Knowledge database retrieval method based on feature extraction

Abstract

The invention relates to the technical field of database retrieval, in particular to a knowledge database retrieval method based on feature extraction. The method comprises the following steps of obtaining a document to be retrieved, performing local hash de-duplication processing and extracting multi-modal features, performing feature spectrum clustering processing and associated entity analysis on the extracted features, generating a feature associated entity structure, performing document topic modeling and semantic direction positioning, and extracting a semantic topic of the document. The method comprises the following steps: performing word vector retrieval on a to-be-retrieved document and a knowledge database by utilizing a document semantic topic, generating word vector retrieval data, performing sentence vector retrieval, generating a final document retrieval result, integrating word vectors and the document retrieval result, and executing feature retrieval of the knowledge database through an intelligent retrieval engine, thereby realizing efficient knowledge query. The knowledge database retrieval method is more intelligent and more accurate.

Description

Technical Field

[0001] The present invention relates to the technical field of database retrieval, and in particular to a knowledge database retrieval method based on feature extraction. Background Art

[0002] In modern information retrieval systems, with the increasing diversity and complexity of data, traditional retrieval methods based on keyword matching can no longer meet users' needs for accuracy and depth. Especially in multimodal documents, data not only contains text, but also contains multiple formats such as images and audio, making feature extraction and association analysis extremely complicated. This multimodal feature cannot be effectively retrieved through simple keyword matching, resulting in a large amount of potential information in the knowledge database not being fully utilized and unable to provide users with accurate query results. In current retrieval technology, although deep learning models such as word vectors and sentence vectors have been significantly improved, they have the problem of insufficient matching accuracy when facing the association and conversion between different modalities. This problem is particularly reflected in the cross-modal understanding and semantic direction positioning of features, resulting in the model being unable to accurately extract the topic and semantic association of the document when facing complex document structures, thereby limiting the intelligence level of the retrieval engine. In addition, the joint retrieval strategy of word vectors and sentence vectors has not been fully optimized, which is prone to low computational efficiency and inaccurate results when processing large-scale data. Summary of the invention

[0003] Based on this, it is necessary to provide a knowledge database retrieval method based on feature extraction to solve at least one of the above technical problems.

[0004] To achieve the above purpose, a knowledge database retrieval method based on feature extraction includes the following steps:

[0005] Step S1: Obtain the document to be retrieved; perform local hash deduplication processing on the document to be retrieved, and perform multimodal feature extraction to obtain multimodal document features;

[0006] Step S2: performing feature spectrum clustering processing on multimodal document features to generate a document multidimensional feature spectrum; performing associated entity analysis on the document multidimensional feature spectrum to generate a feature associated entity structure;

[0007] Step S3: Modeling the document topic for the multimodal document features according to the feature-associated entity structure to obtain a topic-feature matrix; positioning the topic-feature matrix in a semantic direction to obtain the document semantic topic;

[0008] Step S4: Perform multi-way parallel search on the documents to be searched and the knowledge database based on the document semantic theme to obtain a candidate word vector set; perform word vector search on the candidate word vector set to generate word vector search data;

[0009] Step S5: Based on the document semantic theme, multi-level semantic matching is performed on the word vector retrieval data and the knowledge database to obtain a set of candidate sentences; sentence vector retrieval is performed on the candidate sentence set to generate a document retrieval result;

[0010] Step S6: construct a search engine for the word vector search data and the document search results to obtain an intelligent search engine; use the intelligent search engine to perform intelligent search on the knowledge database to execute a feature search method for the knowledge database.

[0011] The present invention ensures that redundant information is eliminated in the process of extracting multimodal features of documents by performing local hash deduplication processing on the documents to be retrieved, and the obtained multimodal features are more accurate and unique, which improves the feature expression efficiency of retrieval. The feature spectrum clustering process systematically groups the extracted multimodal document features, combines the associated entity analysis, generates a closely related feature entity structure, effectively optimizes the organization and expression of feature data, and is convenient for subsequent modeling and semantic processing. The document theme modeling and semantic direction positioning performed according to the associated entity structure effectively capture the core semantics of the document, and ensure that the generated document semantic theme is highly consistent with the document content characteristics, thereby improving the accuracy of semantic matching in the subsequent retrieval process. By performing word vector and sentence vector retrieval on the document semantic theme, combined with the feature data of the knowledge database, it is ensured that the semantics of words and sentences are deeply matched during the retrieval process, and the accuracy and comprehensiveness of the retrieval data are enhanced. The intelligent retrieval engine constructed based on word vectors and document retrieval results has the ability to efficiently retrieve features of the knowledge database, can realize multi-dimensional retrieval optimization of features and semantics, and effectively improves the retrieval performance and response efficiency of the knowledge database. The knowledge database retrieval method based on feature extraction builds an intelligent retrieval engine through precise extraction of multimodal features, feature spectrum clustering analysis, semantic topic modeling and multi-layer vector retrieval integration, realizing efficient processing from local features to global semantics.

[0012] Preferably, step S1 comprises the following steps:

[0013] Step S11: Obtain the document to be retrieved;

[0014] Step S12: Cut the document to be retrieved into fixed lengths to obtain local segments of the document to be retrieved;

[0015] Step S13: performing hash deduplication processing on the local fragments of the document to be retrieved to generate deduplicated document data;

[0016] Step S14: extract multimodal features from the deduplicated document data to obtain multimodal document features.

[0017] The present invention ensures that the document is effectively decomposed into manageable local fragments by cutting the document to be retrieved into fixed lengths, which is convenient for subsequent processing and analysis and improves the pertinence and accuracy of information extraction. The local fragments are subjected to hash deduplication processing to eliminate redundant information, ensure that the data set for subsequent feature extraction is more streamlined, and avoid the negative impact of duplicate data on retrieval efficiency. Multimodal features are extracted from the deduplicated document data, covering a variety of information types such as text and images, enhancing the comprehensive understanding of the document content and improving the support for different information types during retrieval. Through the extraction of multimodal features, the feature space is enriched, the expressiveness of the document in the knowledge database is improved, and more abundant basic data is provided for subsequent retrieval and analysis. The efficiency and accuracy of the knowledge database retrieval method are improved as a whole, making the retrieval results more relevant and practical.

[0018] Preferably, step S2 comprises the following steps:

[0019] Step S21: performing local clustering on the multimodal document features at the feature level to obtain spatial dimension reduction document features;

[0020] Step S22: clustering the feature spectrum of the spatial dimension reduction document features to generate a multi-dimensional feature spectrum of the document;

[0021] Step S23: performing related entity mapping processing on the document multi-dimensional feature spectrum to obtain the related path between entities;

[0022] Step S24: simulate entity connection processing on multimodal document features based on the inter-entity association paths to generate a feature-associated entity structure.

[0023] The present invention achieves spatial dimensionality reduction by locally clustering multimodal document features at the feature level, reduces the complexity of feature dimensions, and enhances the efficiency and operability of subsequent analysis. By generating a document multidimensional feature spectrum through feature spectrum clustering, the relationship structure between features is optimized, making document features more identifiable and improving the understanding and matching capabilities of document content during the retrieval process. The document multidimensional feature spectrum is mapped to associated entities, revealing the association paths between entities inside and outside the document, providing a deeper understanding of semantic relationships, and enhancing the networked representation of knowledge. Multimodal document features are simulated for entity connections based on the association paths between entities to form a feature-associated entity structure, providing an integrated perspective of multidimensional information and enriching the content representation of the knowledge database. The retrieval accuracy of the knowledge database is enhanced, allowing users to obtain more relevant and meaningful results during the query process.

