A method for dynamic updating and real-time retrieval optimization of multimodal knowledge base

Abstract

This invention discloses a method for dynamic updating and real-time retrieval optimization of a multimodal knowledge base, relating to the fields of knowledge engineering, artificial intelligence, and information retrieval. The method includes: mapping heterogeneous data to a vector space of the same dimension, performing hierarchical clustering to generate knowledge clusters; evaluating the consolidation score for each cluster, extracting feature vectors from new data, matching them with the generated knowledge clusters, and calculating prediction errors; triggering a three-way update routing strategy based on the consolidation value and prediction error value of each cluster, and outputting the updated knowledge clusters and knowledge graph; constructing an LSM-tree index and a scene-level hierarchical index, and outputting a query interface supporting real-time retrieval. This invention significantly improves the accuracy of contradictory information identification and the recall rate of new knowledge, effectively suppresses noise interference, reduces query latency and memory usage, and enhances the adaptive evolution capability of the knowledge base.

Description

Technical Field

[0001] This invention relates to the fields of knowledge engineering, artificial intelligence and information retrieval technology, and in particular to a method for dynamic updating and real-time retrieval optimization of a multimodal knowledge base. Background Technology

[0002] With the rapid development of artificial intelligence technology, multimodal knowledge bases are increasingly widely used in fields such as intelligent customer service, medical diagnosis, and video retrieval. In practical applications, knowledge bases need to continuously receive and integrate new information from multiple modalities such as text, images, and audio to achieve dynamic updates and evolution of knowledge. However, during the dynamic update process, accurately distinguishing between real knowledge evolution and temporary noise interference, and maintaining the stability of existing knowledge structures while absorbing new knowledge, have become the core contradictions restricting the performance improvement of knowledge bases. When faced with new information that contradicts existing knowledge, existing systems often adopt simple overwrite, rejection, or uniform learning rate update strategies, making it difficult to achieve a balance between knowledge stability and adaptability. This results in systems that are either overly sensitive to noise or slow to respond to real knowledge evolution.

[0003] To address the dynamic updating problem of multimodal knowledge bases, existing technologies propose a dynamically updated intelligent customer service knowledge base system capable of recognizing multimodal inputs and performing knowledge confusion detection. However, its update mechanism lacks a quantitative assessment of knowledge consolidation, cannot distinguish the degree of conflict between stable and newly added knowledge, and uses fixed TF-IDF and TextCNN models for confusion detection without considering the historical access frequency and stability weights of knowledge points, resulting in excessive sensitivity of the knowledge base to noisy data. Existing technologies also propose a RAG-based retrieval optimization method to improve retrieval efficiency through title enhancement and intent recognition. However, its scenario knowledge base division is based on a static high-frequency word index, which cannot adapt to the continuous changes in knowledge structure and concept drift. Furthermore, the retrieval process relies solely on cosine similarity for matching, ignoring the temporal evolution characteristics and contextual dependencies of knowledge, and lacks an effective update and elimination mechanism when dealing with contradictory or outdated information. In summary, existing technologies generally suffer from deficiencies such as a lack of knowledge consolidation assessment, a single strategy for handling contradictory information, and insufficient adaptability to knowledge evolution. Summary of the Invention

[0004] In view of the above-mentioned problems, the present invention provides a method for dynamic updating and real-time retrieval optimization of a multimodal knowledge base.

[0005] Therefore, the technical problem solved by this invention is: how to accurately identify and process information that contradicts the existing knowledge structure while continuously receiving new information, so as to achieve adaptive evolution of the knowledge base.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a method for dynamic updating and real-time retrieval optimization of a multimodal knowledge base, including: The original multimodal data is vectorized to map the heterogeneous data to a vector space of the same dimension, resulting in feature vectors of the same dimension. Hierarchical clustering is performed based on feature vectors of a unified dimension to generate knowledge clusters and corresponding cluster center vectors and covariance matrices. Based on the generated knowledge clusters, a consolidation score is evaluated for each cluster, and the consolidation value of each cluster is output. Extract the feature vectors of the new data and match them with the generated knowledge clusters. Calculate the prediction error and output the nearest neighbor cluster and the prediction error value. Based on the consolidation value and prediction error value of each cluster, a three-way update routing strategy is triggered to output the updated knowledge clusters and knowledge graph; Based on the updated knowledge clusters and knowledge graph, an LSM-tree index and a scenario-based hierarchical index are constructed, and a query interface that supports real-time retrieval is output.

[0007] As a preferred solution for dynamic updating and real-time retrieval optimization methods of multimodal knowledge bases, the following is provided: The vectorization process of the original multimodal data, which maps heterogeneous data to a vector space of the same dimension to obtain feature vectors of the same dimension, includes: Obtain the raw multimodal data and extract feature vectors for each modality: For text data, a pre-trained language model is used to extract word vectors, and each document is represented as a feature matrix. For image data, a convolutional neural network is used to extract image feature vectors; for video data, it is represented as an image vector sequence matrix. For audio data, acoustic features are extracted using an acoustic feature extraction model to form a speech feature matrix; The feature matrices or feature vectors obtained from each modality are uniformly mapped to a preset vector dimension space.

[0008] As a preferred solution for dynamic updating and real-time retrieval optimization methods of multimodal knowledge bases, the following is provided: The hierarchical clustering based on feature vectors of a unified dimension to generate knowledge clusters and corresponding cluster center vectors and covariance matrices includes: Hierarchical clustering is performed based on feature vectors of a unified dimension to construct a knowledge graph structure and obtain multiple knowledge clusters. For each knowledge cluster, the weighted average of all vectors within the cluster is calculated as the cluster center vector of the knowledge cluster, where the weight coefficient of each vector is used to characterize the importance of the vector within the cluster; Calculate and store the covariance matrix for each knowledge cluster.

[0009] As a preferred solution for dynamic updating and real-time retrieval optimization methods of multimodal knowledge bases, the following is provided: The generated knowledge clusters are used to evaluate the consolidation score for each cluster, and the consolidation value for each cluster is output, including: Obtain the generated knowledge clusters, each containing multiple vectorized data samples; A consolidation score is maintained for each knowledge cluster. The consolidation score is obtained by weighted summation after nonlinear transformation of historical access frequency, existence duration and internal stability. This results in knowledge clusters with higher historical access frequency, existence duration and stronger internal stability receiving higher consolidation scores.

