A knowledge base retrieval recall quantity and matching degree dynamic regulation system

Abstract

This invention relates to the field of knowledge base retrieval technology and discloses a dynamic control system for the recall and matching degree of knowledge base retrieval, including a change aggregation module, a density prediction module, a parameter pre-generation module, a query processing module, a fusion decision module, and a result output module. The change aggregation module statistically analyzes the change quantification records of each semantic cluster; the density prediction module calculates the expected knowledge density feature vector; the parameter pre-generation module inputs the expected knowledge density feature vector into a joint decision model and writes the corrected parameters into a regional parameter cache table; the fusion decision module weights and fuses the real-time and cached density feature vectors based on time decay to output the final retrieval parameters.

Description

Technical Field

[0001] This invention relates to the field of knowledge base retrieval technology, and more specifically, to a dynamic control system for the recall quantity and matching degree of knowledge base retrieval. Background Technology

[0002] In a knowledge base question-answering scenario based on semantic vector retrieval, the system retrieves relevant fragments from the knowledge base based on the semantic vector of the user's query, and filters and outputs the search results according to the matching degree threshold and recall parameters. As the knowledge base content is continuously updated with business development, new documents are added in batches, and old documents are deleted or revised, resulting in a dynamic change in the fragment density of local semantic regions of the knowledge base.

[0003] Existing technologies typically employ joint decision-making models trained on historical samples, which output matching thresholds and recall parameters based on knowledge density feature vectors calculated in real time, in order to achieve adaptive adjustment of retrieval parameters.

[0004] However, when a semantic region experiences a large-scale increase or decrease in fragments within a short period, the joint decision-making model must accumulate a sufficient number of new training samples before it can naturally adapt to the new density distribution. There is a lag window before this accumulation process is complete. During this lag window, the historical training distribution of the joint decision-making model does not match the current actual density distribution. The system continuously outputs recall numbers and matching threshold parameters that do not conform to the actual density, leading to a temporary decline in retrieval quality. Summary of the Invention

[0005] This invention provides a dynamic control system for the recall quantity and matching degree of knowledge base retrieval, which solves the technical problem in related technologies that the retrieval parameters cannot be adaptively adjusted in a timely manner after the knowledge base content changes, resulting in a mismatch between the recall quantity and matching degree thresholds and the actual knowledge density distribution.

[0006] This invention discloses a dynamic control system for knowledge base retrieval recall quantity and matching degree, comprising: The change aggregation module continuously monitors the document change event stream of the knowledge base, extracts the document vectors involved in each change operation and their corresponding semantic clustering identifiers, and counts the number of new, deleted and modified fragments for each semantic cluster within a preset time window, generating a quantitative statistical record of changes for each semantic cluster. The density prediction module filters the change quantification statistics records, retains semantic clusters whose change volume exceeds the preset significance threshold, and calculates the expected knowledge density feature vector of each significant change semantic cluster by combining historical density baseline parameters and change volume data through the density change prediction function. The parameter pre-generation module inputs the expected knowledge density feature vector into the joint decision-making model, pre-calculates the corrected matching degree threshold and corrected recall quantity parameters corresponding to the affected semantic clusters, and writes them into the regional parameter cache table with the semantic cluster identifier as the index. The query processing module obtains the user's query text, generates a query semantic vector and performs an initial retrieval, obtains a candidate fragment set and a matching score distribution, and determines the semantic clustering identifier of the hit; The fusion decision module queries the regional parameter cache table. If the semantic clustering identifier has a corrected parameter record, it calculates the fusion weight based on the time decay coefficient and the proportion of newly accumulated training samples. It performs weighted fusion on the real-time calculated knowledge density feature vector and the cached expected knowledge density feature vector, and inputs the fusion density feature vector into the joint decision model to obtain the final parameters. If there is no record of corrected parameters, the real-time knowledge density feature vector is directly input into the joint decision model to obtain the final parameters. The results output module filters and reorders the candidate fragment set according to the final matching degree threshold and recall quantity parameters, outputs the retrieval results, and records the actual knowledge density feature vector and decision effect of this query as new training samples, which are then added to the sample accumulation pool of the hit semantic cluster. When the accumulation quantity reaches the sufficient coverage threshold, the correction parameter record of the semantic cluster in the region parameter cache table is cleared.