[0024] Preferably, step S24 comprises the following steps:

[0025] Perform path conflict point detection on the associated paths between entities and generate associated path conflict points;

[0026] According to the preset path weight distribution data, weights are distributed to the associated paths between entities to obtain a weighted path set;

[0027] Performing simulated entity mapping processing on multimodal document features based on a weighted path set to generate an initial simulated entity connection model;

[0028] According to the initial simulation entity connection model, the associated path conflict points are corrected to generate a feature associated entity structure.

[0029] The present invention detects conflict points in the association paths between entities, identifies inconsistencies in the paths, ensures that subsequent analysis is based on an accurate and reliable foundation, and improves the integration of the knowledge graph. The association paths between entities are allocated according to preset path weights, which optimizes the relative importance of the paths, makes information transmission more targeted and effective, and enhances the characterization ability of features. Based on the weighted path set, multimodal document features are simulated for entity mapping, and an initial simulated entity connection model is generated, which provides an intuitive expression of entity relationships and facilitates subsequent in-depth analysis and mining. By correcting the conflict points of the association paths, the accuracy and consistency of the feature-associated entity structure are ensured, potential error propagation is eliminated, and the reliability of the final model is improved. The quality and efficiency of the knowledge database retrieval method are improved, so that users can obtain more accurate and practical retrieval results in the information retrieval process.

[0030] Preferably, step S3 comprises the following steps:

[0031] Step S31: Perform graph convolutional network analysis on the feature-associated entity structure to obtain entity semantic representation;

[0032] Step S32: performing non-negative matrix decomposition on the entity semantic representation to obtain a topic-feature matrix;

[0033] Step S33: performing multidimensional scaling analysis on the topic-feature matrix to obtain a low-dimensional semantic coordinate system;

[0034] Step S34: performing density peak clustering on the low-dimensional semantic coordinate system to obtain the document semantic topic representation.

[0035] The present invention analyzes the feature-associated entity structure through a graph convolutional network, captures the complex relationship between entities, extracts richer semantic information, and improves the depth and accuracy of entity representation. The entity semantic representation is processed by non-negative matrix decomposition to generate a topic-feature matrix, which effectively separates the relationship between the topic and the feature, and enhances the interpretability and organization of the information. Through multidimensional scaling analysis, the topic-feature matrix is ​​mapped to a low-dimensional space, which simplifies the complexity of the data, facilitates subsequent processing and analysis, and improves the computational efficiency of the algorithm. The low-dimensional semantic coordinate system is analyzed using density peak clustering to obtain a clear document semantic topic representation, which promotes the understanding and classification of document content and improves the relevance of retrieval results. The semantic understanding ability in the knowledge database retrieval process is enhanced, ensuring that the retrieval results obtained by users are more accurate and practical, and improving the user experience.

[0036] Preferably, step S4 comprises the following steps:

[0037] Step S41: perform query word expansion on the document semantic topic to obtain a query word vector representation;

[0038] Step S42: Perform distributed semantic modeling on the query word vector representation and the knowledge database to obtain a word vector index;

[0039] Step S43: performing multi-path parallel retrieval on the query word vector representation according to the word vector index to obtain a candidate word vector set;

[0040] Step S44: Calculate the cosine similarity of the candidate word vector set to generate a similarity sorted list;

[0041] Step S45: reorder the candidate word vector set based on the similarity sorting list to generate word vector retrieval data.

[0042] The present invention enriches the semantic information of query words and improves the comprehensiveness and accuracy of retrieval by expanding query words for document semantic topics. Distributed semantic modeling is performed on query word vector representation and knowledge database, which enhances the understanding of the relationship between words and ensures the semantic accuracy of word vector index. Multi-way parallel retrieval of query word vectors based on word vector index significantly improves the retrieval speed and optimizes the responsiveness of the system on large-scale data sets. By calculating cosine similarity of candidate word vector sets, the similarity between word vectors is quantified, ensuring the relevance and effectiveness of subsequent sorting results. The candidate word vector set is reordered based on the similarity sorting list to ensure that the finally generated word vector retrieval data meets the user's query intent, thereby improving the relevance of the retrieval results and user satisfaction.

[0043] Preferably, step S45 includes the following steps:

[0044] Perform multi-dimensional sorting evaluation on the candidate word vector set based on the similarity sorting list to obtain a multi-dimensional sorting score matrix;

[0045] According to the multi-dimensional sorting score matrix, the candidate word vector set is adjusted in ascending order of score to generate a dynamically sorted candidate word set;

[0046] Multimodal feature fusion is performed on the dynamically sorted candidate word set to obtain word vector retrieval data.

[0047] The present invention conducts multi-dimensional sorting evaluation on the candidate word vector set, comprehensively considers the influence of multiple features, and improves the accuracy and comprehensiveness of the sorting results. The candidate word vector set is adjusted in ascending order according to the multi-dimensional sorting scoring matrix, ensuring that the most relevant candidate words are displayed first, and improving the fit between the search results and user needs. Multimodal feature fusion is performed on the dynamic sorting candidate word set, integrating information of different modes, and enhancing the semantic richness and representation ability of word vector retrieval data. The dynamic sorting mechanism realizes real-time update and feedback, so that users can obtain more timely and accurate results when searching, and improves the overall user experience and satisfaction. Through multi-dimensional scoring and dynamic adjustment, a highly adaptable retrieval mechanism is formed, which can flexibly adjust the sorting strategy according to different query scenarios and user needs.

[0048] Preferably, step S5 comprises the following steps:

[0049] Step S51: Perform syntactic dependency analysis on the document semantic topics to obtain a set of grammatically structured sentences;

[0050] Step S52: performing multi-level semantic matching on the knowledge database according to the grammatically structured sentence set to obtain a candidate sentence set;

[0051] Step S53: labeling the candidate sentence set with semantic roles to generate a semantically structured candidate set;

[0052] Step S54: constructing a semantic graph for the semantically structured candidate set to obtain a sentence relationship network;

[0053] Step S55: Perform sentence vector retrieval on word vector retrieval data and knowledge database based on the sentence relationship network to generate document retrieval results.

[0054] The present invention performs syntactic dependency analysis on the semantic topics of the document, clarifies the grammatical relationship within the sentence, enhances the understanding of the text structure, and thus improves the accuracy of semantic matching. In the multi-level semantic matching process, the rich information of the knowledge database is used to accurately screen out candidate sentences with high relevance to the query, ensuring the relevance and effectiveness of the retrieval results. The candidate sentences are labeled with semantic roles to clarify the semantic roles of each component in the sentence. The generated semantic structured candidate set provides more in-depth semantic information, laying a solid foundation for subsequent analysis. By constructing a relationship network between sentences, a semantic graph is formed, which intuitively shows the correlation between sentences, helps to discover potential semantic connections, and improves the depth and breadth of retrieval. Sentence vector retrieval is performed based on the relationship network between sentences, and rich semantic information is used to ensure the accuracy and completeness of the retrieval results, thereby improving the user's retrieval experience.