[0010] As a preferred solution for dynamic updating and real-time retrieval optimization methods of multimodal knowledge bases, the following is provided: The process of extracting feature vectors from new data, matching them with generated knowledge clusters, calculating prediction errors, and outputting nearest neighbor clusters and prediction error values ​​includes: Feature vectors of new data are extracted and matched with knowledge clusters in the existing knowledge base. The comprehensive similarity between the feature vectors and each knowledge cluster is calculated. The comprehensive similarity is obtained by weighted summation of cosine similarity and Mahalanobis distance-based similarity, where Mahalanobis distance is calculated using the covariance matrix of each knowledge cluster.

[0011] As a preferred solution for dynamic updating and real-time retrieval optimization methods of multimodal knowledge bases, the following is provided: The process of extracting feature vectors from new data, matching them with generated knowledge clusters, calculating prediction errors, and outputting nearest neighbor clusters and prediction error values ​​also includes: Select the knowledge cluster with the highest comprehensive similarity as the nearest neighbor cluster; The prediction error is calculated based on the comprehensive similarity of the nearest neighbor clusters. The prediction error ranges from 0 to 1. The larger the prediction error, the greater the deviation between the new data and the existing knowledge base.

[0012] As a preferred solution for dynamic updating and real-time retrieval optimization methods of multimodal knowledge bases, the following is provided: The three-way update routing strategy, based on the consolidation value and prediction error value of each cluster, is triggered to output the updated knowledge clusters and knowledge graph, including: In response to the prediction error value being lower than the first threshold, it is determined that the new data is consistent with the existing knowledge, and a smooth incremental update is performed: with a learning rate that is negatively correlated with the consolidation value, the cluster centers of the nearest neighbor clusters are adjusted towards the feature vector direction of the new data, and the consolidation value of the corresponding cluster is increased by a fixed increment. If the prediction error exceeds the second threshold and the consolidation value exceeds the third threshold, it is determined that the new data represents the true knowledge evolution, and a cluster splitting operation is performed: a new cluster is created with the feature vector of the new data as the center, an initial consolidation value is assigned, and an evolutionary relationship edge is established between the original nearest neighbor cluster and the new cluster in the knowledge graph; if the prediction error exceeds the fourth threshold, the cluster center of the original nearest neighbor cluster is adjusted at a high learning rate.

[0013] As a preferred solution for dynamic updating and real-time retrieval optimization methods of multimodal knowledge bases, the following is provided: The process of triggering a three-way update routing strategy based on the consolidation value and prediction error value of each cluster, and outputting the updated knowledge clusters and knowledge graph, also includes: If the prediction error is higher than the second threshold and the consolidation value is lower than the third threshold, the new data is determined to be potential noise, and a conservative strategy is implemented: the update is rejected or only updated at a very low learning rate, and the new data is marked as a state to be verified and stored in a temporary buffer; if multiple consecutive inputs produce prediction errors higher than the second threshold, a re-evaluation is triggered.

[0014] As a preferred solution for dynamic updating and real-time retrieval optimization methods of multimodal knowledge bases, the following is provided: Based on the updated knowledge clusters and knowledge graph, an LSM-tree index and a scenario-level hierarchical index are constructed, outputting a query interface that supports real-time retrieval, including: A multi-level inverted index is constructed using a log structure merging tree. Newly inserted feature vectors first enter the smallest level. When the level capacity reaches the upper limit, merging with the next level is triggered. Two types of inverted lists are provided for different vector space sizes.

[0015] As a preferred solution for dynamic updating and real-time retrieval optimization methods of multimodal knowledge bases, the following is provided: The process of constructing an LSM-tree index and a scenario-level hierarchical index based on the updated knowledge clusters and knowledge graph, and outputting a query interface that supports real-time retrieval, also includes: The knowledge base is divided into multiple sub-bases according to scenarios. High-frequency words are extracted for each scenario and quantified into scenario indexes. During a query, the similarity between the query vector and each scenario index is calculated, and the scenario sub-base with the highest matching degree is selected as the search scope, thus narrowing the search space to a single scenario.

[0016] The beneficial effects of this invention are as follows: By introducing knowledge consolidation assessment and prediction error calculation, this invention realizes a cognitive bionic update mechanism for the knowledge base, enabling the system to generate a stronger prediction error response to consolidated knowledge structures, much like the human cognitive system. This allows for more effective identification and absorption of truly new knowledge while suppressing interference from noisy data. A three-way update routing strategy is adopted, dynamically selecting the update mode based on the combination of knowledge consolidation and prediction error. Compared to the single update strategy of existing technologies, this significantly improves the robustness and adaptability of the knowledge base in complex dynamic environments. Combined with the LSM-tree index structure and vector residual compression technology, while ensuring real-time retrieval performance, the accumulation of historical stability weights enables the knowledge base to remember the trajectory of knowledge evolution, providing a technical foundation for the source analysis of contradictory information and the temporal evolution of knowledge graphs. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is an overall flowchart of a dynamic updating and real-time retrieval optimization method for a multimodal knowledge base provided by the present invention. Detailed Implementation

[0019] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0020] Example 1, referring to Figure 1 This is the first embodiment of the present invention, which provides a method for dynamic updating and real-time retrieval optimization of a multimodal knowledge base, including: S1: Vectorize the original multimodal data to map the heterogeneous data to a vector space of the same dimension, thus obtaining a feature vector of the same dimension. S2: Perform hierarchical clustering based on feature vectors of a unified dimension to generate knowledge clusters and corresponding cluster center vectors and covariance matrices; S3: Based on the generated knowledge clusters, evaluate the consolidation score for each cluster and output the consolidation value for each cluster; S4: Extract the feature vector of the new data, match it with the generated knowledge cluster, calculate the prediction error, and output the nearest neighbor cluster and the prediction error value; S5: Based on the consolidation value and prediction error value of each cluster, trigger the three-way update routing strategy and output the updated knowledge clusters and knowledge graph; S6: Based on the updated knowledge clusters and knowledge graph, construct an LSM-tree index and a scenario-level hierarchical index, and output a query interface that supports real-time retrieval.

