vectors are concatenated into a unified-dimensional text surface feature vector, generating a set of text surface feature vectors.

[0059] Step S124: Construct a feature fusion layer for the cross-modal semantic enhancement model. Input the set of semantic label feature vectors and the set of text surface feature vectors into the feature fusion layer. The feature fusion layer calculates the correlation weights between the feature vectors through an attention mechanism and performs weighted fusion of the feature vectors according to the correlation weights to generate a fused feature vector set.

[0060] Step S1241: Build the basic architecture of the feature fusion layer of the cross-modal semantic enhancement model. The basic architecture includes an input layer, an attention calculation layer, a weight allocation layer, and a fusion output layer. Each layer is connected through a data transmission channel.

[0061] The input layer receives a set of semantic label feature vectors and a set of text surface feature vectors, and the data transmission channel uses tensor concatenation. The attention calculation layer contains a multi-head self-attention module, the weight allocation layer uses a fully connected network to implement weight mapping, and the fusion output layer outputs the fusion result through residual connections and layer normalization. All layers are connected via the ReLU activation function to ensure effective gradient propagation.

[0062] Step S1242: Input the semantic label feature vector set and the text surface feature vector set into the input layer of the feature fusion layer, perform dimension transformation processing on the feature vectors with inconsistent dimensions, and generate a feature vector set with unified dimensions.

[0063] The semantic label feature vector has dimension AF, and the text surface feature vector has dimension AG. If AF ≠ AG, the two are mapped to the same dimension AH (e.g., AI dimension) through the linear transformation matrix of the input layer. For example, the semantic label vector is transformed by W1·x1+b1, and the text surface vector is transformed by W2·x2+b2 (W is the weight matrix, and b is the bias term), generating a set of feature vectors with uniform dimension.

[0064] Step S1243: Input the set of feature vectors with uniform dimensions into the attention calculation layer, calculate the semantic similarity between the semantic label feature vector and the text surface feature vector, and generate a semantic similarity matrix.

[0065] In the attention calculation layer, the semantic label vector is denoted as Q (query), and the text surface vector is denoted as K (key). The original similarity score is obtained by calculating Q·KT / √d (where d is the vector dimension), and then normalized by the Softmax function to generate a semantic similarity matrix. Each element (i,j) in the matrix represents the similarity between the i-th semantic label vector and the j-th text surface vector.

[0066] Step S1244: Based on the semantic similarity matrix, the attention calculation layer calculates the attention score of each feature vector. The attention score represents the degree of association between the feature vector and the core semantics of the document, and generates a set of feature attention scores.

[0067] Using the semantic similarity matrix as weights, the text surface vector V (values) is weighted and summed to obtain the attention output. Simultaneously, the contribution of each input vector to the output is calculated as an attention score; for example, the "[amount]" tag vector has a high similarity to the "D-yuan" text vector, and therefore receives a high attention score. The scores of all vectors are recorded as a feature attention score set.

[0068] Step S1245: Input the feature attention score set into the weight allocation layer, assign corresponding feature weights based on the attention scores, and generate a feature weight set. The scores and weights are positively correlated.

[0069] The weight allocation layer maps the attention scores to the [0,1] interval using a sigmoid function, with higher scores receiving higher weights. For example, feature vectors with score AJ are assigned weight AK, and feature vectors with score AL are assigned weight AM. A set of feature weights corresponding one-to-one with each feature vector is generated.

[0070] Step S1246: Perform weighted fusion processing on the set of feature vectors with uniform dimensions and the set of feature weights. Multiply each feature vector by its corresponding weight and sum them to generate a preliminary fused feature vector.

[0071] For each dimension of the unified feature vector x, perform element-wise multiplication of x·w (where w is the corresponding weight). Then, sum all the weighted vectors according to their positions to obtain the preliminary fused feature vector. For example, the semantic label vector (AN, AO) multiplied by the weight AP is (AN*AP, AO*AP), and the text surface vector (AQ, AR) multiplied by the weight AS is (AQ*AS, AR*AS). After summing, the preliminary fused vector is (AN*AP+AQ*AS, AO*AP+AR*AS).

[0072] Step S1247: Perform feature normalization processing on the preliminary fused feature vector to transform the value range of the feature vector to a unified interval, eliminate the dimensional differences between different feature dimensions, and generate a normalized fused feature vector.

[0073] The L2 normalization method is used to calculate the L2 norm of the initial fused vector. Then, each element is divided by this norm to make the vector magnitude 1. For example, the L2 norm of the above vector is √[(AN*AP+AQ*AS)²+(AO*AP+AR*AS)²]. After normalization, each element of the vector is in the interval [-1,1], generating a normalized fused feature vector.

[0074] Step S1248: Input the normalized fusion feature vector into the fusion output layer for noise reduction; then, based on the importance score of the feature dimension on the training set, retain the feature dimensions with an importance score higher than the preset threshold to generate a simplified fusion feature vector; integrate all simplified fusion feature vectors to generate a fusion feature vector set covering all semantic segments of the document.

[0075] The fusion output layer uses a Dropout operation (ratio AT) to denoise the normalized vector, randomly discarding some feature dimensions to prevent overfitting. Importance scores (e.g., Gini importance) for each feature dimension are pre-calculated using the training set, and a threshold AU is set. Dimensions with scores higher than AU are retained; for example, the importance AV of the "amount value" dimension is retained, while the importance AW of the "font color" dimension is discarded. The simplified vectors of all semantic fragments are then arranged sequentially to generate a fused feature vector set.

[0076] Step S125: Configure the encoding parameters of the cross-modal semantic enhancement model, which include the number of encoding layers, hidden layer dimensions, number of attention heads, and activation function type, and generate a model encoding parameter set.

[0077] The cross-modal semantic enhancement model employs a Transformer encoder structure, configured with AX encoding layers, AY hidden layers, AZ attention heads (each head with BA dimension), BB intermediate layers in the feedforward neural network, and GELU activation function. The ε parameter of LayerNorm is set to BC, the dropout ratio of the residual connections is set to BD, and a set of model encoding parameters is generated.

[0078] Step S126: Input the fused feature vector set into the encoding layer of the cross-modal semantic enhancement model, perform semantic deepening encoding processing based on the model encoding parameter set, capture the deep semantic associations and semantic implicit information of the text, and generate a deep semantic feature vector set.

[0079] The fused feature vector set is input into the Transformer encoder through an embedding layer (containing positional encoding). Each encoder layer contains a multi-head self-attention sublayer and a feedforward neural network sublayer. The self-attention sublayer learns long-distance dependencies between feature vectors, such as the calculation relationship between "tax-exclusive amount" and "tax amount"; the feedforward sublayer captures complex semantic patterns through nonlinear transformations, such as the association rules between "tax rate" and "VAT invoice". After encoding by the AX layer, the output feature vector contains deep semantic information, forming a deep semantic feature vector set.

[0080] Step S127: Construct the feature integration layer of the cross-modal semantic enhancement model. Input the text surface feature vector set and the deep semantic feature vector set into the feature integration layer, perform feature dimension alignment and cross-dimensional information fusion processing, and generate a preliminary enhanced semantic representation.

[0081] The feature integration layer first performs a linear transformation on the text surface feature vector (dimensional AF) and the deep semantic feature vector (dimensional AY) to unify them to BE dimension. A gating mechanism (such as the update gate of GRU) is then used to adaptively fuse the two features: z = σ(Wz·[x;y] + bz), h = z⊙y + (1-z)⊙x, where x is the surface vector, y is the deep vector, z is the gating vector, σ is the Sigmoid function, and Wz and bz are learnable parameters. The output is the fused preliminary enhanced semantic representation.

[0082] Step S128: The preliminary enhanced semantic representation is subjected to noise reduction and normalization processing. Then, based on the semantic boundary information in the semantic pre-annotation results, the processed enhanced semantic representation is divided into segment enhanced semantic representations corresponding to each semantic segment unit, and a set of segment enhanced semantic representations is generated.

[0083] A wavelet thresholding denoising algorithm is used to process the preliminary enhanced semantic representation to remove high-frequency noise. L2 normalization is applied to normalize the vectors, ensuring consistency across dimensions. Based on the boundary coordinates in the semantic pre-annotation results, the continuous enhanced semantic representation is segmented into vectors corresponding to semantic segments. For example, the "total price and tax" segment corresponds to a BE-dimensional segment enhanced semantic representation vector. All segment vectors constitute the segment enhanced semantic representation set.

