Interactive health science popularization question and answer guiding system for constructing medical knowledge graph

Abstract

The invention relates to the technical field of knowledge fusion, and discloses an interactive health science popularization question and answer guidance system for constructing a medical knowledge graph, comprising: a guidance instruction generation module extracting attribute variation dimensions between to-be-aligned data and candidate nodes and generating a guidance instruction; the feedback analysis module converts the feedback data into a feedback correction vector; the atlas fusion actuator determines an initial fusion weight according to the coordinate distance, determines a weight correction coefficient based on the topological centrality of the target node, constructs a weighted feature update function by using a feedback correction vector and a local topological relation to calculate a feature increment, and completes feature convergence fusion; according to the method, a priori constraint force balance feedback feature pointing value provided by a map structure is utilized, it is ensured that heterogeneous knowledge fusion conforms to medical ontology logic, and the global stability of science popularization map evolution is guaranteed.

Description

Technical Field

[0001] This invention relates to an interactive health science popularization question-and-answer guidance system for constructing a medical knowledge graph, belonging to the field of knowledge fusion technology. Background Technology

[0002] Current medical knowledge fusion typically establishes entity mapping relationships to build a structured knowledge base. When processing input data from non-professional users, there are semantic differences between colloquial expressions and professional terms. In cases of semantic overlap, a single text feature points to multiple medical entities. Industry practices involve increasing matching depth or introducing external information. However, in scenarios with high feature density, this can easily trigger topological overshoot, causing heterogeneous data to erroneously collapse to non-target nodes, thus affecting the stability of the knowledge graph.

[0003] Even with layered design to optimize the knowledge storage topology, logical conflicts cannot be eliminated if the dynamic interaction process lacks accurate detection and feedback compensation for semantic deviations. For example, the Chinese invention patent CN120162419A discloses a verifiable question-and-answer system that integrates a layered medical atlas with dynamic retrieval. This system uses a multi-layered structure to achieve vertical retrieval from overall disease topics to detailed explanations. Although the solution has the advantage of multi-source data association, its core logic is still based on static matching. It relies on the original accuracy of user input. When faced with vague or cognitively biased popular science consultations, it cannot identify the dimension of maximum variation that causes semantic conflicts. It lacks the ability to establish closed-loop control by using feedback correction vectors and local topological constraints. When dealing with hesitant or vague inputs, the lack of necessary damping adjustment causes pulse-like jumps, resulting in a sharp drop in entity alignment accuracy. It cannot ensure that the integration of heterogeneous knowledge conforms to the logic of medical ontology, exhibiting topological structural fragility.

[0004] Therefore, the technical problem to be solved by this invention is how to guide the system to actively detect and control the convergence of fuzzy semantic features by constructing an interactive health science Q&A based on medical knowledge graphs, thereby solving the lack of topological stability in the heterogeneous data fusion process. Summary of the Invention

[0005] To address the problems raised in the background art, the technical solution of the present invention is as follows: An interactive health science popularization question-and-answer guidance system for constructing a medical knowledge graph, comprising: The guidance instruction generation module is used to calculate the attribute entropy difference between the data to be aligned and the candidate nodes in the medical knowledge graph on the semantic attribute set, extract the attribute dimension with the largest attribute entropy difference as the attribute variation dimension, and generate attribute guidance instructions. The feedback parsing module is used to obtain interactive feedback data in response to attribute guidance instructions, and convert the interactive feedback data into feedback correction vectors in the feature space based on a preset feature mapping matrix. The graph fusion executor, connected to the guidance instruction generation module and the feedback parsing module, determines the initial fusion weights based on the coordinate distances between the feature vectors of the data to be aligned and each candidate node. When performing feature fusion, the graph fusion executor extracts the topological centrality of the target node in the medical knowledge graph and determines the corresponding weight correction coefficient based on the topological centrality. The graph fusion executor uses the weight correction coefficient to numerically adjust the initial fusion weights and establishes a weighted feature update function composed of the feedback correction vector and the local topological connectivity of the medical knowledge graph. By calculating the feature increment of the feature vector under the weighted feature update function, the graph fusion executor enables the feature vector to converge toward the target node under the combined effect of the feature orientation value generated by the interactive feedback data and the structured prior constraint value generated by the medical knowledge graph.

[0006] Preferably, the graph fusion executor is further used to extract a subgraph representing the logical relationship of the medical ontology based on the association topological path of the candidate nodes in the medical knowledge graph; the graph fusion executor inputs the feedback correction vector into the subgraph, calculates the association conflict entropy between the feedback correction vector and the subgraph topology, calculates the confidence weight of the feedback correction vector based on the association conflict entropy, and uses the confidence weight to perform deviation compensation on the update direction of the feature vector.