[0007] Furthermore, the historical density baseline parameters include the number of historical effective candidate segments, the historical average semantic distance, and historical distribution morphology parameters; the density change prediction function maps the increase in the number of segments to three components: the increase in the number of effective candidate segments, the reduction rate of the average semantic distance, and the shift in distribution morphology. These three components are respectively superimposed on the corresponding components of the historical density baseline parameters to form the expected knowledge density feature vector. : ; in, The number of valid candidate segments in history. This represents an increase in the number of valid candidate fragments. The historical average semantic distance. This represents the average semantic distance reduction rate. These are historical distribution morphological parameters. This represents the offset of the distribution pattern.

[0008] Furthermore, the number of effective candidate fragments increased. The average semantic distance reduction rate is determined by the product of the net increase in fragments and the efficiency coefficient. The net increase in fragments is the number of new additions minus the number of deletions, and the efficiency coefficient is the average proportion of valid candidate fragments in the historical samples to the total number of fragments. The offset is determined by multiplying the difference between the average semantic distance of the newly added segments and the distance to the cluster centers by a new segment ratio coefficient, where the new segment ratio coefficient is the ratio of the number of newly added segments to the total number of segments after the change; the distribution pattern offset is... The new proportion is determined by multiplying the difference between the estimated semantic distribution skewness of the new segment and the historical distribution morphological parameters by the new proportion coefficient.

[0009] Furthermore, the joint decision-making model is a multi-output regression model trained with historical samples. Its input layer receives a three-dimensional knowledge density feature vector, and its output layer outputs the joint predicted value of the matching degree threshold and the recall quantity parameter. It adopts the mean squared error loss function, uses the actual density feature vector in the historical query and the corresponding feedback confirmation parameter combination as training samples, and uses the Adam optimization algorithm for training.

[0010] Furthermore, the fusion weights in the fusion decision module The calculation formula is: ; in, This represents the elapsed time from the time the change occurred to the current time. The time decay rate, The time decay coefficient, Represents an exponential function; This represents the number of new training samples accumulated for this semantic clustering. To fully cover the threshold; and For the weighting coefficients, satisfying ; The range of values ​​is , The larger the value, the higher the dependence on the cached expected knowledge density feature vector.

[0011] Furthermore, the fusion decision module is based on fusion weights. Knowledge density feature vectors for real-time computing The expected knowledge density feature vector of the cache Perform weighted fusion to obtain the fusion density feature vector. : ; in, This refers to the knowledge density feature vector calculated in real time based on the matching score distribution of the candidate fragment set, and... With the same dimension, each component corresponds to the number of valid candidate segments, the average semantic distance, and the distribution morphology parameters, respectively.

[0012] Furthermore, the aforementioned The three components are calculated in real time from the matching score distribution of the candidate fragment set in the following ways: the number of effective candidate fragments component is the number of candidate fragments whose matching score exceeds the preset validity threshold; the average semantic distance component is the mean of the Euclidean distance between the semantic vector of each candidate fragment and the query semantic vector; and the distribution morphology parameter component is the skewness statistic of the candidate fragment matching score sequence.

[0013] Furthermore, when the accumulated number reaches the full coverage threshold, the result output module clears the corresponding semantic clustering correction parameter record from the regional parameter cache table, marking that the joint decision model has completed the natural adaptation to the new density distribution of the region, and subsequent queries are restored to the decision path of directly inputting the real-time knowledge density feature vector into the joint decision model.

[0014] Furthermore, the change aggregation module extracts the document vectors involved in each change operation, categorizes and aggregates the change events according to semantic clustering identifiers, and counts the number of newly added, deleted, and modified fragments for each semantic cluster within a preset time window.

[0015] This invention discloses a method for dynamically adjusting the recall and matching degree of knowledge base retrieval, used to execute the aforementioned dynamic adjustment system for the recall and matching degree of knowledge base retrieval. This includes: Step 1, continuously monitoring the document change event stream of the knowledge base, extracting the document vectors and semantic clustering identifiers involved in the change operations, and counting the number of new, deleted, and modified fragments of each semantic cluster within a preset time window to generate a quantitative change statistics record; Step 2: Filter semantic clusters whose change amount exceeds the preset significance threshold, and calculate the expected knowledge density feature vector of each significant change semantic cluster by combining historical density baseline parameters and change amount data through the density change prediction function. Step 3: Input the expected knowledge density feature vector into the joint decision model, pre-calculate the corrected matching degree threshold and corrected recall quantity parameters, and write them into the region parameter cache table using the semantic clustering identifier as the index. Step 4: Obtain the user query text, generate the query semantic vector and perform the initial retrieval, obtain the candidate fragment set and matching score distribution, and determine the semantic clustering identifier of the hit; Step 5: Query the regional parameter cache table. If a corrected parameter record exists for the matched semantic clustering identifier, calculate the fusion weight and perform a weighted fusion of the real-time knowledge density feature vector and the cached expected knowledge density feature vector. Input the fused density feature vector into the joint decision model to obtain the final parameters. If no corrected parameter record exists, directly input the real-time knowledge density feature vector into the joint decision model to obtain the final parameters. Step 6: Filter and reorder the candidate fragment set according to the final matching degree threshold and recall quantity parameter, output the retrieval results, record the actual knowledge density feature vector and decision effect of this query as new training samples, and add them to the sample accumulation pool of the hit semantic cluster. When the accumulation quantity reaches the sufficient coverage threshold, clear the correction parameter record of the corresponding semantic cluster in the region parameter cache table.