[0084] Step S129: Integrate the fragment-enhanced semantic representation set to generate a document-enhanced semantic representation that integrates surface text features and deep semantic features covering the entire document text.

[0085] The set of enhanced semantic representations is concatenated according to the semantic fragment order in the document to form a document-level enhanced semantic representation sequence. At the same time, document-level meta-features (such as the embedding vector of the document type "VAT invoice") are added to the beginning of the sequence, and finally a document-level enhanced semantic representation containing surface visual features of the text, deep semantic relationships and document type information is generated, with a dimension of (number of fragments + 1) × BE.

[0086] Step S130: Perform bidirectional semantic interaction processing based on document-enhanced semantic representation and dynamic business knowledge network to generate the association and interaction results of document semantics and business knowledge.

[0087] The dynamic business knowledge network of the enterprise financial and tax management system includes knowledge such as financial and tax regulations (e.g., the Provisional Regulations on Value-Added Tax), enterprise accounting system, and historical reimbursement rules. Through two-way interaction, it links the "tax rate" in the document with the "tax rate applicable to the sale of goods by general taxpayers" in the knowledge network, and at the same time feeds back the "abnormal deduction amount" in the document to the knowledge network to update the risk identification rules.

[0088] Step S131: Obtain the latest version data of the dynamic business knowledge network, which includes business entity nodes, business relationship edges and business attribute information, and generate a business knowledge network data set.

[0089] The latest snapshot of the dynamic business knowledge network is obtained through a knowledge graph interface, including entity nodes (such as "VAT rate", "invoice type", "reimbursement standard"), relation edges (such as "applicable", "included", "restricted"), and attribute information (such as "tax rate value: BF%", "invoice type: special / ordinary", "daily catering reimbursement limit: BG yuan"). The above data is stored in the form of (entity, relation, entity, attribute) quadruples to generate a business knowledge network data set.

[0090] Step S132: Perform semantic feature extraction processing on the business knowledge network data set, extract the semantic description features of each business entity node, the association type features of each business relationship edge, and the attribute value features of each business attribute information, and generate a business knowledge semantic feature set.

[0091] For the entity node "Value Added Tax Rate", the word vector of its semantic description "tax type and tax rate levied on the sale of goods or the provision of processing, repair and maintenance services" is extracted as a semantic description feature; for the "Applicable" relation edge, its type code "TAX_APPLY" is extracted as an association type feature; for the attribute "Tax Rate Value: BF%", the numerical feature BF% and the unit feature "percentage" are extracted as attribute value features. All features constitute a set of business knowledge semantic features.

[0092] Step S133: Input the document enhanced semantic representation and the set of business knowledge semantic features into the bidirectional semantic interaction module to build a semantic interaction channel and realize the bidirectional transmission and interaction of document semantic features and business knowledge semantic features.

[0093] The bidirectional semantic interaction module includes a document-to-knowledge channel and a knowledge-to-document channel. The document-to-knowledge channel transmits entity vectors (such as "BF%") from the document's enhanced semantic representation to the knowledge network for matching; the knowledge-to-document channel transmits rule vectors (such as "BF% tax rate applicable scope") from the knowledge network to the document feature space. Interoperability of bidirectional features is achieved through a shared semantic space, for example, by uniformly adopting the BERT embedding space.

[0094] Step S134: In the bidirectional semantic interaction module, the mapping process from document semantic features to business knowledge semantic features is performed. Based on semantic similarity association rules, the semantic features in the document enhanced semantic representation are mapped to the entity nodes or relation edges corresponding to the business knowledge network, generating the semantic mapping result from document to knowledge.

[0095] Step S1341: Extract the document-enhanced semantic representation and business knowledge semantic feature set from the semantic interaction channel of the bidirectional semantic interaction module, and generate the interaction feature set.

[0096] Extract entity feature subsequences (such as "tax rate BF%" and "special invoice") from the document-to-knowledge channel to enhance the semantic representation of the document, and extract entity node features (such as "value-added tax rate" and "invoice type") from the business knowledge semantic feature set from the knowledge-to-document channel, and merge them into an interaction feature set.

[0097] Step S1342: Construct a semantic similarity association rule base. The rule base includes semantic equivalence association rules, semantic inclusion association rules, and semantic related association rules. Each rule clarifies the conditions for semantic feature matching and the association type.

[0098] Semantic equivalence association rule: If the literal similarity between a document feature and a knowledge entity is ≥ BH and the contextual semantic similarity is ≥ BI, then they are considered equivalent associations. For example, "BF% tax rate" is equivalent to "VAT rate BF%". Semantic inclusion association rule: If the semantic scope of a document feature is completely included by the knowledge entity, such as "restaurant invoice" being included in "expense invoice". Semantic relevance association rule: If there is an indirect semantic relationship between the two, such as "input tax amount" being related to "deduction period".

[0099] Step S1343: Compare the document-enhanced semantic representation in the interaction feature set with the business knowledge semantic feature set one by one, and determine the semantic matching relationship between the two based on the rules in the semantic similarity association rule base.

[0100] The document feature “tax rate BF%” is compared one by one with the knowledge entities “value-added tax rate” and “corporate income tax rate”, and the literal similarity (BJ) and contextual semantic similarity (BK) are calculated. If both meet the threshold conditions of the semantic equivalence association rule, they are determined to be equivalent matching relationships.

[0101] Step S1344: For document semantic features and business knowledge semantic features that conform to the semantic equivalence association rules, establish a one-to-one equivalence mapping relationship, label the mapping type as equivalence mapping, and generate a set of equivalence mapping relationships.

[0102] Establish an equivalent mapping relationship between "tax rate BF%" and "value-added tax rate BF%", and record the mapping type as "equivalent" and the mapping strength as "high". All equivalent matching pairs constitute the set of equivalent mapping relationships.

[0103] Step S1345: For document semantic features and business knowledge semantic features that conform to the semantic inclusion association rules, establish inclusion mapping relationships, mark the inclusion direction and inclusion scope, and generate a set of inclusion mapping relationships.

[0104] The document feature "high-speed rail ticket" and the knowledge entity "transportation expense invoice" satisfy the inclusion association rule. An inclusion mapping relationship is established, labeled with the direction "high-speed rail ticket → transportation expense invoice" and the scope "transportation vehicle class". A set of inclusion mapping relationships is generated.

[0105] Step S1346: Establish relevant mapping relationships between document semantic features and business knowledge semantic features that conform to semantic relevance association rules, mark the degree of relevance and the basis for association, and generate a set of relevant mapping relationships.

[0106] The document feature "Input Tax Amount BL Yuan" is related to the knowledge entity "Deduction Period BM Days". A relevant mapping relationship is established, and the degree of relevance is marked as "Medium". Based on the principle that "Input Tax Amount must be certified within the deduction period", a set of relevant mapping relationships is generated.

[0107] Step S1347: Integrate the set of equivalent mapping relations, the set of included mapping relations, and the set of related mapping relations to generate a preliminary document to a knowledge mapping set.

[0108] The three types of mapping relationships mentioned above are merged, duplicates are removed, and a preliminary document-to-knowledge mapping set containing all document feature and knowledge feature mappings is generated.

[0109] Step S1348: For the initial document-to-knowledge mapping set, retain the mapping relationship with the highest semantic similarity score for the semantic features of the same document, remove other mapping relationships, and generate a set of deduplicated mapping relationships; normalize the semantic similarity score and use it as the mapping confidence of the corresponding mapping relationship to generate a set of mapping relationships with confidence.

[0110] For the document feature "invoice type", if the initial mapping set contains "equivalent → special invoice" (score BN) and "related → ordinary invoice" (score BO), retain the equivalent mapping with the highest score and remove the related mapping. Normalize BN to the [0,1] interval as the confidence level BP to generate a set of mapping relationships with confidence.

[0111] Step S1349: Associate and bind the set of mapping relationships with confidence with the entity nodes and relationship edge information of the business knowledge network, label the knowledge network elements corresponding to each mapping relationship, and generate the semantic mapping result from document to knowledge.

[0112] The mapping relationship “tax rate BF% → VAT rate BF% (confidence level BP)” is bound to the entity node ID “EN-BQ” and the relation edge ID “RE-BR” in the knowledge network, and the specific elements of the knowledge network corresponding to the mapping are marked to generate the semantic mapping result from document to knowledge.

[0113] Step S135: Perform the fusion processing of business knowledge semantic features into document semantic features. Select features from the set of business knowledge semantic features that have a similarity to the document enhanced semantic representation higher than a preset threshold, and perform feature concatenation or weighted summation with the document enhanced semantic representation to generate the knowledge-enhanced document semantic representation.