[0007] Preferably, the guidance instruction generation module is specifically used to identify feature dimensions whose attribute entropy difference exceeds a preset deviation threshold, and to use the feature dimension as the attribute variation dimension.

[0008] Preferably, when the graph fusion executor calculates the weight correction coefficient, it obtains the hierarchical depth value of the target node in the medical knowledge graph and calculates the weight correction coefficient using a preset positive correlation mapping function. By increasing the weight distribution of deep nodes in the fusion calculation, the global update stability of the medical knowledge graph is maintained.

[0009] Preferably, the system further includes: a dynamic graph maintenance module, connected to the graph fusion executor, used to mark the target node as a new topological anchor point to perform incremental updates after feature convergence is completed, and to update the topological centrality of the affected local region in real time.

[0010] Preferably, when the feedback parsing module generates the feedback correction vector, it performs semantic projection processing on the unstructured interactive feedback data and calculates the projection intensity of the interactive feedback data on each orthogonal basis vector in the feature space.

[0011] Preferably, the graph fusion executor is used to iteratively update the feature vector according to a preset update cycle, and in each update cycle, calculate the feature fusion correction amount at the current time based on the coordinate distance and weight correction coefficient at the current time.

[0012] Preferably, the graph fusion executor is also used to send a conflict trigger signal to the guidance instruction generation module when the association conflict entropy exceeds a preset logical difference threshold, so that the guidance instruction generation module can regenerate the attribute guidance instruction for another attribute variation dimension.

[0013] Preferably, the system further includes: an interactive guidance interface, connected to the guidance instruction generation module, used to convert attribute guidance instructions into natural language guidance text and output it to the interactive terminal.

[0014] Preferably, the graph fusion executor computes the semantic feature bias values ​​representing semantic alignment conflicts. When doing so, the following quantitative calculation rules shall be followed: ,in, These are partial values ​​for semantic features; The semantic distance between the feature vector and the target node coordinates; Let be the topological centrality of the target node.

[0015] Compared with the prior art, the beneficial effects of the present invention are: 1. In interactive health science popularization of medical knowledge graph, the orthogonal difference attributes between each candidate alignment node are extracted and a single-dimensional guided detection instruction is constructed. The maximum variation dimension that causes the current entity alignment conflict is identified. According to the feedback signal, the projection operation is performed in the feature space to drive the semantic features of the text to be fused to undergo logical collapse towards the target alignment node. The above process transforms the large-scale high-dimensional semantic disambiguation task into a dynamic feedback control process based on information entropy difference, bypassing the performance bottleneck of static semantic similarity calculation method in high-density semantic overlap area, and eliminating the semantic ambiguity of medical slang and professional terminology.

[0016] 2. By constructing a logical consensus subgraph on the topological path of the medical knowledge graph using target alignment nodes, user feedback features are input as perturbations to calculate conflict entropy. The confidence weight of the feedback signal is determined based on the magnitude of the conflict entropy, enabling automatic filtering of noise in the interaction channel. By utilizing a structured truth constraint mechanism based on the existing graph, node misalignment due to user subjective cognitive biases is avoided, ensuring that the heterogeneous knowledge fusion process conforms to the causal logic of the medical ontology and guaranteeing the global topological stability of the science popularization graph evolution process. A quasi-static fusion mechanism based on topological elastic potential energy is introduced into the feedback fusion actuator, mapping the feedback signal into a feature space orientation vector. The topological damping coefficient is calculated in combination with the candidate node level depth, so that the fusion position is determined by the balance between the feedback traction force and the restoring force of the existing connection relationship in the graph. Displacement compensation is achieved by damping adjustment, making the entity alignment process a robust approximation based on entropy reduction rather than a pulsed jump, eliminating the topological fragility of the fusion process, and achieving adaptive adaptation to dense data environments without relying on massive samples.

[0017] 3. Based on historical fusion records, knowledge momentum is allocated to graph nodes. During the initial semantic mapping, knowledge momentum is used to perform logical displacement processing on candidate alignment nodes, increasing the spatial distinguishability between high-frequency path nodes and noise nodes. Spatiotemporal coupling momentum balance logic is used to reduce the frequency of dependence on external detection, reducing real-time computation overhead in large-scale graph environments, enabling the system to have the characteristic of co-evolution and self-organization as the fusion frequency increases. Combined with interaction delay fingerprints to discriminate cognitive load, the mapping weight is dynamically adjusted according to the deviation between the delay and the preset standard response window. When the feedback confidence is found to be lower than the threshold, the secondary orthogonal axis is reset to reconstruct the detection dimension. By reusing implicit temporal information, the system can penetrate the ambiguity of expression without increasing the number of interaction rounds. Lateral residual information is used to reposition the detection focus, improving the semantic convergence efficiency of the interactive guidance system when handling hesitant inputs. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the overall architecture and data flow closed-loop logic of the interactive guidance system of the present invention; Figure 2 This is a comparison of the semantic feature convergence trajectory and optimization rate under different topological damping coefficients in this invention; Figure 3 This is a comparison chart of the system entity alignment accuracy and noise resistance stability under Gaussian noise interference environment of the present invention; Figure 4 This is a sequence diagram of the multi-module collaborative interaction of the science popularization question-and-answer guidance and graph feature fusion task of this invention. Detailed Implementation

