Adaptive normalization-based multi-source heterogeneous document knowledge base construction method and system

Abstract

The application relates to a multi-source heterogeneous document knowledge base construction method and system based on adaptive normalization, and belongs to the field of knowledge engineering. The method comprises the following steps: collecting original document data of different modes from multiple heterogeneous data sources; pre-processing and mode feature extraction are performed on the original document data to obtain mode feature vectors corresponding to the original document data; the mode feature vectors of different modes are dynamically mapped to a unified shared semantic vector space to generate unified document vectors; local reasoning is respectively performed on the mode feature vectors based on mode-specific AI agents; a document set with similar semantics, conflicts or repetitions is identified; online aggregation and deduplication operations are performed on the document set; a knowledge graph is constructed based on cross-modal semantic association results; and a model parameter of the AI agent is automatically triggered to update in real time. The application constructs a knowledge base system with adaptive, real-time and self-learning capabilities.

Description

Technical Field

[0001] This application relates to the technical field of knowledge engineering, and in particular to a method and system for constructing a multi-source heterogeneous document knowledge base based on adaptive normalization. Background Technology

[0002] As enterprises accelerate their digital transformation, knowledge base construction faces the severe challenge of multi-source heterogeneous document data. This multi-source heterogeneous document data includes text documents, image documents, audio documents, video documents, and professional tool documents. These documents are scattered in origin, have heterogeneous formats, and weak semantic connections. As a result, a lot of manpower is needed to filter, merge, and organize them during the knowledge base construction process. This is not only inefficient and has a high information loss rate, but also generally suffers from content conflicts and semantic similarities.

[0003] To improve the efficiency of knowledge summarization, existing technologies attempt to introduce parsing engines, big data processing platforms, or simple machine learning algorithms, but there are still insurmountable technical bottlenecks: First, cross-modal association capabilities are insufficient. Traditional rule-based or static model-based solutions cannot achieve deep semantic matching and dynamic deduplication of different types of documents such as text, images, and audio, and documents with the same or similar semantics are difficult to automatically aggregate. Second, there is a lack of adaptive processing mechanisms. Fixed parsing logic and model parameters cannot adapt to the dynamic changes of different business scenarios, and manual intervention is required when new formats or new business scenarios appear. Third, model maintenance depends on expert experience. Parsing rules and parameters need to be continuously and manually updated, otherwise misidentification or omissions are likely to occur, resulting in high maintenance costs.

[0004] These technical bottlenecks severely restrict the efficiency and effectiveness of building large-scale enterprise knowledge bases. Therefore, how to construct a knowledge base system with adaptive, real-time, and self-learning capabilities has become an urgent technical problem to be solved. Summary of the Invention

[0005] To construct a knowledge base system with adaptive, real-time, and self-learning capabilities, this application provides a method and system for constructing a multi-source heterogeneous document knowledge base based on adaptive normalization.

[0006] Firstly, this application provides a method for constructing a multi-source heterogeneous document knowledge base based on adaptive normalization, employing the following technical solution:

[0007] Collect raw document data of different modalities from multiple heterogeneous data sources;

[0008] The original document data is preprocessed and modal features are extracted to obtain modal feature vectors corresponding to each original document data.

[0009] The modal feature vectors of different modalities are dynamically mapped to a unified shared semantic vector space. The weights of each modality are dynamically calculated and adjusted according to the current semantic distribution and business scenario. The modal feature vectors are then weighted and aggregated to generate a unified document vector.

[0010] Based on modality-specific AI agents, local inference is performed on the modality feature vectors to obtain local analysis results and inference confidence for each modality. The local analysis results and inference confidence are shared through a unified message bus.

[0011] Based on the unified document vector, vector similarity nearest neighbor retrieval and graph algorithms are used in the shared semantic vector space to obtain cross-modal semantic association results and identify sets of documents that are semantically similar, conflicting, or duplicated.

[0012] Perform online aggregation and deduplication operations on the document collection;

[0013] A knowledge graph is constructed based on the cross-modal semantic association results, and newly added original document data is updated to the knowledge graph in an incremental manner.

[0014] By combining the local analysis results, the inference confidence, the cross-modal semantic association results, and the system performance monitoring indicators, the active learning process is automatically triggered, and the model parameters of the AI ​​agent are updated in real time in conjunction with manual annotation feedback.

[0015] By adopting the above technical solutions, and through adaptive normalization and multimodal fusion technologies, deep semantic association and dynamic integration of multi-source heterogeneous documents are achieved, significantly improving the accuracy and processing efficiency of cross-type document merging. The online aggregation and deduplication mechanism effectively reduces information redundancy and conflicts, incremental updates to the knowledge graph ensure real-time iteration of knowledge assets, active learning and adaptive model optimization reduce reliance on expert maintenance, and full-process automation greatly reduces manual costs. The system is adaptable to different business scenarios, providing an efficient, intelligent, and low-loss solution for knowledge base construction, and enhancing knowledge retrieval and decision support capabilities.

[0016] Furthermore, the modalities include text, images, audio, video, and professional tool formats. The preprocessing and modal feature extraction of the original document data to obtain the modal feature vectors corresponding to each original document data includes:

[0017] Full-text extraction is performed on text documents, OCR technology is used to extract text content from image documents, speech-to-text transcription and keyframe extraction are performed on audio or video documents, and hierarchical structure and object attributes are extracted from professional tool format documents to obtain preprocessed data.

[0018] The preprocessed data is then cleaned and denoised.

[0019] For each cleaned and denoised data, a modal feature vector is extracted using a feature extraction tool corresponding to the data modality.

[0020] By adopting the above technical solutions, differentiated preprocessing and dedicated models are used to extract features from multimodal documents, accurately transforming heterogeneous data into structured vectors. OCR, speech-to-text and other technologies break through modality barriers, cleaning and denoising ensure data purity, and dedicated models improve the accuracy of feature representation, significantly improving the accuracy and efficiency of knowledge base construction.

[0021] Furthermore, based on the unified document vector, the method utilizes vector similarity nearest neighbor retrieval and graph algorithms within the shared semantic vector space to obtain cross-modal semantic association results and identify sets of documents with semantically similar, conflicting, or repetitive elements, including:

[0022] For all unified document vectors in the shared semantic vector space, an adaptive kernel density estimation algorithm is used to calculate the semantic density value of the local region where each unified document vector is located;

[0023] Filter out the core vector clusters whose semantic density value is not lower than the preset density threshold, and remove isolated vectors whose semantic density value is lower than the preset density threshold;

[0024] For each unified document vector within the core vector cluster, extract the corresponding modality tracing information and core fields of the local analysis results, calculate the semantic overlap of the core fields of the local analysis results between unified document vectors of different modalities, retain vector pairs with semantic overlap higher than a preset overlap threshold, and construct a modality semantic consistency matrix.

[0025] For the modality semantic consistency matrix, the initial similarity is obtained by calculating the vector spacing using the cosine similarity algorithm and the mean matrix similarity is calculated. The initial similarity is then calibrated by combining the dynamic weights of each modality with the mean matrix similarity to obtain the corrected similarity.