[0055] Preferably, step S52 includes the following steps:

[0056] The grammatically structured sentence set is processed by a bidirectional long short-term memory network to obtain a context-aware sentence vector representation;

[0057] Perform multi-level hierarchical encoding on the knowledge database to generate a hierarchical sentence vector space;

[0058] Build a multi-level index structure for the hierarchical sentence vector space to obtain an efficient sentence vector retrieval tree;

[0059] The context-aware sentence vector representation is used to perform multi-level semantic matching on the efficient sentence vector retrieval tree to obtain a set of candidate sentences.

[0060] The present invention processes a sentence set through a bidirectional long short-term memory network, and the generated context-aware sentence vector representation effectively captures the context information in the sentence, thereby improving the accuracy of the semantic representation. The knowledge database is multi-level hierarchical encoded, and the hierarchical sentence vector space formed can flexibly represent semantic relationships at different levels, thereby improving the flexibility and accuracy of retrieval. The constructed multi-level index structure provides an efficient query path for the sentence vector retrieval tree, significantly reduces the retrieval time, and improves the response speed of the system. Multi-level semantic matching is performed using context-aware sentence vector representation, thereby ensuring the relevance and accuracy of candidate sentences and optimizing the final retrieval results. By generating a set of candidate sentences, users are provided with a variety of retrieval results, meeting different retrieval requirements, and improving the user experience.

[0061] Preferably, step S6 comprises the following steps:

[0062] Step S61: Perform deep semantic embedding on the word vector retrieval data to obtain a fast word vector retrieval structure;

[0063] Step S62: extracting semantic relations from the document retrieval results to obtain a sentence vector retrieval module;

[0064] Step S63: Designing a joint search strategy for the fast word vector search structure and the sentence vector search module to obtain an intelligent search engine;

[0065] Step S64: Use the intelligent search engine to perform intelligent search on the knowledge database to execute the feature search method of the knowledge database.

[0066] The present invention performs deep semantic embedding of word vector retrieval data to form a fast word vector retrieval structure, which significantly speeds up the retrieval speed and optimizes the user experience. The semantic relationship extraction of the document retrieval results and the generated sentence vector retrieval module provide a deeper semantic understanding capability, which helps to capture the complex relationship between sentences. The designed upper and lower joint retrieval strategy effectively integrates semantic information at different levels, improves the accuracy and relevance of the retrieval, and meets diverse retrieval needs. The use of an intelligent retrieval engine to search the knowledge database enhances the execution effect of the feature retrieval method and ensures that the returned results are more accurate and relevant. By integrating the fast word vector retrieval structure and the sentence vector retrieval module, the intelligent retrieval engine formed provides efficient retrieval capabilities for the overall system and promotes the rapid acquisition of information. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 A schematic diagram of the steps of a knowledge database retrieval method based on feature extraction;

[0068] Figure 2 for Figure 1 Detailed implementation steps of step S2 in the flowchart;

[0069] Figure 3 for Figure 1 Detailed implementation steps of step S3 in FIG.

[0070] The realization of the purpose, functional features and advantages of the present invention will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0071] The technical method of the present invention is described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by technicians in this field without creative work are within the scope of protection of the present invention.

[0072] In addition, the accompanying drawings are only schematic illustrations of the present invention and are not necessarily drawn to scale. The same reference numerals in the figures represent the same or similar parts, and their repeated description will be omitted. Some of the block diagrams shown in the accompanying drawings are functional entities and do not necessarily correspond to physically or logically independent entities. The functional entities can be implemented in software form, or implemented in one or more hardware modules or integrated circuits, or implemented in different networks and / or processor methods and / or microcontroller methods.

[0073] It should be understood that, although the terms "first", "second", etc. may be used herein to describe various units, these units should not be limited by these terms. These terms are used only to distinguish one unit from another unit. For example, without departing from the scope of the exemplary embodiments, the first unit may be referred to as the second unit, and similarly the second unit may be referred to as the first unit. The term "and / or" used herein includes any and all combinations of one or more of the listed associated items.

[0074] To achieve this, please refer to Figures 1 to 3 , a knowledge database retrieval method based on feature extraction, comprising the following steps:

[0075] Step S1: Obtain the document to be retrieved; perform local hash deduplication processing on the document to be retrieved, and perform multimodal feature extraction to obtain multimodal document features;

[0076] Step S2: performing feature spectrum clustering processing on multimodal document features to generate a document multidimensional feature spectrum; performing associated entity analysis on the document multidimensional feature spectrum to generate a feature associated entity structure;

[0077] Step S3: Modeling the document topic for the multimodal document features according to the feature-associated entity structure to obtain a topic-feature matrix; positioning the topic-feature matrix in a semantic direction to obtain the document semantic topic;

[0078] Step S4: Perform multi-way parallel search on the documents to be searched and the knowledge database based on the document semantic theme to obtain a candidate word vector set; perform word vector search on the candidate word vector set to generate word vector search data;

[0079] Step S5: Based on the document semantic theme, multi-level semantic matching is performed on the word vector retrieval data and the knowledge database to obtain a set of candidate sentences; sentence vector retrieval is performed on the candidate sentence set to generate a document retrieval result;

[0080] Step S6: construct a search engine for the word vector search data and the document search results to obtain an intelligent search engine; use the intelligent search engine to perform intelligent search on the knowledge database to execute a feature search method for the knowledge database.

[0081] The present invention ensures that redundant information is eliminated in the process of extracting multimodal features of documents by performing local hash deduplication processing on the documents to be retrieved, and the obtained multimodal features are more accurate and unique, which improves the feature expression efficiency of retrieval. The feature spectrum clustering process systematically groups the extracted multimodal document features, combines the associated entity analysis, generates a closely related feature entity structure, effectively optimizes the organization and expression of feature data, and is convenient for subsequent modeling and semantic processing. The document theme modeling and semantic direction positioning performed according to the associated entity structure effectively capture the core semantics of the document, and ensure that the generated document semantic theme is highly consistent with the document content characteristics, thereby improving the accuracy of semantic matching in the subsequent retrieval process. By performing word vector and sentence vector retrieval on the document semantic theme, combined with the feature data of the knowledge database, it is ensured that the semantics of words and sentences are deeply matched during the retrieval process, and the accuracy and comprehensiveness of the retrieval data are enhanced. The intelligent retrieval engine constructed based on word vectors and document retrieval results has the ability to efficiently retrieve features of the knowledge database, can realize multi-dimensional retrieval optimization of features and semantics, and effectively improves the retrieval performance and response efficiency of the knowledge database. The knowledge database retrieval method based on feature extraction builds an intelligent retrieval engine through precise extraction of multimodal features, feature spectrum clustering analysis, semantic topic modeling and multi-layer vector retrieval integration, realizing efficient processing from local features to global semantics.