[0019] The following embodiments are intended to explain the present invention, and not to limit the scope of protection of the present invention.

[0020] An interactive health science Q&A guidance system for constructing a medical knowledge graph includes a guidance instruction generation module, a feedback parsing module, and a graph fusion executor. These components are connected via a data bus to achieve controlled convergence from unstructured text to knowledge graph entities. The guidance instruction generation module identifies feature conflicts and generates probe instructions by calculating attribute entropy differences. The feedback parsing module converts interactive information into feedback correction vectors using a feature mapping matrix. The graph fusion executor calculates feature increments and performs feature convergence under topological constraints. When processing science Q&A text input by non-professional users, the ambiguity of natural language often leads to data to be aligned mapping to multiple professional medical entities, resulting in semantic overlap. To address this challenge, the guidance instruction generation module acquires the data to be aligned and searches the medical knowledge graph for entities with a semantic distance less than a preset threshold. Multiple candidate nodes, The preset threshold is determined by the local feature density of the medical ontology library. That is, in the local feature space where the node distribution is relatively dense, the system sets the threshold to a smaller value to narrow the initial screening range. The guidance instruction generation module calculates the attribute entropy difference between the data to be aligned and each candidate node in the semantic attribute set. , This refers to the attribute entropy difference; in the specific calculation procedure, the system extracts the common attributes of candidate nodes. For each attribute dimension, calculate the information distribution probability of the data to be aligned in each dimension. Using the formula Calculate the information entropy, and select the attribute dimension with the largest information difference among candidate nodes as the attribute variation dimension. For information entropy, The system represents the information distribution probability. For the input science text "chest discomfort," the system identifies angina and reflux esophagitis as candidate nodes, and calculates their attribute entropy difference on the pain trigger dimension. for The attribute entropy difference is higher than the frequency of attacks. Based on this, the system determines the dimension of the pain trigger as an attribute deterioration and outputs detection instructions for that dimension to the interactive terminal.

[0021] To address cognitive noise in user feedback, the feedback parsing module executes a feature space mapping procedure; the feedback parsing module acquires interactive feedback data in response to attribute guidance commands and utilizes a preset feature mapping matrix. Convert the data into a feedback correction vector in the feature space. , The feature mapping matrix, Feedback correction vector; feature mapping matrix By using the preset Perform singular value decomposition on the medical science popularization corpus. After dimensionality reduction, it is constructed to establish a linear mapping relationship between natural language vocabulary and components of the high-dimensional vector space; feature mapping matrix. The offline calibration procedure selects 10,000 medical science popularization corpora containing descriptions of respiratory system pathological features to construct an original sample set. A word vector model is used to map the semantic feature terms in the sample set into a 256-dimensional word vector space. Singular value decomposition is then performed on the obtained sample feature matrix to extract the semantic features of the preceding words. The projection operator is composed of orthogonal basis vectors. When the cumulative energy contribution rate of the singular values ​​reaches 90%, the parameters of the projection operator are locked and determined as the eigenmap matrix. This matrix is ​​used to convert the received unstructured interactive feedback data into a feedback correction vector in the feature space. .

[0022] The feedback parsing module generates the feedback correction vector. During the process, the interaction latency of user feedback is recorded simultaneously. , For interaction latency; the system compares Compared with the preset response benchmark The degree of deviation is used to calculate the feedback confidence coefficient. ,in In response to the benchmark, For feedback confidence coefficient; feedback confidence coefficient The computation depends on the interaction latency Compared with response benchmark Deviation analysis, response benchmark The calibration procedure selects historical consultation texts from specific specialized institutions as input sources and statistically analyzes the time distribution from displaying attribute guidance instructions to receiving feedback signals for texts with an average word length of 10 to 12 Chinese characters. The arithmetic mean of the time distribution is set as the response benchmark. When the real-time monitored interaction latency Deviation from Response Benchmark When the amplitude increases, the feedback confidence coefficient is reduced using a linear interpolation function. The numerical value is used as a weight adjustment factor in the weighted feature update function to incorporate the topological centrality provided by the medical knowledge graph. The update direction of the jointly constrained feature vectors, while ensuring that the fusion of heterogeneous knowledge conforms to the logic of medical ontology, achieves automatic filtering of noise in the interaction channel. This feedback confidence coefficient Used to quantify the credibility of user feedback, and decreases linearly with increasing latency deviation; the feedback parsing module adjusts the feedback correction vector. With feedback confidence coefficient The data is passed to the graph fusion executor, which uses implicit temporal information as an auxiliary criterion to improve the reliability of system decision-making in complex semantic environments.