[0021] It should be noted that through steps S1-S6, driven by the cognitive consolidation mechanism, the stable measurement of knowledge clusters is combined with the dynamic response to prediction errors. This achieves smooth absorption of consistent information, structural splitting of real knowledge evolution, and active suppression of noise interference, significantly improving the robustness and adaptability of the knowledge base in complex dynamic environments. Combined with LSM-tree indexing and scenario-based hierarchical retrieval optimization, query latency and memory consumption are significantly reduced while ensuring high-concurrency write performance. Ultimately, this forms a continuously evolving, real-time responsive, and resource-efficient intelligent knowledge management solution.

[0022] Example 2, refer to Figure 1 As an embodiment of the present invention, based on the previous embodiment, a method for dynamic updating and real-time retrieval optimization of a multimodal knowledge base is provided, including: In this embodiment, step S1 above performs vectorization processing on the original multimodal data, mapping the heterogeneous data to a vector space of the same dimension to obtain a feature vector of the same dimension, including: Heterogeneous multimodal data are uniformly converted into matrix representation.

[0023] Specifically, for text data, pre-trained language models with the Transformer architecture, such as BERT and GPT, are used to extract word vectors, representing each document as a matrix. ,in For word count, For vector dimensions.

[0024] For image data, convolutional neural networks, such as ResNet and ViT, are used to extract image feature vectors, while videos are represented as a sequence matrix of image vectors. , The number of frames.

[0025] For audio data, acoustic features are extracted using MFCC or Wav2Vec to form a speech grid vector matrix. , This represents the number of time steps.

[0026] Map all vectors to the same dimensional space to ensure consistency in cross-modal similarity calculation.

[0027] In this embodiment, step S2 above, which involves hierarchical clustering based on feature vectors of a unified dimension to generate knowledge clusters and corresponding cluster center vectors and covariance matrices, includes: Based on vectorized multimodal data, a hierarchical clustering algorithm is used to construct a knowledge graph structure.

[0028] Each knowledge cluster (Schema) From cluster center vector The representation, calculated as a weighted average of all vectors within the cluster, is expressed as: in, The number of vectors within the cluster. These are weighting coefficients, reflecting the importance of the vector.

[0029] Simultaneously record the covariance matrix of each cluster. This is used for subsequent Mahalanobis distance calculation, improving the accuracy of similarity judgment.

[0030] In this embodiment, in step S3 above, based on the generated knowledge clusters, a consolidation score is evaluated for each cluster, and the consolidation value of each cluster is output, including: For each knowledge cluster Maintain a consolidation score Taking into account three dimensions: (1) Historical visit frequency This records the cumulative number of times the cluster has been retrieved or updated. (2) Duration of existence The time elapsed since the cluster was created, in hours; (3) Internal stability , is defined as the reciprocal of the average distance from the vector within the cluster to the center, and its value ranges from [0,1].

[0031] The formula for calculating consolidation is: in, Representing knowledge clusters The consolidation score, i.e., the consolidation value; This is an adjustable weighting parameter. A higher consolidation score indicates a more stable knowledge cluster and a stronger resistance to interference from new information.

[0032] It should be noted that the weighting parameters in the consolidation calculation formula need to be adjusted according to the application scenario. For rapidly changing fields, such as news and social media, it is recommended that... Emphasis should be placed on frequency of visits rather than cumulative time. For stable fields, such as law and medicine, it is recommended... We should emphasize the long-term stability of knowledge.

[0033] In this embodiment, the feature vector of the new data is extracted in step S4 above, matched with the generated knowledge cluster, the prediction error is calculated, and the nearest neighbor cluster and prediction error value are output, including: When new multimodal data Upon input, its feature vector is first extracted. .

[0034] Retrieve nearest neighbor knowledge clusters from the existing knowledge base. Matching is performed using a weighted combination of cosine similarity and Mahalanobis distance, expressed as: in , is the balance coefficient.

[0035] The prediction error is defined as: Prediction error The range of values ​​is The larger the value, the greater the deviation between the new information and existing knowledge.

[0036] In this embodiment, step S5 above triggers a three-way update routing strategy based on the consolidation value and prediction error value of each cluster, and outputs the updated knowledge clusters and knowledge graph, including: Based on prediction error and consolidation The combination of these employs three different update strategies: Route A: Low prediction error update (consistency information) Triggering conditions: ,in , which is the low error threshold.

[0037] At this point, the new information is highly consistent with the existing schema, and a smooth incremental update is performed. The cluster center update formula is: in, The learning rate is expressed by the formula: The higher the consolidation, the lower the learning rate, thus maintaining the stability of knowledge.

[0038] Simultaneously update the consolidation level; the updated consolidation level... Represented as: in, To enhance the stability of the cluster by using a fixed increment.

[0039] Route B: High prediction error + high consolidation update (contradictory information + stable knowledge) Triggering conditions: and ,in This represents a high error threshold. , which is the high consolidation threshold.

[0040] At this point, the system is confident that the old knowledge has been sufficiently consolidated, and that the new information represents environmental changes rather than noise. Perform the knowledge cluster splitting operation: create a new cluster. The center is Initial consolidation Simultaneously, establish within the knowledge graph and The "evolutionary relationship" edges between them have a weight of To record the trajectory of knowledge evolution.

[0041] like Then a high learning rate is adopted. A significant adjustment to the original cluster is represented as: in, This indicates the new cluster center after adjustment. This indicates the original cluster center before the update.

[0042] Route C: High prediction error + low consolidation updates (contradictory information + unstable knowledge) Triggering conditions: and At this point, existing knowledge is not yet fully consolidated, and high bias in new information is likely noise or catastrophic interference. The system adopts a conservative strategy, refusing to update or assigning an extremely low learning rate. Mark the new data as "pending verification" and store it in a temporary buffer.

[0043] If the subsequent consecutive If each input produces a high prediction error, a re-evaluation is triggered, which may initiate the splitting mechanism of route B.

