Big model-based knowledge base intelligent management system and method

Abstract

The application belongs to the technical field of database management, and particularly relates to a knowledge base intelligent management system and method based on a large model. The system collects question and answer texts and index information of different target subjects through a data collection module, generates a first text hierarchical sequence and a directed index graph through word segmentation preprocessing and directed graph neural network mapping, extracts text and index embedding representations in combination with a feature distillation model through a feature extraction module, obtains the similarity overlap of different target subjects through a three-dimensional overlap model, constructs an initial knowledge base based on a tree database through a conflict simulation module, detects index conflicts within and between target subjects by using a Bayesian simulation algorithm, generates a conflict risk list, outputs a resolution strategy in combination with a conflict resolution adversarial model through a conflict resolution module, and iteratively optimizes the knowledge base to meet the conflict threshold and user demand through a feedback adjustment module. The application realizes intelligent management of cross-subject question and answer indexes, and improves the knowledge base conflict resolution efficiency and cross-subject query accuracy.

Description

Technical Field

[0001] This invention belongs to the field of database management technology, and in particular relates to an intelligent management system and method for knowledge bases based on large models. Background Technology

[0002] In the advancement of healthcare informatization, the demand for cross-hospital collaborative diagnosis and treatment is increasing. This necessitates the use of a unified knowledge base to enable shared querying of question-and-answer information across hospitals. However, due to differences in treatment characteristics and management models, different hospitals exhibit significant individual characteristics in their database structures, configuration parameters, and indexing logic, posing challenges to cross-hospital information index management. Specifically, there is no unified standard for index structures, and varying departmental naming rules lead to incompatibility between key-value pairs when matching text and indexes. Semantic ambiguity exists, with the same term having different connotations in different hospitals, causing index positioning errors. Independent knowledge update mechanisms mean that when one hospital adds new information, other hospitals' indexes are not synchronized, leading to query failures or errors. Furthermore, the lack of an automatic cross-hospital index conflict verification mechanism makes manual troubleshooting inefficient, affecting the accuracy, response speed, and collaborative service capabilities of cross-hospital queries. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention proposes a knowledge base intelligent management system and method based on a large model. The system collects question-and-answer text and index information from different target subjects through a data acquisition module. After word segmentation preprocessing and directed graph neural network mapping, a first text hierarchical sequence and a directed index graph are generated. A feature extraction module extracts text and index embedding representations using a feature distillation model and obtains the similarity overlap of different target subjects through a three-dimensional overlap model. A conflict simulation module constructs an initial knowledge base based on a tree-structured database and uses a Bayesian simulation algorithm to detect index conflicts within and between target subjects, generating a conflict risk list. A conflict resolution module outputs resolution strategies using a conflict resolution adversarial model, and a feedback adjustment module iteratively optimizes the knowledge base to meet conflict thresholds and user requirements. This application achieves intelligent management of cross-subject question-and-answer indexes, improving the efficiency of knowledge base conflict resolution and the accuracy of cross-subject queries.

[0004] To achieve the above objectives, the present invention provides the following technical solution:

[0005] A knowledge base intelligent management system based on a large model includes: a data acquisition module, a feature extraction module, and a conflict simulation module;

[0006] The data acquisition module is used to collect question and answer text information and question and answer index information corresponding to different target subjects, and to perform word segmentation preprocessing on the collected text information, and to perform directed mapping processing on the index information using a directed graph neural network to obtain the first text hierarchical sequence and directed index graph corresponding to different target subjects; the directed index graph is constructed from the sub-subjects and required index paths corresponding to different search information under the corresponding target subject;

[0007] The feature extraction module, based on the first text hierarchical sequence and directed index graph corresponding to different target subjects, combined with a preset feature distillation extraction model, obtains the first text embedding representation and the first index embedding representation corresponding to different target subjects; based on the first text embedding representation and the first index embedding representation corresponding to different target subjects, combined with a preset three-dimensional overlap model, it obtains the similarity overlap of different target subjects; the similarity overlap of target subjects includes content overlap, index structure overlap, semantic representation overlap, and comprehensive overlap.

[0008] The conflict simulation module constructs an initial question-and-answer knowledge base based on question-and-answer text information and question-and-answer index information of different target subjects, combined with a tree-structured database. Based on the initial question-and-answer knowledge base and the similarity and overlap of different target subjects, it uses a Bayesian function-based simulation question-and-answer algorithm to detect and identify question-and-answer index conflicts between different sub-subjects within a target subject and between target subjects with different degrees of overlap, thus obtaining a target subject conflict risk list. The target subject conflict risk list includes the index conflict risk and probability corresponding to different sub-subjects within a target subject, as well as the index conflict risk and probability corresponding to target subjects with different degrees of overlap.

[0009] Specifically, the knowledge base intelligent management system also includes: a conflict resolution module and a feedback adjustment module;

[0010] The conflict resolution module, based on the target subject's conflict risk list and combined with a conflict resolution adversarial model with preset conflict resolution rules, obtains index conflict risk resolution strategies for different sub-subjects within the target subject and index conflict risk resolution strategies for target subjects with different degrees of overlap. The index conflict risk resolution strategies include correcting unified standard index keywords and directed index paths across sub-subjects or across subjects.

[0011] The feedback adjustment module feeds back the index conflict risk mitigation strategies for different sub-subjects within the target subject and the index conflict risk mitigation strategies for target subjects with different degrees of overlap to the initial question-and-answer knowledge base. It adjusts the initial index path in the initial question-and-answer knowledge base and then performs question-and-answer index conflict simulation, detection, identification, and conflict mitigation adjustment until the target subject's conflict risk list meets the corresponding index conflict risk threshold and the question-and-answer information meets the corresponding user's retrieval needs. At the same time, the index conflict risk mitigation strategy that meets the index conflict risk threshold and user retrieval needs is configured into the initial question-and-answer knowledge base to obtain the conflict-corrected question-and-answer knowledge base.

[0012] Specifically, the process of obtaining the first text embedding representation corresponding to the target subject includes:

[0013] Based on the question-and-answer text information corresponding to different target subjects, special symbols are filtered and mixed Chinese and English boundary annotations are performed by combining regular expressions with Chinese and English mixed boundary rules to obtain the preprocessed question-and-answer text information corresponding to different target subjects.

[0014] Based on the preprocessed question-and-answer text information corresponding to different target subjects and combined with the three-level hierarchical rules, we obtain the hierarchical question-and-answer text information of different target subjects;

[0015] Based on the question-and-answer text information after being layered according to different target subjects, combined with the preset entity relationship extraction algorithm, entity word segmentation is performed on the text corresponding to each layer under different target subjects, and the directed mapping relationship between the corresponding layer and different layers is extracted to obtain the first text embedding representation corresponding to different target subjects.

[0016] Specifically, the three-level hierarchical rule includes the topic layer, the sub-question layer, and the keyword layer. In the topic layer, based on the question and answer text information corresponding to each target subject, the co-occurrence frequency of keywords extracted in the search text and the response text is calculated using the TextRank algorithm. Core words are then selected. Based on the selected core word sequence, BioBERT is used to generate a target keyword vector cluster to obtain the topic name corresponding to the current target subject's question and answer information and assign the root node ID associated with the text storage address.

[0017] In the sub-question layer, sub-question sentence vectors are generated by combining the search text under each topic with Sentence-BERT. After clustering by the DBSCAN algorithm, the sub-question keyword sequence and sub-question answer priority are obtained. Sub-node IDs are assigned based on the sub-question answer priority corresponding to the sub-question keyword sequence, and the corresponding topic root node is associated with them.

[0018] In the keyword layer, based on the response text corresponding to each sub-question, the TF-IDF value of the keyword corresponding to each response text is obtained through a statistical algorithm. After comparison and verification with the UMLS terminology database, the keywords are sorted in descending order according to their TF-IDF values, and keyword IDs are assigned according to the sorting results and bound to the corresponding sub-question nodes.

[0019] Specifically, the construction process of the feature distillation extraction model includes:

[0020] Based on the first text hierarchical sequence corresponding to different target subjects, the teacher sub-model in the feature distillation extraction model performs hierarchical text feature extraction along the corresponding nodes of the topic layer, sub-question layer, and keyword layer, specifically as follows:

[0021] Based on the first text extraction network, the root node text of the topic layer is used to extract topic feature vectors to obtain the first topic vector;

[0022] Based on the second text extraction network, features are extracted from each priority sub-question in the sub-question layer to generate a hierarchical sub-question vector. Then, based on the hierarchical sub-question vector and the first topic vector, a cross-attention network is used to obtain the topic-oriented priority association weight vector of the sub-question.

[0023] Based on the third text extraction network, feature extraction is performed on the keyword layer corresponding to each sub-question according to the ranking result to obtain the keyword feature vector of the answer text corresponding to each sub-question. Based on the keyword information of the answer text corresponding to each sub-question, combined with the hierarchical sub-question vector and the cross-attention network, the association weight vector of the answer text keywords oriented towards the sub-question is obtained.

[0024] Based on the first topic vector, hierarchical sub-question vectors, sub-question-oriented topic priority association weight vectors, keyword feature vectors of the answer text corresponding to each sub-question, and keyword association weight vectors of the answer text oriented to the sub-question, a first text embedding representation of the question and answer text corresponding to the target subject is constructed through a graph neural network.

[0025] Based on the sub-subjects and demand index paths corresponding to different retrieval information under different target subjects in the directed index graph, combined with the GraphSAGE algorithm, the index information and access edge weight information features are extracted along the retrieval path corresponding to the directed index graph to obtain the first retrieval path features.

[0026] Based on the first retrieval path features, a convolutional attention network is used to extract the retrieval direction and frequency of access paths corresponding to different retrieval needs under the same subject, and the retrieval path direction and frequency of access paths for different subjects facing the same retrieval needs, so as to obtain the first index embedding representation corresponding to the target subject.

[0027] Based on the first text embedding representation and the first index embedding representation, the feature space is aligned by combining the feature alignment algorithm with the retrieval target requirement as the label to obtain the aligned first text embedding representation and the first index embedding representation.

[0028] The aligned first text embedding representation and the first index embedding representation are input into the text distillation branch and the index distillation branch of the student sub-model in the feature distillation extraction model, respectively. Feature distillation is performed according to the contribution of the corresponding features to the retrieval target requirements and the contribution threshold to obtain the first text embedding representation and the first index embedding representation after feature distillation.