[0023] In large-scale dense medical knowledge graphs, the high coupling between nodes can easily lead to topological overshoot during feature convergence. To address this issue, the graph fusion executor employs a quasi-static fusion operator based on topological elastic potential. The graph fusion executor uses the feature vectors of the data to be aligned... Calculate the initial fusion weights based on the coordinate distance between the candidate nodes and the target nodes. ,in For feature vectors, The initial fusion weights are used; simultaneously, the executor extracts the topological centrality of the target node in the medical knowledge graph. This serves as a quantitative indicator to measure the importance of the node in the global structure. Topological centrality; the graph fusion actuator utilizes topological centrality. For the initial fusion weights Adjustments are made and a weighted feature update function is constructed; during the update iteration, the executor calculates the semantic feature bias values. Its calculation formula is ,in These are partial values ​​for semantic features. This represents the semantic distance between the feature vector and the target node coordinates.

[0024] The graph fusion actuator adjusts the weighting coefficients. Calculate displacement correction This correction is used to drive the feature vector to perform a small displacement towards the target node, where This is the weighting adjustment factor. For displacement correction; displacement correction Follow the formula below: ,in The target position vector is pointed to by the feedback. The preset topological damping coefficient; Topological damping coefficient The value of is positively correlated with the hierarchical depth of the target node in the medical ontology library; that is, for basic pathological nodes at deeper levels, the system sets a larger topological damping coefficient to maintain their stability. The quantization selection follows the principle of depth value based on the target node level. The dynamic mapping rules, during the system initialization phase, calculate the path steps from the target node to the root node by traversing the local topological paths of the medical knowledge graph, in order to establish the hierarchical depth value. Substitute into the linear adjustment formula The topological damping coefficient under the current operating condition is calculated. ,in The reference damping value is set to 1.2. This is the hierarchical gain coefficient, set to 0.5. The path length from the target node to the root node is set such that it generates a large topological damping coefficient for deeper-level baseline pathology nodes. Thus, in calculating the displacement correction amount The system compresses the single-step displacement amplitude of the feature vector by increasing the denominator component, driving the data to be aligned to converge smoothly towards the target node and suppressing displacement overshoot in dense regions of the topology. For first-level disease classification nodes, the system uses the topological damping coefficient... Set as For peripheral symptom nodes, the topological damping coefficient will be... Set as .

[0025] To prevent erroneous feedback from causing local topological distortions in the knowledge graph, the graph fusion executor is also configured with path consensus arbitration logic; after receiving the feedback signal, the system extracts the target node's position in the medical knowledge graph. The topological path of the order association is used to construct a logical consensus subgraph; the feedback parsing module calculates the feedback correction vector. The conflict entropy associated with the subgraph topology is determined by the fact that if the conflict entropy exceeds a preset difference threshold. The graph fusion executor determines that there is a logical deviation in the current feedback and sends a trigger signal to the guidance instruction generation module to initiate the re-probing of secondary difference attributes. This constraint mechanism based on the existing graph structured truth realizes the automatic filtering of noise in the interaction channel, ensures that the fusion of heterogeneous knowledge conforms to the logic of medical ontology, and guarantees the global stability of the evolution process of popular science graph.

[0026] Example 1: In a specific online health consultation data processing scenario, the system receives a user's colloquial description of a squeezing sensation in the chest. The corresponding initial semantic feature map points to multiple candidate nodes, such as acute myocardial infarction and reflux esophagitis, within the local feature space of the medical knowledge graph. Because the symptom features of these candidate nodes have semantic overlap in the feature space, simply relying on static semantic similarity matching cannot achieve definite entity alignment. To eliminate logical deadlocks in the feature space, the instruction generation module extracts the orthogonal difference feature set between each candidate node and calculates the attribute entropy difference of the candidate nodes in dimensions such as pain triggers, duration of attack, and relief methods. , This is the attribute entropy difference; its specific calculation procedure follows the information entropy calculation rules, namely... ,in For information entropy, This represents the information distribution probability of the data to be aligned along a specific dimension; the calculation results show the attribute entropy difference in the duration dimension. for The attribute entropy difference in the dimension of pain properties for Based on this calculation, the instruction generation module selects duration as the attribute variation dimension and outputs a query to the interactive terminal asking whether the symptom duration exceeds [a certain value]. A single-dimensional guided probe command lasting one minute.