[0044] It should be noted that this mechanism effectively prevents the system from being misled by noise when knowledge is not stable, reflecting the "conservative bias" characteristic of cognitive systems.

[0045] Furthermore, route A's low learning rate employs exponential decay: ,in The initial learning rate, This is the decay coefficient. This ensures that highly consolidated knowledge clusters are updated slowly, maintaining stability. Route B's high learning rate employs a step function: when... hour, ;when hour, ;when hour, The larger the prediction error, the higher the learning rate, thus accelerating knowledge evolution.

[0046] In this embodiment, step S6 above, based on the updated knowledge clusters and knowledge graph, constructs an LSM-tree index and a scene-level index, and outputs a query interface that supports real-time retrieval, including: To support real-time insertion and querying, a multi-level inverted index is constructed using a log-structured merged tree (LSM-tree).

[0047] Let the index level be Size ratio of adjacent levels ,Right now .

[0048] The new vector is first inserted at the smallest index. ,when Reaching capacity limit When a record is entered, triggering and The merge operation uses a merge sort strategy, with a time complexity of [missing information - likely OCR errors]. .

[0049] It should be noted that this structure reduces the amortized cost of insertion operations to [missing information]. ,in The total number of vectors is significantly better than that of a traditional B-tree. Worst-case scenario.

[0050] Furthermore, the number of levels in an LSM-tree and hierarchical ratio A balance needs to be struck between data size and query frequency. For small-scale systems ( ),suggestion , Medium-sized ( ),suggestion , Large-scale ( ),suggestion , Hierarchical proportions The smaller the value, the higher the merging frequency, resulting in better query performance but increased insertion overhead. The larger the value, the opposite is true. Experiments show that... Achieving the best balance in most scenarios.

[0051] Furthermore, low error threshold and high error threshold It can adaptively adjust based on historical data distribution. (Statistics on past data) The prediction error distribution of the next update is used to calculate the 25th percentile. 75th percentile as This adaptive mechanism enables the system to adapt to the characteristics of different datasets. For example, in highly homogeneous datasets, It might drop to 0.15, while in heterogeneous datasets it might rise to 0.45. Every... The threshold is recalculated with each update to ensure the system continuously adapts to data evolution.

[0052] Two types of inverted lists are designed for different application scenarios.

[0053] Exact inverted lists are used in scenarios with small vector spaces, such as English vocabulary (approximately 180,000 words), where each vector in the dictionary... The corresponding inverted list contains only matrix IDs that exactly match the vector. Hash lookup is used to accelerate retrieval; the hash function... , This is the size of the hash table. Collisions are resolved using a linked list approach, with an average search time of [missing value]. .

[0054] Fuzzy inverted lists are useful for scenarios with huge vector spaces, such as audio features, and can potentially achieve [a certain level of efficiency / efficiency]. A combination of similar vectors. Vector quantization techniques are used to map similar vectors to the same representative vector. .

[0055] The quantification process uses K-means clustering, and the number of clusters... , This represents the total number of vectors.

[0056] Each vector Store as a signature Among them, residual norm .

[0057] It should be noted that this compression method reduces storage overhead from Dimensions dropped to (Residuals can be represented by half-precision floating-point numbers), saving approximately 50% of memory. During querying, retrieval... The nearest neighbor vectors are represented by an inverted list, and the original vectors are reconstructed using residuals to perform accurate similarity calculations.

[0058] Furthermore, the residuals in the fuzzy inverted list Different storage precisions can be used. 16-bit half-precision (FP16) has a compression rate of approximately 50%, with some precision loss. 8-bit fixed-point number (INT8) compression rate of 75%, with precision loss. 4-bit quantization achieves a compression rate of 87.5%, with minimal precision loss. The choice of quantization method depends on the application's accuracy requirements: FP16 is used for high-precision scenarios (medical, financial); INT8 is used for general scenarios (recommendation, search); and 4-bit quantization is used for low-precision-tolerance scenarios (initial screening, pre-filtering). Experiments show that INT8 maintains a retrieval accuracy of over 98% in most scenarios while significantly reducing memory usage.

[0059] Drawing inspiration from RAG search optimization principles, the knowledge base is divided into sub-bases based on different scenarios. .

[0060] Extract a list of high-frequency words from each scenario library. Jieba word segmentation was used to filter stop words, and the top 20 words were sorted by TF-IDF. High-frequency words were then vectorized into scene indexes. During a query, the query vector is calculated first. The scene library with the highest cosine similarity to each scene index is selected. This narrows the search scope to this scenario, reducing query time. Experiments show that when At that time, the retrieval speed increased by about 8 times, and the accuracy rate remained above 95%.

[0061] Example 3, referring to Tables 1-6, is an embodiment of the present invention, providing a method for dynamic updating and real-time retrieval optimization of a multimodal knowledge base. To verify the beneficial effects of the present invention, a simulation experiment is conducted for scientific demonstration.

[0062] Scenario 1: Dynamic updates of the intelligent customer service knowledge base A power company deployed an intelligent customer service system. The initial knowledge base contained 5,000 frequently asked questions and answers, covering scenarios such as fault reporting and business inquiries. After three months of operation, the system accumulated a large amount of user interaction data.

[0063] Step 1: Initialization and Consolidation Establishment The 5000 question-answer pairs were converted into 768-dimensional vectors using the BERT model, and initial knowledge clusters were constructed using K-means clustering (K=100). The consolidation of each cluster was initialized to... , This represents the number of samples within a cluster. After three months of operation, the consolidation of the frequently accessed "Fault Reporting" cluster reached 8.5, while that of the low-frequency "Special Services" cluster was only 2.3.

[0064] Step 2: Dynamic handling of new problems A user asked, "How do I apply for a charging station for new energy vehicles?" The system extracted the vector. The nearest neighbor cluster found was "Business Consulting" ( However, the prediction error (Above the threshold of 0.6). Due to the consolidation exceeding 5.0, route B was triggered. The system determined that this was an emerging business type and created a new cluster, "New Energy Service," which initially included this issue. Over the next three days, 12 more similar issues entered this cluster, rapidly increasing the consolidation to 3.8.