[0029] Specifically, the process of obtaining the target subject conflict risk list includes:

[0030] Based on the question-and-answer text information and question-and-answer index information of different target subjects, an initial question-and-answer knowledge base is constructed through a tree-structured database. The initial question-and-answer knowledge base is associated with the nodes of the topic layer, sub-question layer, and keyword layer, as well as the sub-subjects, retrieval units, and path information of the directed index graph, forming a hierarchical index structure.

[0031] Extract the index features of different sub-subjects within the target subject and the similarity overlap data of different target subjects from the initial question-answering knowledge base, and use them as input parameters for conflict detection. The index features include a set of index keywords and a directed path sequence.

[0032] Based on the conflict detection between different sub-sub ...

[0033] Based on the conflict detection between target subjects under different overlap degrees, high overlap subject pairs are screened based on comprehensive overlap degree. The first text embedding representation and the first index embedding representation of the high overlap subject pairs are similar to identify feature pairs that are semantically related but have different indices. Comprehensive overlap degree is used as a weighting factor to adjust the prior probability of the Bayes function. The cross-subject retrieval scenario is simulated by a simulated question answering algorithm and the cross-subject conflict probability is calculated by combining the likelihood function. The index conflict risk entries between subjects are generated by arranging them in descending order of conflict probability.

[0034] Based on the intra-subject index conflict risk entries and inter-subject index conflict risk entries, supplementary conflict feature descriptions and conflict impact scope are generated to create a target subject conflict risk list.

[0035] Intelligent knowledge base management methods based on large models include:

[0036] Collect question-and-answer text information and question-and-answer index information corresponding to different target subjects, perform word segmentation preprocessing on the collected text information, and perform directed mapping processing on the index information in combination with a directed graph neural network to obtain the first text hierarchical sequence and directed index graph corresponding to different target subjects;

[0037] A directed index graph is constructed from the sub-subjects and index paths corresponding to different search information under the target subject;

[0038] Based on the first text hierarchical sequence and directed index graph corresponding to different target subjects, combined with the preset feature distillation extraction model, the first text embedding representation and the first index embedding representation corresponding to different target subjects are obtained.

[0039] Based on the first text embedding representation and the first index embedding representation corresponding to different target subjects, combined with the preset three-dimensional overlap model, the similarity overlap of different target subjects is obtained; the similarity overlap of target subjects includes content overlap, index structure overlap, semantic representation overlap, and comprehensive overlap.

[0040] An initial question-and-answer knowledge base is constructed based on question-and-answer text information and question-and-answer index information of different target subjects, combined with a tree-structured database. Based on the initial question-and-answer knowledge base and the similarity and overlap of different target subjects, a simulated question-and-answer algorithm using Bayesian functions is used to detect and identify question-and-answer index conflicts between different sub-subjects within the target subject and between target subjects with different overlap, thereby obtaining a list of target subject conflict risks.

[0041] The target subject conflict risk list includes the index conflict risk and probability corresponding to different sub-subjects within the target subject, as well as the index conflict risk and probability between target subjects under different degrees of overlap.

[0042] Specifically, intelligent knowledge base management methods also include:

[0043] Based on the target subject conflict risk list, and combined with the conflict resolution adversarial model with preset conflict resolution rules, the index conflict risk resolution strategies for different sub-subjects within the target subject and the index conflict risk resolution strategies for target subjects with different degrees of overlap are obtained; the index conflict risk resolution strategies include correcting the unified standard index keywords and the directed index paths across sub-subjects or across subjects;

[0044] The index conflict risk resolution strategies for different sub-subjects within the target subject and the index conflict risk resolution strategies for target subjects with different degrees of overlap are fed back to the initial question-answering knowledge base. The initial index path in the initial question-answering knowledge base is adjusted, and question-answering index conflict simulation, detection, identification, and conflict resolution are performed after the adjustment until the target subject's conflict risk list meets the corresponding index conflict risk threshold and the question-answering information meets the corresponding user's retrieval needs. At the same time, the index conflict risk resolution strategy that meets the index conflict risk threshold and user retrieval needs is configured into the initial question-answering knowledge base to obtain the conflict-corrected question-answering knowledge base.

[0045] Compared with the prior art, the beneficial effects of the present invention are:

[0046] This invention addresses the shortcomings of existing technologies by employing a data acquisition module to perform structured, layered processing of question-and-answer text and index information. Combined with a directed graph neural network, a hierarchical indexing system is constructed, enabling multi-granular semantic representation and association mapping of knowledge base data. The feature extraction module utilizes a feature distillation extraction model to perform deep feature alignment and compression on text and index information. A three-dimensional overlap model quantifies the multi-dimensional similarity relationships between different target subjects, providing accurate metrics for conflict detection. The conflict simulation module, based on a tree-structured database and a Bayesian simulation question-and-answer algorithm, simulates potential index conflicts between sub-subjects and across subjects during the knowledge base construction phase. A conflict list is generated through probabilistic risk assessment, significantly improving the knowledge base's pre-detection capabilities. The conflict resolution module dynamically optimizes the index path and keyword standards by combining an adversarial model, while the feedback adjustment module forms a closed-loop optimization mechanism, enabling the knowledge base to gradually converge to a conflict-free stable state through iterative adjustments. This application realizes intelligent management of the entire knowledge base process, from data collection, feature modeling, conflict warning to self-optimization adjustment, effectively solving the problem of low retrieval efficiency caused by semantic ambiguity and index redundancy in traditional knowledge bases. At the same time, the conflict prediction mechanism ensures the accuracy and consistency of cross-domain knowledge association, providing a scalable technical framework for the construction and maintenance of large-scale knowledge bases. Attached Figure Description

[0047] Figure 1 This is a module diagram of the knowledge base intelligent management system based on a large model, as described in Embodiment 1 of the present invention.

[0048] Figure 2 This is a flowchart of the knowledge base intelligent management method based on a large model, which is an embodiment of the present invention. Detailed Implementation

[0049] Example 1

[0050] Please see Figure 1 The present invention provides an embodiment of a knowledge base intelligent management system based on a large model, comprising: a data acquisition module, a feature extraction module, a conflict simulation module, a conflict resolution module, and a feedback adjustment module;

[0051] The data acquisition module is used to collect question-and-answer text information and question-and-answer index information corresponding to different target subjects. It performs word segmentation preprocessing on the collected text information and performs directed mapping processing on the index information using a directed graph neural network to obtain the first text hierarchical sequence and directed index graph corresponding to different target subjects. It should be further noted that the question-and-answer text information corresponding to different target subjects in this embodiment includes, but is not limited to, the search questions raised by users for the target subject and the subject's responses. For example, in a general hospital, the question "How should diabetic patients control their diet?" and the responses including dietary restrictions and recommended foods; in a children's hospital, the question "How to physically reduce fever in children?" and the responses including wiping with warm water and reducing clothing.

[0052] It should be further explained that one specific implementation method for word segmentation preprocessing of the collected text information in this embodiment is as follows:

[0053] Based on the collected question-and-answer text information corresponding to different target subjects, the special symbols in the text are scanned by regular expressions to perform replacement operations, removing non-Chinese and English characters while retaining numbers and medical unit symbols. At the same time, the boundaries of mixed Chinese and English text are marked according to character encoding features to obtain cleaned and standardized text.

[0054] Based on standardized text, the text is scanned sentence by sentence using a pre-set text segmentation model. The bidirectional maximum matching algorithm is used to match words in the built-in dictionary. For unmatched words, a hidden Markov model is used to calculate the probability of the character sequence to determine the segmentation boundary and obtain preliminary segmentation results.

[0055] Based on the preliminary word segmentation results, the BioBERT model is used to identify entities in the segmented sequence, label medical entity types and extract entity words. Then, the entity words are matched with the UMLS terminology database for string matching and semantic similarity calculation. Non-standard terms are replaced with corresponding standard terms to obtain a word segmentation sequence containing standardized entities.

[0056] Based on the word segmentation sequence containing standardized entities, the co-occurrence frequency of words is calculated using the TextRank algorithm to determine keywords. The co-occurrence network is used to analyze the association strength between different keywords, and a directed mapping table of keywords is generated to obtain the structured word segmentation preprocessing results.

[0057] It should be further explained that the purpose of this process in this embodiment is to transform the original text into standardized data that can be used for hierarchical feature extraction through standardization and structured extraction, providing a consistent lexical basis for the subsequent construction of the topic layer, sub-question layer and keyword layer, while ensuring the comparability of text features between different target subjects through entity standardization.

[0058] It should be further explained that the text segmentation model in this embodiment adopts the jieba segmentation framework that integrates medical domain dictionaries, and improves the accuracy of medical terminology recognition by dynamically adjusting the dictionary weights; the BioBERT model is used for entity recognition after being fine-tuned by medical corpus, and the UMLS terminology database provides a unified medical terminology standard across institutions; the TextRank algorithm counts the number of word co-occurrences through a sliding window, with the window size set to 5 words.

[0059] It should be further explained that the first text hierarchical sequence in this embodiment includes, but is not limited to, a topic layer, a sub-question layer, and a keyword layer. The topic layer is the core topic of the question and answer information, such as "diabetes diet management" and "child fever care". The sub-question layer is the specific sub-questions around the topic, such as "fruits that diabetic patients can eat" and "correct sites for physical cooling". The keyword layer is the key terms in the answers to the sub-questions, such as "low-sugar fruits", "glycemic index", "armpit", and "forehead". Some examples are given here, and the rest are specifically organized and preprocessed by those skilled in the art based on the collected data.

[0060] It should be further explained that, in this embodiment, the directed index graph is constructed from the sub-subjects and index paths corresponding to different retrieval information under the corresponding target subject;

[0061] It should be further explained that one method of obtaining the directed index graph in this embodiment includes:

[0062] Based on the question-and-answer index information of different target subjects, invalid indexes are filtered out using regular expressions, and sub-subject nodes and target keyword nodes (such as specific question-and-answer entries) are identified. Node attributes and edge attributes are extracted to complete the graph structure initialization. It should be further noted that the question-and-answer index information in this embodiment includes, but is not limited to, the target keyword ID, sub-subject affiliation, and access path records; sub-subject nodes include, but are not limited to, department, disease unit, doctor type, and function and skill level; node attributes include, but are not limited to, node type and historical access volume; edge attributes include, but are not limited to, the number of search jumps and path generation time; invalid indexes include, but are not limited to, empty paths and duplicate IDs.