[0016] The beneficial effects of this invention are as follows: This invention proactively counts the changes in each semantic cluster when a change event occurs through a change aggregation module. A density prediction module predicts the density changes of the affected semantic regions and generates expected knowledge density feature vectors. Then, a parameter pre-generation module writes the corrected parameters into the regional parameter cache table before the joint decision-making model completes incremental adaptation. This solves the technical problems of lag window period and inaccurate recall and matching degree threshold parameters in the joint decision-making model after large-scale knowledge base updates. It achieves the technical effect of obtaining retrieval parameters that match the actual density within the lag window period and smoothly transitioning back to a pure data-driven decision-making mode through time decay weighted fusion as new samples accumulate. Attached Figure Description

[0017] Figure 1 This is a flowchart of the method for dynamically adjusting the recall quantity and matching degree of knowledge base retrieval provided in the embodiments of the present invention; Figure 2 This is a sub-flowchart of the dynamic control method for knowledge base retrieval recall quantity and matching degree provided in the embodiments of the present invention for calculating the expected knowledge density feature vector for clustering with significant changes; Figure 3 This is a sub-flowchart of the knowledge base retrieval recall quantity and matching degree dynamic control method provided in this embodiment of the invention, which calculates the final matching degree threshold and recall quantity parameters based on the hit status of the regional parameter cache table. Detailed Implementation

[0018] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0019] At least one embodiment of the present invention discloses a dynamic control system for knowledge base retrieval recall quantity and matching degree, the method of which is described below. Figures 1-3 The system includes: The change aggregation module is configured to acquire document change event streams and generate quantitative statistical records of changes for each semantic cluster: The change aggregation module continuously monitors the document change event stream of the knowledge base, extracts the document vectors involved in each change operation and their corresponding semantic clustering identifiers, and categorizes and aggregates the change events according to the semantic clustering identifiers. Within a preset time window, it counts the number of newly added, deleted, and modified fragments for each semantic cluster, generating quantitative change statistics records for each semantic cluster.

[0020] The density prediction module is configured to calculate the expected knowledge density feature vector for clustering with significant changes. The density prediction module filters the change quantification statistics records and retains semantic clusters whose change amount exceeds the preset significance threshold. For each semantic cluster that meets the conditions, the following processing is performed.

[0021] Obtain the historical density baseline parameters of the semantic cluster before the change, including the number of historical valid candidate segments, historical average semantic distance, and historical distribution morphology parameters.

[0022] Based on the current statistical change data, the expected knowledge density features after the changes are calculated using a density change prediction function. The density change prediction function maps the increase in fragment additions to three components: the increase in the number of valid candidate fragments, the average semantic distance reduction rate, and the distribution pattern shift. These three components are then superimposed on the corresponding components of the historical density baseline parameters to form the expected knowledge density feature vector for this semantic clustering. Its definition is: ; in, This represents the number of valid candidate segments in history. This represents an increase in the number of valid candidate fragments; The historical average semantic distance. This represents the average semantic distance reduction rate. These are historical distribution morphological parameters. This represents the offset of the distribution pattern.

[0023] Furthermore, the three components of the density change prediction function are calculated as follows. Increase in the number of valid candidate fragments. The average semantic distance reduction rate is determined by the product of the net increase in fragments and the efficiency coefficient, where the net increase in fragments is the number of new fragments minus the number of deleted fragments, and the efficiency coefficient is the average proportion of valid candidate fragments in the historical samples to the total number of fragments. The offset is determined by multiplying the difference between the average semantic distance of the newly added segments and the distance to the cluster centers by a scaling factor, which is defined as the ratio of the number of newly added segments to the total number of segments after the change. Distribution morphology offset. The new proportion is determined by multiplying the difference between the estimated semantic distribution skewness of the new segment and the historical distribution morphological parameters by the new proportion coefficient.