[0065] Step 3: Noise suppression A user inputs the meaningless text "asdfghjkl", and the system calculates... However, the nearest neighbor cluster has "other problems". (Low consolidation). Triggering route C, the system refuses to update and marks the data as "suspected noise". Continuous monitoring revealed that this type of data did not reappear, verifying the correctness of the rejection policy.

[0066] Performance evaluation: Compared with the traditional fixed learning rate method, this method improves the accuracy of the knowledge base from 82% to 91%, reduces the false inclusion rate of noisy data from 15% to 3%, and improves the recall rate of new knowledge recognition from 68% to 89%.

[0067] Scenario 2: Multimodal Medical Imaging Knowledge Base A hospital is building a diagnostic assistance system that includes multimodal images such as CT, MRI, and X-ray. The initial dataset contains 100,000 labeled images.

[0068] Step 1: Multimodal Vectorization CT images were processed using ResNet-50 to extract 2048-dimensional feature vectors, MRI images were processed using 3D-CNN to extract 1024-dimensional features, and X-ray images were processed using DenseNet to extract 512-dimensional features. An autoencoder was used to uniformly map all modalities to a 768-dimensional space to ensure consistency in cross-modal similarity calculations. 500 knowledge clusters were constructed, each representing a pathological feature pattern.

[0069] Step 2: Knowledge Evolution of Rare Cases After one year of system operation, the consolidation of a certain cluster of "pulmonary nodules" reached 12.3 (containing 8000 samples). A new, rare "ground-glass opacity" image was input, and the prediction error... Due to high consolidation and high prediction error, cluster splitting of route B was triggered. The system created a new cluster "Ground-glass nodules" and established an evolutionary relationship edge (weight 0.78) with "Lung nodules" in the knowledge graph, recording the detailed evolution of pathological features. In the following two months, this new cluster accumulated 45 samples, with a consolidation of 4.2, becoming an independent diagnostic reference category.

[0070] Step 3: Handling data drift caused by equipment upgrade The hospital replaced its CT scanner with a new one, and the changes in imaging parameters caused a systematic shift in the image feature vectors of the same pathological condition. The system detected a cluster of "fractures" ( New samples continue to be generated. The prediction error was moderate. After 7 days of continuous monitoring, the system determined that this was a change in equipment rather than a pathological change, and activated the "cluster center drift" mode. The cluster centers are gradually adjusted using a moderate learning rate to adapt them to new device characteristics while maintaining the continuity of historical diagnostic knowledge.

[0071] Performance evaluation: The system's accuracy in identifying rare cases increased from 54% to 82%, the adaptation time after the equipment upgrade was shortened from 30 days to 7 days, and the misdiagnosis rate was reduced by 23%.

[0072] Scenario 3: Real-time Video Content Retrieval System A video platform is building a real-time content review and recommendation system that needs to handle 1,000 video clip upload and retrieval requests per second.

[0073] Step 1: Building an LSM-tree index A 5-layer LSM-tree structure is adopted. Capacity 10,000 entries, hierarchy ratio The video frames are processed using the ViT model to extract 512-dimensional feature vectors, and each video is represented as follows: matrix( (For keyframes). The new video vector is first inserted. The average insertion time is 0.8ms. When full, trigger and The merging process employs multi-threaded parallel merging, with a merging time of approximately 150ms, during which new data insertion is not blocked.

[0074] Step 2: Optimize the fuzzy inverted list Due to the enormous feature vector space of videos (theoretically) (Several combinations) are used, employing a fuzzy inverted list. Historical vectors are clustered using K-means (K=50000), with each vector stored as a signature. Residual Using 16-bit half-precision floating-point storage, memory usage was reduced from 512×4=2048 bytes to 512×2+4=1028 bytes, a compression rate of 49.8%. (Search during query) The inverted list of nearest neighbor vectors is used to calculate the exact similarity through residual reconstruction, with an average query time of 12ms and a recall rate of 96.5%.

[0075] Step 3: Accelerate Scene Layering The video library is divided into 10 scenarios: news, entertainment, sports, education, etc. 20 high-frequency visual concepts are extracted from each scenario (such as "football" and "goal" corresponding to the sports scenario), and are vectorized into scenario indices. When the user queries "wonderful goals in the World Cup", the system first matches the "sports" scenario (similarity 0.92), and only retrieves from the 500,000 videos in this scenario instead of the 5 million videos in the entire library. The retrieval time is reduced from 120 ms to 15 ms, and the speed is increased by 8 times.

[0076] Step 4: Quick response to hot events During a major sports event, videos related to "champion celebration" increased sharply. The system detected the rapid formation of a new cluster in the "sports" scenario, accumulating 2,000 samples within 3 hours, and the consolidation degree increased from 0 to 5.8 rapidly. Trigger the "hot event" mode, automatically increase the weight of this cluster in the recommendation algorithm, and make the relevant content be displayed preferentially. After the event ended, the new growth rate of this cluster decreased, and the system gradually reduced its recommendation weight to achieve the natural attenuation of the hot event.

[0077] Effect evaluation: The system supports 1,200 concurrent queries per second, with an average response time of 18 ms. The memory occupancy is reduced by 52% compared with the traditional method. The recognition delay of hot events is shortened from 2 hours to 15 minutes, and the user satisfaction is increased by 34%.

[0078] Scenario 4: Cross-language document knowledge base An international organization constructs a knowledge base of policy documents in 10 languages including Chinese, English, French, German, Japanese, etc., with a total of 2 million documents.

[0079] Step 1: Cross-language vector space alignment Adopt the multilingual BERT (mBERT) model to map documents in different languages to a unified 768-dimensional vector space. For each language, 1,000 high-frequency words are extracted, and their vector representations are calculated to construct an alignment matrix between languages. For example, the cosine similarity of the vectors of the Chinese word "policy" and the English word "policy" reaches 0.94, ensuring the accuracy of cross-language retrieval.