[0026] Calculate the semantic distribution variance of the core vector cluster, dynamically generate a distance threshold, and filter out the vectors to be matched whose vector spacing is less than the distance threshold;

[0027] Using the vector to be matched as a node, and the weighted value of the semantic overlap and the corrected similarity as the edge weight, an initial association graph is constructed, and the relationship of the initial association graph is strengthened.

[0028] If the edge weight is greater than or equal to the approximation threshold and less than the repetition threshold, it is determined to be semantically approximate;

[0029] If the edge weight is greater than or equal to an approximate threshold and the core fields of the local analysis results are contradictory, then it is determined to be a semantic conflict;

[0030] If the edge weight is greater than or equal to the repetition threshold and the modal semantic consistency matrix matching degree is higher than the preset matching threshold, then it is determined to be semantic repetition;

[0031] Wherein, the approximate threshold is less than the repetition threshold;

[0032] Collect documents that are semantically similar, conflicting, or repetitive.

[0033] By adopting the above technical solutions, through multiple refined processes such as density clustering, semantic consistency verification and similarity calibration, core association vectors are accurately selected and isolated data is eliminated. Combined with dynamic thresholds and graph algorithms to strengthen associations, semantically similar, conflicting and duplicate documents are efficiently distinguished.

[0034] Furthermore, the online aggregation and deduplication operation on the document collection includes:

[0035] For each document in the document set, the unified document vector is combined with the corresponding local analysis results to calculate the semantic entropy weight;

[0036] For documents with similar or repetitive semantics:

[0037] Compare the similarity T between any two documents with the preset high-level similarity threshold T1, medium-level similarity threshold T2, and low-level similarity threshold T3:

[0038] When T > T1, if the absolute value of the semantic entropy weight difference between the two documents is less than or equal to the entropy weight threshold, the two documents are determined to be completely duplicate documents, and the duplicates are automatically removed while the original document links are retained; if the absolute value of the semantic entropy weight difference between the two documents is greater than the entropy weight threshold, the two documents are determined to be suspected duplicate documents and are pushed for manual review.

[0039] When T2 < T ≤ T1, if the absolute value of the semantic entropy weight difference between the two documents is less than or equal to the entropy weight threshold and the overlap of the core semantic fields is greater than or equal to the preset overlap, then the two documents are determined to be semantically similar documents, and the two documents are automatically aggregated into a related document group and a semantic mapping is established; if the absolute value of the semantic entropy weight difference between the two documents is greater than the entropy weight threshold or the overlap of the core semantic fields is less than the preset overlap, then the two documents are determined to be locally similar documents, independent entries are retained and the association relationship is marked;

[0040] When T3≤T≤T2, if the absolute value of the semantic entropy weight difference between the two documents is less than or equal to the entropy weight threshold and the cross-modal semantic consistency check passes, then the two documents are determined to be potentially related documents and included in the weak association link of the knowledge graph; if the consistency check fails, the association is excluded.

[0041] The order is T1 > T2 > T3;

[0042] For documents with semantic conflicts:

[0043] If the confidence levels of the inferences of both parties are respectively and The semantic entropy weights are respectively and Then calculate the priority Y: In the formula, and These are the weighting coefficients. + =1, i=1 or 2;

[0044] The document with higher priority is used as the benchmark document;

[0045] If the priority difference between two documents is less than the preset difference, it is determined that the conflict cannot be automatically resolved, a conflict label is generated and pushed to the visualization interface for manual review.

[0046] By adopting the above technical solutions, based on semantic entropy weights and hierarchical thresholds, document types such as complete duplication and semantic similarity are finely distinguished, realizing automated aggregation and deduplication and retention of the original link. Semantic conflicts are intelligently adjudicated through priority formulas, and difficult conflicts are pushed to manual review, effectively reducing information redundancy and contradictions, and improving the purity and structure level of knowledge base content.

[0047] Further, the calculation of semantic entropy weights includes:

[0048] The semantic entropy weight H is calculated as follows:

[0049]

[0050] In the formula, This represents the weight percentage of the i-th core semantic field; For scene adaptation factor, 0 < <1; The semantic complexity coefficient is 0 < <1.

[0051] Furthermore, the construction of the knowledge graph involves incrementally updating the knowledge graph with newly added original document data, including:

[0052] Extract entity-relation-entity triples from the newly added original document data and add modal origination tags. Construct an anchoring matrix based on the cross-modal semantic association results and retain valid triples with multimodal support or high confidence.

[0053] By using global coarse alignment and local association path verification, newly added entities are mapped to unique identifiers in the knowledge graph, while low-confidence entities are temporarily stored in a confirmation database.

[0054] By detecting indirect contradiction propagation paths through graph neural networks, low-confidence contradictions are automatically corrected or high-confidence contradictions are marked for manual review.

[0055] Add timestamps to entities, relations, and triples, and construct evolutionary trajectories by combining cross-modal associations.

[0056] By adopting the above technical solutions, the accuracy of knowledge graph construction is ensured by screening effective triples and accurately aligning entities. Contradictions are detected and corrected by using graph neural networks, and evolutionary trajectories are constructed by combining timestamps to achieve incremental knowledge updates and traceability. Furthermore, low-confidence content is temporarily stored and reviewed to balance automation and accuracy, thereby improving the structuring level of the knowledge base and the reliability and traceability of knowledge assets.

[0057] Furthermore, the method also includes:

[0058] The system constructs a three-layer view of document association network, knowledge graph evolution, and deduplication decision through a visualization panel, which respectively displays document clustering relationships, graph update trajectory, and automatic processing basis, and supports node highlighting, version backtracking, and manual review entry.

[0059] It supports graphical configuration of visual parameters, and configuration changes are synchronized to the system policy library in real time;

[0060] It offers multi-dimensional filtering capabilities; clicking on a node allows you to view the original content of the corresponding document, partial analysis results, and inference confidence.

[0061] By adopting the above technical solutions, a three-layer visualization view intuitively presents document associations, graph evolution, and deduplication logic, along with features such as node highlighting and version backtracking. This reduces the difficulty of manual judgment, and the graphical configuration and multi-dimensional filtering support flexible adaptation to various scenarios. Linked viewing of document details facilitates review and verification. This improves system usability and transparency, reduces operational costs, and helps efficiently manage the entire knowledge base construction process.

[0062] Secondly, this application provides a multi-source heterogeneous document knowledge base construction system based on adaptive normalization, which adopts the following technical solution:

[0063] The data acquisition module is used to collect raw document data of different modalities from multiple heterogeneous data sources;

[0064] The modal feature vector extraction module is used to preprocess and extract modal features from the original document data to obtain modal feature vectors corresponding to each original document data.

[0065] The modal feature vector aggregation module is used to dynamically map the modal feature vectors of different modalities to a unified shared semantic vector space. It dynamically calculates and adjusts the weights of each modality according to the current semantic distribution and business scenario, and performs weighted aggregation of each modal feature vector to generate a unified document vector.

[0066] The local inference module is used to perform local inference on the modality feature vectors based on the modality-specific AI agent, obtain local analysis results and inference confidence of each modality, and share the local analysis results and inference confidence through a unified message bus.