[0027] User feedback indicates that the duration of symptoms has exceeded [a certain period]. Within minutes, the feedback parsing module obtains the interactive feedback data and uses the feature mapping matrix. Transform it into a feedback correction vector in the feature space ,in The feature mapping matrix, This serves as the feedback correction vector; to ensure automatic filtering of interaction channel noise, the feedback parsing module synchronously records the interaction delay of user feedback. for seconds, of which This is the interaction latency; based on this latency and the preset response benchmark... The system calculates the feedback confidence coefficient based on the degree of deviation. for ,in In response to the benchmark, The feedback confidence coefficient is used; the graph fusion actuator obtains the feedback correction vector. Then, the topological centrality of the target node in the medical knowledge graph is extracted. ,in To measure topological centrality, and to prevent overshoot in the graph structure due to single feedback bias, a topological damping coefficient is introduced into the actuator during the feedback fusion process. ,in This is the topological damping coefficient; for acute pathological nodes at higher levels, the actuator will use the topological damping coefficient. Set as And combined with the feedback confidence coefficient Adjust the mapping weights of the feature vectors.

[0028] The graph fusion executor calculates feature increments using a weighted feature update function. Under the action of this function, the semantic feature vector gradually converges towards the acute myocardial infarction node, balancing the feature orientation values ​​generated by interactive feedback and the structured prior constraint values ​​generated by the medical knowledge graph structure. The displacement correction during the convergence process... Follow the formula Perform calculations, where This is the displacement correction amount. The target position vector is pointed to by the feedback. The semantic feature vector to be fused is used as the basis for the system alignment. When the topological energy between the shifted feature vector and the target node reaches a local minimum, the system completes the entity alignment of the data to be aligned with the target disease node. The dynamic graph maintenance module marks the target node as a new topological anchor point, performs incremental updates, and updates the topological centrality of the affected local region in real time. The convergence trajectory of the eigenvectors in the feature space is affected by the topological damping coefficient. It exhibits a robust approximation state due to constraints.

[0029] Example 2: In the experiment verifying the fusion performance and topological stability of the medical knowledge graph, the test environment consisted of a computing server with floating-point computing capabilities and storage... The system consists of an in vitro database of standard medical science popularization texts. The experimental data comes from a manually annotated medical ontology mapping set. To simulate the cognitive biases that non-professional users may have during interactive feedback, the experimental system adds a feedback correction vector to the feature space during the feedback parsing stage. Active superposition signal-to-noise ratio is Gaussian white noise, For feedback correction vectors; regarding the core parameter response benchmark The settings, in which To meet the baseline, the technical trade-off lies in balancing the real-time nature of interactive responses with the accuracy of feedback confidence assessment. As the semantic complexity of guided probing commands (i.e., the number of attribute dimensions involved) increases, the system adjusts the linear interpolation function to reduce the probability of random guessing. The numerical value; under the experimental conditions, for complex detection commands involving third-order or higher-order associated paths, the response benchmark Set as Seconds are used as the feedback confidence coefficient. The base timescale for calculation, where The test was conducted to evaluate the confidence level. The experimental groups included the present invention sample group, a control group, and a partially missing control group. The present invention sample group employed a complete technical solution consisting of a guidance instruction generation module, a feedback parsing module, and a graph fusion actuator. The control group used a static alignment method based on text string similarity. The partially missing control group removed the topological elastic potential energy mechanism from the graph fusion actuator. The experiment involved inputting data into the system... The study analyzed respiratory symptom texts with highly semantically overlapping features, and observed the entity alignment accuracy and topological conflict entropy of each sample group under noise interference and feature vector convergence. Table 1 shows the performance comparison data of the sample group and the control group under different experimental conditions.

[0030] Table 1: Performance Comparison Data of the Sample Group and Control Group of the Invention under Different Experimental Conditions See Table 1, in superposition Under the objective condition of Gaussian white noise, the entity alignment accuracy of the control group was reduced from... Reduce to And the topological conflict entropy reaches This indicates that it cannot effectively suppress cognitive noise in the interaction channel; in contrast, the sample group of this invention uses the guidance instruction generation module to extract the maximum variation dimension and combines it with the feedback confidence coefficient output by the feedback parsing module. Weighted adjustment is applied to the convergence path to maintain alignment accuracy under the same noise conditions. And the topological conflict entropy is only By comparing the data of the sample group of this invention with that of the partially missing control group, it can be found that after removing the topological elastic potential energy mechanism, the entity alignment accuracy of the sample group drops significantly, and the topological conflict entropy decreases from... Rise to .