[0080] Step 2: Knowledge tracking of policy evolution A cluster of environmental protection policies (mainly Chinese documents, ) was established in 2020. The newly released English policy documents in 2023 generated a high prediction error, and the system triggered Route B to create a new cluster "carbon neutrality policy". An evolutionary edge of "environmental protection policy" → "carbon neutrality policy" with a weight of 0.72 and a timestamp of 2023-06 is established in the knowledge graph. When retrieving content related to "environmental protection" later, the system automatically associates the two clusters to provide a complete view of policy evolution.

[0081] Step 3: Maintenance of translation version consistency Different language versions of the same policy should be grouped into the same cluster. The system detected newly uploaded French documents and existing English documents. (Low prediction error), trigger route A, using The cluster centers are fine-tuned with a low learning rate to better cover multilingual representations. Language version relationships are also marked in the metadata, supporting the "view other language versions" function.

[0082] Performance evaluation: Cross-language retrieval accuracy reached 88% (compared to 62% for traditional methods), policy evolution tracking completeness was 95%, translation version association accuracy was 99%, and users' efficiency in obtaining cross-language information was improved by 3.2 times.

[0083] On a test set containing 1 million documents, the method of this invention was compared with three baseline methods: (1) Fixed learning rate update ( (2) Simple TF-IDF retrieval; (3) Standard RAG method. Evaluation metrics include: knowledge base accuracy, new knowledge recall, noise suppression rate, and query response time. The results show that the method of the present invention improves accuracy by 11% compared to baseline 1, 18% compared to baseline 2, and 7% compared to baseline 3; new knowledge recall is improved by 21%, 15%, and 9% respectively; noise suppression rate is improved by 12%, 25%, and 8% respectively; and query response time is reduced by 35%, 52%, and 18% respectively. The overall performance is significantly better than existing methods.

[0084] The study was conducted across five different domains (news, healthcare, law, e-commerce, and social media), with 100,000 samples in each domain. The news domain, with its most frequently updated knowledge, was the most suitable for this method, utilizing low-frequency data. The value (0.2) adapts quickly with an accuracy of 89%; legal knowledge is the most stable and has a high accuracy. The value (0.5) maintains consistency with an accuracy of 94%; social media noise is the most prevalent, and the rejection strategy of route C achieves a noise suppression rate of 92%. This demonstrates the cross-domain adaptability of our method.

[0085] To fully verify the effectiveness of the technical solution of this invention, a standardized experimental environment was built. The hardware platform used an Intel Xeon E5-2680 v4 processor (2.4GHz, 28 cores, 56 threads), 128GB DDR4-2400 memory, a 2TB NVMe SSD (read / write speeds of 3500 / 3000 MB / s), and an NVIDIA Tesla V100 GPU (32GB VRAM) for deep learning model inference. The software environment consisted of an Ubuntu 20.04 LTS operating system, Python 3.8.10, the PyTorch 1.12.0 deep learning framework, the Faiss 1.7.2 vector retrieval library, and Elasticsearch 7.15.0 as a baseline for comparison.

[0086] The experimental dataset comprises three scales: a small-scale dataset (100,000 multimodal records for rapid validation), a medium-scale dataset (1 million records, simulating medium-sized enterprise applications), and a large-scale dataset (10 million records, simulating internet-level applications). Data sources include public datasets (Wikipedia, Common Crawl) and synthetic data (simulating domain-specific knowledge evolution). Vector dimensions are uniformly set to 768 dimensions (BERT-base output dimension), and the initial number of knowledge clusters is set to... ,in The size of the dataset.

[0087] Comparison of core performance indicators: Table 1 shows the experimental results of the method of this invention and three comparative methods on key performance indicators (based on a medium-sized dataset of 1 million records): Table 1 Comparison of Experimental Results for Key Performance Indicators

[0088] As shown in Table 1, the method of this invention achieves optimal performance across all key metrics. The accuracy of contradictory information identification reaches 89.6%, an improvement of 11.3 percentage points compared to the fixed learning rate method, 23.9 percentage points compared to the TF-IDF method, and 7.5 percentage points compared to the standard RAG method. This is mainly attributed to the accuracy of the knowledge consolidation assessment mechanism and the calculation of prediction error. The new knowledge recall rate reaches 89.2%, significantly higher than other methods, demonstrating that the cluster splitting strategy of route B can effectively capture real knowledge evolution. The noise suppression rate is 82.7%, reflecting the effectiveness of the conservative strategy of route C.

[0089] In terms of performance, the average query latency is only 87ms, a further reduction of 8.4% compared to the standard RAG method, thanks to the optimization of the scenario-based hierarchical index and LSM-tree structure. Insertion throughput reaches 3,420 rows / second, a 20% improvement over the standard RAG method, demonstrating the advantages of LSM-tree in high-concurrency write scenarios. Memory usage is only 5.8GB, saving 52.8% compared to the fixed learning rate method and 36.3% compared to the standard RAG method, demonstrating the significant effect of vector signature compression technology. The knowledge graph consistency score is 0.876, indicating that the system can maintain a high-quality knowledge structure.

[0090] Verification of the effectiveness of the three-way update routing strategy: Table 2 shows the detailed performance of the three routing strategies when processing different types of input data (based on statistics from 100,000 update operations): Table 2. Detailed performance of routing strategies when handling different types of input data.

[0091] Table 2 shows the experimental results, verifying the rationality and effectiveness of the three-way update routing strategy. Route A processed 62.3% of the input data, which was highly consistent with existing knowledge. The system adopted smooth incremental updates, achieving an accuracy of 94.2% and a fast processing speed (23ms), while steadily improving knowledge consolidation (+0.10). Route B processed 18.5% of the input, which were genuine knowledge evolution signals. Although the processing time was longer (156ms, due to cluster splitting and graph updates), the accuracy still reached 87.5%, successfully capturing environmental changes. Route C processed 19.2% of the input, which was potentially noisy data. The system adopted a rejection strategy, achieving an accuracy of 91.8% (i.e., 91.8% of rejection decisions were correct), and the fastest processing speed (18ms), effectively preventing noise pollution.