[0114] Set a similarity threshold BS, and select features from the business knowledge semantic feature set that have a similarity ≥ BS with the document augmented semantic representation, such as "VAT rate calculation rules" and "input tax deduction conditions". Concatenate the above knowledge feature vectors with the document augmented semantic representation vector (adding a knowledge feature dimension) to generate the knowledge-enhanced document semantic representation.

[0115] Step S136: Perform relation completion processing on the semantic mapping result from document to knowledge. Based on the inherent relation logic of the business knowledge network, supplement the missing relation between the semantic features of the document and the features of the business knowledge to generate a complete semantic mapping result.

[0116] In the knowledge network, there is an inherent "permission" relationship between "value-added tax special invoice" and "deductible input tax". If the mapping result from document to knowledge is missing this relationship, it will be automatically completed based on the knowledge network logic to generate a complete semantic mapping result.

[0117] Step S137: Perform feature selection processing on the knowledge-enhanced document semantic representation, set a threshold based on the correlation strength between the features and the business knowledge features, retain the document semantic features with a correlation strength higher than the threshold, and generate a simplified knowledge-enhanced semantic representation.

[0118] The association strength (such as mutual information value) between each feature dimension and business knowledge features in the knowledge-enhanced semantic representation of the document is calculated. A threshold BT is set, and dimensions with an association strength ≥ BT, such as "tax amount calculation" and "tax rate matching", are retained to generate a simplified knowledge-enhanced semantic representation.

[0119] Step S138: Integrate the complete semantic mapping result with the simplified knowledge-enhanced semantic representation, label the association type, association strength and interaction path of the document semantic features and business knowledge features, and generate bidirectional semantic interaction intermediate results.

[0120] Associate the mapping relationships in the complete semantic mapping result (e.g., "tax rate BF% → VAT rate BF%") with the corresponding feature dimensions in the simplified knowledge-enhanced semantic representation, and label the association type as "equivalent", association strength as "BP", and interaction path as "document entity → knowledge entity". Perform this operation on all interaction features to generate bidirectional semantic interaction intermediate results.

[0121] Step S139: Perform semantic consistency integration processing on the intermediate results of bidirectional semantic interaction to avoid semantic conflicts and generate semantic interaction integration results.

[0122] The system checks for semantic conflicts in the intermediate results of bidirectional semantic interaction. For example, the same document feature might map to two contradictory knowledge entities (e.g., "tax rate BF%" maps to both "value-added tax rate" and "business tax rate"). By calculating the confidence level of the conflicting mappings, mappings with high confidence are retained, conflicts are eliminated, and a semantic interaction integration result is generated.

[0123] Step S1310: Structure the semantic interaction integration results to generate association interaction results that include the association type, association strength, interaction path and enhanced semantic representation of document semantics and business knowledge.

[0124] The semantic interaction integration results are organized according to the structure of "document features - knowledge features - association type - association strength - interaction path - enhanced semantic representation", stored in a structured data format (such as JSON), and the association interaction results are generated.

[0125] Step S140: Perform document semantic asset construction processing based on the association interaction results to generate structured semantic assets containing entity association networks, event logic chains, and core viewpoint networks.

[0126] Based on the results of the interaction, an entity association network of "taxpayer identification number-company name-invoice amount-tax rate" is constructed from the VAT invoice documents. The event logic chain of "invoice issuance-input deduction-tax declaration" is sorted out, and core viewpoints such as "accurate tax calculation" and "compliance of tax rate application" are extracted to form a structured semantic asset.

[0127] Step S141: Extract the knowledge-enhanced semantic representation of the document from the associated interaction results, parse the entity semantic features, relational semantic features and event semantic features, and generate a set of core semantic elements of the document.

[0128] Extract entity semantic features (such as "company name", "taxpayer identification number", "amount D yuan" and "tax rate BF%)", relational semantic features (such as "company name-issue-invoice" and "amount-applicable-tax rate"), and event semantic features (such as "invoice issuance event" and "tax calculation event") from the knowledge-enhanced document semantic representation, and integrate them to generate a set of core semantic elements of the document.

[0129] Step S142: Based on the entity semantic features in the core semantic element set of the document, identify various business entities in the document, extract the attribute information and semantic description of each business entity, and generate a document entity set.

[0130] The business entities identified in the document include "Invoice Entity" (attributes: invoice code, invoice number, invoice date), "Enterprise Entity" (attributes: name, taxpayer identification number, address), "Amount Entity" (attributes: amount excluding tax, tax amount, total price including tax), and "Tax Rate Entity" (attributes: tax rate value BF%, applicable tax type). The attribute information and semantic descriptions of each entity are extracted to generate a document entity set.

[0131] Step S143: Analyze the relationships between different entities in the document entity set, and combine the relationship types and relationship strengths in the relationship interaction results to construct a relationship network between entities. The entity relationship network uses entities as nodes and relationship as connecting edges to generate an initial entity relationship network.

[0132] Analyze the relationships between entities: "Enterprise Entity - Issuing - Invoice Entity" (Relationship Strength BU), "Invoice Entity - Containing - Amount Entity" (Relationship Strength BV), "Amount Entity - Applicable - Tax Rate Entity" (Relationship Strength BW). Construct an initial entity relationship network with entities as nodes, relationships as edges, and edge weights as relationship strengths.

[0133] Step S144: Extract event semantic features from the core semantic element set of the document, identify business events in the document, determine the event subject, event action, event time, event location and event result of each business event, and generate a document event set.

[0134] Extract semantic features from the events to identify "Invoice Issuance Events" (Subject: Enterprise Entity, Action: Issuance, Time: Invoice Date, Location: Enterprise Registered Location, Result: Generating Invoice Entity) and "Tax Calculation Events" (Subject: System, Action: Calculation, Time: When Invoicing, Location: System Backend, Result: Generating Tax Amount Attribute). Generate a set of document events containing the above information.

[0135] Step S145: Based on the event time characteristics and event relationships in the document event set, sort the events according to the order of their occurrence time, identify the causal logic between the events, construct the event logic chain, and generate the initial event logic chain.

[0136] For example, step S1451: Extract the event time features of each event from the document event set. The event time features include the event occurrence time, event duration, and event end time, and generate an event time feature set.

[0137] Extract the occurrence time (invoice date), duration (instantaneous), and end time (invoice date) of the "Invoice Issuance Event"; and the occurrence time (invoice date), duration (instantaneous), and end time (invoice date) of the "Tax Calculation Event". Generate a set of event time features.

[0138] Step S1452: Standardize the time information in the event time feature set. Based on the standardized event time set, sort the events in the document event set according to the order of their occurrence time to generate a time-sorted event list.

[0139] The event timestamps are standardized to YYYY-MM-DD format, and the events are sorted in chronological order of occurrence. If events occur at the same time, they are sorted by event type priority (e.g., "Invoice Issuance Event" takes precedence over "Tax Calculation Event"), generating a time-sorted event list.

[0140] Step S1453: Analyze the semantic relationships between adjacent events in the time-sorted event list, extract causal connectors, conditional connectors, and progressive connectors between events, identify the causal logical relationships between events, and generate a set of event causal relationships.

[0141] Analyze adjacent events "invoice issuance event" and "tax calculation event", extract the causal conjunction "based on...calculate...", identify the causal logical relationship "invoice issuance event is a prerequisite for tax calculation event", and generate a set of event causal relationships.

[0142] Step S1454: Based on the time-ordered event list and the event causal relationship set, construct the basic framework of the event logic chain. The framework takes the time sequence as the axis and the causal relationship as the connecting link to generate a preliminary event logic framework.

[0143] Using the timeline as the horizontal axis, sort the events by time and use directed arrows to represent causal relationships (such as "invoice issuance event" → "tax calculation event") to generate a preliminary event logic framework.

[0144] Step S1455: Based on the event co-occurrence information, temporal relationships and causal connectives in the core semantic element set of the document, infer the causal relationships or logical dependencies between events that are not explicitly stated through a rule engine or statistical model, and generate a supplementary set of logical relationships.

[0145] The rule engine infers that the "tax calculation event" depends on the "tax rate determination event" (not explicitly stated in the documentation), generating a supplementary logical relationship "tax rate determination event → tax calculation event", thus forming a supplementary logical relationship set.