[0031] For parameters Boundary verification tests show that, This is the topological damping coefficient, set below the lower limit of the working window. At that time, the update direction of the semantic feature vector is relative to the feedback correction vector containing noise. This results in an overresponse, causing frequent displacement overshoots in the system's feature space, and the topological conflict entropy rises to [a certain level]. ;when Set to be higher than the maximum number of working windows At that time, although the topological conflict entropy remained at The low position, but due to the displacement correction amount The single-step increment is too small, causing the feature vector to deviate from the preset value. If the semantic potential minimum point of the target node cannot be reached within the next iteration cycle, the entity alignment accuracy decreases due to the convergence rate limitation. , This is the displacement correction amount.

[0032] Example 3: This example combines Figures 1 to 4 This section describes an interactive health science Q&A guidance system for constructing a medical knowledge graph, such as... Figure 1 As shown, the system receives the data to be aligned, i.e., the user's science Q&A text. The guidance instruction generation module extracts the attribute variation dimension between the data to be aligned and the candidate nodes based on the candidate nodes provided by the medical knowledge graph and the topological structure, and generates attribute guidance instructions. These instructions are transmitted to the interactive terminal or the user side. The user outputs the instructions and inputs the interactive feedback data generated by the feedback into the feedback parsing module, which converts the unstructured data into a feedback correction vector in the feature space. The graph fusion executor, as the core processing unit, determines the initial fusion weights based on the coordinate distance between the feature vector to be aligned and the candidate nodes, and determines the weight correction coefficients based on the topological centrality of the target node. At the same time, it constructs a weighted feature update function using the feedback correction vector and the local topological relationship of the medical knowledge graph, and balances the feature pointing value and the prior constraint by calculating the feature increment, ultimately achieving feature convergence fusion output.

[0033] like Figure 2As shown, a feature convergence performance curve is constructed with the number of iterations as the x-axis and semantic distance as the y-axis. The figure shows the convergence trajectory under three different topological damping coefficients K: when K=0.5 (low damping), the semantic distance decreases sharply in the early stage but tends to level off later; when K=15.0 (high damping), the semantic distance decreases slowly and linearly with the increase of the number of iterations; and when K=5.0 (using the parameters of this invention), the curve shows a smooth decreasing trend that balances convergence speed and accuracy. Figure 3 As shown, a bar chart comparing the entity alignment accuracy of the complete system and the partially missing control group under different interference conditions is constructed. In a noise-free environment, the alignment accuracy of the complete system is slightly higher than that of the partially missing control group. However, under the condition of introducing 20dB Gaussian noise, the accuracy of the partially missing control group decreases, while the complete system of this invention maintains a high entity alignment accuracy, directly demonstrating the effectiveness of the topological elastic potential energy mechanism in suppressing interactive channel noise. Figure 4 As shown, this time-series process covers the entire process from the user terminal inputting popular science Q&A text to the final display of matched medical entities. After receiving the data to be aligned, the guidance instruction generation module retrieves and returns a set of candidate nodes from the medical knowledge graph. It extracts the attribute variation dimension by calculating the attribute entropy difference and generates attribute guidance instructions. The interactive guidance interface outputs natural language guidance text to the user terminal and submits interactive feedback. The feedback parsing module transmits the interactive feedback data and performs feature mapping transformation to transmit the feedback correction vector. The graph fusion executor then obtains the topological centrality, calculates the initial fusion weights, and executes the feature convergence loop. Finally, after completing the entity alignment, the alignment result is returned to the interactive guidance interface for display.

[0034] Example 4: In the initial construction of a respiratory disease knowledge system, the system acquires raw question-and-answer text containing non-standardized medical corpus, executes the initial state definition procedure to establish the physical pathway for semantic preprocessing; the system calls Jieba (a Chinese language learning platform). Word segmentation algorithms segment the input text, extract semantic features, and utilize... The model is based on the continuous bag-of-words system. The architecture maps the segmented words to The word vectors are 3D, where the sliding window size during model training is set to 1. The number of iteration rounds is set to To expose the mapping mechanism in the feature space, the system executes a mapping on the feature mapping matrix. The offline calibration procedure utilizes data stored in the in vitro database. Using standard medical science popularization texts as a benchmark, the singular value decomposition of the sample feature matrix is ​​calculated, i.e. Obtain the projection weights from the original text space to the medical ontology space, where The feature mapping matrix; in this calibration procedure, when the cumulative energy contribution rate of singular values ​​reaches... When the calibration is complete, the system determines that the calibration has converged and locks the current feature mapping matrix. .