[0092] Impact analysis of knowledge consolidation mechanism: Table 3 shows the impact of knowledge consolidation on system update performance (comparison with and without consolidation assessment mechanism): Table 3. Impact of Knowledge Consolidation on System Update Performance

[0093] Table 3 clearly demonstrates the importance of the knowledge consolidation assessment mechanism. During the unconsolidated period (0-1 day), the performance difference between the two methods was minimal (only 0.8%), as the knowledge had not yet formed a stable structure. However, the difference widened significantly with increasing consolidation time. During the fully consolidated period (4-7 days), the system incorporating consolidation assessment achieved an update success rate of 85.7%, 23.6 percentage points higher than the system without this mechanism. During the highly consolidated period (8-14 days), the performance improvement reached 25.5%. This validates the application value of cognitive neuroscience theory in the dynamic updating of knowledge bases: consolidated knowledge structures can generate stronger prediction error signals, thus more effectively identifying and absorbing truly new knowledge while suppressing noise interference.

[0094] LSM-tree index structure performance analysis: Table 4 compares the performance of different indexing schemes on large-scale datasets (10 million records): Table 4. Performance Comparison of Different Indexing Schemes on Large-Scale Datasets

[0095] Table 4 shows that the LSM-tree combined with vector residual compression scheme adopted in this invention performs excellently in various indicators. Insertion throughput reaches 3,420 rows / second, slightly lower than uncompressed LSM-tree (3,580 rows / second), but still significantly higher than traditional methods (1,240 rows / second) and B+ trees (980 rows / second), representing a performance improvement of 2.8-3.5 times. Query latency is only 87ms, a further reduction of 8.4% compared to uncompressed LSM-tree and 38.7% compared to traditional inverted indexes. Memory usage is only 5.8GB, a reduction of 52.8% compared to traditional inverted indexes and 36.3% compared to uncompressed LSM-trees. This is thanks to vector signature technology, which stores only the vector residuals and norms, rather than the complete vector. Index building time is 38 minutes, 55.3% faster than traditional methods, supporting rapid deployment.

[0096] Verification of the acceleration effect of scene hierarchical index: Table 5 shows the performance of scene hierarchical index under different numbers of scenes (based on 1 million records): Table 5. Performance Comparison of Scene Hierarchical Index under Different Number of Scenes

[0097] Table 5 verifies the effectiveness of the scene-based hierarchical index. When the number of scenes is 10, the query latency decreases from 142ms to 18ms, achieving a speedup of 7.9 times, while maintaining a scene recognition accuracy of 95.2%. This is the optimal configuration point, balancing retrieval speed and accuracy. As the number of scenes increases to 20, the speedup further improves to 11.8 times, but the accuracy drops to 91.5% because overly fine scene segmentation leads to blurred boundaries. When the number of scenes reaches 50, the accuracy drops to 85.3%, and the speedup actually decreases to 9.5 times due to the increased overhead of scene recognition itself. Experiments show that for datasets with millions of scenes, 10-20 scenes are the optimal choice.

[0098] Cross-domain adaptability testing: Table 6 shows the adaptability performance of the method of the present invention in five different domains: Table 6 Adaptability performance in different fields

[0099] Table 6 demonstrates the cross-domain adaptability of the method of this invention. The news media field has the most frequent knowledge updates; by setting higher... Value (0.6) and lower The value (0.2), emphasizing access frequency rather than time accumulation, achieved an accuracy of 89.3% and a recall rate as high as 91.2%. Legal document knowledge is the most stable, employing high... Maintaining consistency with a value of (0.5), the system achieved an accuracy of 94.1% and a noise suppression rate of 91.7%. In the social media domain, where noise is most prevalent, the rejection strategy of route C resulted in a noise suppression rate of 92.1%, effectively preventing spam from contaminating the knowledge base. In the medical diagnostics domain, where precision is paramount, the system achieved an accuracy of 93.7%, meeting clinical application needs. In the e-commerce recommendation domain, requiring high concurrency processing, the system maintained an accuracy of 87.8% and a recall rate of 89.7%, while supporting over 3000 queries per second.

[0100] Long-term stability testing: The system underwent a 30-day continuous operation test, processing 100,000 new data entries daily. The trends in knowledge base size, query performance, and memory usage were monitored. Results showed that the number of knowledge clusters steadily increased from an initial 1000 to 1850, a growth rate of 85%, consistent with the natural evolution of knowledge. The average query latency slowly increased from 85ms on day 1 to 92ms on day 30, an increase of only 8.2%, demonstrating the good scalability of the LSM-tree structure. Memory usage increased from 5.8GB to 7.2GB, a growth rate of 24.1%, far lower than the data volume growth rate (300%), indicating the continued effectiveness of vector compression technology. The knowledge graph consistency score consistently remained above 0.87, demonstrating excellent long-term system stability.

[0101] Based on the above experimental results, the technical solution of this invention is significantly superior to existing methods in multiple dimensions: (1) the accuracy of contradictory information identification is improved by 7.5-23.9 percentage points; (2) the recall rate of new knowledge is improved by 14.4-36.0 percentage points; (3) the noise suppression rate is improved by 8.1-26.9 percentage points; (4) the query latency is reduced by 8.4-60.1%; (5) the insertion throughput is increased by 20-249%; (6) the memory usage is reduced by 36.3-63.1%; and (7) the consistency of the knowledge graph is improved by 9.1-46.5%. The experiment fully verifies the effectiveness of the core innovations such as cognitive consolidation mechanism, prediction error-driven three-way update strategy, LSM-tree index structure, vector residual compression technology, and scene hierarchical index, providing reliable technical support for building a large-scale intelligent knowledge management system.

[0102] 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, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for dynamic updating and real-time retrieval optimization of a multimodal knowledge base, characterized in that, include: The original multimodal data is vectorized to map the heterogeneous data to a vector space of the same dimension, resulting in a feature vector of the same dimension. Hierarchical clustering is performed based on feature vectors of a unified dimension to generate knowledge clusters and corresponding cluster center vectors and covariance matrices. Based on the generated knowledge clusters, a consolidation score is evaluated for each cluster, and the consolidation value of each cluster is output. Extract the feature vectors of the new data and match them with the generated knowledge clusters. Calculate the prediction error and output the nearest neighbor cluster and the prediction error value. Based on the consolidation value and prediction error value of each cluster, a three-way update routing strategy is triggered to output the updated knowledge clusters and knowledge graph; Based on the updated knowledge clusters and knowledge graph, an LSM-tree index and a scenario-based hierarchical index are constructed, and a query interface that supports real-time retrieval is output.