[0067] The association module is used to obtain cross-modal semantic association results based on the unified document vector and within the shared semantic vector space using vector similarity nearest neighbor retrieval and graph algorithms, and to identify sets of documents that are semantically similar, conflicting, or duplicated.

[0068] The document processing module is used to perform online aggregation and deduplication operations on the document collection;

[0069] The knowledge graph update module is used to construct a knowledge graph based on the cross-modal semantic association results and update the knowledge graph with newly added original document data in an incremental manner.

[0070] The active learning module is used to automatically trigger the active learning process by combining the local analysis results, the inference confidence, the cross-modal semantic association results, and the system performance monitoring indicators, and to update the model parameters of the AI ​​agent in real time by combining manual annotation feedback.

[0071] Thirdly, this application provides an electronic device that adopts the following technical solution:

[0072] An electronic device, comprising:

[0073] At least one processor;

[0074] Memory;

[0075] At least one computer program, wherein the at least one computer program is stored in the memory and configured to be executed by the at least one processor, the at least one computer program being configured to: perform a method for constructing a multi-source heterogeneous document knowledge base based on adaptive normalization as described in any one of the first aspects.

[0076] Fourthly, this application provides a computer-readable storage medium, which adopts the following technical solution:

[0077] A computer-readable storage medium storing a computer program that can be loaded by a processor and executed as described in any one of the first aspects, a method for constructing a multi-source heterogeneous document knowledge base based on adaptive normalization.

[0078] For a detailed description of the second to fourth aspects of the present invention and their various implementations, please refer to the detailed description in the first aspect and its various implementations; and for a detailed description of the beneficial effects of the second to fourth aspects and their various implementations, please refer to the beneficial effect analysis in the first aspect and its various implementations, which will not be repeated here.

[0079] In summary, this application includes at least one of the following beneficial technical effects:

[0080] 1. By using adaptive normalization and multimodal fusion technology, deep semantic association and dynamic integration of multi-source heterogeneous documents are achieved, significantly improving the accuracy and processing efficiency of cross-type document merging;

[0081] 2. The online aggregation and deduplication mechanism effectively reduces information redundancy and conflicts, and incremental updates to the knowledge graph ensure real-time iteration of knowledge assets;

[0082] 3. Active learning and adaptive model optimization reduce reliance on expert maintenance;

[0083] 4. Full-process automation significantly reduces labor costs. The system is adaptable to different business scenarios, providing an efficient, intelligent, and low-loss solution for knowledge base construction, and enhancing knowledge retrieval and decision support capabilities. Attached Figure Description

[0084] Figure 1 This is a flowchart illustrating the method for constructing a multi-source heterogeneous document knowledge base based on adaptive normalization in this application embodiment.

[0085] Figure 2 This is a structural block diagram of the multi-source heterogeneous document knowledge base construction system based on adaptive normalization in the embodiments of this application.

[0086] Figure 3 This is a structural block diagram of the electronic device in the embodiments of this application. Detailed Implementation

[0087] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0088] Furthermore, the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article, unless otherwise specified, generally indicates that the preceding and following related objects have an "or" relationship.

[0089] This application discloses a method for constructing a multi-source heterogeneous document knowledge base based on adaptive normalization. (Refer to...) Figure 1 This process is performed by electronic devices, which can be servers or terminal devices. The server can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud computing services. Terminal devices can be smartphones, tablets, desktop computers, etc., but are not limited to these. These include:

[0090] Step S101: Collect raw document data of different modalities from multiple heterogeneous data sources.

[0091] Specifically, electronic devices connect to local file systems, network storage, cloud drives, and third-party collaboration platforms through a unified scheduling layer, supporting various document formats such as Word, PDF, JPG, MP3, MP4, PSD, DWG, and VSDX.

[0092] Electronic devices utilize automated plugins or adapters to capture data in real time. Examples include monitoring local file changes via File SystemWatcher, reading Word / Excel documents using the Office SDK, extracting PSD file layer information using the Photoshop API, and parsing DWG drawing attributes using the AutoCAD API. For cloud storage or third-party platforms, secure data collection is ensured through authenticated SDK calls. Collected data is transmitted to middleware (such as Kafka) via encrypted channels, and data traceability information, including source path, collection time, file size, and format type, is recorded.

[0093] Step S102: Preprocess the original document data and extract modal features to obtain the modal feature vectors corresponding to each original document data.

[0094] Step S1021: Perform full-text extraction on text documents, extract text content from image documents using OCR technology, perform speech-to-text transcription and keyframe extraction on audio or video documents, and extract hierarchical structure and object attributes from professional tool format documents to obtain preprocessed data.

[0095] Specifically, electronic devices employ differentiated processing for documents of different modalities. For image documents, text content is extracted while visual features are preserved. For audio or video documents, keyframes, including video or audio keyframes, are extracted. For professional tool formats such as PSD and DWG, layer names and pixel dimensions of PSDs, and construction types and coordinate information of DWGs can be extracted using dedicated parsing libraries.

[0096] Step S1022: Clean and denoise the preprocessed data.

[0097] Specifically, cleaning and denoising includes removing invalid information, such as placeholders in text, watermarks in images, and background noise in audio; it also includes standardizing formats, such as unifying image size to 1024×768 pixels and audio sampling rate to 16kHz; and it includes correcting data errors, such as typos in text and color distortion in images. Cleaning and denoising improves the purity of feature data and reduces the interference of invalid information on subsequent analysis.

[0098] Step S1023: For each cleaned and denoised data, use the feature extraction tool corresponding to the data mode to extract the mode feature vector.

[0099] Specifically, the BERT (Bidirectional Encoder Representations from Transformers) / RoBERTa (Robustly Optimized BERTPretraining Approach) model is used to extract text modal feature vectors for text documents, the CLIP (Contrastive Language-Image Pretraining) / ViT (Vision Transformer) model is used to extract visual modal feature vectors for image documents, the Wav2Vec2 (self-supervised speech representation learning framework 2.0) / AudioCLIP (audio-contrastive language-image pretraining model) model is used to extract audio modal feature vectors for audio documents, and a custom metadata extraction method is used to generate professional tool modal feature vectors for professional tool format documents.

[0100] Electronic devices transform heterogeneous documents into structured feature vectors, laying the data foundation for cross-modal association.

[0101] Step S103: Dynamically map the modal feature vectors of different modalities to a unified shared semantic vector space, dynamically calculate and adjust the weights of each modality according to the current semantic distribution and business scenario, and aggregate the feature vectors of each modality by weighting to generate a unified document vector.

[0102] Specifically, electronic devices use a multimodal Transformer to map modal feature vectors of different modalities to a unified shared semantic vector space, ensuring that vectors such as text, images, and audio have comparable semantic scales.

[0103] For common business scenarios, electronic devices predefine initial base weights for each modality. When the system starts, it loads the corresponding base weight template based on the business scenario selected by the user, which serves as the initial value for dynamic weight adjustment. For example, in a technical document scenario, the text modality weight is 0.6, the image modality weight is 0.3, and the audio modality weight is 0.1; or in a video tutorial scenario, the video modality weight is 0.5, the audio modality weight is 0.3, and the text modality weight is 0.2.