[0146] Step S1456: Integrate the supplementary logical relationship set into the preliminary event logic framework, improve the logical connections between events, and generate a complete event logic framework.

[0147] Insert the "Tax Rate Determination Event" between the "Invoice Issuance Event" and the "Tax Amount Calculation Event", and add a directed arrow "Invoice Issuance Event → Tax Rate Determination Event → Tax Amount Calculation Event" to generate a complete event logic framework.

[0148] Step S1457: Add event attribute information to each event in the complete event logic framework. The attribute information includes the event subject, event action, event result and event associated entity. Generate an attribute-enhanced event logic framework, correct logical contradictions and time sequence deviations in the attribute-enhanced event logic framework, and generate an initial event logic chain.

[0149] Add the following attributes to the "Tax Rate Determination Event": Subject (System), Action (Determination), Result (Selected Tax Rate BF%), and Associated Entity (Tax Rate Entity). Check and correct logical inconsistencies (such as incorrect time sequence) and generate the initial event logic chain.

[0150] Step S146: Extract concise knowledge to enhance semantic representation from the associated interaction results, identify the semantic features of the core viewpoints in the document, extract viewpoint expressions that can represent the core ideas of the document, and generate a set of core viewpoints of the document.

[0151] Identify the semantic features of core viewpoints from the simplified knowledge-enhanced semantic representation, such as "the tax rate is applied correctly," "the tax amount is calculated correctly," and "the invoice information is complete." Extract these viewpoint expressions to generate a set of core viewpoints for the document.

[0152] Step S147: Based on the semantic features of viewpoints in the core viewpoint set of the document and the business knowledge association logic in the associated interaction results, construct the supporting and progressive relationships between the core viewpoints to generate a core viewpoint network.

[0153] Analyze the relationships between core viewpoints: "Complete invoice information" supports "Correct application of tax rate," and "Correct application of tax rate" supports "Incorrect tax calculation." Construct a core viewpoint network with viewpoints as nodes and supporting relationships as directed edges.

[0154] Step S148: Perform network completion processing on the initial entity association network. Based on path analysis algorithms or association rule mining, discover and add indirect association relationships between entities that are not directly expressed in the document but can be inferred from existing relationships, and generate an optimized entity association network.

[0155] The path analysis algorithm reveals that "Enterprise Entity" and "Tax Rate Entity" can be indirectly linked through "Invoice Entity" (Enterprise Entity - Issuing - Invoice Entity - Applicable - Tax Rate Entity). By adding the indirect link "Enterprise Entity - Applicable - Tax Rate Entity", an optimized entity association network is generated.

[0156] Step S149: Analyze the initial event logic chain, infer and add implicit causal or conditional relationships between events that are not explicitly stated based on the causal reasoning model or temporal logic rules; at the same time, verify and adjust the event order based on standardized event time to generate a complete event logic chain.

[0157] Based on a causal reasoning model, it is inferred that the "invoice information entry event" is a prerequisite for the "invoice issuance event," thus adding an implicit causal relationship. The event time sequence is verified to ensure that the "tax rate determination event" precedes the "tax amount calculation event," generating a complete event logic chain.

[0158] Step S1410: Integrate and optimize the entity association network, the complete event logic chain, and the core viewpoint network, annotate the semantic relationships between each part, and generate a structured semantic asset containing the entity association network, the event logic chain, and the core viewpoint network.

[0159] The system optimizes the association between "amount entities" in the entity association network and "tax calculation events" in the complete event logic chain, labeling them as "amount entities are the result data of tax calculation events"; it also associates "tax calculation is correct" in the core viewpoint network with "tax calculation events," labeling them as "viewpoints are generated based on event results." All relationships are then integrated to generate structured semantic assets.

[0160] Step S150: Input the structured semantic assets into the intelligent document application engine to drive multiple application modules to collaboratively execute document understanding application operations. At the same time, collect semantic feedback data during the application operation process and input the semantic feedback data back into the cross-modal semantic enhancement model to complete parameter optimization.

[0161] By inputting structured semantic assets into the intelligent document application engine of the enterprise's financial and tax management system, the system drives the collaborative work of modules such as invoice verification, tax calculation, and risk warning. At the same time, it collects user feedback on the calculation results to optimize the cross-modal semantic enhancement model.

[0162] Step S151: Input the structured semantic assets into the application scheduling module of the intelligent document application engine, parse the entity association network, event logic chain and core viewpoint network in the structured semantic assets, and generate application scheduling instructions.

[0163] The application scheduling module parses the structured semantic assets, identifies the current document as a value-added tax invoice, and requires verification, tax calculation, and archiving operations. It then generates application scheduling instructions that include the module call order, input parameters, and priorities.

[0164] Step S152: Based on the application scheduling instruction, drive the document classification and archiving module of the intelligent document application engine to perform operations, determine the document classification category and archiving path according to the business entities and core viewpoints in the structured semantic assets, and generate document classification and archiving results.

[0165] The document classification and archiving module determines the classification category as "compliant VAT special invoice" based on the "invoice type: special invoice" and the core concept "compliance" in the entity association network, and the archiving path is " / financial and tax data / input invoices / year / month", generating the document classification and archiving results.

[0166] Step S153: Drive the key information extraction module of the intelligent document application engine to perform operations, extract key entities, core events and core viewpoints from the document based on the entity association network and event logic chain in the structured semantic assets, and generate a set of key information of the document.

[0167] The key information extraction module extracts "company name", "taxpayer identification number", "amount D yuan" and "tax rate BF%" from the entity association network, "invoice issuance event (time: invoice date)" from the event logic chain, and "tax amount calculation is correct" from the core viewpoint network, generating a set of key information for the document.

[0168] Step S154: Drive the risk warning module of the intelligent document application engine to perform operations, combine risk knowledge in structured semantic assets and dynamic business knowledge network, identify potential risk semantic units in the document, and generate risk warning information.

[0169] The risk warning module compares the "tax rate BF%" in the structured semantic assets with the "applicable tax rate range for the industry" in the dynamic business knowledge network. If BF% is not within the range, it is identified as a potential risk semantic unit and a risk warning message of "tax rate abnormal, review recommended" is generated.

[0170] Step S155: Integrate the document classification and archiving results, the set of key document information, and the risk warning information to generate intelligent document application results and output them to the user interface.

[0171] The document classification and archiving results (path, category), key information sets (entities, events, opinions), and risk warning information (warning content, suggestions) are integrated into an application result in HTML format and displayed on the user interface.

[0172] Step S156: Collect operation log data of each module of the intelligent document application engine during the execution of operations. The operation log data includes module start time, number of documents processed, semantic feature call records and application result generation time, and generate application process feedback data; collect user interaction feedback data on the intelligent document application results. The interaction feedback data includes user adjustment operations on classification results, modification operations on key information and confirmation operations on risk warning information, and generate user interaction feedback data.

[0173] Collect operation logs for each module: verification module startup time, number of invoices processed, entity features called (such as "invoice code"), and result generation time; user interaction feedback: adjust the category to "invoices pending review", modify the key information "amount D yuan" to "amount BN yuan", and confirm the risk warning information "tax rate abnormal". Generate application process feedback data and user interaction feedback data.

[0174] Step S157: Integrate application process feedback data and user interaction feedback data, extract semantic deviation information, feature missing information and association error information, and generate semantic feedback data.

[0175] Integrate feedback data, extract semantic deviation information (such as "incorrect identification of amount entity"), feature missing information (such as "incorrect identification of invoice remarks column information"), and association error information (such as "incorrect association between enterprise entity and tax rate entity"), and generate semantic feedback data.

[0176] Step S158: Input the semantic feedback data as part of the training data into the parameter optimization module of the cross-modal semantic enhancement model. With the goal of minimizing the difference between the application results and the feedback data, adjust the feature fusion weights, encoding layer parameters and attention mechanism parameters of the cross-modal semantic enhancement model to complete the update and iteration of model parameters.

[0177] The semantic feedback data is labeled as training samples and input into the parameter optimization module. Through the backpropagation algorithm, the attention weights of the feature fusion layer, the layer normalization parameters of the Transformer encoder, and the projection matrix of the multi-head attention are adjusted to minimize the difference between the application results (such as the amount recognition result) and the user feedback data (the modified amount BN yuan), thus completing the model parameter update.

[0178] In the above embodiments, data collection and data authorization strictly comply with relevant laws and regulations. During the data collection phase, only legally owned data within the enterprise is collected, such as the enterprise's own VAT invoices, expense reimbursement forms, and other financial and tax documents. For data involving the enterprise's trade secrets and sensitive information, explicit written authorization from the enterprise is obtained before collection to ensure the legality and compliance of the data source. The system acquires data through an encrypted transmission channel to prevent data leakage during transmission.