[0063] Based on the initial graph structure, node IDs are converted into discrete features through one-hot encoding, and then mapped into continuous vectors by the embedding layer. Linear transformations are performed on the edge features, such as the normalized value of the number of jumps and the direction label, to obtain the initial feature vector of the node and the feature vector of the edge.

[0064] Based on the feature vectors of nodes and edges, a direction-aware aggregation layer of a directed graph neural network is constructed. Specifically, an inbound edge mask matrix and an outbound edge mask matrix are generated based on the adjacency matrix of the directed graph. In the inbound edge mask matrix, an element of 1 indicates that there is an edge from node j to node i, i.e., node j→ node i, and the rest of the elements are 0. In the outbound edge mask matrix, an element of 1 indicates that there is an edge from node i to node j, i.e., node i→ node j, and the rest of the elements are 0. The dimensions of the mask are matched with the dimensions of the node feature matrix by expanding the matrix dimensions.

[0065] For each node i, all neighboring nodes j pointing to i are selected through the incoming edge mask matrix. The number of jumps in the edge features is extracted to obtain the edge weight. Specifically, the number of jumps of a single edge is divided by the sum of the number of jumps of all incoming edges of node i. The feature vector of the neighboring node j is multiplied by the corresponding edge weight and then summed to obtain the incoming edge aggregate feature vector.

[0066] For each node i, all neighboring nodes j pointed to by node i are filtered out by the outgoing edge mask matrix, i.e., node i → node j. The cosine similarity of the feature vectors of node i and neighboring node j is calculated. The feature vector of neighboring node j is multiplied by the corresponding similarity and then summed to obtain the outgoing edge aggregated feature vector.

[0067] In this embodiment, the associations between nodes in the directed graph have a clear directionality. For example, A→B and B→A represent different semantics, where A and B represent different nodes. This embodiment uses a binary filtering mechanism to accurately isolate neighboring nodes of incoming and outgoing edges, avoiding interference from irrelevant node features in the aggregation process. The matrix dimension expansion is to adapt to the high-dimensional feature vectors of nodes, such as those containing semantic, attribute, and other multi-dimensional information, so that the 0 / 1 markings of the mask can act on each dimension feature. Element-level operations are used to achieve effective mapping of directional filtering in high-dimensional space. This process can strictly limit the neighbors participating in incoming edge aggregation through the incoming edge mask. The scope ensures that the calculation of incoming edge weights is based on real incoming edge jump data; the outgoing edge mask can accurately filter the neighbors pointed to by the current node, ensuring that the calculation of outgoing edge similarity focuses on real outgoing edge associations; the combination of the two thoroughly distinguishes the feature sources of incoming and outgoing edges, avoiding feature pollution caused by directional confusion, making the node aggregation features more consistent with the topological structure and semantic associations of the directed graph, providing accurate directional constraints for subsequent incoming edge weighting and outgoing edge similarity fusion, and ultimately improving the modeling accuracy of directional dependencies between nodes in the directed graph neural network, enhancing the directionality and specificity of feature aggregation; for example, in the directed graph of a hospital question-and-answer database... In the index graph, let the sub-sub-nodes be Internal Medicine (A1) and Surgery (A2), and the target keyword nodes be "diabetes medication consultation" (K1) and "fracture surgery procedure" (K2). In the inbound edge mask matrix, because there are retrieval paths A1→K1 (Internal Medicine points to diabetes consultation) and A2→K2 (Surgery points to fracture procedure), the matrix elements (K1, A1) = 1 and (K2, A2) = 1, with the remaining positions being 0. The inbound neighbors of K1 are only A1, and the inbound neighbors of K2 are only A2. In the outbound edge mask matrix, the elements (A1, K1) = 1 and (A2, K2) = 1. The rest are 0, used to limit the outgoing neighbors of A1 to only K1, and the outgoing neighbors of A2 to only K2; by expanding the mask matrix from the number of nodes × the number of nodes to a 128-dimensional dimension matching node features, each feature dimension can be filtered by the directional mask, ensuring that when K1 aggregates incoming edge features, it only includes the department attribute features of A1, such as "chronic disease diagnosis and treatment", and when A1 aggregates outgoing edge features, it only associates the keyword features of K1, such as "insulin use", avoiding the mixing of features from surgery or other irrelevant nodes, and achieving accurate feature aggregation for targeted retrieval paths such as "internal medicine → diabetes consultation" and "surgery → fracture process".

[0068] The initial feature vector of node i is concatenated with the feature vector of the incoming or outgoing edge, and then input into a two-layer perceptron. After Sigmoid activation, the incoming edge association score and outgoing edge association score are obtained, which are used to measure the association strength between the node features and the two types of edge features.

[0069] Multiply the aggregated feature vector of incoming edges by the association score of incoming edges to obtain the weighted incoming edge features; multiply the aggregated feature vector of outgoing edges by the association score of outgoing edges to obtain the weighted outgoing edge features; and sum the two types of weighted features element-wise to obtain the fused feature vector.

[0070] The fused feature vector is input into the ReLU activation function to suppress negative features, retain effective feature information, and output the current layer node feature vector with dimension D, thus completing the single-layer orientation-aware feature aggregation.

[0071] Based on the feature vector of the current layer node, three layers of direction-aware aggregation are repeatedly executed. Each layer retains the independent aggregation path of the incoming or outgoing edge. Finally, the multi-layer features are concatenated into 64-dimensional global node features through a fully connected layer.

[0072] Based on global node features, high-frequency retrieval paths are extracted using the PageRank path mining algorithm. Bidirectional LSTM is used to encode the path sequence, and the directional features and access frequency features of the path are fused to generate a path feature vector. The path sequence, for example, is: sub-subject → retrieval target keyword → sub-subject.

[0073] Global node features and path feature vectors are fused through residual connections, and a directed indexed graph containing node hierarchy relationships, directed edge weights, and path semantic vectors is output using the NetworkX graph structure serialization tool.

[0074] For example, taking a general hospital question-and-answer system as an example, the sub-sub-nodes are "Internal Medicine" and "Endocrinology", and the target keyword nodes for retrieval are "diabetic diet" and "insulin use". The index information shows that there are multiple jumps between "Internal Medicine → Diabetic Diet" and "Endocrinology → Insulin Use", and there are also reverse jumps between "Diabetic Diet → Endocrinology". When processing with a directed graph neural network, "Internal Medicine" and "Endocrinology" are first encoded one-hot and then converted into high-dimensional vectors through an embedding layer. In the edge features, the weight of "Internal Medicine → Diabetic Diet" is calculated according to the jump frequency, and the direction is marked as 1. The weight of "Diabetic Diet → Endocrinology" is calculated according to the jump frequency, and the direction is marked as -1.

[0075] When the direction-aware aggregation layer calculates the features of the "Endocrinology" node, the features of the inbound neighbor "diabetic diet" are aggregated through corresponding weights, and the features of the outbound neighbor "insulin use" are aggregated through high semantic similarity. After multiple aggregations, a low-dimensional global feature is generated.

[0076] The high-frequency path "Internal Medicine → Diabetic Diet → Endocrinology" is encoded by LSTM and then fused with node features to generate a directed index graph containing directed edges and path vectors such as "Weight corresponding to Internal Medicine → Diabetic Diet" and "Weight corresponding to Endocrinology → Insulin Use".

[0077] The feature extraction module, based on the first text hierarchical sequence and directed index graph corresponding to different target subjects, combined with a preset feature distillation extraction model, obtains the first text embedding representation and the first index embedding representation corresponding to different target subjects; based on the first text embedding representation and the first index embedding representation corresponding to different target subjects, combined with a preset three-dimensional overlap model, it obtains the similarity overlap of different target subjects; the similarity overlap of target subjects includes content overlap, index structure overlap, semantic representation overlap, and comprehensive overlap.

[0078] It should be further explained that the process of obtaining the first text embedding representation corresponding to the target subject in this embodiment includes:

[0079] Based on the question-and-answer text information corresponding to different target subjects, special symbols are filtered and mixed Chinese and English boundary annotations are performed by combining regular expressions with Chinese and English mixed boundary rules to obtain the preprocessed question-and-answer text information corresponding to different target subjects.

[0080] It should be further explained that, based on the question-and-answer text information of different target subjects, a multi-level regular expression rule set is constructed in conjunction with Chinese and English mixed boundary recognition logic to perform special symbol filtering and boundary annotation, specifically as follows:

[0081] Based on a preset general special character regular expression pattern, punctuation marks, special characters and meaningless placeholders in the text are matched and removed through replacement operations to obtain preliminary cleaned text;

[0082] Furthermore, for symbols unique to the medical field, a professional symbol regularity pattern is constructed to match drug dosage symbols, test indicator unit symbols, and formula symbols. Through conditional judgment, professional symbols with semantic value are retained, and redundant symbols are removed to obtain domain-cleaned text.

[0083] Then, the Chinese-English boundary recognition process was initiated. A bilingual corpus was built based on a Chinese word segmentation dictionary and an English word list. Continuous sequences of Chinese words and English words were identified through regular expression matching, and boundary markers were inserted at the boundary positions.

[0084] For sequences containing a mixture of numbers and text, regular expressions are used to match combinations of numbers with Chinese units and combinations of numbers with English units, inserting markers at the boundaries between numbers and text.

[0085] For nested Chinese and English structures, patterns of English abbreviations followed by Chinese explanations and patterns of Chinese terms containing English roots are identified through recursive matching. Markers are inserted at the nesting boundaries to finally obtain preprocessed text with boundary annotations.

[0086] It should be further explained that this embodiment reduces noise interference by systematically filtering irrelevant symbols, and provides a clear text structure for subsequent three-level hierarchical and entity segmentation by accurately marking the boundaries between Chinese and English, numbers and text, thus ensuring the accurate extraction of key information such as medical terminology.

[0087] Based on the preprocessed question-and-answer text information corresponding to different target subjects and combined with the three-level hierarchical rules, we obtain the hierarchical question-and-answer text information of different target subjects;

[0088] It should be further explained that the process of obtaining the first index embedding representation corresponding to the target subject in this embodiment includes:

[0089] Based on the feature data of each sub-main node, the keyword node of the search target requirement, and the directed edges in the directed index graph, the GraphSAGE algorithm is used to perform average pooling aggregation on the features of the incoming and outgoing neighbors of each node. The features of the incoming edges are weighted by the edge weights, and the features of the outgoing edges are weighted by the semantic similarity of the nodes. After three layers of aggregation, the local path features of the nodes are obtained.