[0104] For each modality feature vector mapped to the shared semantic vector space, an adaptive kernel density estimation algorithm is used to calculate the semantic density value of the local region where all feature vectors in each modality are located.

[0105] The formula for calculating the semantic density value D is: ;

[0106] In the formula, x is the target vector. Let x be the neighborhood vector, n be the number of samples in the neighborhood of the target vector x, h be the bandwidth parameter, d be the vector dimension (which can be 1024), and K be the kernel function. Where, u= , Let be the Euclidean norm of the vector.

[0107] Furthermore, the electronic device normalizes the semantic density values ​​of each modality to obtain the semantic adaptation coefficients of each modality. Then, the base weights are multiplied by the semantic adaptation coefficients to obtain the preliminarily adjusted weights.

[0108] Next, the electronic device extracts the core semantic keywords from the currently processed document set and determines the relevance of each modality to the core keywords. For example, when the core keyword is "material parameters," the text modality has the highest relevance, assigned a relevance coefficient of 1.2; the image modality is second, with a relevance coefficient of 1; and the audio modality has the lowest relevance, with a relevance coefficient of 0.8. The initial adjusted weights are multiplied by the relevance coefficients to obtain the final dynamic weights for each modality.

[0109] When a new document enters the system, the above steps are repeated, and the modal weights are updated in real time based on the semantic distribution of the new document and the keywords of the current scenario. If the system detects a change in business scenario, the basic weight template for the corresponding scenario is reloaded before proceeding with the subsequent adjustment process.

[0110] Furthermore, the electronic device generates a unified document vector by weighted summation of the feature vectors of each modality.

[0111] Step S104: Based on the modality-specific AI agent, perform local inference on the modality feature vectors respectively to obtain the local analysis results and the inference confidence of each modality. Share the local analysis results and inference confidence through a unified message bus.

[0112] Specifically, the electronic devices deploy dedicated AI agents for each modality, with each AI agent focusing on local reasoning for its own modality.

[0113] The text AI agent performs topic classification, named entity recognition, and sentiment analysis; the image AI agent performs object detection and image tag generation; the audio AI agent performs speaker recognition and keyword extraction; the video AI agent performs video content parsing and summary generation; and the professional tool document AI agent performs layout layer relationship recognition and design element classification. After inference, the system outputs local analysis results and inference confidence.

[0114] Electronic devices upload the local analysis results, modal feature vectors, and inference confidence of each agent to the shared model via a unified message bus.

[0115] Step S105: Based on the unified document vector, vector similarity nearest neighbor retrieval and graph algorithms are used in the shared semantic vector space to obtain cross-modal semantic association results and identify sets of documents that are semantically similar, conflicting, or duplicated.

[0116] Step S0151: For all unified document vectors in the shared semantic vector space, use the adaptive kernel density estimation algorithm to calculate the semantic density value of the local region where each unified document vector is located.

[0117] Specifically, electronic devices employ existing adaptive kernel density estimation algorithms to calculate the semantic density values ​​of local regions for each unified document vector in a shared semantic space. These semantic density values ​​accurately quantify the degree of semantic association between each vector and its neighboring vectors. The more similar vectors within a local region and the more concentrated their distribution, the larger the calculated semantic density value, indicating that the region is more likely to contain similar documents. Conversely, the smaller the semantic density value, the fewer the locally similar vectors, and the more likely they are isolated, unrelated vectors.

[0118] Step S1052: Select the core vector clusters whose semantic density value is not lower than the preset density threshold, and remove isolated vectors whose semantic density value is lower than the preset density threshold.

[0119] Specifically, electronic devices preset a density threshold, then compare the semantic density value with the preset density threshold, eliminate unrelated documents, and thus retain a set of documents of the same type.

[0120] Step S1053: For the unified document vector within each core vector cluster, extract the corresponding modality tracing information and the core fields of the local analysis results, calculate the semantic overlap of the core fields of the local analysis results between unified document vectors of different modalities, retain vector pairs with semantic overlap higher than the preset semantic overlap, and construct a modality semantic consistency matrix.

[0121] Specifically, modal source information is the set of types of a unified document vector, such as those from text and image modalities, and core fields of local analysis results, such as the project name of text and the construction type of images.

[0122] Electronic devices construct a modal semantic consistency matrix M(m×m), where m is the number of documents within a cluster, and the matrix elements are... This indicates the semantic overlap of the core fields between document i and document j, with a value ranging from 0 to 1, calculated based on the ratio of the number of overlapping fields to the total number of fields.

[0123] Then, the electronic device presets an overlap threshold, for example, 0.8, and retains vector pairs with semantic overlap greater than the preset overlap threshold to obtain the filtered modal semantic consistency matrix.

[0124] Step S1054: For the modality semantic consistency matrix, the cosine similarity algorithm is used to calculate the vector spacing to obtain the initial similarity and to calculate the mean matrix similarity. The initial similarity is calibrated by combining the dynamic weights of each modality and the mean matrix similarity to obtain the corrected similarity.

[0125] Specifically, the initial similarity is calculated using the cosine similarity algorithm to determine the uniform document vector distance between the two documents, and the row mean of the modality semantic consistency matrix M is calculated to obtain the matrix similarity mean, which is used to represent the average semantic consistency between document i and other documents in the cluster.

[0126] Furthermore, since cosine similarity is calculated on the entirety of a unified document vector, but a unified document vector is a weighted aggregation of feature vectors from different modalities, if the core fields of two documents are inherently inconsistent, even if a small vector angle results in a high cosine similarity, it may still be a spurious similarity. Moreover, cosine similarity cannot reflect the contribution of the core modality in the current scenario. In contrast, the mean matrix similarity can quantify the overlap of core fields across different modalities in two documents. Using it for calibration can filter out cases where vectors are similar but semantic cores are inconsistent. The dynamic weight summation of modalities can anchor the weight proportion of the core modality in the current scenario, making the similarity align with the needs of the scenario.

[0127] If the initial similarity is The corrected formula is: In the formula, The mean of matrix similarity. and The weights are derived from experience. This is the sum of the modal dynamic weights of the successfully matched modalities between the two documents.

[0128] Step S1055: Calculate the semantic distribution variance of the core vector cluster, dynamically generate a distance threshold, and filter out vectors to be matched whose vector spacing is less than the distance threshold.

[0129] Specifically, the distance threshold is ,in The variance represents the semantic distribution. The smaller the variance, the smaller the distance threshold, which makes the selection of matching vectors more rigorous and more likely to select semantically similar vectors.

[0130] Step S1056: Using the vector to be matched as nodes, and the weighted value of semantic overlap and the corrected similarity as edge weights, construct an initial association graph, and strengthen the relationships in the initial association graph:

[0131] ① If the edge weight is greater than or equal to the approximation threshold and less than the repetition threshold, it is determined to be semantically approximate.

[0132] ② If the edge weight is greater than or equal to the approximate threshold and the core fields of the local analysis results are contradictory, then it is determined to be a semantic conflict.

[0133] ③ If the edge weight is greater than or equal to the repetition threshold and the modal semantic consistency matrix matching degree is higher than the preset matching threshold, then it is determined to be semantic repetition.