[0179] The cross-modal semantic enhancement model comprises modules such as an input layer, a feature fusion layer, an encoding layer, and a feature integration layer. The input layer receives a set of semantic label feature vectors and a set of text surface feature vectors, and transforms feature vectors with inconsistent dimensions. The feature fusion layer includes an attention calculation layer and a weight allocation layer, which calculate the correlation weights between feature vectors through an attention mechanism and perform weighted fusion. The encoding layer adopts a Transformer encoder structure, containing multiple multi-head self-attention sublayers and feedforward neural network sublayers. Parameters for each layer, such as the number of encoding layers, hidden layer dimensions, and the number of attention heads, are configured according to the training data and task requirements.

[0180] During model training, a large-scale labeled dataset is first prepared, all of which comes from legally authorized corporate financial and tax documents. The training process is divided into multiple stages, with a small learning rate initially used, which is gradually adjusted as training progresses. The model parameters are continuously optimized through the backpropagation algorithm, enabling the model to accurately capture the deep semantic relationships of the text. During training, the data is anonymized to remove sensitive information that could identify the company, ensuring that data usage complies with privacy protection requirements.

[0181] In specific application scenarios, the model's input consists of semantic pre-annotation results and original document text features, while the output is an enhanced semantic representation of the document that integrates surface-level and deep semantic features. After feature extraction and transformation, the input data is processed through various layers of the model to ultimately obtain a representation that accurately reflects the document's semantics. The model is integrated with enterprise financial and tax management scenarios, optimized for the characteristics of financial and tax documents, such as the recognition and processing of professional terminology, ensuring that the model can accurately understand document semantics in the financial and tax domain.

[0182] Furthermore, a comprehensive authorization management mechanism has been established. Before using the system, enterprise users must sign a data authorization agreement, clearly defining the scope and purpose of data use. The system strictly monitors data usage to ensure that data is used only for document understanding and related financial and tax business processing, and not for any other unauthorized purposes. Simultaneously, the system regularly reviews data authorization status, promptly processing expired or changed licenses to ensure the legality and compliance of data use.

[0183] Furthermore, for data involving personal information, such as personal information in expense reports, the system employs anonymization technology to remove or replace sensitive personal information, protecting individual privacy. Regarding data storage, the system uses encrypted storage to prevent unauthorized access and leakage. Data is backed up regularly to ensure its security and integrity. Through these measures, the entire data processing workflow complies with legal and regulatory requirements, protecting the data security and privacy of both businesses and individuals.

[0184] Prior to step S130, this embodiment of the application also includes a detailed implementation process for the construction and dynamic update mechanism of a dynamic business knowledge network. This network serves as the core infrastructure for bidirectional interaction between document semantics and business knowledge, and its construction and update methods are as follows: First, an initial version of a dynamic business knowledge network was constructed. This initial version's data sources included multiple channels: The first source was publicly available financial and tax regulations. By accessing the official website's regulatory database interface, the latest versions of laws and regulations were regularly collected, and the clauses, definitions, scope of application, and penalty rules in the regulatory documents were analyzed. The second source was the company's internal financial management system. Documents such as the "Expense Reimbursement Management Measures," "Procurement Payment Process Specifications," and "Contract Review Guidelines" were exported from the Enterprise Resource Planning system, and rules regarding approval processes, limits, and voucher requirements were extracted. The third source was historical business data. Invoices, contracts, expense reports, and other documents processed over the past three years were extracted from the historical database of the company's financial and tax management system, along with corresponding review results, risk cases, and user feedback records, forming an experience knowledge base.

[0185] Then, knowledge extraction and fusion are performed on the collected source data to construct the basic triples of the knowledge network. For regulatory documents, a rule extraction method based on dependency parsing is used to identify core elements in clauses such as "the taxpayer's tax payable for the sale of goods, labor, services, intangible assets, and real estate is the balance after deducting the input tax from the output tax of the current period," and extract triple relationships such as "taxpayer-generated-output tax," "taxpayer-generated-input tax," and "output tax-deduction-input tax." For internal enterprise policy documents, a rule extraction method based on text matching is used to identify rules such as "a single invoice amount exceeding 5,000 yuan requires the approval of the finance manager," and extract triple relationships such as "invoice amount-exceeds-threshold 5,000 yuan" and "exceeds the threshold-triggers-finance manager approval." For historical business data, a statistical association mining method is used to calculate entity co-occurrence frequency and event association strength from reviewed documents. For example, given the relationship "Restaurant Invoice - Occurrence - Reimbursement Form A" and "Reimbursement Form A - Rejected - Reason: Incomplete Invoice Header," the empirical knowledge triple "Restaurant Invoice - Common Mistakes - Incomplete Header" is extracted. The triples extracted from these three sources are then subjected to entity alignment and relationship fusion. For instance, "Taxpayer" in regulations and "Supplier" in internal company policies are mapped to the same entity class, forming a unified business knowledge network data set stored in a graph database. Each entity node includes node type, node identifier, and attribute list, and each relationship edge includes relationship type, direction, and weight.

[0186] Next, semantic representations are generated for entity nodes and relation edges in the business knowledge network, constructing vectorized knowledge representations that can be directly called by deep learning models. For each entity node, such as the "VAT rate" node, its name, description, and attribute values ​​are concatenated into a text sequence, input into a pre-trained language model (such as BERT) for encoding, and the output is a fixed-dimensional vector as the semantic feature vector of that node. For each relation edge, such as the "applicable" relation, its relation type is concatenated with the vector of the associated entity node, and then mapped to the semantic feature vector of the relation edge through a fully connected layer. At the same time, the attribute information of each entity node (such as the tax rate value "13%)) is encoded separately. Numerical attributes are directly normalized into numerical vectors, and enumerated attributes are converted into one-hot encoded vectors. All entity node vectors, relation edge vectors, and attribute vectors together constitute the business knowledge semantic feature set, which is stored in a vector database to support efficient similarity retrieval.

[0187] Next, a dynamic update triggering mechanism for the business knowledge network is established, which includes three triggering modes: The first is a timed triggering mode, which automatically checks the external regulatory database for updates at a fixed time each week (e.g., 3 AM every Monday). It determines whether a new version exists by comparing the regulatory version number or release date; if so, an incremental update process is initiated. The second is an event-triggered mode, where the system administrator uploads a new policy document through the management interface when internal financial regulations change, triggering the knowledge network update process. The third is a feedback triggering mode, where the semantic feedback data collected in step S150 accumulates to a certain level (e.g., 100 new user confirmations of risk warning information), and contains information inconsistent with the current knowledge network, automatically triggering the knowledge network optimization update process.

[0188] Next, when an external regulatory update is detected, such as a new VAT rate adjustment announcement issued by the State Taxation Administration, the system first obtains the updated document and extracts the newly added or modified triples using the extraction method in step S202. Then, the new triples are compared with the existing knowledge network to identify conflict points. If the new regulation stipulates that "the original 13% tax rate is adjusted to 11%", while the existing knowledge network still retains the triple "VAT rate - value is -13%", the system initiates a conflict resolution process: marking the original triple as "to be discarded", setting the effective deadline to the old regulation's expiration date, and simultaneously adding the new triple "VAT rate - value is -11%", setting the effective start date to the new regulation's effective date. During the transition period, both versions of the triples coexist, but the associated confidence weights are dynamically adjusted according to the date. For internal policy changes, such as changing "a single invoice amount exceeding 5,000 yuan requires financial manager approval" to "an invoice exceeding 10,000 yuan requires financial manager approval", the system directly updates the attribute values ​​of the corresponding triples and records the change history.

[0189] After the semantic feedback data collected in step S150 is accumulated, the feedback data is analyzed and processed. For example, if a user repeatedly identifies a risk warning message about "abnormal tax rate" as a "false alarm," and the feedback data shows that these false alarms are concentrated on invoices from a specific supplier, the system extracts common features from this feedback data and finds that the supplier's invoices are usually subject to a "tax exemption" policy. However, the current knowledge network's rule descriptions regarding tax exemption policies are not precise enough, leading to misjudgments. The system initiates a knowledge optimization process: extracting a new rule pattern "supplier type: agricultural producer - applicable - tax exemption policy" from the feedback data and matching it with the existing knowledge network. If this rule does not exist in the existing network, it is considered as a candidate for a new triple. After a confidence assessment (calculating the frequency of the rule's occurrence in the feedback data), if the confidence exceeds a preset threshold, the triple is added to the knowledge network. Simultaneously, the system adjusts the weights of existing risk identification rules, reducing their priority when dealing with agricultural producers.