[0090] Based on local path features, a convolutional attention network is used to perform convolution operations on the direction markers and access frequency sequences of different retrieval paths under the same subject to extract direction-sensitive features and frequency trend features. At the same time, comparative convolution is performed on the path direction and frequency features of different subjects facing the same retrieval target to obtain cross-subject path difference features.

[0091] After concatenating local path features, direction-sensitive features, frequency trend features, and cross-subject path difference features, a feature alignment algorithm is used to map the features to the retrieval target requirement tags in the feature space, so that the index features of different subjects are in a unified dimensional space, and the aligned multi-dimensional path features are obtained.

[0092] Based on the aligned multi-dimensional path features, the index distillation branch of the input feature distillation extraction model is used to calculate the contribution score of each feature to the retrieval target requirement. Core features with scores higher than the threshold are retained and compressed to a fixed dimension through linear transformation to obtain the first index embedding representation.

[0093] It should be further explained that the purpose of the first index embedding representation in this embodiment is to provide structured feature support for the calculation of index structure overlap in the three-dimensional overlap model, to provide quantified vector representation for index path comparison, conflict detection and cross-subject index alignment between different target subjects, and at the same time, as the core output of the index branch in the feature distillation model, it supports the accurate matching and index path optimization of subsequent retrieval needs.

[0094] It should be further explained that in this embodiment, the GraphSAGE algorithm is used to aggregate the features of the node's neighbors to capture local structural information; the convolutional attention network strengthens the feature representation of path direction and frequency through convolution operations, and the attention mechanism highlights the features of key paths; the feature alignment algorithm ensures that the features of different subjects are comparable in the same space; the contribution score calculation is used to filter the features most critical to the retrieval target and improve the effectiveness of the representation.

[0095] It should be further explained that the three-level hierarchical rule in this embodiment includes a topic layer, a sub-question layer, and a keyword layer; it should be further explained that in the topic layer of this embodiment, based on the question and answer text information corresponding to each target subject, the co-occurrence frequency of keywords extracted in the search text and the response text is calculated by the TextRank algorithm, core words are screened, and the search target keyword vector is generated by combining the screened core word sequence with BioBERT to obtain the topic name corresponding to the question and answer information of the current target subject and assign the root node ID associated with the text storage address;

[0096] It should be further explained that the specific process of implementing the theme layer in this embodiment includes:

[0097] Based on question-and-answer text information from different target subjects, a dynamic word graph model is constructed using the TextRank algorithm. Words after word segmentation are used as nodes, and the word spacing threshold is determined statistically through the distribution of sentence lengths in the domain text. For example, in this embodiment, 1 / 3 of the average sentence length is used. Undirected edges are established between words whose word spacing is ≤ the word spacing threshold. An adaptive iterative strategy is used to calculate node weights. Initial iterations are performed for 10 rounds, and iteration stops when the weight change rate is ≤ 1% for three consecutive rounds. The top 20 words by weight are extracted as initial keywords, and nouns, verbs, and technical terms are retained through part-of-speech filtering, while function words and stop words are filtered out.

[0098] Based on the initial keywords, a co-occurrence matrix is ​​constructed to count the number of times each keyword pair co-occurs in the search and response texts. Combined with mutual information values, low semantic relevance co-occurrence pairs are filtered out. Keywords are sorted in descending order by the weighted sum of co-occurrence frequency and mutual information value, and keywords corresponding to a weighted sum ≥ 10% of the total number of texts are selected to form a core word sequence. At the same time, domain-irrelevant words are removed through domain dictionary verification. For example, the medical domain matches the MeSH vocabulary. It should be further noted that the mutual information value is the ratio of the joint probability to the marginal probability of the keyword pair.

[0099] Each word in the core word sequence is input into the pre-trained BioBERT model. The hidden states from the third-to-last layer to the last layer of the BioBERT model are fused together. The weights of each layer are assigned through the attention mechanism to generate a comprehensive word vector.

[0100] Based on comprehensive word vectors, dimensionality reduction is performed by improving the principal component analysis algorithm. The dimensionality reduction dimension is dynamically determined by the variance explanation rate threshold, which retains the core semantic features while reducing redundant information.

[0101] It should be further explained that the implementation process of dimensionality reduction using the improved principal component analysis algorithm in this embodiment includes:

[0102] For the set of word vectors generated by the BioBERT model from the core word sequence, calculate the covariance matrix between the vectors, and extract the eigenvalues ​​of the covariance matrix and the corresponding eigenvectors through eigenvalue decomposition.

[0103] The eigenvalues ​​are sorted in descending order, and the corresponding eigenvectors are sorted synchronously to obtain an ordered principal component sequence.

[0104] Calculate the cumulative variance explained rate of the ordered principal component sequence, starting from the first principal component and accumulating until the cumulative variance explained rate reaches the preset variance explained rate threshold. Record the number of principal components at this point as the target dimensionality reduction dimension.

[0105] Based on the target dimensionality reduction dimension, a projection matrix is ​​constructed by selecting a corresponding number of preceding feature vectors. The original word vectors are then multiplied with the projection matrix to obtain the dimensionality-reduced word vectors.

[0106] It should be further noted that the preset variance explanation rate threshold in this embodiment is determined through domain semantic retention verification to ensure that the word vectors after dimensionality reduction retain core semantic features and remove redundant dimensions.

[0107] It should be further explained that the process of obtaining the domain semantic retention in this embodiment includes: extracting sample word vectors from a domain standard corpus, reducing the dimensionality by a variance explanation rate threshold, calculating a semantic retention metric by combining the mean cosine similarity, evaluating the entity matching accuracy through a domain entity linking task, constructing a comprehensive score that integrates the two, and finally selecting the minimum variance explanation rate threshold that meets the comprehensive score. This achieves the quantitative determination of the domain semantic retention, provides a domain adaptability basis for the retention of core semantic features during the dimensionality reduction process, and ensures that the dimensionality-reduced word vectors are adapted to the specific domain knowledge structure.

[0108] Based on the dimensionality-reduced word vectors, an improved K-means clustering algorithm is adopted, wherein the number of clusters is automatically determined by the silhouette coefficient maximization criterion, specifically: the value range is 1 / 6 to 1 / 4 of the length of the core word sequence;

[0109] Vector similarity is calculated using a hybrid metric method that primarily uses Euclidean distance and secondarily uses cosine similarity. The K-means++ algorithm is used to initialize cluster centers, and the iteration stops when the distance the cluster centers have moved is less than or equal to a preset threshold.

[0110] Words with similarity greater than or equal to a preset threshold are grouped into the same cluster. The three words with the highest weight in each cluster and covering 80% of the semantics within the cluster are selected as the cluster core words. The consistency within the cluster is verified twice by cosine similarity of word vectors.

[0111] Based on cluster core words, a bidirectional LSTM model is used to learn word order dependencies and generate candidate topic names, for example, "diabetes blood sugar control". Semantic alignment correction is performed by combining a topic thesaurus annotated by domain experts. A unique root node ID is assigned to each topic, which is associated with the distributed storage address of the corresponding question and answer text. The mapping relationship between the root node and the text is recorded in the meta-database to achieve efficient association between the topic layer and the underlying text. It should be further noted that the distributed storage address in this embodiment includes, but is not limited to, the primary replica and backup replica paths.

[0112] In this embodiment, the domain adaptability and semantic accuracy of topic extraction are improved by using dynamic word graph construction, multi-dimensional core word selection, word vector generation that integrates multi-layer features, and adaptive clustering strategies. The generated topic names are both representative and structured, laying a precise top-level framework for subsequent hierarchical knowledge organization.

[0113] Based on the core words of the cluster, the topic candidate names are generated through semantic concatenation. Ambiguous names are then corrected by manual verification, and the topic name is finally determined. For example, "diabetes", "blood sugar" and "control" are concatenated into "diabetes blood sugar control".

[0114] Based on the unique identifier of the topic name in the target subject knowledge base, a root node ID is assigned and associated with the storage path of the corresponding question and answer text, such as server directory + text ID, to complete the construction of the topic layer.

[0115] For example, taking the question "How can diabetic patients control their blood sugar through diet? The answers include staple food choices, sugar intake restrictions, etc." as an example, TextRank extracts keywords such as "diabetes", "diet", "blood sugar" and "control". The co-occurrence frequency statistics show that "diabetes-blood sugar" and "diet-control" co-occurred the most. After generating vectors using BioBERT, they are clustered into one cluster, and the final topic name is determined as "diabetes diet and blood sugar control".

[0116] It should be further explained that in this embodiment, the sub-question layer generates sub-question sentence vectors based on the search text under each topic and Sentence-BERT. After clustering by the DBSCAN algorithm, the sub-question keyword sequence and sub-question answer priority are obtained. Sub-node IDs are assigned based on the sub-question answer priority corresponding to the sub-question keyword sequence, and the corresponding topic root node is associated with them.

[0117] It should be further explained that the specific implementation process of the sub-problem layer in this embodiment includes:

[0118] Based on the set of search texts under each topic, the search texts are mapped to sentence vectors through the Sentence-BERT model, and then the sentence vector matrix is ​​formed after L2 normalization.

[0119] Clustering is performed using the DBSCAN algorithm based on the sentence vector matrix. The optimal neighborhood radius is verified by the Silhouette Score and the minimum number of samples is determined by combining neighborhood data. The neighborhood range of each point is calculated, and points with density connection are divided into the same cluster. Core points, boundary points and noise points are identified to form the initial sub-problem cluster.

[0120] For the search text within each cluster, word frequency statistics are performed, and keywords with high TF-IDF values ​​are extracted to form a candidate keyword sequence. Nouns, verbs, and adjectives are retained through part-of-speech filtering, and medical entity terms are selected as the final keyword sequence in combination with domain dictionary matching. The average sentence vector of each sub-question cluster is calculated as the cluster center, and the correlation strength between each cluster center and the topic vector is measured by cosine similarity. Sub-question clusters that reach the preset correlation strength are included in the priority ranking process.

[0121] A comprehensive scoring function is constructed to evaluate the priority of sub-questions. This function comprehensively considers the retrieval frequency (i.e., the number of recent retrievals), the timeliness factor (i.e., the weight of the latest response time), and the professionalism score (i.e., the confidence level of the doctor's annotation). Among them, the retrieval frequency accounts for the main weight, while the timeliness factor and the professionalism score account for the secondary and auxiliary weights, respectively. The sub-questions are ranked in descending order of scores to determine their priority.