[0082] In the embodiment of the present invention, reference Figure 1 The above is a schematic diagram of the steps of a knowledge database retrieval method based on feature extraction of the present invention. In this example, the knowledge database retrieval method based on feature extraction includes the following steps:

[0083] Step S1: Obtain the document to be retrieved; perform local hash deduplication processing on the document to be retrieved, and perform multimodal feature extraction to obtain multimodal document features;

[0084] In this embodiment, the process of obtaining the document to be retrieved includes extracting text data from a specified data source (such as a document database, a web crawler, or a file system). First, a crawler tool (such as Scrapy) is used to crawl relevant documents from a specified web page or API interface and store them locally. Each document is formatted to ensure its structural consistency, such as converting different formats (PDF, DOCX, etc.) into plain text format. Then, local hash deduplication processing is performed. The document content is calculated using a hash algorithm (such as SHA-256) to generate a unique hash value to exclude duplicate documents. Finally, multimodal feature extraction is performed on the document to be retrieved, and feature extraction tools (such as TF-IDF or BERT models) are used to analyze the document content, extract features of text, images, and other media content, and form a multimodal document feature vector.

[0085] Step S2: performing feature spectrum clustering processing on multimodal document features to generate a document multidimensional feature spectrum; performing associated entity analysis on the document multidimensional feature spectrum to generate a feature associated entity structure;

[0086] In this embodiment, the extracted feature vectors are standardized to ensure that each feature dimension has the same scale. Subsequently, a spectral clustering algorithm (e.g., a spectral clustering method based on the Laplacian matrix) is used to calculate the similarity matrix between features, and the data is mapped to a low-dimensional space through feature spectrum decomposition. Next, an associated entity analysis is performed, and natural language processing tools (such as SpaCy or Stanford NLP) are used to identify entities in the document (such as names, place names, organization names, etc.), and an association relationship map between entities is constructed, and finally a feature-associated entity structure is generated to represent the relationship and similarity between features.

[0087] Step S3: Modeling the document topic for the multimodal document features according to the feature-associated entity structure to obtain a topic-feature matrix; positioning the topic-feature matrix in a semantic direction to obtain the document semantic topic;

[0088] In this embodiment, key entities are extracted from the feature-associated entity structure as preliminary candidates for the document topic. Subsequently, the words extracted from the document are mapped to the topic through the document-word matrix construction. Next, semantic direction positioning is performed, and the words are mapped to a high-dimensional space using word embedding technology (such as Word2Vec or FastText), and the semantic similarity between different topics is calculated, and finally the semantic topic of the document is determined to facilitate subsequent retrieval operations.

[0089] Step S4: Perform multi-way parallel search on the documents to be searched and the knowledge database based on the document semantic theme to obtain a candidate word vector set; perform word vector search on the candidate word vector set to generate word vector search data;

[0090] In this embodiment, the semantic topic of the document is converted into a query word vector representation. The subject word is vectorized using a pre-trained word embedding model (such as GloVe or BERT) to generate a query vector. Next, a word vector index structure is constructed, for example, an efficient vector index is constructed using the FAISS library to support fast retrieval. The query word vector is similar to the word vector in the knowledge database (such as cosine similarity), the relevant word vectors are screened out, and word vector retrieval data is generated as input for subsequent sentence vector retrieval.

[0091] Step S5: Based on the document semantic theme, multi-level semantic matching is performed on the word vector retrieval data and the knowledge database to obtain a set of candidate sentences; sentence vector retrieval is performed on the candidate sentence set to generate a document retrieval result;

[0092] In this embodiment, the retrieved related word vectors are aggregated to generate a sentence-level vector representation. Sentence embedding technology (such as Sentence-BERT) is used to convert the aggregated word vectors into sentence vectors. Then, the knowledge database is searched using the constructed sentence vector index, and a similarity calculation method (such as Euclidean distance or Manhattan distance) is used to find the candidate sentences most similar to the sentence vector. Finally, the search results are integrated to generate document retrieval results, including sentences related to the original document topic.

[0093] Step S6: construct a search engine for the word vector search data and the document search results to obtain an intelligent search engine; use the intelligent search engine to perform intelligent search on the knowledge database to execute a feature search method for the knowledge database.

[0094] In this embodiment, an inverted index or vector index is used as an index structure to quickly access and retrieve data. Then, the word vector retrieval data is integrated with the document retrieval results to form a knowledge graph of the retrieval engine, which supports multi-level queries. Subsequently, a machine learning algorithm (such as a random forest or a support vector machine) is used to train the retrieval model and optimize the retrieval strategy to improve retrieval accuracy. Finally, the intelligent retrieval engine is used to perform intelligent retrieval on the knowledge database, and a feature retrieval method is executed to ensure that relevant information can be quickly and accurately returned during the retrieval process.

[0095] Preferably, step S1 comprises the following steps:

[0096] Step S11: Obtain the document to be retrieved;

[0097] Step S12: Cut the document to be retrieved into fixed lengths to obtain local segments of the document to be retrieved;

[0098] Step S13: performing hash deduplication processing on the local fragments of the document to be retrieved to generate deduplicated document data;

[0099] Step S14: extract multimodal features from the deduplicated document data to obtain multimodal document features.

[0100] In this embodiment, the acquisition process of the document to be retrieved starts from a legal and compliant literature database or public resource, and selects a data source that complies with intellectual property laws and regulations, such as the National Library, university open resources, or authorized professional databases. When each document is extracted, its format is first verified to ensure that common file types such as PDF, Word, and TXT can be processed. For PDF and Word documents, PDFMiner or Apache POI library is used for parsing to extract plain text content. All acquired documents are stored in a local database, and metadata, including document ID, title, author, acquisition time, and other information, are recorded for subsequent retrieval and tracking. The document is read as a string. A fixed length, such as 512 characters, is set to divide the document into multiple local fragments. A cutting program is written in Python to traverse the document string and cut according to the set length. In order to avoid cutting into the middle of a sentence, the cutting program will first find the nearest period or line break to ensure the integrity of each fragment. After cutting, the local fragments are stored in a list, and a unique ID is generated for each fragment for subsequent processing and indexing. After cutting, all local fragments are stored in the database in JSON format to facilitate subsequent deduplication processing. Calculate the SHA-256 hash value for each fragment to ensure uniqueness. Use the hashlib library in Python to write a hash calculation function that takes each local fragment as input and generates the corresponding hash value. Create a collection to store the generated hash values. Check whether the hash value of the current fragment already exists in the collection to determine whether the fragment is repeated. Duplicate fragments will be discarded, and only unique fragment data will be retained to form a deduplicated document dataset. After processing, the deduplicated fragments are stored in a new database table with corresponding metadata for use in the subsequent feature extraction stage. Extract text features from the deduplicated local fragments. Use the TF-IDF algorithm to generate a word frequency vector representation, and combine the Scikit-learn library to implement text vectorization. For related image data, use the OpenCV library to read the image files associated with the text, resize the images to 224x224 pixels, and extract image feature vectors through a pre-trained convolutional neural network (such as VGG16 or ResNet). If the document contains audio data, use the Librosa library for audio signal processing and extract Mel spectrum features. All extracted features (text features, image features, audio features) are integrated into a multimodal feature vector and stored in the database for subsequent topic modeling and retrieval operations.

[0101] Preferably, step S2 comprises the following steps:

[0102] Step S21: performing local clustering on the multimodal document features at the feature level to obtain spatial dimension reduction document features;

[0103] Step S22: clustering the feature spectrum of the spatial dimension reduction document features to generate a multi-dimensional feature spectrum of the document;

[0104] Step S23: performing related entity mapping processing on the document multi-dimensional feature spectrum to obtain the related path between entities;

[0105] Step S24: simulate entity connection processing on multimodal document features based on the inter-entity association paths to generate a feature-associated entity structure.