[0134] Among them, the approximate threshold is less than the repetition threshold.

[0135] Specifically, electronic devices preset an approximation threshold and a repetition threshold. For example, the approximation threshold is 0.7 and the repetition threshold is 0.9. When the edge weight is greater than or equal to the approximation threshold and less than the repetition threshold, the two documents are semantically similar based on the combined semantic overlap and similarity. Similarly, if the edge weight is greater than or equal to the repetition threshold and the modal semantic consistency matrix matching degree is higher than the preset matching threshold, then it is considered semantic repetition. The preset matching threshold can be 0.95.

[0136] Step S1057: Collect documents that are semantically similar, conflicting, or duplicated.

[0137] Step S106: Perform online aggregation and deduplication operations on the document collection.

[0138] Step S11: For each document in the document collection, calculate the semantic entropy weights by combining the unified document vector with the corresponding local analysis results.

[0139] Specifically, when the electronic device extracts the core fields of the local analysis results by the AI ​​agent, it simultaneously obtains the inference confidence of each field and then calculates the field weight ratio:

[0140] ;

[0141] In the formula, k is the field index, and n is the number of core fields. Let the inference confidence of the i-th field be denoted as follows: For example, if the confidences of the three core fields are 0.9, 0.8, and 0.7 respectively, then the weight percentage of the first core semantic field is... =0.9 / (0.9+0.8+0.7)=0.375.

[0142] Semantic entropy weights : ;

[0143] In the formula, in the formula, This represents the weight percentage of the i-th core semantic field; For scene adaptation factor, 0 < <1; is the semantic complexity coefficient.

[0144] Based on the information entropy theory, it reflects the uniformity of the weight distribution of core semantic fields. The more uniform the distribution, that is, the more important each field is, the greater the entropy value; the more concentrated the distribution, that is, the more dominant a certain field is, the smaller the entropy value.

[0145] This reflects the richness and hierarchical depth of a document's semantics. Electronic devices utilize dedicated AI agents for each modality to extract core semantic fields and nesting levels from the document, count the number of fields and the maximum hierarchical depth, and then transform them into a value less than 0 using a linear normalization formula. For values ​​less than 1, the more core fields and the deeper the hierarchy, the better. The larger the value, the more automatically it will be recalculated in real time when a new document is added. It is obtained automatically by the system and requires no manual intervention.

[0146] Scenario adaptation factor: This factor balances the weights of semantic distribution uniformity and complexity, dynamically adjusting based on the business scenario. Users select a business scenario during initialization, and the system loads the corresponding scenario. The base value is then automatically fine-tuned based on metrics such as cross-modal semantic consistency verification pass rate and deduplication accuracy, using scenario adaptation scores. The value fluctuates within a reasonable range, and new base values ​​are automatically loaded and adapted when switching scenes.

[0147] Step S12: Compare the similarity T between any two documents with the preset high-level similarity threshold T1, medium-level similarity threshold T2, and low-level similarity threshold T3:

[0148] (1) When T > T1, if the absolute value of the difference in semantic entropy weight between the two documents is less than or equal to the entropy weight threshold, the two documents are determined to be completely duplicate documents, and the duplicates are automatically removed while the original document links are retained; if the absolute value of the difference in semantic entropy weight between the two documents is greater than the entropy weight threshold, the two documents are determined to be suspected duplicate documents and are pushed for manual review.

[0149] Specifically, the electronic device presets T1, T2, and T3, with T1 > T2 > T3, and also presets an entropy weight threshold, such as 0.2.

[0150] If T > T1, it indicates that the similarity between the two documents is very high. Furthermore, if the absolute value of the semantic entropy weight difference is less than or equal to the entropy weight threshold, they can be essentially identified as duplicate documents. Automatic deduplication is performed while preserving the original document links. For example, if document A is stored at " / projectA / constructionplan.docx" and document B is stored at " / backup / constructionplan_v2.docx", after deduplication, the path of document A is retained, and the path of document B is associated.

[0151] (2) When T2<T≤T1, if the absolute value of the semantic entropy weight difference between the two documents is less than or equal to the entropy weight threshold and the overlap of the core semantic fields is greater than or equal to the preset overlap, then the two documents are determined to be semantically similar documents, and the two documents are automatically aggregated into a related document group and a semantic mapping is established; if the absolute value of the semantic entropy weight difference between the two documents is greater than the entropy weight threshold or the overlap of the core semantic fields is less than the preset overlap, the two documents are determined to be locally similar documents, independent entries are retained and the association relationship is marked.

[0152] (3) When T3≤T≤T2, if the absolute value of the difference in semantic entropy weight between the two documents is less than or equal to the entropy weight threshold and the cross-modal semantic consistency check passes, then the two documents are determined to be potentially related documents and included in the weak association link of the knowledge graph; if the consistency check fails, then the association is excluded.

[0153] Step S13: For documents with semantic conflicts:

[0154] If the confidence levels of the inferences of both parties are respectively and The semantic entropy weights are respectively and Then calculate priority. : In the formula, and These are the weighting coefficients. + =1, i=1 or 2. The document with higher priority is used as the benchmark document. If the priority difference between the two documents is less than the preset difference, it is determined that the conflict cannot be automatically resolved, a conflict label is generated and pushed to the visualization interface for manual review.

[0155] Specifically, the priority weighs the inference confidence and semantic entropy weight of each document. The more concentrated a certain field is, i.e., the smaller the semantic entropy weight, the clearer the content and the more suitable it is as a benchmark document. Documents with higher inference confidence are also more suitable as benchmark documents. Through comprehensive calculation, a comprehensive priority is obtained. The higher the priority, the more suitable it is as a benchmark document.

[0156] For example, if It is 0.6. The value is 0.4, for document A. =0.15, =0.9, then Y=0.6×(1-0.15)+0.4×0.9=0.87, document B =0.2, =0.8, then Y=0.8. Document A is used as the baseline file.

[0157] However, if the priorities of the two documents differ significantly, a conflict marker is generated and the documents are manually reviewed.

[0158] Step S107: Construct a knowledge graph based on cross-modal semantic association results, and update the knowledge graph with newly added original document data in an incremental manner.

[0159] Step S1071: Extract entity-relation-entity triples from the document and add modal origination tags. Construct an anchoring matrix based on the cross-modal semantic association results and retain valid triples with multimodal support or high confidence.

[0160] Specifically, electronic devices employ different extraction tools to extract triples for each modality of document. For example, text documents use OpenIE 5.0 (Open Information Extraction 5.0) to extract triples, image documents are extracted based on a combination of object detection results and OCR text, and documents from specialized tools are constructed based on attribute extraction results. Modality source tags identify the modality origin of the triples.

[0161] Furthermore, the electronic device constructs an anchoring matrix, where each row represents a triple and the columns represent the number of supported modes and the inference confidence. Triples that support ≥2 modes or have a confidence level ≥0.8 are selected.

[0162] Step S1072: Through global coarse alignment and local association path verification, the newly added entities are mapped to the unique identifier of the knowledge graph, and the low-confidence entities are temporarily stored in the confirmation database.