[0122] Assigning continuously increasing child node IDs based on the priority of sub-problems, and establishing a mapping relationship between each child node ID and the corresponding topic root node ID to complete the construction of the sub-problem layer.

[0123] It should be further explained that this embodiment organizes the search text under the topic into a set of structured sub-questions through semantic clustering, and the extracted keyword sequence provides a focus direction for the subsequent keyword layer. Priority sorting ensures that high-frequency and important sub-questions are processed first, thereby improving the practicality and retrieval efficiency of the knowledge base.

[0124] For example, taking the topic of "diabetes" as an example, DBSCAN clustering identified sub-problem clusters such as "diet control", "exercise management" and "drug therapy", with keyword sequences of {diet, carbohydrates, blood sugar}, {exercise, intensity, frequency}, and {insulin, oral medication, side effects}, respectively. The priority ranking shows that "drug therapy" received the highest priority due to its high search frequency and the inclusion of content with high professional scores.

[0125] It should be further explained that in this embodiment, the keyword layer obtains the TF-IDF value of the keyword corresponding to each answer text through a statistical algorithm based on the answer text corresponding to each sub-question. After comparison and verification with the UMLS terminology database, the keywords are sorted in descending order according to their TF-IDF values, and keyword IDs are assigned according to the sorting results and bound to the corresponding sub-question nodes.

[0126] It should be further explained that the specific implementation process of the keyword layer in this embodiment includes:

[0127] Based on the set of response texts for each subquestion, the text is decomposed into a word sequence through word segmentation preprocessing, stop words are filtered out and lemmatization is performed; based on the preprocessed word sequence, a word frequency statistics dictionary within the subquestion is constructed, and the number of times each word appears in the current response text is counted to obtain the word frequency (TF) value; based on the set of response texts for all subquestions, the number of texts containing each word is counted, and the inverse document frequency (IDF) value is calculated, where IDF = log(total number of texts / (number of texts containing the word + 1));

[0128] Multiply the TF value and IDF value of each word to obtain the TF-IDF value, forming a preliminary keyword list;

[0129] Based on the initial keyword list, each keyword is semantically matched with medical terms in the UMLS terminology database. Cosine similarity is used to calculate the similarity between the keyword and each term in the terminology database. A similarity threshold is set, and keywords that reach the threshold are marked as medical terms.

[0130] For keywords that do not match UMLS terms, try matching again through stemming and word transformation. Keywords that still do not match are retained but marked as non-medical terms.

[0131] Sort all keywords by TF-IDF values ​​in descending order, prioritizing medical terms. If two or more keywords have the same TF-IDF value, medical terms will be ranked before non-medical terms.

[0132] Based on the sorting results, assign a unique, incremental keyword ID to each keyword, and establish a mapping relationship between the keyword ID and the corresponding sub-problem node ID to complete the keyword layer construction.

[0133] It should be further explained that this embodiment quantifies the importance of keywords in the response text by using TF-IDF values, combines UMLS terminology database verification to ensure the accuracy of medical terminology, constructs a hierarchical keyword system, provides accurate terminology indexes for subsequent knowledge retrieval, and improves the accuracy and efficiency of semantic matching.

[0134] For example, taking the sub-problem "drug treatment of diabetes" as an example, the response text is processed to extract the keywords "insulin", "metformin", "side effects" and "blood sugar control". Among them, "insulin" and "metformin" are completely matched with the UMLS terminology database, "blood sugar control" is matched after word form transformation, and "side effects" is not matched but is retained. After sorting by TF-IDF value, the keyword IDs are assigned in sequence and bound to the sub-problem node.

[0135] This application achieves deep structured processing of medical question-and-answer data by constructing a multi-level knowledge representation system. In the data preprocessing stage, a word segmentation algorithm that integrates a domain dictionary is used in conjunction with BioBERT entity recognition, which effectively solves the problems of the diversity and non-standard expression of professional terminology in medical texts, ensuring the accuracy and consistency of terminology expression. Through a progressive feature extraction using a three-tiered hierarchical architecture—topic layer, sub-question layer, and keyword layer—a hierarchical knowledge organization from macro-topics to micro-terms is achieved. The topic layer employs the TextRank algorithm combined with improved clustering methods to extract core topics, ensuring a clear top-level structure for the knowledge system. The sub-question layer forms a structured sub-question network based on density clustering and multiple priority evaluation. The keyword layer constructs a precise terminology index system through TF-IDF statistics and terminology standardization. A direction-aware aggregation mechanism using a directed graph neural network accurately captures path dependencies and semantic associations in medical knowledge retrieval by distinguishing between inbound edge weight aggregation and outbound edge semantic aggregation. The feature distillation model achieves efficient knowledge compression and transfer through a teacher-student architecture, preserving deep semantic features while significantly improving computational efficiency. This process, through the organic combination of standardization, hierarchical modeling, graph neural network representation, and feature distillation, constructs a medical knowledge management system with both semantic depth and structural rationality. It provides a high-quality knowledge organization framework for clinical decision support systems, effectively solving problems such as terminology ambiguity, structural chaos, and low retrieval efficiency in traditional medical knowledge bases, and significantly improving the accuracy and usability of knowledge services.

[0136] Based on the question-and-answer text information after being layered according to different target subjects, combined with the preset entity relationship extraction algorithm, entity word segmentation is performed on the text corresponding to each layer under different target subjects, and the directed mapping relationship between the corresponding layer and different layers is extracted to obtain the first text embedding representation corresponding to different target subjects.

[0137] In this embodiment, it should be further explained that the construction process of the feature distillation extraction model includes:

[0138] Based on the first text hierarchical sequence corresponding to different target subjects, the teacher sub-model in the feature distillation extraction model performs hierarchical text feature extraction along the corresponding nodes of the topic layer, sub-question layer, and keyword layer, specifically as follows:

[0139] Based on the first text extraction network, the root node text of the topic layer is used to extract topic feature vectors to obtain the first topic vector. It should be further explained that, based on the root node text of the topic layer, the word sequence is cleaned through preprocessing, word embeddings are generated through the BioBERT embedding layer, the Transformer encoder is used to capture contextual relationships, and after global pooling and feature enhancement layer transformation, and then L2 normalization, the first topic vector is obtained, realizing the accurate representation of the core semantics of the topic. Its purpose is to provide a benchmark vector for the calculation of the association weight between the sub-question layer and the topic, ensuring that the sub-question feature extraction revolves around the core of the topic, and enhancing the semantic consistency and topic focus of the hierarchical features.

[0140] Based on the second text extraction network, features are extracted from each priority sub-question in the sub-question layer to generate a hierarchical sub-question vector. Then, based on the hierarchical sub-question vector and the first topic vector, a cross-attention network is used to obtain the topic-oriented priority association weight vector of the sub-question.

[0141] It should be further explained that this embodiment, based on the text of each priority sub-question in the sub-question layer, forms a purified sequence through word segmentation and filtering, which is then embedded by Sentence-BERT and encoded and pooled by a Transformer encoder to obtain a hierarchical sub-question vector. This vector is then input into a cross-attention network with the first topic vector to calculate the association weights. After weighted fusion and linear transformation, a priority association weight vector of sub-questions oriented towards the topic is obtained, which accurately quantifies the semantic association strength between sub-questions and the topic. In the application, by quantifying the association weights between sub-questions and the topic, priority guidance is provided for subsequent keyword layer feature extraction, ensuring that keyword selection focuses on sub-questions that are strongly related to the topic, and improving the topic consistency and information focus of the hierarchical features.

[0142] Based on the third text extraction network, feature extraction is performed on the keyword layer corresponding to each sub-question according to the ranking result to obtain the keyword feature vector of the answer text corresponding to each sub-question. Based on the keyword information of the answer text corresponding to each sub-question, combined with the hierarchical sub-question vector and the cross-attention network, the association weight vector of the answer text keywords oriented towards the sub-question is obtained.

[0143] It should be further explained that, based on the keyword layer ranking results corresponding to each sub-question, this embodiment generates word vectors for keyword nodes through the embedding layer of the third text extraction network. The pre-trained BioBERT model is used to obtain the initial word embeddings. For out-of-vocabulary words, character-level bidirectional LSTM is used to generate sub-word embeddings and concatenate them. After linear transformation, they are mapped to a unified dimension. The mapped keyword vectors are residually connected with the sub-question layer vectors, that is, the keyword vectors are added to the projection of the sub-question layer vectors. Through three Transformer encoding layers, each layer contains a multi-head self-attention mechanism and a feedforward neural network, feature extraction is performed. The feature vectors output by each layer are concatenated according to the dimension to obtain a text feature matrix. This matrix contains low-level semantic features, mid-level structural features, and high-level association features. The text feature matrix is ​​used as the query, and the layered sub-question vectors are used as the key and value, which are input into the cross-attention network. By calculating the attention score and normalizing it with softmax, the association weight vector of the answer text keywords to the sub-question is obtained. This vector represents the semantic association strength between the keywords and the sub-question.

[0144] Based on the first topic vector, hierarchical sub-question vectors, sub-question-oriented topic priority association weight vectors, keyword feature vectors of the answer text corresponding to each sub-question, and keyword association weight vectors of the answer text oriented to the sub-question, a first text embedding representation of the question and answer text corresponding to the target subject is constructed through a graph neural network.

[0145] It should be further explained that the feature extraction process of the convolutional attention network in this embodiment includes:

[0146] Based on the first retrieval path features, node identifiers, edge weights, and access frequencies are extracted from the path sequence. The node identifiers are converted into vectors through one-hot encoding and concatenated with the edge weights and frequency normalization values ​​to form a path feature sequence.

[0147] The path feature sequence is input into the convolutional layer of the third text extraction network. Multiple convolutional kernels are used to perform convolution operations. The associated features of the path segments are extracted through a sliding window to calculate local feature maps and output multi-scale local feature maps.

[0148] The channel attention mechanism is applied to the multi-scale local feature map to calculate the importance weight of each channel feature. Specifically, the weight is the output of the global average pooling result of the feature map after passing through a fully connected layer and sigmoid activation. The channel features are then weighted and summed according to their weights to obtain the channel-enhanced features.

[0149] The channel enhancement features are input into the spatial attention module to calculate the spatial association weights of different retrieval paths under the same subject. Specifically, the spatial attention weights are obtained by normalizing through softmax based on the product of path direction consistency and access frequency, and the spatial features are weighted and fused.