[0024] The parameter pre-generation module is configured to input the expected knowledge density feature vector into the joint decision model, generate corrected parameters, and write them to the regional parameter cache table. The parameter pre-generation module will generate the expected knowledge density feature vector. The input joint decision model pre-calculates the corrected matching threshold and corrected recall parameters corresponding to the affected semantic clustering regions. The joint decision model is a multi-output regression model trained on historical samples. Its input layer receives a three-dimensional knowledge density feature vector, and its output layer outputs the joint predicted value of the matching threshold and recall parameters. A mean squared error loss function is used, and the training samples are the actual density feature vectors from historical queries combined with corresponding manually labeled or confirmed parameters. The Adam optimization algorithm is used for training. The corrected parameters are stored in a region parameter cache table indexed by the semantic clustering identifier for retrieval in subsequent query stages.

[0025] The query processing module is configured to obtain the user's query text, generate a query semantic vector, perform an initial retrieval, and determine the semantic clustering identifier that matches the query. The query processing module obtains the user's query text, generates a query semantic vector through semantic encoding, performs an initial retrieval on the knowledge base, obtains a set of candidate segments and their matching score distribution, and determines the main semantic clustering identifiers that are matched in this query.

[0026] The fusion decision module is configured to calculate the final matching threshold and recall quantity parameters based on the hit status of the regional parameter cache table. The fusion decision module queries the regional parameter cache table to determine whether there is a corrected parameter record for the hit semantic clustering identifier, and processes it according to the following two paths respectively.

[0027] If the matched semantic clustering identifier has a corrected parameter record in the region parameter cache table, then the fusion weights are calculated, weighted fusion is performed, and finally, the final parameters are output through the joint decision model. Fusion weights The time decay coefficient is determined by both the time decay coefficient and the proportion of newly accumulated training samples, and the calculation formula is as follows: ; in, This represents the elapsed time from the time the change occurred to the current time. The time decay rate, The time decay coefficient, Represents an exponential function; This represents the number of new training samples accumulated for this semantic clustering. To fully cover the threshold, The proportion of newly accumulated training samples; and For the weighting coefficients, satisfying . The range of values ​​is , The larger the value, the higher the dependence on the cached expected knowledge density feature vector.

[0028] Based on fusion weight Knowledge density feature vectors for real-time computing Compared with the expected knowledge density feature vector in the cache Perform weighted fusion to obtain the fusion density feature vector. : ; in, This refers to the knowledge density feature vector calculated in real time based on the matching score distribution of the candidate fragment set, and... They are all three-dimensional vectors with the same dimension, and each component corresponds to the number of valid candidate segments, the average semantic distance, and the distribution morphology parameter, respectively. The three components are calculated in real time from the matching score distribution of the candidate segment set in the following ways: the effective candidate segment quantity component is the number of candidate segments whose matching score exceeds a preset validity threshold; the average semantic distance component is the mean of the Euclidean distance between the semantic vector of each candidate segment and the query semantic vector; and the distribution morphology parameter component is the skewness statistic of the candidate segment matching score sequence. The fused density feature vector is then used. Input the joint decision model to obtain the final matching threshold and recall parameters.

[0029] If the semantic clustering identifier does not have a correction parameter record in the region parameter cache table, then the knowledge density feature vector is directly passed. Input the joint decision model to obtain the final matching threshold and recall parameters.

[0030] The results output module is configured to filter candidate fragments according to the final parameters and output the search results, while updating the training samples and managing the region parameter cache table. The output module filters and re-ranks the candidate fragment set based on the final matching threshold and recall parameters, and outputs the retrieval results. Simultaneously, the actual knowledge density feature vector and corresponding decision performance of this query are recorded as new training samples and added to the sample accumulation pool of the hit semantic clusters. The number of new training samples accumulated by the affected semantic clusters... Reaching the full coverage threshold At that time, the correction parameter record for the semantic clustering is cleared from the regional parameter cache table, marking that the joint decision model has completed the natural adaptation to the new density distribution of the region, and subsequent queries are restored to the pure data-driven decision path.