[0035] To address the topology stability issue in large-scale dense data environments, the system performs optimizations targeting the topology damping coefficient. An adaptive calibration procedure is established to define the quantitative logical relationship between the damping factor and the target node level depth. The topological damping coefficient is used; the graph fusion actuator obtains the hierarchical depth value of the target node in the medical knowledge graph. The topological damping coefficient is calculated according to the following linear growth formula. : ,in The reference damping value is set to [value]. , This is the hierarchical gain coefficient, and its value is set to... , This represents the path length from the target node to the root node; when the system receives a feedback correction vector involving the root-level disease classification node... At that time, due to Increase, and the calculated topological damping coefficient Increase to This limits the amount of displacement correction. The single-step increment suppresses overshoot of the eigenvector in the high-frequency connection region, where... For feedback correction vector, This is the displacement correction amount; during the graph evolution process, the system executes a process judgment quantization procedure and uses objective measurement tools to obtain the topological conflict entropy after each iteration. ,in Let the topological conflict entropy be the topological conflict entropy; when the topological conflict entropy is... continuous The reduction in accuracy in the next iteration is less than the preset accuracy threshold. When the system determines that the feature vector has reached the semantic potential minimum point of the target node, it locks the entity alignment result. After the fusion is completed, the dynamic graph maintenance module uses a hash mapping algorithm to perform anonymization labeling on the feature vectors of the newly added entities, protecting the user's original privacy data while maintaining the stability of the graph structure.

[0036] Example 5: In the scenario of deploying the guidance system to a respiratory specialty medical institution, in order to establish a system benchmark for business semantic distribution, the system executes a pre-deployment calibration procedure to extract the institution's existing [data / data]. The system retrieves historical consultation texts and generates a local word frequency statistical distribution for that scenario. This statistical distribution is then used to map the feature matrix. Perform linear offset correction to ensure that the directional deviation of the projection operator in the specialized context is less than [value missing]. ,in This is the feature mapping matrix; simultaneously, the average interaction latency under this environment is obtained through multipath sampling, and the response baseline is reset accordingly. For an average word length of to The system will respond to the baseline text input of one Chinese character. Calibrated as seconds, of which To establish a response baseline, a baseline for assessing feedback confidence that is consistent with the cognitive load of this user group is created.

[0037] In the initial connection weight synchronization procedure of the knowledge graph, the system acquires the medical ontology subgraph of the target deployment environment and performs static topology analysis. The graph connection weights are initialized by calculating the association probability density between each pathological node in the subgraph. This process uses objective measurement tools to obtain the degree centrality distribution of the nodes and maps it to the topological damping coefficient in the initial state. ,in This represents the topological damping coefficient; based on the differences in density between different specialized subgraphs, the system uses an adaptive calculation formula to determine the baseline damping value. Perform gain adjustment to correct the displacement at the initial moment. It is dimensionally equivalent to the average semantic radius of the subgraph, where The reference damping value, This is the displacement correction amount, thus completing the baseline field strength modeling of the topological elastic potential field of the graph before the first round of convergence iteration of the eigenvector.

[0038] Example 6: In a multi-level triage scenario, to establish the geometric boundaries of the feature space and suppress numerical overflow during computation, the system executes a global topology normalization calibration procedure, utilizing the feature mapping matrix... Under the action A linear stretching process is performed on the dimensional vector space to limit the magnitude of each feature component to a unit interval. Inside, among which The feature mapping matrix is ​​used; the system sets the initial semantic search radius by calculating the average Euclidean distance between candidate nodes in the medical knowledge graph. , This is the semantic search radius; for nodes with more than [a certain number of connections], [this is the radius of the search]. In high-weight node regions, the system utilizes an automatic step-size adjustment algorithm to adjust the displacement correction. The initial gain coefficient is calibrated as This provides a geometric benchmark with a deterministic step size for the subsequent feature convergence process, where This is the displacement correction amount.

[0039] When the interactive feedback data received by the guidance instruction generation module remains in the fuzzy range and the feedback confidence coefficient is... Below the preset smoothing threshold At that time, the system initiates a logical isolation identification procedure to cope with non-random interference in the interaction channel, in which... To provide feedback confidence coefficients, the system will use the current semantic feature vector to be fused. Stored in a logically isolated buffer and marked as pending review. The semantic feature vector drives the guidance instruction generation module to extract secondary difference attributes orthogonal to the variation dimension of the preceding attributes, generating multidimensional composite probing instructions for cross-validation; if the topological conflict entropy after secondary probing... Still at the difference threshold The system automatically switches to a flexible alignment method to drive the semantic feature vectors to be fused. Topological centrality within nearest neighbor The highest stable node performs a small displacement, allowing the eigenvectors to reach a local energy equilibrium state without disrupting the stability of the local structure. For topological conflict entropy, Let be the topological centrality.