[0150] Based on spatial fusion features, feature subsets that distinguish between different retrieval targets of the same subject and the same retrieval target of different subjects are distinguished. For the former, path direction entropy (measuring direction diversity) is calculated, and for the latter, path overlap (the proportion of intersection paths) is calculated. The entropy value and overlap are used as auxiliary features and concatenated to the spatial fusion features.

[0151] The concatenated features are compressed in dimension by global max pooling, then input into a fully connected layer for feature aggregation, and output a feature vector of fixed dimension.

[0152] The output vector is L2 normalized to obtain the first index embedding representation corresponding to the target entity.

[0153] Based on the first text embedding representation and the first index embedding representation, the feature space is aligned by combining the feature alignment algorithm with the retrieval target requirement as the label to obtain the aligned first text embedding representation and the first index embedding representation.

[0154] The aligned first text embedding representation and the first index embedding representation are respectively input into the text distillation branch and the index distillation branch of the student sub-model in the feature distillation extraction model. Feature distillation is performed according to the contribution of the corresponding features to the retrieval target requirements and the contribution threshold to obtain the first text embedding representation and the first index embedding representation after feature distillation.

[0155] It should be further explained that the process of implementing the text distillation branch and the index distillation branch in this embodiment includes:

[0156] Based on the aligned first text embedding representation and first index embedding representation, the text distillation branch and index distillation branch of the student sub-model in the feature distillation extraction model are respectively input. The text distillation branch calculates the attention contribution of features and the semantic vector of the retrieval target, filters the core dimensions according to the cumulative contribution threshold, and performs dimensionality reduction through principal component analysis to obtain the first text embedding representation after feature distillation. The index distillation branch calculates the comprehensive contribution by weighting the matching accuracy and access frequency, filters the key dimensions according to the quantile threshold, and performs fusion through a multilayer perceptron to obtain the first index embedding representation after feature distillation. This achieves the filtering of redundant features with low contribution to the retrieval target demand and retains the core features.

[0157] Based on the hierarchical text features and retrieval path features extracted from the topic layer, sub-question layer, and keyword layer by the teacher model, the soft label distillation loss is obtained by calculating the KL divergence between the intermediate features and the student model and amplifying the distribution differences.

[0158] Based on the entity boundaries and retrieval path weights labeled by the teacher model, the hard label distillation loss is obtained by calculating the intersection-union ratio (when the text branch does not meet the standard) and mean square error (when the index branch does not meet the standard) between the student model prediction results and the labels.

[0159] Based on the first text embedding representation of similar texts from different subjects and the first index embedding representation of the same retrieval requirement, the domain alignment loss is obtained by calculating the cosine similarity and forcing it to meet the corresponding requirements.

[0160] Based on soft label distillation loss, hard label distillation loss, and domain alignment loss, a comprehensive loss function is obtained by weighted summation and integration of the contributions of each loss component. This enables the joint optimization of feature transfer and cross-subject feature alignment during the knowledge distillation process.

[0161] The feature distillation extraction model is obtained by training based on the constructed comprehensive loss function combined with the preset training period and loss function threshold.

[0162] This application achieves multi-level accurate representation and efficient compression of medical knowledge through a feature distillation extraction model. The teacher sub-model employs a hierarchical feature extraction architecture. At the topic level, BioBERT embedding and Transformer encoding capture global semantics, generating a first topic vector with strong topic representation capabilities. At the sub-question level, Sentence-BERT and cross-attention mechanisms are combined to establish semantic association weights between sub-questions and topics. At the keyword level, residual connections and multi-layer Transformer encoding are used to construct a fine-grained term association network. This hierarchical processing ensures that knowledge representation maintains topic consistency while reflecting fine-grained semantic differences. The convolutional attention network extracts local path features through multi-scale convolutional kernels, combining channel attention and spatial attention mechanisms to effectively capture the directional sensitivity and access pattern features of the retrieval path. The feature alignment algorithm maps text and index features to a unified space, enabling the student model to simultaneously learn semantic associations and retrieval path features. The knowledge distillation process is jointly optimized through soft and hard labels, achieving feature compression while preserving the teacher model's knowledge. KL divergence loss ensures deep feature transfer, intersection-union ratio and mean squared error loss strengthen entity boundary learning, and domain alignment loss improves cross-subject feature consistency. The resulting feature distillation extraction model generates highly discriminative text and index embedding representations, providing accurate feature support for subsequent conflict detection. Simultaneously, feature compression significantly reduces computational overhead, enabling the system to handle large-scale medical knowledge bases. This technical solution, through the organic combination of hierarchical representation, attention mechanisms, and distillation learning, achieves collaborative modeling of medical knowledge from global topics to local terms, laying a feature foundation for intelligent knowledge base management.

[0163] The conflict simulation module constructs an initial question-and-answer knowledge base based on the question-and-answer text information and question-and-answer index information of different target subjects combined with a tree-structured database. Based on the initial question-and-answer knowledge base and the similarity and overlap of different target subjects, it uses a Bayesian function simulation question-and-answer algorithm to detect and identify question-and-answer index conflicts between different sub-subjects within the target subject and between target subjects with different overlap, thereby obtaining a list of target subject conflict risks.

[0164] In this embodiment, it should be further explained that the process of obtaining the target subject conflict risk list includes:

[0165] Based on the question-and-answer text information and question-and-answer index information of different target subjects, an initial question-and-answer knowledge base is constructed through a tree-structured database. The initial question-and-answer knowledge base is associated with the nodes of the topic layer, sub-question layer, and keyword layer, as well as the sub-subjects, retrieval units, and path information of the directed index graph, forming a hierarchical index structure.

[0166] Extract the index features of different sub-subjects within the target subject and the similarity overlap data of different target subjects from the initial question-answering knowledge base, and use them as input parameters for conflict detection. The index features include a set of index keywords and a directed path sequence.

[0167] Based on the conflict detection between different sub-sub ...

[0168] It should be further explained that, in this embodiment, based on the conflict detection between target subjects under different degrees of overlap, high-overlap subject pairs are screened based on comprehensive overlap. The similarity of the first text embedding representation and the first index embedding representation of the high-overlap subject pairs is calculated to identify feature pairs that are semantically related but have different indices. The comprehensive overlap is used as a weighting factor to adjust the prior probability of the Bayesian function. The cross-subject retrieval scenario is simulated by a question-answering simulation algorithm and the cross-subject conflict probability is calculated by combining the likelihood function. The index conflict risk entries between subjects are generated by arranging them in descending order of conflict probability.

[0169] Based on the index features and semantic representation of highly overlapping subject pairs, the simulated question answering algorithm extracts the topic layer text, sub-question layer text, and keyword layer terms of the two subjects, and combines them to generate diverse retrieval query samples. It simulates cross-subject retrieval paths. For example, starting from the topic root node of subject one, it visits the corresponding node according to the index path, records the node matching results and path jump differences with subject two, and collects conflict events in the retrieval process. For example, when there is divergence in the index path or ambiguity in the keyword mapping, it is marked as a potential conflict event, and the conflict characteristics are recorded to obtain simulated conflict data of cross-subject retrieval scenarios.

[0170] Based on simulated conflict data and historical cross-entity conflict data, a cross-entity conflict feature vector is formed by concatenating conflict features and comprehensive overlap features. Reference samples are extracted from similar conflict cases in the historical database. The feature similarity between the current feature vector and the reference samples is calculated and the most similar subset is selected. The likelihood value is obtained by fitting the probability distribution through kernel density estimation. The comprehensive overlap is used as a weighting factor to adjust the prior probability of the Bayesian function. The cross-entity conflict probability is obtained by combining the likelihood value with the Bayesian formula.

[0171] This process is based on the features of highly overlapping subject pairs. The simulated question-answering algorithm simulates cross-subject retrieval scenarios to collect conflict events, constructs conflict feature vectors, calculates feature similarity by combining historical data and fits a likelihood function, adjusts the prior probability by comprehensively considering the overlap, and obtains the cross-subject conflict probability by calculating the Bayesian formula. This enables a quantitative assessment of cross-subject index conflicts, provides a basis for generating inter-subject conflict risk entries, and improves the compatibility of cross-subject index structure and retrieval accuracy.

[0172] Based on the intra-subject index conflict risk entries and inter-subject index conflict risk entries, supplementary conflict feature descriptions and conflict impact scope are generated to create a target subject conflict risk list.

[0173] The target subject conflict risk list includes the index conflict risk and probability of index conflict for different sub-subjects within the target subject, as well as the index conflict risk and probability of index conflict between target subjects under different degrees of overlap;

[0174] The conflict resolution module, based on the target subject's conflict risk list and combined with a conflict resolution adversarial model with preset conflict resolution rules, obtains index conflict risk resolution strategies for different sub-subjects within the target subject and index conflict risk resolution strategies for target subjects with different degrees of overlap. The index conflict risk resolution strategies include correcting unified standard index keywords and directed index paths across sub-subjects or across subjects.

[0175] It should be further explained that the construction process of the conflict resolution adversarial model in this embodiment includes:

[0176] Based on the conflict types in the target subject conflict risk list, three types of preset conflict resolution rules are constructed: Keyword conflict resolution rules use authoritative domain dictionaries as a benchmark, calculate the frequency of ambiguous keywords in the search text and their cosine similarity with the topic vector, and take the terms with the highest frequency or higher similarity as standard keywords; Path conflict resolution rules should delete the edge with the lowest access frequency in the closed loop when there is a path loop conflict, calculate the product of the path length and access frequency for cross-subject path divergence and take the path with the largest value as the benchmark path, and take the weighted average of the edge weights of different sub-subjects (the weight is the proportion of the sub-subject's search volume) as the unified weight for weight contradiction conflicts; Priority adaptation rules require higher cross-subject compatibility for subject pairs with high comprehensive overlap, and prioritize the retention of high-priority sub-problem index features for conflicts within sub-subjects, thus obtaining preset conflict resolution rules; In this embodiment, in addition to the above-mentioned conflict resolution rules, when new conflicts and rules occur, those skilled in the art use text feature encoding algorithms to update the obtained conflict features and corresponding resolution strategies into the conflict resolution adversarial model;

[0177] A conflict resolution adversarial model is constructed based on preset rules. The generator adopts a Transformer encoder-decoder structure, which takes conflict features and preset rule features as input and captures the correlation through a self-attention mechanism. The decoder outputs candidate resolution strategies. The discriminator adopts a multilayer perceptron structure, which takes candidate strategies, initial conflict features and user retrieval demand vector as input, calculates conflict reduction rate, rule compliance, demand matching degree, and outputs the probability of strategy effectiveness.