[0190] Next, a version management and rollback mechanism for the knowledge network is established. Each time the knowledge network is updated, the system automatically creates a new version number and saves the previous version as a backup. Version information includes the version number, update time, update type (scheduled / event-based / feedback), a summary of the update content, and the scope of the update's impact (which entities and relationships are involved). When an update is found to cause a significant decrease in document comprehension accuracy (judged by monitoring the frequency and type of user feedback in step S150), the system can trigger an automatic rollback, restoring to the previous stable version, and recording the rollback event in the audit log for administrator review.

[0191] Next, the dynamically updated knowledge network and the cross-modal semantic enhancement model are collaboratively optimized. When the knowledge network is updated, the system recalculates the semantic feature vectors of affected entity nodes and relation edges, updating the vector database. Simultaneously, the incremental training process of the cross-modal semantic enhancement model is triggered: document samples related to the updated content are selected from historical data, steps S130 to S150 are re-executed using the updated knowledge network to generate new semantic feedback data. This data is then mixed with the original training data to fine-tune the model, enabling it to adapt to changes in the knowledge network. This collaborative mechanism ensures the synchronous evolution of the knowledge network and the understanding model.

[0192] In step S150, semantic feedback data is collected during the application operation process, and the specific implementation method of inputting the semantic feedback data back into the dynamic business knowledge network to complete the knowledge update is also included: First, effective information for knowledge network updates is extracted from semantic feedback data. Semantic feedback data includes user adjustments to the application results; for example, a user might change the risk warning information "tax rate abnormal" to "tax rate applicable correctly," adding the reason "This supplier is a small-scale taxpayer, subject to a 3% tax rate." The system parses the user's note text and extracts key information such as "supplier type - small-scale taxpayer - applicable tax rate 3%." Simultaneously, features of the documents themselves are extracted from the feedback data, such as the "supplier name" field on invoices. By tracing the supplier's historical transaction records through the entity association network, it is found that invoices issued by this supplier in the past six months were all marked as "small-scale taxpayers." This information is then integrated into a candidate knowledge update entry, including new knowledge content, supporting evidence, and a confidence score.

[0193] Next, the candidate knowledge update items undergo confidence assessment and conflict detection. Confidence assessment considers multiple factors: the frequency of the item's appearance in feedback data (e.g., the number of times the same type of feedback has appeared in the past week), the strength of supporting evidence (e.g., the consistency of historical transaction records), and user role weight (feedback from financial managers carries a higher weight than from ordinary accountants). Conflict detection compares the candidate item with the existing knowledge network. If a direct conflict exists (e.g., the existing network stipulates "a tax rate of 3% applies to small-scale taxpayers," while the candidate item suggests "a tax rate of 5% applies to small-scale taxpayers"), the conflict resolution process is initiated. The resolution process first checks the status of new regulations. If there is indeed a policy adjustment, it is handled according to the effective date mechanism in step S205; if the policy remains unchanged, the confidence of the candidate item is reduced, and it is marked as "pending manual review" and pushed to the knowledge administrator interface.

[0194] Candidate knowledge update entries that pass the confidence assessment and are conflict-free will be formally incorporated into the knowledge network. The system generates unique entity nodes or relation edge identifiers for new knowledge, such as creating a new entity node "Small-scale taxpayer collection rate," with attributes including "collection rate value 3%", "applicable scenarios", "policy basis", and "effective date". Simultaneously, it establishes associations between new nodes and existing nodes, such as "Small-scale taxpayer - Applicable - Small-scale taxpayer collection rate" and "Small-scale taxpayer collection rate - Origin - Implementation Rules of the Provisional Regulations on Value-Added Tax". After the update is completed, the system records detailed information about the update, including the update type "feedback-driven update", a list of document IDs that triggered the feedback, user confirmation records, and a comparison of the knowledge network differences before and after the update.

[0195] The system broadcasts knowledge network update events to relevant application modules. After an update is complete, the system sends a knowledge update notification to each application module via a message queue. The notification includes the affected entity types, relationship types, and effective time. Upon receiving the notification, the document classification and archiving module updates its classification rule base; the risk warning module reloads its risk identification rules to ensure that subsequent document processing uses the latest knowledge. Simultaneously, the system records update events in the audit log, forming a complete knowledge evolution trajectory for subsequent auditing and analysis.

[0196] Through the above mechanism, the dynamic business knowledge network can not only acquire knowledge from external regulations and corporate systems, but also learn and optimize autonomously from users' actual operational feedback, realizing the continuous evolution and dynamic updating of knowledge. This, together with the parameter optimization of the cross-modal semantic enhancement model, forms a dual-drive to jointly improve the accuracy and practicality of document understanding.

[0197] Figure 2 The diagram illustrates the hardware structure of an intelligent document understanding system 100 combining natural language processing and deep learning, as provided in an embodiment of the present invention. Figure 2 As shown, the intelligent document understanding system 100 combining natural language processing and deep learning may include a processor 110, a machine-readable storage medium 120, a bus 130, and a communication unit 140.

[0198] Machine-readable storage medium 120 can store data and / or instructions. In some embodiments, machine-readable storage medium 120 can store data acquired from an external terminal. In some embodiments, machine-readable storage medium 120 can store data and / or instructions used by the intelligent document understanding system 100 combining natural language processing and deep learning to execute or use in order to complete the exemplary methods described in this invention. In a specific implementation, one or more processors 110 execute the computer-executable instructions stored in machine-readable storage medium 120, enabling processor 110 to execute the intelligent document understanding method combining natural language processing and deep learning as described in the above method embodiments. Processor 110, machine-readable storage medium 120, and communication unit 140 are connected via bus 130, and processor 110 can be used to control the sending and receiving actions of communication unit 140. The specific implementation process of processor 110 can be found in the various method embodiments executed by the intelligent document understanding system 100 combining natural language processing and deep learning described above, and their implementation principles and technical effects are similar, so they will not be repeated here.

[0199] Furthermore, embodiments of the present invention also provide a readable storage medium containing computer-executable instructions. When a processor executes the computer-executable instructions, the above-described intelligent document understanding method combining natural language processing and deep learning is implemented.

[0200] It should be noted that, in order to simplify the description of this invention and thus aid in the understanding of one or more embodiments, the foregoing description of the embodiments of this invention sometimes combines multiple features into a single embodiment, drawing, or description thereof. Similarly, it should be noted that, in order to simplify the description of this invention and thus aid in the understanding of one or more embodiments, the foregoing description of the embodiments of this invention sometimes combines multiple features into a single embodiment, drawing, or description thereof.

Claims

1. An intelligent document understanding method combining natural language processing and deep learning, characterized in that, The method includes: Semantic pre-annotation is performed on the target unstructured document collection to generate semantic pre-annotation results containing document semantic boundaries, candidate semantic labels, and label confidence. The semantic pre-annotation results are fused with the original text features of the document and input into the cross-modal semantic enhancement model to perform semantic deepening encoding processing, resulting in a document-enhanced semantic representation that integrates surface text features and deep semantic features; Based on document-enhanced semantic representation and dynamic business knowledge network, bidirectional semantic interaction processing is performed to generate the association and interaction results between document semantics and business knowledge; Based on the results of related interactions, document semantic asset construction processing is performed to generate structured semantic assets containing entity association networks, event logic chains, and core viewpoint networks; Structured semantic assets are input into the intelligent document application engine, driving multiple application modules to collaboratively execute document understanding application operations. At the same time, semantic feedback data is collected during the application operation process, and the semantic feedback data is fed back into the cross-modal semantic enhancement model to complete parameter optimization.