[0163] Specifically, global coarse alignment calculates the name similarity between the new entity and existing entities in the knowledge graph, while local association path verification checks whether the association paths of the new entity and existing entities are consistent. For example, if the new entity is "Bridge Body" and the existing entity "Bridge Project" has the association path "Bridge Body - Belongs to - Bridge Project," which is consistent with the path of similar entities in the knowledge graph, then alignment is confirmed. Entities with similarity between 0.5 and 0.8 are temporarily stored in a confirmation database, alignment suggestions are generated, and then pushed for manual review.

[0164] Step S1073: Detect indirect contradiction propagation paths through graph neural networks, automatically correct low-confidence contradictions or mark high-confidence contradictions for manual review.

[0165] For example, the document "Document A: Bridge Main Girder - Material - Q355 Steel" and "Document B: Bridge Main Girder - Material - Q235 Steel" were linked through the path "Bridge Main Girder - Belongs to - Bridge Project," and a conflict propagation was detected. Low-confidence conflicts (where the confidence of the conflict triplet is <0.7) are automatically corrected using high-confidence triplets; high-confidence conflicts (where the confidence of both parties is ≥0.7) are marked with a conflict type, such as "Material Attribute Conflict," and are then sent for manual review.

[0166] Step S1074: Add timestamps to entities, relations, and triples, and construct evolutionary trajectories by combining cross-modal associations.

[0167] For example, a new document is added, and the triple "Bridge Project X - Contains - Steel Box Girder Main Beam" is extracted. The confidence level is 0.92, and the text + image modality is supported. Through global coarse alignment and path verification, the similarity with the existing entity "Bridge Project X" is 1.0, and it is aligned to the existing entity. No conflict is detected, and the timestamp "2024-01-15" is added to complete the incremental update.

[0168] Step S108: By combining local analysis results, inference confidence, cross-modal semantic association results, and system performance monitoring indicators, the active learning process is automatically triggered, and the model parameters of the AI ​​agent are updated in real time by combining manual annotation feedback.

[0169] For example, the performance indicators of electronic equipment real-time monitoring systems, such as extraction speed, similarity distribution, deduplication rate, and inference confidence, can automatically trigger a manual annotation process for samples with an inference confidence of <0.6, such as pushing them to an annotation platform for experts to annotate semantic association types and conflict solutions.

[0170] Furthermore, manually labeled data is used as training samples, and TensorFlow Serving (TensorFlow model service framework) is used to achieve online fine-tuning of AI agent models, updating core parameters such as parsing rules, semantic matching thresholds, and modality weight calculation logic.

[0171] Furthermore, the performance of the statistical model changes every quarter. If the deduplication accuracy drops by ≥5%, a full model update is triggered. If the processing error of a certain type of document increases, the feature extraction model of that modality is adjusted accordingly.

[0172] The system enables continuous adaptive optimization of the model, reduces reliance on manual updates by experts, and ensures that system performance continues to improve with business iterations through a closed-loop learning mechanism.

[0173] Furthermore, the above method also includes:

[0174] Step S21: Construct a three-layer view of document association network, knowledge graph evolution and deduplication decision through a visualization panel, which displays document clustering relationships, graph update trajectory and automatic processing basis, and supports node highlighting, version backtracking and manual review entry.

[0175] Specifically, for the document association network view: documents are represented by nodes, semantic relationships are represented by edges, node size maps to document weight, and zooming, panning, and node highlighting are supported.

[0176] Knowledge graph evolution view: Displays the addition, modification, and deletion trajectory of entities and relationships, with a timeline controlling version backtracking and different colors marking update types.

[0177] Deduplication Decision View: Displays the basis for automatic processing, including similarity scores, semantic entropy weights, core field comparisons, and decision rules.

[0178] Step S22: Supports graphical configuration of visual parameters, and configuration changes are synchronized to the system policy library in real time.

[0179] Specifically, it supports adjusting parameters such as similarity threshold, entropy weight threshold, scene adaptation factor, and various weights. Configuration changes are synchronized to the system policy library in real time without requiring a service restart. It also supports saving configuration schemes for different business scenarios and loading them directly when needed.

[0180] Step S23: Provides multi-dimensional filtering function. Clicking on a node allows you to view the original content, partial analysis results, and inference confidence of the corresponding document.

[0181] To better implement the above method, embodiments of this application also provide a multi-source heterogeneous document knowledge base construction system based on adaptive normalization, referring to... Figure 2 The system for building a multi-source heterogeneous document knowledge base based on adaptive normalization includes:

[0182] The data acquisition module 201 is used to acquire raw document data of different modalities from multiple heterogeneous data sources;

[0183] The modal feature vector extraction module 202 is used to preprocess the original document data and extract modal features to obtain the modal feature vectors corresponding to each original document data.

[0184] The modal feature vector aggregation module 203 is used to dynamically map modal feature vectors of different modalities to a unified shared semantic vector space. It dynamically calculates and adjusts the weights of each modality according to the current semantic distribution and business scenario, and performs weighted aggregation of each modal feature vector to generate a unified document vector.

[0185] The local inference module 204 is used to perform local inference on the modality feature vectors based on the modality-specific AI agent, obtain the local analysis results and the inference confidence of each modality, and share the local analysis results and inference confidence through a unified message bus.

[0186] The association module 205 is used to obtain cross-modal semantic association results based on a unified document vector and in a shared semantic vector space using vector similarity nearest neighbor retrieval and graph algorithms, and to identify sets of documents that are semantically similar, conflicting, or duplicated.

[0187] Document processing module 206 is used to perform online aggregation and deduplication operations on document collections;

[0188] The knowledge graph update module 207 is used to construct a knowledge graph based on cross-modal semantic association results and update the knowledge graph with newly added original document data in an incremental manner.

[0189] The active learning module 208 is used to automatically trigger the active learning process by combining local analysis results, inference confidence, cross-modal semantic association results, and system performance monitoring indicators, and to update the model parameters of the AI ​​agent in real time by combining manual annotation feedback.

[0190] The various variations and specific examples of the methods in the foregoing embodiments are also applicable to the adaptive normalization-based multi-source heterogeneous document knowledge base construction system of this embodiment. Through the foregoing detailed description of the adaptive normalization-based multi-source heterogeneous document knowledge base construction method, those skilled in the art can clearly understand the implementation method of the adaptive normalization-based multi-source heterogeneous document knowledge base construction system of this embodiment. Therefore, for the sake of brevity, it will not be described in detail here.

[0191] To better implement the above methods, embodiments of this application provide an electronic device, referring to... Figure 3 The electronic device 300 includes a processor 301, a memory 303, and a display screen 305. The memory 303 and the display screen 305 are both connected to the processor 301, such as via a bus 302. Optionally, the electronic device 300 may also include a transceiver 304. It should be noted that in practical applications, the transceiver 304 is not limited to one type, and the structure of this electronic device 300 does not constitute a limitation on the embodiments of this application.

[0192] Processor 301 may be a CPU (Central Processing Unit), a general-purpose processor, a DSP (Digital Signal Processor), an ASIC (Application Specific Integrated Circuit), an FPGA (Field Programmable Gate Array), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. It can implement or execute the various exemplary logic blocks, modules, and circuits described in conjunction with the disclosure of this application. Processor 301 may also be a combination that implements computational functions, such as including one or more microprocessor combinations, a combination of a DSP and a microprocessor, etc.