[0106] As an example of the present invention, refer to Figure 2 As shown, in this example, step S2 includes:

[0107] Step S21: performing local clustering on the multimodal document features at the feature level to obtain spatial dimension reduction document features;

[0108] In the present embodiment, the number of clusters K is set, and the K value is usually selected to be 10, which is adjusted according to the characteristics of the data set. Using the Scikit-learn library in Python, the multimodal feature data is input into the K-means algorithm, and multiple iterations are performed to optimize the clustering results. During each iteration, the distance from each eigenvector to each cluster center is calculated, feature allocation is performed, and the cluster center position is updated. After clustering is completed, each eigenvector is assigned to its corresponding cluster to form a local feature set. The results of each cluster will be recorded, including the cluster center and the eigenvector list belonging to the cluster, and finally these cluster data are stored in the database in JSON format for subsequent processing.

[0109] Step S22: clustering the feature spectrum of the spatial dimension reduction document features to generate a multi-dimensional feature spectrum of the document;

[0110] In the present embodiment, principal component analysis (PCA) is used to reduce the dimension of local clustering results, reducing the feature dimension to 2D or 3D space to ensure easier visualization and clustering. Using the PCA module in Scikit-learn, input local clustering feature data, set the target dimension, and extract the principal component. After completing the dimensionality reduction, the spectral clustering algorithm is applied, and the SpectralClustering method is used to cluster the features after dimensionality reduction. In the specific implementation, the similarity matrix of the features is calculated, and the feature spectrum is clustered by the Laplace matrix to identify the potential association between different features. Finally, the multidimensional feature spectrum of the document is generated, and each spectrum corresponds to different clustering results. The data will be stored in the database in a tabular form for subsequent analysis and retrieval.

[0111] Step S23: performing related entity mapping processing on the document multi-dimensional feature spectrum to obtain the related path between entities;

[0112] In this embodiment, key entities in the feature spectrum are identified, and mapping relationships between features and entities are constructed. Using a graph-based structure, edges between features and entities are established, and the weights of edges are calculated based on feature similarity. In this process, a graph database (such as Neo4j) is used for entity mapping management, and the relationship between features and entities is inserted through the Cypher query language. The association path between entities is determined by the similarity between features, and the calculated path will be represented in the form of a graph, with nodes as entities and edges as feature similarities. Each path will be marked for subsequent analysis, and the resulting association path between entities will be stored in the graph database for query and retrieval.

[0113] Step S24: simulate entity connection processing on multimodal document features based on the inter-entity association paths to generate a feature-associated entity structure.

[0114] In this embodiment, the association path between entities is extracted from the graph database, and the starting node and the ending node of the path are obtained. For each path, the comprehensive feature representation of the connection path is calculated by aggregating the feature vectors on the path. This operation uses a linear weighted method to weight the feature vectors on the path to obtain the connection feature vector of the simulated entity. The generated feature vector is combined with the original multimodal document features, and a feature-associated entity structure is generated through a matching algorithm. This structure will be stored in a relational database for subsequent knowledge retrieval and query, and provide structured data support for downstream retrieval processes.

[0115] Preferably, step S24 comprises the following steps:

[0116] Perform path conflict point detection on the associated paths between entities and generate associated path conflict points;

[0117] According to the preset path weight distribution data, weights are distributed to the associated paths between entities to obtain a weighted path set;

[0118] Performing simulated entity mapping processing on multimodal document features based on a weighted path set to generate an initial simulated entity connection model;

[0119] According to the initial simulated entity connection model, the associated path conflict points are corrected to generate a feature associated entity structure.

[0120] In this embodiment, all associated paths are extracted from the graph database. Using a graph analysis tool (such as the NetworkX library), the paths are analyzed and a similarity matrix between paths is constructed. By calculating the number of nodes shared on each path, paths with similar features are identified. On this basis, the criteria for conflict points are defined. For example, when two paths share nodes exceeding a set threshold, it is considered that there is a conflict. These conflict points are recorded in a new data structure for subsequent processing. After the detection is completed, the generated associated path conflict points will be stored in a table form for subsequent use. The preset path weight standard is obtained from historical data or domain experts to establish a weight distribution model. For example, using a frequency-based distribution method, the frequency of each path appearing in the document feature is calculated, and normalized according to the frequency to form a weight value. In this process, a Python script is used to read the path data, calculate the weight of each path, and associate it with the path information. Finally, all paths and their corresponding weights will be stored in the database to form a weighted path set. The weighted path set is loaded and mapped with multimodal document features. To this end, the feature vectors on the path are weighted combined using a weighted summation method. In the specific implementation, each path is traversed, and each feature vector is multiplied by the corresponding weight to calculate the comprehensive feature representation of each path. The NumPy library is used for vector operations to ensure computational efficiency. After the feature representations of all paths are generated, the initial simulated entity connection model is constructed and represented as a high-dimensional feature vector set. Finally, the generated model will be stored in the database for subsequent retrieval. The initial simulated entity connection model and the detected path conflict points are loaded. The preset correction rules are used, for example, conflicting paths are selected based on path weights, and paths with higher weights are retained first. By analyzing the conflict points, graph algorithms (such as the Dijkstra algorithm) are applied to replan the path to ensure the validity of the connection path. In this process, the update function of the graph database is used to adjust the data structure of the conflicting path. Finally, the generated corrected feature-associated entity structure will be stored for subsequent knowledge database retrieval to ensure the accuracy and completeness of the retrieval results.

[0121] Preferably, step S3 comprises the following steps:

[0122] Step S31: Perform graph convolutional network analysis on the feature-associated entity structure to obtain entity semantic representation;

[0123] Step S32: performing non-negative matrix decomposition on the entity semantic representation to obtain a topic-feature matrix;

[0124] Step S33: performing multidimensional scaling analysis on the topic-feature matrix to obtain a low-dimensional semantic coordinate system;

[0125] Step S34: performing density peak clustering on the low-dimensional semantic coordinate system to obtain the document semantic topic representation.

[0126] As an example of the present invention, refer to Figure 3 As shown, in this example, step S3 includes:

[0127] Step S31: Perform graph convolutional network analysis on the feature-associated entity structure to obtain entity semantic representation;

[0128] In this embodiment, the feature-associated entity structure is converted into a graph, in which nodes represent entities and edges represent the associations between entities. The PyTorch Geometric library is used to build a graph convolutional network model, and node features and adjacency matrices are input. By defining a graph convolutional layer, the node features are updated using local neighborhood aggregation to generate a new node representation. The parameters of multiple convolutional layers are set to ensure the propagation and learning of features at multiple levels. During the model training process, the cross entropy loss function is used, and the optimizer uses Adam to adjust the model parameters through back propagation. After the training is completed, the obtained entity semantic representation will be used to generate the subject-feature matrix in the subsequent steps.