[0031] The fundamental reason why existing retrieval systems experience inaccuracies in recall and matching threshold parameters after large-scale knowledge base updates lies in the lag window between the historical training distribution of the joint decision-making model and the current actual density distribution. The joint decision-making model must accumulate sufficient new samples before it can naturally adapt to the new density distribution. This implementation addresses this by having a change aggregation module proactively count changes when a change event occurs, a density prediction module predicting density changes in the affected semantic regions and generating expected knowledge density feature vectors, and a parameter pre-generation module writing corrected parameters into the region parameter cache table before the joint decision-making model completes incremental learning. This ensures that the retrieval system can still obtain recall and matching threshold parameters that match the actual density within the lag window, directly eliminating the negative impact of this window on retrieval quality. The time-decay weighted fusion of the fusion decision-making module gradually increases the fusion weight as new samples accumulate. By continuously reducing the number of data points, the system smoothly transitions back to a pure data-driven decision-making mode, avoiding abrupt conflicts between the corrected parameters and the natural learning results of the joint decision-making model. This enables continuous and stable control of the recall quantity and matching degree in scenarios with dynamic knowledge base updates.

[0032] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A dynamic control system for knowledge base retrieval recall quantity and matching degree, characterized in that, include: The change aggregation module continuously monitors the document change event stream of the knowledge base, extracts the document vectors involved in each change operation and their corresponding semantic clustering identifiers, and counts the number of new, deleted and modified fragments for each semantic cluster within a preset time window, generating a quantitative statistical record of changes for each semantic cluster. The density prediction module filters the change quantification statistics records, retains semantic clusters whose change volume exceeds the preset significance threshold, and calculates the expected knowledge density feature vector of each significant change semantic cluster by combining historical density baseline parameters and change volume data through the density change prediction function. The parameter pre-generation module inputs the expected knowledge density feature vector into the joint decision-making model, pre-calculates the corrected matching degree threshold and corrected recall quantity parameters corresponding to the affected semantic clusters, and writes them into the regional parameter cache table with the semantic cluster identifier as the index. The query processing module obtains the user's query text, generates a query semantic vector and performs an initial retrieval, obtains a candidate fragment set and a matching score distribution, and determines the semantic clustering identifier of the hit; The fusion decision module queries the regional parameter cache table. If a corrected parameter record exists for the matched semantic clustering identifier, the fusion weight is calculated based on the time decay coefficient and the proportion of newly accumulated training samples. The real-time calculated knowledge density feature vector and the cached expected knowledge density feature vector are then weighted and fused. The fused density feature vector is then input into the joint decision model to obtain the final parameters. If no corrected parameter record exists, the real-time knowledge density feature vector is directly input into the joint decision model to obtain the final parameters. The results output module filters and reorders the candidate fragment set according to the final matching degree threshold and recall quantity parameters, outputs the retrieval results, and records the actual knowledge density feature vector and decision effect of this query as new training samples, which are then added to the sample accumulation pool of the hit semantic cluster. When the accumulation quantity reaches the sufficient coverage threshold, the correction parameter record of the semantic cluster in the region parameter cache table is cleared.

2. The knowledge base retrieval recall and matching degree dynamic control system according to claim 1, characterized in that, The historical density baseline parameters include the number of historical effective candidate segments, the historical average semantic distance, and historical distribution morphology parameters. The density change prediction function maps the increase in the number of segments to three components: the increase in the number of effective candidate segments, the reduction rate of the average semantic distance, and the shift in the distribution morphology. These three components are respectively superimposed on the corresponding components of the historical density baseline parameters to form the expected knowledge density feature vector. : ； in, The number of valid candidate segments in history. This represents an increase in the number of valid candidate fragments. The historical average semantic distance. This represents the average semantic distance reduction rate. These are historical distribution morphological parameters. This represents the offset of the distribution pattern.

3. The knowledge base retrieval recall and matching degree dynamic control system according to claim 2, characterized in that, The increase in the number of valid candidate fragments The average semantic distance reduction rate is determined by the product of the net increase in fragments and the efficiency coefficient. The net increase in fragments is the number of new additions minus the number of deletions, and the efficiency coefficient is the average proportion of valid candidate fragments in the historical samples to the total number of fragments. The offset is determined by multiplying the difference between the average semantic distance of the newly added segments and the distance to the cluster centers by a new segment ratio coefficient, where the new segment ratio coefficient is the ratio of the number of newly added segments to the total number of segments after the change; the distribution pattern offset is... The new proportion is determined by multiplying the difference between the estimated semantic distribution skewness of the new segment and the historical distribution morphological parameters by the new proportion coefficient.