[0178] By defining the generator loss (-log discriminator effective probability plus rule constraint loss) and the discriminator loss (-[log effective policy probability + log(1-ineffective policy probability)]), the Adam optimizer is used for iterative training until the discriminator effective probability of the generator output policy stabilizes above the preset threshold, thus completing the model training.

[0179] Based on the trained model, the generator outputs multiple candidate resolution strategies. The top three strategies are selected by ranking them by effective probability using a discriminator. Each strategy is then verified using preset rules, including but not limited to keyword compliance with the domain dictionary, path elimination of loops and access frequency meeting the standard, and weight within a reasonable range. Strategies that do not meet the rules are eliminated. The remaining strategies are simulated and applied to the initial question-answering knowledge base. The conflict reduction rate and user retrieval demand satisfaction rate are calculated through a conflict simulation module. The strategy with the highest conflict reduction rate and demand satisfaction rate that reaches the preset value is selected to obtain the final index conflict risk resolution strategy.

[0180] The feedback adjustment module feeds back the index conflict risk mitigation strategies for different sub-subjects within the target subject and the index conflict risk mitigation strategies for target subjects with different degrees of overlap to the initial question-and-answer knowledge base. It adjusts the initial index path in the initial question-and-answer knowledge base and then performs question-and-answer index conflict simulation, detection, identification, and conflict mitigation adjustment until the target subject's conflict risk list meets the corresponding index conflict risk threshold and the question-and-answer information meets the corresponding user's retrieval needs. At the same time, the index conflict risk mitigation strategy that meets the index conflict risk threshold and user retrieval needs is configured into the initial question-and-answer knowledge base to obtain the conflict-corrected question-and-answer knowledge base.

[0181] This application employs a data acquisition module to perform deep cleaning and structuring of multi-source heterogeneous medical question-and-answer texts. It utilizes an adaptive word segmentation algorithm that integrates domain knowledge with standardized terminology mapping to achieve accurate conversion of the original text into a standardized medical entity sequence, effectively addressing the issues of diverse professional terminology and inconsistent expression in medical texts. Based on a direction-aware aggregation mechanism of a directed graph neural network, a hierarchical index graph with medical domain characteristics is constructed through dual-path feature fusion of incoming edge weight aggregation and outgoing edge semantic similarity aggregation. This allows the knowledge organization structure to simultaneously reflect the access patterns and semantic association strength of the retrieval path. The feature extraction module innovatively adopts a three-level hierarchical feature distillation architecture. At the topic level, core topic vectors are extracted using a dynamic word graph model and an improved clustering algorithm. At the sub-question level, hierarchical question representations are generated based on density clustering and multi-dimensional priority scoring. At the keyword level, a precise terminology index system is constructed by combining TF-IDF statistics and a standardized terminology database. Furthermore, a cross-attention mechanism is used to achieve cross-level semantic association modeling, forming a text embedding representation that combines topic consistency with fine-grained searchability. The three-dimensional overlap model provides a quantitative evaluation benchmark for cross-subject knowledge conflict detection through multi-dimensional similarity calculation of content overlap, index structure overlap, and semantic representation overlap. Content overlap uses the Jaccard similarity algorithm to capture the repetition of term sets; index structure overlap uses a graph edit distance algorithm to analyze path topological differences; semantic representation overlap uses cosine similarity to measure deep semantic connections; and comprehensive overlap achieves overall similarity assessment through weighted fusion. The conflict simulation module constructs an initial knowledge base based on a tree-structured database and combines it with a Bayesian network to build a simulated question-answering algorithm. It calculates the probability of conflict risk through likelihood estimation of prior probability and feature similarity, accurately identifying terminological ambiguities between sub-subjects and path contradictions across subjects. The conflict resolution adversarial model adopts a generative adversarial network architecture. The generator learns preset resolution rules based on a Transformer structure to generate candidate strategies, and the discriminator evaluates the effectiveness of the strategies through a multilayer perceptron, forming a rule-guided adversarial optimization mechanism that ensures that the resolution scheme conforms to domain norms and meets retrieval requirements. The feedback adjustment module establishes a closed-loop optimization mechanism. By iteratively applying conflict resolution strategies and reassessing conflict risks, it dynamically adjusts the knowledge base index structure until a stable, low-conflict state is reached, achieving adaptive evolution of the knowledge base. The entire technical solution, through the collaborative process of data preprocessing, feature extraction, conflict detection, intelligent resolution, and feedback optimization, constructs a self-checking and self-healing medical knowledge management system. This significantly improves the accuracy and efficiency of knowledge retrieval, effectively reduces the conflict risks of cross-source knowledge integration, and provides reliable technical support for intelligent knowledge services in medical institutions.This system is particularly suitable for complex medical scenarios involving multidisciplinary collaboration. It can automatically identify and resolve semantic gaps and retrieval path contradictions between different specialty terminology systems. By standardizing index keywords and optimizing cross-subject retrieval paths, it enables the secure sharing and efficient utilization of medical knowledge, ultimately providing high-quality knowledge support for clinical decision support and patient services.

[0182] Example 2

[0183] Please see Figure 2 Another embodiment of the present invention provides: a knowledge base intelligent management method based on a large model, comprising the following steps:

[0184] S1. Collect question and answer text information and question and answer index information corresponding to different target subjects, perform word segmentation preprocessing on the collected text information, and perform directed mapping processing on the index information in combination with a directed graph neural network to obtain the first text hierarchical sequence and directed index graph corresponding to different target subjects.

[0185] A directed index graph is constructed from the sub-subjects and index paths corresponding to different search information under the target subject;

[0186] S2. Based on the first text hierarchical sequence and directed index graph corresponding to different target subjects, combined with the preset feature distillation extraction model, the first text embedding representation and the first index embedding representation corresponding to different target subjects are obtained.

[0187] S3. Based on the first text embedding representation and the first index embedding representation corresponding to different target subjects, combined with the preset three-dimensional overlap model, the similarity overlap of different target subjects is obtained; the similarity overlap of target subjects includes content overlap, index structure overlap, semantic representation overlap, and comprehensive overlap.

[0188] S4. Based on the question and answer text information and question and answer index information of different target subjects, an initial question and answer knowledge base is constructed by combining a tree database. Based on the initial question and answer knowledge base and the similarity and overlap of different target subjects, a simulated question and answer algorithm using Bayesian functions is used to detect and identify question and answer index conflicts between different sub-subjects within the target subject and between target subjects with different overlap, thereby obtaining a list of target subject conflict risks.

[0189] The target subject conflict risk list includes the index conflict risk and probability corresponding to different sub-subjects within the target subject, as well as the index conflict risk and probability between target subjects under different degrees of overlap.

[0190] S5. Based on the target subject conflict risk list, and combined with the conflict resolution adversarial model of the preset conflict resolution rules, obtain the index conflict risk resolution strategies corresponding to different sub-subjects within the target subject and the index conflict risk resolution strategies corresponding to the target subjects under different overlaps; the index conflict risk resolution strategies include correcting the unified standard index keywords and the directed index paths across sub-subjects or across subjects;

[0191] S6. Feed back the index conflict risk resolution strategies corresponding to different sub-subjects within the target subject and the index conflict risk resolution strategies corresponding to target subjects with different degrees of overlap to the initial question-and-answer knowledge base. Adjust the initial index path in the initial question-and-answer knowledge base and then perform question-and-answer index conflict simulation, detection, identification, and conflict resolution adjustment until the target subject conflict risk list meets the corresponding index conflict risk threshold and the question-and-answer information meets the corresponding user's retrieval needs. At the same time, configure the index conflict risk resolution strategy that meets the index conflict risk threshold and user retrieval needs to the initial question-and-answer knowledge base to obtain the conflict-corrected question-and-answer knowledge base.

[0192] Example 3

[0193] An electronic device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement a knowledge base intelligent management method based on a large model.

[0194] A computer-readable storage medium storing computer instructions that, when executed, produce a knowledge base intelligent management method based on a large model.

[0195] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments under the guidance of the present invention without departing from the spirit and scope of the claims. All of these variations are within the protection scope of the present invention.

[0196] If the technical solution disclosed herein involves personal information, the product using this technical solution has clearly informed the user of the personal information processing rules and obtained the user's voluntary consent before processing the personal information. If the technical solution disclosed herein involves sensitive personal information, the product using this technical solution has obtained the user's separate consent before processing the sensitive personal information, and also meets the requirement of "express consent". For example, at personal information collection devices such as cameras, clear and prominent signs are set up to inform users that they have entered the scope of personal information collection and that personal information will be collected. If an individual voluntarily enters the collection scope, it is deemed that they have agreed to the collection of their personal information; or on the personal information processing device, with clear signs / information informing users of the personal information processing rules, authorization is obtained from the individual through pop-up information or by asking the individual to upload their personal information; wherein, the personal information processing rules may include information such as the personal information processor, the purpose of personal information processing, the processing method, and the types of personal information processed.