[0193] Bus 302 may include a pathway for transmitting information between the aforementioned components. Bus 302 may be a PCI (Peripheral Component Interconnect) bus or an EISA (Extended Industry Standard Architecture) bus, etc. Bus 302 may be divided into address bus, data bus, control bus, etc.

[0194] The memory 303 may be a ROM (Read Only Memory) or other type of static storage device capable of storing static information and instructions, RAM (Random Access Memory) or other type of dynamic storage device capable of storing information and instructions, or an EEPROM (Electrically Erasable Programmable Read Only Memory), CD-ROM (Compact Disc Read Only Memory) or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital universal optical discs, Blu-ray discs, etc.), magnetic disk storage media or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but not limited thereto.

[0195] The memory 303 is used to store application code that executes the solution of this application, and its execution is controlled by the processor 301. The processor 301 is used to execute the application code stored in the memory 303 to implement the content shown in the foregoing method embodiments.

[0196] Figure 3 The electronic device 300 shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments of this application.

[0197] This application also provides a computer-readable storage medium storing a computer program. When executed by a processor, the program implements the multi-source heterogeneous document knowledge base construction method based on adaptive normalization provided in the above embodiments. Through adaptive normalization and multimodal fusion technology, it achieves deep semantic association and dynamic integration of multi-source heterogeneous documents, significantly improving the accuracy and processing efficiency of cross-type document merging. The online aggregation and deduplication mechanism effectively reduces information redundancy and conflicts. Incremental updates to the knowledge graph ensure real-time iteration of knowledge assets. Active learning and adaptive model optimization reduce reliance on expert maintenance. Full-process automation significantly reduces manual costs. The system adapts to different business scenarios, providing an efficient, intelligent, and low-loss solution for knowledge base construction, and enhancing knowledge retrieval and decision support capabilities.

[0198] In this embodiment, the computer-readable storage medium can be a tangible device that holds and stores instructions used by an instruction execution device. The computer-readable storage medium can be, but is not limited to, an electrical storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any combination thereof. Specifically, the computer-readable storage medium can be a portable computer disk, a hard disk, a USB flash drive, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory stick, floppy disk, optical disk, magnetic disk, mechanical encoding device, or any combination thereof.

[0199] The computer program in this embodiment includes program code for performing all the aforementioned methods. The program code may include instructions corresponding to the method steps provided in the above embodiments. The computer program can be downloaded from a computer-readable storage medium to various computing / processing devices, or downloaded to an external computer or external storage device via a network (e.g., the Internet, a local area network, a wide area network, and / or a wireless network). The computer program can be executed entirely on the user's computer as a standalone software package.

[0200] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

[0201] Additionally, it should be understood that relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. The terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

Claims

1. A method for constructing a multi-source heterogeneous document knowledge base based on adaptive normalization, characterized in that, include: Collect raw document data of different modalities from multiple heterogeneous data sources; The original document data is preprocessed and modal features are extracted to obtain modal feature vectors corresponding to each original document data. The modal feature vectors of different modalities are dynamically mapped to a unified shared semantic vector space. The weights of each modality are dynamically calculated and adjusted according to the current semantic distribution and business scenario. The modal feature vectors are then weighted and aggregated to generate a unified document vector. Based on modality-specific AI agents, local inference is performed on the modality feature vectors to obtain local analysis results and inference confidence for each modality. The local analysis results and inference confidence are shared through a unified message bus. Based on the unified document vector, vector similarity nearest neighbor retrieval and graph algorithms are used in the shared semantic vector space to obtain cross-modal semantic association results and identify sets of documents that are semantically similar, conflicting, or duplicated. Perform online aggregation and deduplication operations on the document collection; A knowledge graph is constructed based on the cross-modal semantic association results, and newly added original document data is updated to the knowledge graph incrementally. Combining the local analysis results, the inference confidence, the cross-modal semantic association results, and system performance monitoring indicators, an active learning process is automatically triggered, and the model parameters of the AI ​​agent are updated in real time in conjunction with manual annotation feedback. Based on the unified document vector, the method utilizes vector similarity nearest neighbor retrieval and graph algorithms within the shared semantic vector space to obtain cross-modal semantic association results and identify sets of documents with similar, conflicting, or duplicate semantics, including: For all unified document vectors in the shared semantic vector space, an adaptive kernel density estimation algorithm is used to calculate the semantic density value of the local region where each unified document vector is located; Filter out the core vector clusters whose semantic density value is not lower than the preset density threshold, and remove isolated vectors whose semantic density value is lower than the preset density threshold; For each unified document vector within the core vector cluster, extract the corresponding modality tracing information and core fields of the local analysis results, calculate the semantic overlap of the core fields of the local analysis results between unified document vectors of different modalities, retain vector pairs with semantic overlap higher than a preset overlap threshold, and construct a modality semantic consistency matrix. For the modality semantic consistency matrix, the initial similarity is obtained by calculating the vector spacing using the cosine similarity algorithm and the mean matrix similarity is calculated. The initial similarity is then calibrated by combining the dynamic weights of each modality with the mean matrix similarity to obtain the corrected similarity. Calculate the semantic distribution variance of the core vector cluster, dynamically generate a distance threshold, and filter out the vectors to be matched whose vector spacing is less than the distance threshold; Using the vector to be matched as a node, and the weighted value of the semantic overlap and the corrected similarity as the edge weight, an initial association graph is constructed, and the relationship of the initial association graph is strengthened. If the edge weight is greater than or equal to the approximation threshold and less than the repetition threshold, it is determined to be semantically approximate; If the edge weight is greater than or equal to an approximate threshold and the core fields of the local analysis results are contradictory, then it is determined to be a semantic conflict; If the edge weight is greater than or equal to the repetition threshold and the modal semantic consistency matrix matching degree is higher than the preset matching threshold, then it is determined to be semantic repetition; wherein, the approximate threshold is less than the repetition threshold; Collect documents that are semantically similar, conflicting, or repetitive.

2. The method for constructing a multi-source heterogeneous document knowledge base based on adaptive normalization according to claim 1, characterized in that, The modalities include text, images, audio, video, and professional tool formats. The preprocessing and modal feature extraction of the original document data to obtain modal feature vectors corresponding to each original document data includes: Full-text extraction is performed on text documents, OCR technology is used to extract text content from image documents, speech-to-text transcription and keyframe extraction are performed on audio or video documents, and hierarchical structure and object attributes are extracted from professional tool format documents to obtain preprocessed data. The preprocessed data is then cleaned and denoised. For each cleaned and denoised data, a modal feature vector is extracted using a feature extraction tool corresponding to the data modality.