2. The intelligent document understanding method combining natural language processing and deep learning according to claim 1, characterized in that, The semantic pre-annotation process performed on the target unstructured document set generates semantic pre-annotation results containing document semantic boundaries, candidate semantic labels, and label confidence scores, including: Read the original data of various documents in the target unstructured document collection, parse the text format, encoding type and content structure of the documents, convert them into standardized document text in a unified format, and generate a standardized document text collection; Semantic boundary recognition processing is performed on each document text in the standardized document text set. Based on semantic pause features, syntactic structure features and discourse cohesion features in natural language processing, the document text is divided into multiple independent semantic segment units, and a document semantic segment set is generated. Extract the core word sequence and syntactic dependency relations in each semantic segment unit. Based on the preset semantic tag system, match the corresponding candidate semantic tags for each semantic segment unit. The candidate semantic tags include entity tags, relation tags, event tags and opinion tags, and generate a set of semantic segment tag candidates. Construct a semantic tag matching degree calculation model, input the core word sequence, syntactic dependency relation and candidate semantic tags of the semantic segment unit, perform tag matching degree analysis, and generate the matching degree score corresponding to each candidate semantic tag; The matching score is normalized to generate the tag confidence score for each candidate semantic tag, forming a set of semantic fragment tag confidence scores. Based on the label confidence, the candidate semantic labels of each semantic segment unit are sorted in descending order to generate a sorted list of semantic labels. The candidate semantic label with the highest label confidence in the sorted list is extracted as the main label, and the remaining candidate semantic labels are used as auxiliary labels to generate a set of main and auxiliary labels for the semantic segment. The boundary location information, primary and secondary label information, and label confidence information of each semantic segment unit are associated and integrated to generate the semantic pre-annotation intermediate result of a single document; Perform format unification processing on the intermediate results of semantic pre-annotation for all documents, standardize the description format of semantic boundaries, the naming convention of tags, and the representation of confidence, and generate a set of semantic pre-annotated documents; The document semantic pre-annotation set is associated and bound with the corresponding standardized document text set to generate semantic pre-annotation results that include document semantic boundaries, candidate semantic labels, and label confidence.

3. The intelligent document understanding method combining natural language processing and deep learning according to claim 1, characterized in that, The process of fusing semantic pre-annotation results with the original text features of the document and inputting the result into a cross-modal semantic enhancement model for semantic deepening encoding processing yields a document-enhanced semantic representation that integrates surface and deep semantic features. This includes: Extract the main label features, auxiliary label features, and label confidence features of each semantic segment unit from the semantic pre-annotation results, convert them into semantic label feature vectors in vector format, and generate a set of semantic label feature vectors; For each document text in the standardized document text set, perform text surface feature extraction processing to extract lexical features, syntactic features and discourse structure features. Lexical features include part-of-speech distribution, word length distribution and word frequency. Syntactic features include syntactic dependency structure and sentence type. Discourse structure features include paragraph connection mode and sentence association type, generating a text surface feature set. Convert the text surface feature set into a vector format text surface feature vector to generate a text surface feature vector set; A feature fusion layer for a cross-modal semantic enhancement model is constructed by inputting a set of semantic label feature vectors and a set of text surface feature vectors into the feature fusion layer. The feature fusion layer calculates the correlation weights between the feature vectors through an attention mechanism and performs weighted fusion of the feature vectors based on the correlation weights to generate a fused feature vector set. Configure the encoding parameters of the cross-modal semantic enhancement model, which include the number of encoding layers, hidden layer dimensions, number of attention heads, and activation function type, and generate a model encoding parameter set; The fused feature vector set is input into the encoding layer of the cross-modal semantic enhancement model. Based on the model's encoding parameter set, semantic deepening encoding processing is performed to capture the deep semantic associations and implicit semantic information of the text, and generate a deep semantic feature vector set. A feature integration layer for a cross-modal semantic enhancement model is constructed. The set of text surface feature vectors and the set of deep semantic feature vectors are input into the feature integration layer. Feature dimension alignment and cross-dimensional information fusion processing are performed to generate a preliminary enhanced semantic representation. The initial enhanced semantic representation is subjected to noise reduction and normalization. Then, based on the semantic boundary information in the semantic pre-annotation results, the processed enhanced semantic representation is divided into segment enhanced semantic representations corresponding to each semantic segment unit, generating a set of segment enhanced semantic representations. By integrating fragment-enhanced semantic representation sets, a document-enhanced semantic representation that integrates surface text features and deep semantic features covering the entire document text is generated.

4. The intelligent document understanding method combining natural language processing and deep learning according to claim 1, characterized in that, The bidirectional semantic interaction processing based on document-enhanced semantic representation and dynamic business knowledge network generates an association and interaction result between document semantics and business knowledge, including: Obtain the latest version of the dynamic business knowledge network data, which includes business entity nodes, business relationship edges and business attribute information, and generate a business knowledge network data set. Semantic feature extraction processing is performed on the business knowledge network dataset to extract the semantic description features of each business entity node, the association type features of each business relationship edge, and the attribute value features of each business attribute information, thereby generating a set of business knowledge semantic features. The document enhanced semantic representation and the set of business knowledge semantic features are input into the bidirectional semantic interaction module to build a semantic interaction channel and realize the bidirectional transmission and interaction of document semantic features and business knowledge semantic features. In the bidirectional semantic interaction module, the mapping process from document semantic features to business knowledge semantic features is performed. Based on semantic similarity association rules, the semantic features in the document enhanced semantic representation are mapped to the corresponding entity nodes or relation edges in the business knowledge network, generating a semantic mapping result from document to knowledge. The process involves fusing business knowledge semantic features into document semantic features. Features with a similarity higher than a preset threshold to the document-enhanced semantic representation are selected from the set of business knowledge semantic features. These features are then concatenated or weighted and summed with the document-enhanced semantic representation to generate a knowledge-enhanced document semantic representation. Perform relation completion processing on the semantic mapping results from documents to knowledge. Based on the inherent relation logic of the business knowledge network, supplement the missing relation between the semantic features of the documents and the features of the business knowledge to generate complete semantic mapping results. Feature selection processing is performed on the knowledge-enhanced document semantic representation. A threshold is set based on the correlation strength between the features and the business knowledge features. Document semantic features with a correlation strength higher than the threshold are retained to generate a simplified knowledge-enhanced semantic representation. The complete semantic mapping results are associated and integrated with the simplified knowledge-enhanced semantic representation, and the association type, association strength and interaction path of the document semantic features and business knowledge features are labeled to generate bidirectional semantic interaction intermediate results. Perform semantic consistency integration processing on the intermediate results of bidirectional semantic interaction to avoid semantic conflicts and generate semantic interaction integration results; The semantic interaction integration results are structured and organized to generate associated interaction results that include the association type, association strength, interaction path, and enhanced semantic representation of document semantics and business knowledge.

5. The intelligent document understanding method combining natural language processing and deep learning according to claim 1, characterized in that, The step of performing document semantic asset construction processing based on the association interaction results to generate structured semantic assets containing entity association networks, event logic chains, and core viewpoint networks includes: Extract knowledge-enhanced semantic representations of documents from the results of related interactions, analyze the entity semantic features, relational semantic features and event semantic features, and generate a set of core semantic elements of the document; Based on the entity semantic features in the core semantic element set of the document, various business entities in the document are identified, the attribute information and semantic description of each business entity are extracted, and a document entity set is generated. Analyze the relationships between different entities in the document entity set, and combine the relationship types and strengths in the relationship interaction results to construct a relationship network between entities. The entity relationship network uses entities as nodes and relationship as connecting edges to generate an initial entity relationship network. Extract event semantic features from the core semantic element set of the document, identify business events in the document, determine the event subject, event action, event time, event location and event result of each business event, and generate a document event set; Based on the event time characteristics and event relationships in the document event set, events are sorted in chronological order of occurrence, the causal logic between events is identified, an event logic chain is constructed, and an initial event logic chain is generated. Extract concise knowledge from the results of related interactions to enhance semantic representation, identify the semantic features of the core viewpoints in the document, extract viewpoint expressions that can represent the core ideas of the document, and generate a set of the core viewpoints of the document. Based on the semantic features of viewpoints in the core viewpoint set of documents and the business knowledge association logic in the associated interaction results, the supporting and progressive relationships between core viewpoints are constructed to generate a core viewpoint network. The initial entity association network is processed for network completion. Based on path analysis algorithms or association rule mining, indirect association relationships between entities that are not directly expressed in the document but can be inferred from existing relationships are discovered and added to generate an optimized entity association network. The initial event logic chain is analyzed, and implicit causal or conditional relationships between events that are not explicitly stated are inferred and added based on causal reasoning models or temporal logic rules. At the same time, the event order is adjusted based on standardized event time verification to generate a complete event logic chain. Integrate and optimize the entity association network, complete event logic chain, and core viewpoint network, annotate the semantic relationships between each part, and generate a structured semantic asset that includes the entity association network, event logic chain, and core viewpoint network.