4. The knowledge base retrieval recall and matching degree dynamic control system according to claim 1, characterized in that, The joint decision-making model is a multi-output regression model trained on historical samples. Its input layer receives a three-dimensional knowledge density feature vector, and its output layer outputs the joint predicted value of the matching degree threshold and the recall quantity parameter. It adopts the mean squared error loss function, uses the actual density feature vector in historical queries and the corresponding feedback confirmation parameter combination as training samples, and uses the Adam optimization algorithm for training.

5. The knowledge base retrieval recall and matching degree dynamic control system according to claim 1, characterized in that, Fusion weights in the fusion decision module The calculation formula is: ； in, This represents the elapsed time from the time the change occurred to the current time. The time decay rate, The time decay coefficient, Represents an exponential function; This represents the number of new training samples accumulated for this semantic clustering. To fully cover the threshold; and For the weighting coefficients, satisfying ; The range of values ​​is , The larger the value, the higher the dependence on the cached expected knowledge density feature vector.

6. The knowledge base retrieval recall and matching degree dynamic control system according to claim 5, characterized in that, The fusion decision module is based on fusion weights. Knowledge density feature vectors for real-time computing The expected knowledge density feature vector of the cache Perform weighted fusion to obtain the fusion density feature vector. : ； in, This refers to the knowledge density feature vector calculated in real time based on the matching score distribution of the candidate fragment set, and... With the same dimension, each component corresponds to the number of valid candidate segments, the average semantic distance, and the distribution morphology parameters, respectively.

7. The knowledge base retrieval recall and matching degree dynamic control system according to claim 6, characterized in that, The The three components are calculated in real time from the matching score distribution of the candidate fragment set in the following ways: the number of effective candidate fragments component is the number of candidate fragments whose matching score exceeds the preset validity threshold; the average semantic distance component is the mean of the Euclidean distance between the semantic vector of each candidate fragment and the query semantic vector; and the distribution morphology parameter component is the skewness statistic of the candidate fragment matching score sequence.

8. The knowledge base retrieval recall and matching degree dynamic control system according to claim 1, characterized in that, When the accumulated number reaches the full coverage threshold, the result output module clears the corresponding semantic clustering correction parameter record from the regional parameter cache table, marking that the joint decision model has completed the natural adaptation to the new density distribution of the region, and subsequent queries are restored to the decision path of directly inputting the real-time knowledge density feature vector into the joint decision model.

9. The knowledge base retrieval recall and matching degree dynamic control system according to claim 1, characterized in that, The change aggregation module extracts the document vectors involved in each change operation, categorizes and aggregates the change events according to semantic clustering identifiers, and counts the number of newly added, deleted, and modified fragments for each semantic cluster within a preset time window.

10. A method for dynamically adjusting the recall and matching degree of a knowledge base retrieval, used to execute the dynamic adjustment system for the recall and matching degree of a knowledge base retrieval as described in any one of claims 1 to 9, characterized in that, include: Step 1: Continuously monitor the document change event stream of the knowledge base, extract the document vectors and semantic clustering identifiers involved in the change operations, and count the number of new, deleted and modified fragments of each semantic cluster within a preset time window to generate a quantitative change statistics record. Step 2: Filter semantic clusters whose change amount exceeds the preset significance threshold, and calculate the expected knowledge density feature vector of each significant change semantic cluster by combining historical density baseline parameters and change amount data through the density change prediction function. Step 3: Input the expected knowledge density feature vector into the joint decision model, pre-calculate the corrected matching degree threshold and corrected recall quantity parameters, and write them into the region parameter cache table using the semantic clustering identifier as the index. Step 4: Obtain the user query text, generate the query semantic vector and perform the initial retrieval, obtain the candidate fragment set and matching score distribution, and determine the semantic clustering identifier of the hit; Step 5: Query the regional parameter cache table. If a corrected parameter record exists for the matched semantic clustering identifier, calculate the fusion weight and perform a weighted fusion of the real-time knowledge density feature vector and the cached expected knowledge density feature vector. Input the fused density feature vector into the joint decision model to obtain the final parameters. If no corrected parameter record exists, directly input the real-time knowledge density feature vector into the joint decision model to obtain the final parameters. Step 6: Filter and reorder the candidate fragment set according to the final matching degree threshold and recall quantity parameter, output the retrieval results, record the actual knowledge density feature vector and decision effect of this query as new training samples, and add them to the sample accumulation pool of the hit semantic cluster. When the accumulation quantity reaches the sufficient coverage threshold, clear the correction parameter record of the corresponding semantic cluster in the region parameter cache table.

Images