2. The method for dynamic updating and real-time retrieval optimization of a multimodal knowledge base as described in claim 1, characterized in that, The vectorization process of the original multimodal data, which maps heterogeneous data to a vector space of the same dimension to obtain feature vectors of the same dimension, includes: Obtain the raw multimodal data and extract feature vectors for each modality: For text data, a pre-trained language model is used to extract word vectors, and each document is represented as a feature matrix. For image data, a convolutional neural network is used to extract image feature vectors; for video data, it is represented as an image vector sequence matrix. For audio data, acoustic features are extracted using an acoustic feature extraction model to form a speech feature matrix; The feature matrices or feature vectors obtained from each modality are uniformly mapped to a preset vector dimension space.

3. The method for dynamic updating and real-time retrieval optimization of a multimodal knowledge base as described in claim 2, characterized in that, The hierarchical clustering based on feature vectors of a unified dimension to generate knowledge clusters and corresponding cluster center vectors and covariance matrices includes: Hierarchical clustering is performed based on feature vectors of a unified dimension to construct a knowledge graph structure and obtain multiple knowledge clusters. For each knowledge cluster, the weighted average of all vectors within the cluster is calculated as the cluster center vector of the knowledge cluster, where the weight coefficient of each vector is used to characterize the importance of the vector within the cluster. Calculate and store the covariance matrix for each knowledge cluster.

4. The method for dynamic updating and real-time retrieval optimization of a multimodal knowledge base as described in claim 3, characterized in that, The generated knowledge clusters are used to evaluate the consolidation score for each cluster, and the consolidation value for each cluster is output, including: Obtain the generated knowledge clusters, each containing multiple vectorized data samples; A consolidation score is maintained for each knowledge cluster. The consolidation score is obtained by weighted summation after nonlinear transformation of historical access frequency, existence duration and internal stability. This results in knowledge clusters with higher historical access frequency, existence duration and stronger internal stability receiving higher consolidation scores.

5. The method for dynamic updating and real-time retrieval optimization of a multimodal knowledge base as described in claim 4, characterized in that, The process of extracting feature vectors from new data, matching them with generated knowledge clusters, calculating prediction errors, and outputting nearest neighbor clusters and prediction error values ​​includes: Feature vectors of new data are extracted and matched with knowledge clusters in the existing knowledge base. The comprehensive similarity between the feature vectors and each knowledge cluster is calculated. The comprehensive similarity is obtained by weighted summation of cosine similarity and Mahalanobis distance-based similarity, where Mahalanobis distance is calculated using the covariance matrix of each knowledge cluster.

6. The method for dynamic updating and real-time retrieval optimization of a multimodal knowledge base as described in claim 5, characterized in that, The process of extracting feature vectors from new data, matching them with generated knowledge clusters, calculating prediction errors, and outputting nearest neighbor clusters and prediction error values ​​also includes: Select the knowledge cluster with the highest comprehensive similarity as the nearest neighbor cluster; The prediction error is calculated based on the comprehensive similarity of the nearest neighbor clusters. The prediction error ranges from 0 to 1. The larger the prediction error, the greater the deviation between the new data and the existing knowledge base.

7. The method for dynamic updating and real-time retrieval optimization of a multimodal knowledge base as described in claim 6, characterized in that, The three-way update routing strategy, based on the consolidation value and prediction error value of each cluster, is triggered to output the updated knowledge clusters and knowledge graph, including: In response to the prediction error value being lower than the first threshold, it is determined that the new data is consistent with the existing knowledge, and a smooth incremental update is performed: with a learning rate that is negatively correlated with the consolidation value, the cluster centers of the nearest neighbor clusters are adjusted towards the feature vector direction of the new data, and the consolidation value of the corresponding cluster is increased by a fixed increment. If the prediction error exceeds the second threshold and the consolidation value exceeds the third threshold, it is determined that the new data represents the true knowledge evolution, and a cluster splitting operation is performed: a new cluster is created with the feature vector of the new data as the center, an initial consolidation value is assigned, and an evolutionary relationship edge is established between the original nearest neighbor cluster and the new cluster in the knowledge graph; if the prediction error exceeds the fourth threshold, the cluster center of the original nearest neighbor cluster is adjusted at a high learning rate.

8. The method for dynamic updating and real-time retrieval optimization of a multimodal knowledge base as described in claim 7, characterized in that, The process of triggering a three-way update routing strategy based on the consolidation value and prediction error value of each cluster, and outputting the updated knowledge clusters and knowledge graph, also includes: If the prediction error is higher than the second threshold and the consolidation value is lower than the third threshold, the new data is determined to be potential noise, and a conservative strategy is implemented: the update is rejected or only updated at a very low learning rate, and the new data is marked as a state to be verified and stored in a temporary buffer; if multiple consecutive inputs produce prediction errors higher than the second threshold, a re-evaluation is triggered.

9. The method for dynamic updating and real-time retrieval optimization of a multimodal knowledge base as described in claim 8, characterized in that, Based on the updated knowledge clusters and knowledge graph, an LSM-tree index and a scenario-level hierarchical index are constructed, outputting a query interface that supports real-time retrieval, including: A multi-level inverted index is constructed using a log structure merging tree. Newly inserted feature vectors first enter the smallest level. When the level capacity reaches the upper limit, merging with the next level is triggered. Two types of inverted lists are provided for different vector space sizes.

10. The method for dynamic updating and real-time retrieval optimization of a multimodal knowledge base as described in claim 9, characterized in that, The process of constructing an LSM-tree index and a scenario-level hierarchical index based on the updated knowledge clusters and knowledge graph, and outputting a query interface that supports real-time retrieval, also includes: The knowledge base is divided into multiple sub-bases according to scenarios. High-frequency words are extracted for each scenario and quantified into scenario indexes. During a query, the similarity between the query vector and each scenario index is calculated, and the scenario sub-base with the highest matching degree is selected as the search scope, thus narrowing the search space to a single scenario.

Images