Claims

1. A knowledge base intelligent management system based on a large model, characterized in that, include: Data acquisition module, feature extraction module, conflict simulation module, conflict resolution module, feedback adjustment module; The data acquisition module is used to collect question and answer text information and question and answer index information corresponding to different target subjects, and to perform word segmentation preprocessing on the collected text information, and to perform directed mapping processing on the index information in combination with a directed graph neural network to obtain the first text hierarchical sequence and directed index graph corresponding to different target subjects; the directed index graph is constructed from the sub-subjects and demand index paths corresponding to different search information under the corresponding target subject. The feature extraction module obtains the first text embedding representation and the first index embedding representation corresponding to different target subjects by combining the first text hierarchical sequence and the directed index graph corresponding to different target subjects with a preset feature distillation extraction model. Based on the first text embedding representation and the first index embedding representation corresponding to different target subjects, combined with the preset three-dimensional overlap model, the similarity overlap of different target subjects is obtained. The similarity overlap of the target subjects includes content overlap, index structure overlap, semantic representation overlap, and comprehensive overlap. The conflict simulation module constructs an initial question-and-answer knowledge base based on the question-and-answer text information and question-and-answer index information of different target subjects combined with a tree database. Based on the initial question-and-answer knowledge base and the similarity and overlap of different target subjects, it uses a Bayesian function simulation question-and-answer algorithm to detect and identify question-and-answer index conflicts between different sub-subjects within the target subject and between target subjects with different overlap, thereby obtaining a list of target subject conflict risks. The target subject conflict risk list includes the index conflict risk and index conflict risk probability corresponding to different sub-subjects within the target subject, and the index conflict risk and index conflict probability corresponding to target subjects under different degrees of overlap; The conflict resolution module, based on the target subject's conflict risk list and combined with a conflict resolution adversarial model with preset conflict resolution rules, obtains index conflict risk resolution strategies for different sub-subjects within the target subject and index conflict risk resolution strategies for target subjects with different degrees of overlap; the index conflict risk resolution strategies include correcting unified standard index keywords and directed index paths across sub-subjects or across subjects; The feedback adjustment module feeds back the index conflict risk resolution strategies corresponding to different sub-subjects within the target subject and the index conflict risk resolution strategies corresponding to target subjects with different degrees of overlap to the initial question-and-answer knowledge base. It adjusts the initial index path in the initial question-and-answer knowledge base and then performs question-and-answer index conflict simulation, detection, identification, and conflict resolution adjustment until the target subject's conflict risk list meets the corresponding index conflict risk threshold and the question-and-answer information meets the corresponding user's retrieval needs. At the same time, it configures the index conflict risk resolution strategy that meets the index conflict risk threshold and the user's retrieval needs into the initial question-and-answer knowledge base to obtain the conflict-corrected question-and-answer knowledge base.

2. The knowledge base intelligent management system based on a large model as described in claim 1, characterized in that, The process of obtaining the first text embedding representation corresponding to the target subject includes: Based on the question-and-answer text information corresponding to different target subjects, special symbols are filtered and mixed Chinese and English boundary annotations are performed by combining regular expressions with Chinese and English mixed boundary rules to obtain the preprocessed question-and-answer text information corresponding to different target subjects. Based on the preprocessed question-and-answer text information corresponding to different target subjects and combined with the three-level hierarchical rules, we obtain the hierarchical question-and-answer text information of different target subjects; Based on the question-and-answer text information after being layered according to different target subjects, combined with the preset entity relationship extraction algorithm, entity word segmentation is performed on the text corresponding to each layer under different target subjects, and the directed mapping relationship between the corresponding layer and different layers is extracted to obtain the first text embedding representation corresponding to different target subjects.

3. The knowledge base intelligent management system based on a large model as described in claim 2, characterized in that, The three-level hierarchical rule includes a topic layer, a sub-question layer, and a keyword layer. The topic layer calculates the co-occurrence frequency of keywords extracted from the search text and the response text based on the question and answer text information corresponding to each target subject using the TextRank algorithm, performs core word screening, and generates a target keyword vector cluster based on the core word sequence combined with BioBERT to obtain the topic name corresponding to the current target subject's question and answer text information and assigns a root node ID associated with the text storage address. The sub-question layer generates sub-question sentence vectors based on the search text under each topic and Sentence-BERT. After clustering by the DBSCAN algorithm, it obtains the sub-question keyword sequence and sub-question answer priority. Based on the sub-question answer priority corresponding to the sub-question keyword sequence, it assigns sub-node IDs and associates them with the corresponding topic root node. The keyword layer obtains the TF-IDF value of the keyword corresponding to each response text through a statistical algorithm based on the response text corresponding to each sub-question. After comparison and verification with the UMLS terminology database, the keywords are sorted in descending order according to their TF-IDF values, and keyword IDs are assigned according to the sorting results and bound to the corresponding sub-question nodes.

4. The knowledge base intelligent management system based on a large model as described in claim 3, characterized in that, The construction process of the feature distillation extraction model includes: Based on the first text hierarchical sequence corresponding to different target subjects, the teacher sub-model in the feature distillation extraction model performs hierarchical text feature extraction along the corresponding nodes of the topic layer, sub-question layer, and keyword layer, specifically as follows: Based on the first text extraction network, the root node text of the topic layer is used to extract topic feature vectors to obtain the first topic vector; Based on the second text extraction network, features are extracted from each priority sub-problem in the sub-problem layer to generate a hierarchical sub-problem vector. Based on the hierarchical sub-problem vector and the first topic vector, a cross-attention network is combined to obtain the topic-oriented priority association weight vector of the sub-problem. Based on the third text extraction network, the keyword layer corresponding to each sub-question is used to extract features according to the ranking result, so as to obtain the keyword feature vector of the answer text corresponding to each sub-question. Based on the keyword information of the answer text corresponding to each sub-question, combined with the hierarchical sub-question vector and the cross-attention network, the association weight vector of the answer text keywords oriented towards the sub-question is obtained.

5. The knowledge base intelligent management system based on a large model as described in claim 4, characterized in that, The construction process of the feature distillation extraction model also includes: Based on the first topic vector, hierarchical sub-question vector, sub-question-oriented topic priority association weight vector, keyword feature vector of each sub-question corresponding to the answer text, and keyword-oriented association weight vector of the answer text to the sub-question, a graph neural network is used to construct the first text embedding representation of the question and answer text corresponding to the target subject. Based on the sub-subjects and demand index paths corresponding to different retrieval information under different target subjects in the directed index graph, combined with the GraphSAGE algorithm, the index information and access edge weight information features are extracted along the retrieval path corresponding to the directed index graph to obtain the first retrieval path features. Based on the first retrieval path features, a convolutional attention network is used to extract the retrieval direction and access path frequency corresponding to different retrieval needs under the same subject, and the retrieval path direction and access path frequency of different subjects facing the same retrieval needs, to obtain the first index embedding representation corresponding to the target subject.

6. The knowledge base intelligent management system based on a large model as described in claim 5, characterized in that, The construction process of the feature distillation extraction model also includes: Based on the first text embedding representation and the first index embedding representation, the feature space is aligned by combining the feature alignment algorithm with the retrieval target requirement as the label to obtain the aligned first text embedding representation and the first index embedding representation. The aligned first text embedding representation and the first index embedding representation are respectively input into the text distillation branch and the index distillation branch of the student sub-model in the feature distillation extraction model. Feature distillation is performed according to the contribution of the corresponding features to the retrieval target requirements and the contribution threshold to obtain the first text embedding representation and the first index embedding representation after feature distillation.

7. The knowledge base intelligent management system based on a large model as described in claim 6, characterized in that, The process of obtaining the target subject conflict risk list includes: Based on the question-and-answer text information and question-and-answer index information of different target subjects, an initial question-and-answer knowledge base is constructed through a tree-structured database. The initial question-and-answer knowledge base is associated with the nodes of the topic layer, sub-question layer, and keyword layer, as well as the sub-subjects, retrieval units, and path information of the directed index graph, forming a hierarchical index structure. Extract the index features of different sub-subjects within the target subject and the similarity overlap data of different target subjects from the initial question-answering knowledge base, and use them as input parameters for conflict detection. The index features include a set of index keywords and a directed path sequence. Based on the conflict detection between different sub-sub ... Based on the conflict detection between target subjects under different overlap degrees, high overlap subject pairs are screened based on comprehensive overlap degree. The first text embedding representation and the first index embedding representation of the high overlap subject pairs are similar to identify feature pairs that are semantically related but have different indices. Comprehensive overlap degree is used as a weighting factor to adjust the prior probability of the Bayes function. The cross-subject retrieval scenario is simulated by a simulated question answering algorithm and the cross-subject conflict probability is calculated by combining the likelihood function. The index conflict risk entries between subjects are generated by arranging them in descending order of conflict probability. Based on the intra-subject index conflict risk entries and inter-subject index conflict risk entries, supplementary conflict feature descriptions and conflict impact scope are generated to create a target subject conflict risk list.

8. A knowledge base intelligent management method based on a large model, implemented based on the knowledge base intelligent management system based on a large model as described in any one of claims 1-7, characterized in that, include: Collect question-and-answer text information and question-and-answer index information corresponding to different target subjects, perform word segmentation preprocessing on the collected text information, and perform directed mapping processing on the index information in combination with a directed graph neural network to obtain the first text hierarchical sequence and directed index graph corresponding to different target subjects; The directed index graph is constructed from the sub-subjects and index paths corresponding to different search information under the target subject; Based on the first text hierarchical sequence and directed index graph corresponding to different target subjects, combined with the preset feature distillation extraction model, the first text embedding representation and the first index embedding representation corresponding to different target subjects are obtained. Based on the first text embedding representation and the first index embedding representation corresponding to different target subjects, combined with the preset three-dimensional overlap model, the similarity overlap of different target subjects is obtained. The similarity overlap of the target subjects includes content overlap, index structure overlap, semantic representation overlap, and comprehensive overlap. An initial question-and-answer knowledge base is constructed based on question-and-answer text information and question-and-answer index information of different target subjects, combined with a tree-structured database. Based on the initial question-and-answer knowledge base and the similarity and overlap of different target subjects, a simulated question-and-answer algorithm using Bayesian functions is used to detect and identify question-and-answer index conflicts between different sub-subjects within the target subject and between target subjects with different overlap, thereby obtaining a list of target subject conflict risks. The target subject conflict risk list includes the index conflict risk and index conflict probability corresponding to different sub-subjects within the target subject, as well as the index conflict risk and index conflict probability corresponding to target subjects with different degrees of overlap.

9. The intelligent knowledge base management method based on a large model as described in claim 8, characterized in that, The method further includes: Based on the target subject conflict risk list, and combined with the conflict resolution adversarial model of the preset conflict resolution rules, the index conflict risk resolution strategies corresponding to different sub-subjects within the target subject and the index conflict risk resolution strategies corresponding to the target subjects under different overlaps are obtained; the index conflict risk resolution strategies include correcting the unified standard index keywords and the directed index paths across sub-subjects or across subjects; The index conflict risk resolution strategies for different sub-subjects within the target subject and the index conflict risk resolution strategies for target subjects with different degrees of overlap are fed back to the initial question-answering knowledge base. The initial index path in the initial question-answering knowledge base is adjusted, and question-answering index conflict simulation, detection, identification, and conflict resolution are performed after the adjustment until the target subject's conflict risk list meets the corresponding index conflict risk threshold and the question-answering information meets the corresponding user's retrieval needs. At the same time, the index conflict risk resolution strategy that meets the index conflict risk threshold and user retrieval needs is configured into the initial question-answering knowledge base to obtain the conflict-corrected question-answering knowledge base.

Images