3. The method for constructing a multi-source heterogeneous document knowledge base based on adaptive normalization according to claim 1, characterized in that, The online aggregation and deduplication operation on the document collection includes: For each document in the document set, the unified document vector is combined with the corresponding local analysis results to calculate the semantic entropy weight; For documents with similar or repetitive semantics: Compare the similarity T between any two documents with the preset high-level similarity threshold T1, medium-level similarity threshold T2, and low-level similarity threshold T3: When T > T1, if the absolute value of the semantic entropy weight difference between the two documents is less than or equal to the entropy weight threshold, the two documents are determined to be completely duplicate documents, and the duplicates are automatically removed while the original document links are retained; if the absolute value of the semantic entropy weight difference between the two documents is greater than the entropy weight threshold, the two documents are determined to be suspected duplicate documents and are pushed for manual review. When T2 < T ≤ T1, if the absolute value of the semantic entropy weight difference between the two documents is less than or equal to the entropy weight threshold and the overlap of the core semantic fields is greater than or equal to the preset overlap, then the two documents are determined to be semantically similar documents, and the two documents are automatically aggregated into a related document group and a semantic mapping is established; if the absolute value of the semantic entropy weight difference between the two documents is greater than the entropy weight threshold or the overlap of the core semantic fields is less than the preset overlap, then the two documents are determined to be locally similar documents, independent entries are retained and the association relationship is marked; When T3≤T≤T2, if the absolute value of the semantic entropy weight difference between the two documents is less than or equal to the entropy weight threshold and the cross-modal semantic consistency check passes, then the two documents are determined to be potentially related documents and included in the weak association link of the knowledge graph; if the consistency check fails, the association is excluded. The order is T1 > T2 > T3; For documents with semantic conflicts: If the inference confidence levels of the two documents are respectively and The semantic entropy weights are respectively Then calculate priority. In the formula, These are the weighting coefficients. ; The document with higher priority is used as the benchmark document; If the priority difference between two documents is less than the preset difference, it is determined that the conflict cannot be automatically resolved, a conflict label is generated and pushed to the visualization interface for manual review.

4. The method for constructing a multi-source heterogeneous document knowledge base based on adaptive normalization according to claim 3, characterized in that, The calculation of semantic entropy weights includes: The semantic entropy weight Calculation method: ； In the formula, This represents the weight percentage of the i-th core semantic field; For scene adaptation factors, ; The semantic complexity coefficient is... .

5. The method for constructing a multi-source heterogeneous document knowledge base based on adaptive normalization according to claim 1, characterized in that, The construction of the knowledge graph involves incrementally updating the knowledge graph with newly added original document data, including: Extract entity-relation-entity triples from the newly added original document data and add modal origination tags. Construct an anchoring matrix based on the cross-modal semantic association results and retain valid triples with multimodal support or high confidence. By using global coarse alignment and local association path verification, newly added entities are mapped to unique identifiers in the knowledge graph, while low-confidence entities are temporarily stored in a confirmation database. By detecting indirect contradiction propagation paths through graph neural networks, low-confidence contradictions are automatically corrected or high-confidence contradictions are marked for manual review. Add timestamps to entities, relations, and triples, and construct evolutionary trajectories by combining cross-modal associations.

6. The method for constructing a multi-source heterogeneous document knowledge base based on adaptive normalization according to claim 1, characterized in that, The method further includes: The system constructs a three-layer view of document association network, knowledge graph evolution, and deduplication decision through a visualization panel, which respectively displays document clustering relationships, graph update trajectory, and automatic processing basis, and supports node highlighting, version backtracking, and manual review entry. It supports graphical configuration of visual parameters, and configuration changes are synchronized to the system policy library in real time; It offers multi-dimensional filtering capabilities; clicking on a node allows you to view the original content of the corresponding document, partial analysis results, and inference confidence.

7. A system for constructing a multi-source heterogeneous document knowledge base based on adaptive normalization, characterized in that, include: The data acquisition module is used to collect raw document data of different modalities from multiple heterogeneous data sources; The modal feature vector extraction module is used to preprocess and extract modal features from the original document data to obtain modal feature vectors corresponding to each original document data. The modal feature vector aggregation module is used to dynamically map the modal feature vectors of different modalities to a unified shared semantic vector space. It dynamically calculates and adjusts the weights of each modality according to the current semantic distribution and business scenario, and performs weighted aggregation of each modal feature vector to generate a unified document vector. The local inference module is used to perform local inference on the modality feature vectors based on the modality-specific AI agent, obtain local analysis results and inference confidence of each modality, and share the local analysis results and inference confidence through a unified message bus. The association module is used to obtain cross-modal semantic association results based on the unified document vector and within the shared semantic vector space using vector similarity nearest neighbor retrieval and graph algorithms, and to identify sets of documents that are semantically similar, conflicting, or duplicated. The document processing module is used to perform online aggregation and deduplication operations on the document collection; The knowledge graph update module is used to construct a knowledge graph based on the cross-modal semantic association results and update the knowledge graph with newly added original document data in an incremental manner. The active learning module is used to automatically trigger the active learning process by combining the local analysis results, the inference confidence, the cross-modal semantic association results, and the system performance monitoring indicators, and to update the model parameters of the AI ​​agent in real time by combining manual annotation feedback. The associated module is specifically used for: For all unified document vectors in the shared semantic vector space, an adaptive kernel density estimation algorithm is used to calculate the semantic density value of the local region where each unified document vector is located; Filter out the core vector clusters whose semantic density value is not lower than the preset density threshold, and remove isolated vectors whose semantic density value is lower than the preset density threshold; For each unified document vector within the core vector cluster, extract the corresponding modality tracing information and core fields of the local analysis results, calculate the semantic overlap of the core fields of the local analysis results between unified document vectors of different modalities, retain vector pairs with semantic overlap higher than a preset overlap threshold, and construct a modality semantic consistency matrix. For the modality semantic consistency matrix, the initial similarity is obtained by calculating the vector spacing using the cosine similarity algorithm and the mean matrix similarity is calculated. The initial similarity is then calibrated by combining the dynamic weights of each modality with the mean matrix similarity to obtain the corrected similarity. Calculate the semantic distribution variance of the core vector cluster, dynamically generate a distance threshold, and filter out the vectors to be matched whose vector spacing is less than the distance threshold; Using the vector to be matched as a node, and the weighted value of the semantic overlap and the corrected similarity as the edge weight, an initial association graph is constructed, and the relationship of the initial association graph is strengthened. If the edge weight is greater than or equal to the approximation threshold and less than the repetition threshold, it is determined to be semantically approximate; If the edge weight is greater than or equal to an approximate threshold and the core fields of the local analysis results are contradictory, then it is determined to be a semantic conflict; If the edge weight is greater than or equal to the repetition threshold and the modal semantic consistency matrix matching degree is higher than the preset matching threshold, then it is determined to be semantic repetition; wherein, the approximate threshold is less than the repetition threshold; Collect documents that are semantically similar, conflicting, or repetitive.

8. An electronic device, characterized in that, include: At least one processor; Memory; At least one computer program, wherein the at least one computer program is stored in the memory and configured to be executed by the at least one processor, the at least one computer program being configured to: perform the adaptive normalization-based multi-source heterogeneous document knowledge base construction method as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer program is stored that can be loaded by a processor and executed as described in any one of claims 1 to 6, for constructing a multi-source heterogeneous document knowledge base based on adaptive normalization.

Images