[0040] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0041] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. An interactive health science popularization question-and-answer guidance system for constructing a medical knowledge graph, characterized in that, include: The guidance instruction generation module is used to calculate the attribute entropy difference between the data to be aligned and the candidate nodes in the medical knowledge graph on the semantic attribute set, extract the attribute dimension with the largest attribute entropy difference as the attribute variation dimension, and generate attribute guidance instructions. The feedback parsing module is used to obtain interactive feedback data in response to attribute guidance instructions, and convert the interactive feedback data into feedback correction vectors in the feature space based on a preset feature mapping matrix. The graph fusion executor, connected to the guidance instruction generation module and the feedback parsing module, determines the initial fusion weights based on the coordinate distances between the feature vectors of the data to be aligned and each candidate node. When performing feature fusion, the graph fusion executor extracts the topological centrality of the target node in the medical knowledge graph and determines the corresponding weight correction coefficient based on the topological centrality. The graph fusion executor uses the weight correction coefficient to numerically adjust the initial fusion weights and establishes a weighted feature update function composed of the feedback correction vector and the local topological connectivity of the medical knowledge graph. By calculating the feature increment of the feature vector under the weighted feature update function, the graph fusion executor enables the feature vector to converge toward the target node under the combined effect of the feature orientation value generated by the interactive feedback data and the structured prior constraint value generated by the medical knowledge graph.

2. The interactive health science popularization question-and-answer guidance system for constructing a medical knowledge graph according to claim 1, characterized in that, The graph fusion executor is also used to extract subgraphs representing the logical relationships of medical ontology based on the associated topological paths of candidate nodes in the medical knowledge graph. The graph fusion executor inputs the feedback correction vector into the subgraph, calculates the association conflict entropy between the feedback correction vector and the subgraph topology, calculates the confidence weight of the feedback correction vector based on the association conflict entropy, and uses the confidence weight to perform bias compensation on the update direction of the feature vector.

3. The interactive health science popularization question-and-answer guidance system for constructing a medical knowledge graph according to claim 1, characterized in that, The guidance instruction generation module is specifically used to identify feature dimensions whose attribute entropy difference exceeds a preset deviation threshold, and to use the feature dimension as the attribute variation dimension.

4. The interactive health science popularization question-and-answer guidance system for constructing a medical knowledge graph according to claim 1, characterized in that, When the graph fusion executor calculates the weight correction coefficient, it obtains the hierarchical depth value of the target node in the medical knowledge graph and uses a preset positive correlation mapping function to calculate the weight correction coefficient. By increasing the weight distribution of deep nodes in the fusion calculation, the global update stability of the medical knowledge graph is maintained.

5. The interactive health science popularization question-and-answer guidance system for constructing a medical knowledge graph according to claim 1, characterized in that, The system also includes a dynamic graph maintenance module, which is connected to the graph fusion executor. After feature convergence is completed, the target node is marked as a new topological anchor point to perform incremental updates, and the topological centrality of the affected local region is updated in real time.

6. The interactive health science popularization question-and-answer guidance system for constructing a medical knowledge graph according to claim 1, characterized in that, When the feedback parsing module generates the feedback correction vector, it performs semantic projection processing on the unstructured interactive feedback data and calculates the projection intensity of the interactive feedback data on each orthogonal basis vector in the feature space.

7. An interactive health science popularization question-and-answer guidance system for constructing a medical knowledge graph according to claim 1, characterized in that, The graph fusion executor is used to iteratively update the feature vector according to a preset update cycle, and within each update cycle, it calculates the feature fusion correction amount at the current time based on the coordinate distance and weight correction coefficient at the current time.

8. An interactive health science popularization question-and-answer guidance system for constructing a medical knowledge graph according to claim 1, characterized in that, The graph fusion executor is also used to send a conflict trigger signal to the guidance instruction generation module when the association conflict entropy exceeds a preset logical difference threshold, so that the guidance instruction generation module can regenerate the attribute guidance instruction for another attribute variation dimension.

9. An interactive health science popularization question-and-answer guidance system for constructing a medical knowledge graph according to claim 1, characterized in that, The system also includes an interactive guidance interface, connected to the guidance instruction generation module, used to convert attribute guidance instructions into natural language guidance text and output it to the interactive terminal.

10. An interactive health science popularization question-and-answer guidance system for constructing a medical knowledge graph according to claim 1, characterized in that, The graph fusion actuator computes the semantic feature bias values ​​representing semantic alignment conflicts. When doing so, the following quantitative calculation rules shall be followed: ,in, These are partial values ​​for semantic features; The semantic distance between the feature vector and the target node coordinates; Let be the topological centrality of the target node.

Images