[0129] Step S32: performing non-negative matrix decomposition on the entity semantic representation to obtain a topic-feature matrix;

[0130] In this embodiment, the entity semantic representation is organized into a matrix form to ensure that all elements in the matrix are non-negative values. The non-negative matrix decomposition (NMF) algorithm is constructed using Python's NumPy library, and the rank k value of the decomposition is set, that is, the number of topics you want to obtain. The NMF algorithm is executed, and the optimization goal is to minimize the difference between the original matrix and the reconstructed matrix while keeping all matrix elements non-negative. After multiple iterations, the two matrices generated will represent the weights of the topic-feature matrix and the feature, respectively. The final topic-feature matrix will be stored in the database as the basis for subsequent analysis.

[0131] Step S33: performing multidimensional scaling analysis on the topic-feature matrix to obtain a low-dimensional semantic coordinate system;

[0132] In the present embodiment, in the implementation process of multidimensional scaling analysis of the subject-feature matrix, the subject-feature matrix is ​​first imported, and the multidimensional scaling analysis (MDS) module in the scikit-learn library is used. The distance measurement method is set to Euclidean distance, and a random state is used during the analysis to ensure the repeatability of the results. Taking the subject-feature matrix as input, the MDS algorithm will map high-dimensional data to a low-dimensional space by calculating the distance between points, and usually chooses to reduce the data to two or three dimensions. After executing the algorithm, the low-dimensional representation obtained will reflect the relative relationship between the topics and be stored as a low-dimensional semantic coordinate system data structure for subsequent cluster analysis.

[0133] Step S34: performing density peak clustering on the low-dimensional semantic coordinate system to obtain the document semantic topic representation.

[0134] In this embodiment, the local density and distance value of each point are calculated using the density peak clustering algorithm based on the low-dimensional representation data. The density of each point is sorted to determine the cluster center, and then clustering is performed according to the threshold of distance and density. During the clustering process, appropriate parameters are set to ensure the stability and effectiveness of clustering and the distinction between different semantic topics. The clustering results will generate document semantic topic representations, which are stored in the database as the subject information basis for knowledge retrieval, so as to facilitate subsequent document matching and retrieval operations.

[0135] Preferably, step S4 comprises the following steps:

[0136] Step S41: perform query word expansion on the document semantic topic to obtain a query word vector representation;

[0137] Step S42: Perform distributed semantic modeling on the query word vector representation and the knowledge database to obtain a word vector index;

[0138] Step S43: performing multi-path parallel retrieval on the query word vector representation according to the word vector index to obtain a candidate word vector set;

[0139] Step S44: Calculate the cosine similarity of the candidate word vector set to generate a similarity sorted list;

[0140] Step S45: reorder the candidate word vector set based on the similarity sorting list to generate word vector retrieval data.

[0141] In this embodiment, the semantic theme of the document is analyzed, and the core concepts and keywords are extracted by natural language processing technology. Using a word embedding model, such as Word2Vec or GloVe, the vocabulary related to the core concept is found from the pre-trained model. These related words are used as the initial query words, and a synonym library (such as WordNet) or a contextual vocabulary analysis tool is further used to expand multiple synonyms and related words. The expanded vocabulary is calculated by the model to obtain a query word vector representation, which is usually vectorized using a deep learning framework such as TensorFlow or PyTorch. Finally, the generated query word vector representation will be used as the input of the subsequent steps to build the basis for the query. The query word vector is combined with the documents in the knowledge database. A semantic modeling method, such as Latent Semantic Analysis (LSA) or Deep Learning-based embedding technology, is used to construct a word vector index. By using a model pre-trained with a large-scale text corpus, the query word vector is mapped to the semantic space of the existing documents in the knowledge database. In the specific operation, a library such as Scikit-learn or TensorFlow is used for matrix decomposition and semantic similarity calculation. Finally, the generated word vector index will be stored in the database to provide efficient semantic matching capabilities for subsequent retrieval. Match the query word vector representation with the word vector index. Use the GPU-accelerated deep learning framework to implement parallel computing. According to the similarity between the query word vector and the word vector index, the KNN (K-NearestNeighbor) algorithm is used for retrieval. In the specific implementation, efficient retrieval libraries such as Faiss are used to process large-scale vector data, and retrieval parameters are set to ensure speed and accuracy. Finally, the candidate word vector set obtained from the knowledge database will be used for subsequent similarity calculations. The query word vector and the candidate word vector are organized into a matrix format to improve computational efficiency. Use efficient numerical computing libraries such as NumPy to perform pairwise cosine similarity calculations. By performing inner product calculations on each candidate word vector and the query word vector, combined with vector normalization, the similarity score between each pair of vectors is obtained. This process ensures the accuracy and reliability of the similarity score. After the calculation is completed, the generated similarity score will be stored in a list, laying the foundation for the subsequent candidate word re-ranking. Associate the similarity score list with the candidate word vector set. Use efficient sorting algorithms, such as quick sort or merge sort, to sort the candidate word vector set according to the similarity score to ensure that the candidate words with high similarity are ranked first. During the sorting process, ensure that each candidate word vector is consistent with its corresponding similarity score in order to generate accurate sorting results. Finally, the resulting re-sorted candidate word vector set will be organized and saved as word vector retrieval data to facilitate subsequent retrieval system calls and use.

[0142] Preferably, step S45 includes the following steps:

[0143] Perform multi-dimensional sorting evaluation on the candidate word vector set based on the similarity sorting list to obtain a multi-dimensional sorting score matrix;

[0144] According to the multi-dimensional sorting score matrix, the candidate word vector set is adjusted in ascending order of score to generate a dynamically sorted candidate word set;

[0145] Multimodal feature fusion is performed on the dynamically sorted candidate word set to obtain word vector retrieval data.

[0146] In this embodiment, the similarity sort list is converted into a multidimensional scoring matrix. Using the NumPy library in Python, the various feature dimensions of the candidate word vector set are combined to form a matrix. The scores of multiple feature dimensions are calculated for each candidate word vector, including word frequency, context similarity and other relevant features. By normalizing these features, ensure that each dimension is evaluated at the same scale. Then, using a weighted algorithm, the scores of each dimension are merged to form a comprehensive scoring matrix. The multidimensional score of each candidate word will be included in the matrix to facilitate subsequent sorting adjustments. Each score in the scoring matrix is ​​extracted and organized into a one-dimensional array. This array is processed using a quick sorting algorithm to achieve efficient ascending sorting. During the sorting process, ensure that the index of each candidate word vector is consistent and that the sorted array corresponds to the original candidate word vector set. According to the scores after ascending order, the candidate word vector set is reordered to generate a dynamically sorted candidate word set. The dynamic set will automatically adjust according to the comprehensive score of each candidate word, thereby ensuring the relevance and accuracy of the retrieval results. Identify multiple features of each candidate word, including text features, image features and other related information. Use deep learning frameworks such as TensorFlow or PyTorch to build a multimodal feature fusion model. Input the features of each candidate word into the model and process them using feature mapping technology to extract common and unique features. By combining convolutional neural networks (CNN) and recurrent neural networks (RNN), a comprehensive feature representation of the candidate words is obtained. Ultimately, the word vector retrieval data obtained will integrate information from different modalities to form a more representative feature set to support subsequent knowledge database retrieval tasks.