6. The intelligent document understanding method combining natural language processing and deep learning according to claim 2, characterized in that, The process involves performing semantic boundary recognition on each document text in the standardized document text set. Based on semantic pause features, syntactic structure features, and discourse cohesion features in natural language processing, the document text is divided into multiple independent semantic segment units, generating a document semantic segment set, including: Each document text in the standardized document text set is segmented using a word segmentation algorithm from natural language processing to divide the document text into multiple lexical units, generating a document lexical unit set. Semantic pause features are extracted from the document's vocabulary unit set. These features include punctuation distribution, position of modal particles, and frequency of semantic transition words, generating a document semantic pause feature set. Perform syntactic analysis on the document text to identify the syntactic dependency structure, subject-verb-object relationship and modification relationship of each sentence, and generate a set of document syntactic structure features; Extract cohesion features from the document text. These features include the type of conjunctions used, the distribution of referential relationships, and the consistency of paragraph themes. Generate a set of document cohesion features. Construct a semantic boundary recognition model by inputting the document semantic pause feature set, document syntactic structure feature set, and document chapter connection feature set into the model, configuring the model's recognition parameters, and generating a semantic boundary recognition parameter set; Based on the semantic boundary recognition parameter set, the semantic boundary recognition model performs semantic boundary prediction processing on the document text, outputs the potential semantic boundary positions in the document text, and generates a set of predicted semantic boundary positions. Based on the preset minimum semantic unit length threshold, the boundary positions in the predicted semantic boundary position set whose mutual spacing is less than the minimum semantic unit length threshold are removed, and an effective semantic boundary position set is generated. Based on the set of effective semantic boundary locations, the document text is divided into multiple consecutive text segments, each corresponding to an independent semantic expression unit, thus generating an initial set of semantic segments; The semantically incomplete segments in the initial set of semantic segments are merged with adjacent segments to generate a complete set of semantic segments. Add a unique identifier and position index to each semantic fragment in the complete semantic fragment set to generate a document semantic fragment set.

7. The intelligent document understanding method combining natural language processing and deep learning according to claim 3, characterized in that, The feature fusion layer for constructing the cross-modal semantic enhancement model inputs the semantic label feature vector set and the text surface feature vector set into the feature fusion layer. An attention mechanism is used to strengthen the weights of semantically related feature vectors, generating a fused feature vector set, including: The basic architecture for building the feature fusion layer of the cross-modal semantic enhancement model is constructed. The basic architecture includes an input layer, an attention calculation layer, a weight allocation layer, and a fusion output layer. Each layer is connected through a data transmission channel. The semantic label feature vector set and the text surface feature vector set are input into the input layer of the feature fusion layer. Dimension transformation is performed on the feature vectors with inconsistent dimensions to generate a feature vector set with uniform dimensions. A set of feature vectors with uniform dimensions is input into the attention calculation layer to calculate the semantic similarity between the semantic label feature vector and the text surface feature vector, and generate a semantic similarity matrix. Based on the semantic similarity matrix, the attention calculation layer calculates the attention score for each feature vector. The attention score represents the degree of association between the feature vector and the core semantics of the document, and generates a set of feature attention scores. The feature attention score set is input into the weight allocation layer, and corresponding feature weights are assigned based on the attention scores. The scores and weights are positively correlated, thus generating a feature weight set. The set of feature vectors with uniform dimensions and the set of feature weights are weighted and fused. Each feature vector is multiplied by its corresponding weight and then summed to generate a preliminary fused feature vector. The initial fused feature vectors are subjected to feature normalization processing to transform the value range of the feature vectors to a unified interval, eliminate the dimensional differences between different feature dimensions, and generate normalized fused feature vectors. The normalized fusion feature vectors are input into the fusion output layer for noise reduction. Then, based on the importance scores of the feature dimensions on the training set, feature dimensions with importance scores higher than a preset threshold are retained to generate simplified fusion feature vectors. All simplified fusion feature vectors are integrated to generate a set of fusion feature vectors covering all semantic segments of the document.

8. The intelligent document understanding method combining natural language processing and deep learning according to claim 4, characterized in that, In the bidirectional semantic interaction module, the mapping process from document semantic features to business knowledge semantic features is performed. Based on semantic similarity association rules, the semantic features in the document's enhanced semantic representation are mapped to the corresponding entity nodes or relation edges in the business knowledge network, generating a semantic mapping result from document to knowledge, including: Extract the document-enhanced semantic representation and business knowledge semantic feature set from the semantic interaction channel of the two-way semantic interaction module to generate an interaction feature set; Construct a semantic similarity association rule base, which includes semantic equivalence association rules, semantic inclusion association rules, and semantic related association rules. Each rule clarifies the conditions for semantic feature matching and the association type. The document-enhanced semantic representation in the interaction feature set is compared one by one with the business knowledge semantic feature set. Based on the rules in the semantic similarity association rule base, the semantic matching relationship between the two is determined. For document semantic features and business knowledge semantic features that conform to the semantic equivalence association rule, establish a one-to-one equivalence mapping relationship, label the mapping type as equivalence mapping, and generate a set of equivalence mapping relationships; For document semantic features and business knowledge semantic features that conform to semantic inclusion association rules, establish inclusion mapping relationships, mark the inclusion direction and inclusion scope, and generate a set of inclusion mapping relationships; For document semantic features and business knowledge semantic features that conform to semantic relevance association rules, establish relevant mapping relationships, label the degree of relevance and the basis for association, and generate a set of relevant mapping relationships; Integrate the set of equivalent mapping relationships, the set of contained mapping relationships, and the set of related mapping relationships to generate a preliminary document to a knowledge mapping set; For the initial document-to-knowledge mapping set, for the semantic features of the same document, retain the mapping relationship with the highest semantic similarity score, remove other mapping relationships, and generate a set of deduplicated mapping relationships; normalize the semantic similarity score and use it as the mapping confidence of the corresponding mapping relationship to generate a set of mapping relationships with confidence. The set of confidence-based mapping relationships is associated and bound with the entity nodes and relationship edge information of the business knowledge network. The knowledge network elements corresponding to each mapping relationship are labeled to generate semantic mapping results from documents to knowledge.

9. The intelligent document understanding method combining natural language processing and deep learning according to claim 1, characterized in that, The process involves inputting structured semantic assets into an intelligent document application engine to drive multiple application modules to collaboratively execute document understanding application operations. Simultaneously, semantic feedback data is collected during the application operation process, and this data is then fed back into a cross-modal semantic enhancement model to optimize parameters. This includes: The structured semantic assets are input into the application scheduling module of the intelligent document application engine, which parses the entity association network, event logic chain and core viewpoint network in the structured semantic assets and generates application scheduling instructions. Based on application scheduling instructions, the document classification and archiving module of the intelligent document application engine is driven to perform operations. According to the business entities and core viewpoints in the structured semantic assets, the classification category and archiving path of the document are determined, and the document classification and archiving results are generated. The key information extraction module of the intelligent document application engine performs operations, extracting key entities, core events and core viewpoints from the document based on the entity association network and event logic chain in the structured semantic assets, and generating a set of key information of the document; The risk warning module that drives the intelligent document application engine performs operations, combining risk knowledge from structured semantic assets and dynamic business knowledge networks to identify potential risk semantic units in documents and generate risk warning information. Integrate document classification and archiving results, key document information sets, and risk warning information to generate intelligent document application results, which are then output to the user interface. The system collects operation log data from each module of the intelligent document application engine during the execution of operations. The operation log data includes module startup time, number of documents processed, semantic feature call records, and application result generation time, generating application process feedback data. The system also collects user interaction feedback data on the intelligent document application results. The interaction feedback data includes user adjustments to classification results, modifications to key information, and confirmation of risk warning information, generating user interaction feedback data. Integrate application process feedback data and user interaction feedback data, extract semantic deviation information, feature missing information and association error information, and generate semantic feedback data; The semantic feedback data is used as part of the training data and input into the parameter optimization module of the cross-modal semantic enhancement model. With the goal of minimizing the difference between the application results and the feedback data, the feature fusion weights, encoding layer parameters, and attention mechanism parameters of the cross-modal semantic enhancement model are adjusted to complete the update and iteration of the model parameters.

10. An intelligent document understanding system combining natural language processing and deep learning, characterized in that, The system includes a processor and a memory, the memory being connected to the processor. The memory is used to store programs, instructions, or code, and the processor is used to run the programs, instructions, or code in the memory to implement the intelligent document understanding method combining natural language processing and deep learning as described in any one of claims 1-9.

Images