[0147] Preferably, step S5 comprises the following steps:

[0148] Step S51: Perform syntactic dependency analysis on the document semantic topics to obtain a set of grammatically structured sentences;

[0149] Step S52: performing multi-level semantic matching on the knowledge database according to the grammatically structured sentence set to obtain a candidate sentence set;

[0150] Step S53: labeling the candidate sentence set with semantic roles to generate a semantically structured candidate set;

[0151] Step S54: constructing a semantic graph for the semantically structured candidate set to obtain a sentence relationship network;

[0152] Step S55: Perform sentence vector retrieval on word vector retrieval data and knowledge database based on the sentence relationship network to generate document retrieval results.

[0153] In this embodiment, natural language processing tools such as spaCy or NLTK are used to segment the sentences in the document. Then, a dependency syntactic analysis model is used to identify the dependencies between words in the sentence and generate a dependency tree. The analysis will provide the parent node, child node and dependency type of each word. The analysis results are stored in the form of a data structure, including the vocabulary, part-of-speech tagging and dependency relationship of the sentence. Finally, based on the analysis results, a grammatically structured sentence set is constructed to facilitate subsequent multi-level semantic matching. The grammatically structured sentence set is compared with the text in the knowledge database. The sentence is converted into a vector representation using TF-IDF or a word embedding model (such as Word2Vec or BERT). Then, the relevance score is obtained by calculating the similarity between the sentence vector and each entry in the knowledge database. In order to improve the matching effect, a multi-level strategy is adopted to classify the sentences by subject, and high-level and low-level semantic matching is performed respectively. Finally, according to the similarity score, a set of candidate sentences that meet the conditions is screened out for subsequent processing. When semantic role labeling the candidate sentence set, use a pre-trained semantic role labeling model, such as PropBank or FrameNet. First, identify the predicate and its corresponding argument of each sentence and determine the semantic role of each component in the sentence. Decompose each sentence into predicates and corresponding roles to form a structured semantic representation. Use the labeling tool to process each sentence and convert it into a format containing predicate and its argument information to generate a semantic structured candidate set. This process ensures the accuracy and consistency of subsequent analysis. Create a graph structure in which nodes represent predicates in sentences and edges represent semantic relationships between sentences. Use a Graph Database (such as Neo4j) to extract the predicates and arguments in the structured candidate set as nodes, and identify and establish relationship edges between sentences based on the results of semantic role labeling. Label each relationship to indicate the type of semantic connection (such as causal relationship, contrast relationship, etc.). Finally, a relationship network between sentences is formed to support subsequent sentence vector retrieval. Extract the node and edge information in the constructed relationship network. Then, combine the existing word vector retrieval data and compare it with the relevant entries in the knowledge database. A graph convolutional network (GCN) is used to model the relationship between sentences, and the vector representation of sentences is enhanced through a propagation mechanism. Finally, document retrieval results are generated, providing highly relevant sentences for subsequent analysis or display.

[0154] Preferably, step S52 includes the following steps:

[0155] The grammatically structured sentence set is processed by a bidirectional long short-term memory network to obtain a context-aware sentence vector representation;

[0156] Perform multi-level hierarchical encoding on the knowledge database to generate a hierarchical sentence vector space;

[0157] Build a multi-level index structure for the hierarchical sentence vector space to obtain an efficient sentence vector retrieval tree;

[0158] The context-aware sentence vector representation is used to perform multi-level semantic matching on the efficient sentence vector retrieval tree to obtain a set of candidate sentences.

[0159] In this embodiment, each sentence is converted into a word embedding vector, usually using a pre-trained word vector such as GloVe or Word2Vec. These word vectors will be used as the input of the Bi-LSTM network. Next, a model containing two LSTM layers, forward and reverse, is constructed, the forward LSTM processes the order information of the sentence, and the reverse LSTM processes the reverse information of the sentence. Through the cyclic calculation of the time step, the context information of each word is enhanced, and finally the outputs of the two LSTM layers are combined to generate a context-aware sentence vector representation. The sentence vector will capture the dependencies between the words in the sentence to form a semantically rich sentence representation. The sentences in the knowledge database are divided into different levels, usually classified based on themes or semantic similarity. Each layer is processed using a different encoder, for example, the first layer uses a simple TF-IDF encoding, and the second layer uses a deep learning model such as BERT for semantic encoding. After encoding each sentence, a vector representation of the sentence is generated to form a hierarchical sentence vector space. These sentence vectors have a clear hierarchical structure in space, which is convenient for subsequent retrieval and matching. Finally, a multi-level hierarchical coding structure that comprehensively covers the knowledge database is formed. Select an indexing algorithm, such as an inverted index or a tree index structure (e.g., a KD tree). Based on the hierarchical sentence vectors, insert the sentence vectors of each level into the corresponding index structure. For each sentence vector, record its position information in the knowledge database. During the construction process, index each level independently to ensure efficient query. Each index node stores the sentence vector and its related information to improve retrieval efficiency. Finally, an efficient sentence vector retrieval tree is formed to support fast positioning and matching of related sentences. Compare the context-aware sentence vector representation with the nodes in the efficient sentence vector retrieval tree. Use similarity calculation methods, such as cosine similarity, to evaluate the similarity between the context sentence vector and each node sentence vector. Based on the similarity score, select a set of candidate sentences that meet the threshold from the tree. Retrieve multiple levels of indexes to ensure that semantic information at different levels is covered. Finally, a set of candidate sentences is generated for subsequent processing or display, ensuring that the matched sentences have high semantic relevance.

[0160] Preferably, step S6 comprises the following steps:

[0161] Step S61: Perform deep semantic embedding on the word vector retrieval data to obtain a fast word vector retrieval structure;

[0162] Step S62: extracting semantic relations from the document retrieval results to obtain a sentence vector retrieval module;

[0163] Step S63: Designing a joint search strategy for the fast word vector search structure and the sentence vector search module to obtain an intelligent search engine;

[0164] Step S64: Use the intelligent search engine to perform intelligent search on the knowledge database to execute the feature search method of the knowledge database.

[0165] In this embodiment, a pre-trained deep learning model, such as BERT or Word2Vec, is used to convert word vectors into high-dimensional semantic representations. The contextual information of each word will be captured by the trained model to generate a deep semantic embedding of the word vector. This process includes inputting the word vector into the model, forward propagating through a multi-layer neural network, and extracting the deep semantic features of the word. The generated high-dimensional semantic representation will be organized into a fast word vector retrieval structure to facilitate subsequent fast retrieval operations. The construction of this structure will help improve retrieval efficiency and ensure rapid response to user queries. Sentences are extracted from documents and natural language processing is performed to identify key entities and relationships in sentences. Using the named entity recognition (NER) algorithm, important noun phrases are marked as entities. Subsequently, relationship extraction techniques, such as sequence annotation models or graph convolutional networks based on deep learning, are applied to determine the semantic relationships between entities. The generated sentence vector retrieval module will contain structured representations of entities and their relationships, forming a semantic relationship network for use in subsequent retrieval steps. This process ensures that the extracted relationships have sufficient semantic information to facilitate accurate retrieval. Define the hierarchical relationship of the retrieval strategy. The upper-level search performs fast word vector search for keywords or topics, while the lower-level search performs precise matching for sentence vectors. In the implementation process, a connection mechanism is constructed to establish a connection between the two levels by selecting an appropriate similarity calculation method (such as cosine similarity). Whenever a search is performed, the relevant word vectors are first found through the fast word vector search structure, and then mapped to the sentence vector search module for detailed search. This joint search strategy will improve the comprehensiveness and accuracy of the search and provide users with a more intelligent search experience. According to the query words entered by the user, they are converted into word vector representations and input into the fast word vector search structure for preliminary search. Based on the search results, relevant document information is extracted, and then the sentence vector search module is used to conduct in-depth analysis of these documents to obtain more contextual information. By combining the results of the two search structures, the intelligent search engine can comprehensively match the content in the knowledge database. The final search results will be provided to users to meet their information needs in an efficient and accurate manner. This process continuously iterates and optimizes the search strategy to ensure the effective execution of the feature retrieval method of the knowledge database.

[0166] Therefore, the embodiments should be regarded as illustrative and non-restrictive from all points, and the scope of the present invention is limited by the appended claims rather than the above description, and it is therefore intended that all changes falling within the meaning and range of equivalent elements of the application documents are included in the present invention.

[0167] The above description is only a specific embodiment of the present invention, so that those skilled in the art can understand or implement the present invention. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but should conform to the widest scope consistent with the principles and novel features invented herein.

Claims

1. A knowledge database retrieval method based on feature extraction, characterized in that: The following steps are involved: Step S1: Obtain the document to be retrieved; perform local hash deduplication processing on the document to be retrieved, and perform multimodal feature extraction to obtain multimodal document features; Step S2: performing feature spectrum clustering processing on multimodal document features to generate a document multidimensional feature spectrum; performing associated entity analysis on the document multidimensional feature spectrum to generate a feature associated entity structure; Step S3: Document topic modeling is performed on multimodal document features according to the feature-associated entity structure to obtain a topic-feature matrix; Position the semantic direction of the topic-feature matrix to obtain the document semantic topic; Step S4: Perform multi-way parallel search on the documents to be searched and the knowledge database based on the document semantic theme to obtain a candidate word vector set; perform word vector search on the candidate word vector set to generate word vector search data; Step S5: Based on the document semantic theme, multi-level semantic matching is performed on the word vector retrieval data and the knowledge database to obtain a set of candidate sentences; sentence vector retrieval is performed on the candidate sentence set to generate a document retrieval result; Step S6: constructing a search engine for the word vector search data and the document search results to obtain an intelligent search engine; The intelligent retrieval engine is used to perform intelligent retrieval on the knowledge database to implement the feature retrieval method of the knowledge database.

2. The knowledge database retrieval method based on feature extraction according to claim 1 is characterized in that: Step S1 includes the following steps: Step S11: Obtain the document to be retrieved; Step S12: Cut the document to be retrieved into fixed lengths to obtain local segments of the document to be retrieved; Step S13: performing hash deduplication processing on the local fragments of the document to be retrieved to generate deduplication document data; Step S14: extract multimodal features from the deduplicated document data to obtain multimodal document features.

3. The knowledge database retrieval method based on feature extraction according to claim 1 is characterized in that: Step S2 includes the following steps: Step S21: performing local clustering on the multimodal document features at the feature level to obtain spatial dimension reduction document features; Step S22: clustering the feature spectrum of the spatial dimension reduction document features to generate a multi-dimensional feature spectrum of the document; Step S23: performing related entity mapping processing on the document multi-dimensional feature spectrum to obtain the related path between entities; Step S24: simulate entity connection processing on multimodal document features based on the inter-entity association paths to generate a feature-associated entity structure.

4. The knowledge database retrieval method based on feature extraction according to claim 3 is characterized in that: Step S24 includes the following steps: Perform path conflict point detection on the associated paths between entities and generate associated path conflict points; According to the preset path weight distribution data, weights are distributed to the associated paths between entities to obtain a weighted path set; Performing simulated entity mapping processing on multimodal document features based on a weighted path set to generate an initial simulated entity connection model; According to the initial simulation entity connection model, the associated path conflict points are corrected to generate a feature associated entity structure.

5. The knowledge database retrieval method based on feature extraction according to claim 1 is characterized in that: Step S3 includes the following steps: Step S31: Perform graph convolutional network analysis on the feature-associated entity structure to obtain entity semantic representation; Step S32: performing non-negative matrix decomposition on the entity semantic representation to obtain a topic-feature matrix; Step S33: performing multidimensional scaling analysis on the topic-feature matrix to obtain a low-dimensional semantic coordinate system; Step S34: performing density peak clustering on the low-dimensional semantic coordinate system to obtain the document semantic topic representation.

6. The knowledge database retrieval method based on feature extraction according to claim 1 is characterized in that: Step S4 includes the following steps: Step S41: perform query word expansion on the document semantic topic to obtain a query word vector representation; Step S42: Perform distributed semantic modeling on the query word vector representation and the knowledge database to obtain a word vector index; Step S43: performing multi-path parallel retrieval on the query word vector representation according to the word vector index to obtain a candidate word vector set; Step S44: Calculate the cosine similarity of the candidate word vector set to generate a similarity sorted list; Step S45: reorder the candidate word vector set based on the similarity sort list to generate word vector retrieval data.

7. The knowledge database retrieval method based on feature extraction according to claim 6 is characterized in that: Step S45 includes the following steps: Perform multi-dimensional sorting evaluation on the candidate word vector set based on the similarity sorting list to obtain a multi-dimensional sorting score matrix; According to the multi-dimensional sorting score matrix, the candidate word vector set is adjusted in ascending order of score to generate a dynamically sorted candidate word set; Multimodal feature fusion is performed on the dynamically sorted candidate word set to obtain word vector retrieval data.

8. The knowledge database retrieval method based on feature extraction according to claim 1 is characterized in that: Step S5 includes the following steps: Step S51: Perform syntactic dependency analysis on the document semantic topics to obtain a set of grammatically structured sentences; Step S52: performing multi-level semantic matching on the knowledge database according to the grammatically structured sentence set to obtain a candidate sentence set; Step S53: labeling the candidate sentence set with semantic roles to generate a semantically structured candidate set; Step S54: constructing a semantic graph for the semantically structured candidate set to obtain a sentence relationship network; Step S55: Perform sentence vector retrieval on word vector retrieval data and knowledge database based on the sentence relationship network to generate document retrieval results.

9. The knowledge database retrieval method based on feature extraction according to claim 8 is characterized in that: Step S52 includes the following steps: The grammatically structured sentence set is processed by a bidirectional long short-term memory network to obtain a context-aware sentence vector representation; Perform multi-level hierarchical encoding on the knowledge database to generate a hierarchical sentence vector space; Build a multi-level index structure for the hierarchical sentence vector space to obtain an efficient sentence vector retrieval tree; The context-aware sentence vector representation is used to perform multi-level semantic matching on the efficient sentence vector retrieval tree to obtain a set of candidate sentences.

10. The knowledge database retrieval method based on feature extraction according to claim 7, characterized in that: Step S6 includes the following steps: Step S61: Perform deep semantic embedding on the word vector retrieval data to obtain a fast word vector retrieval structure; Step S62: extracting semantic relations from the document retrieval results to obtain a sentence vector retrieval module; Step S63: Designing a joint search strategy for the fast word vector search structure and the sentence vector search module to obtain an intelligent search engine; Step S64: Use the intelligent search engine to perform intelligent search on the knowledge database to execute the feature search method of the knowledge database.

Images