Intelligent archive classification and retrieval system and method based on deep learning

Abstract

The invention discloses an intelligent archive classification and retrieval system and method based on deep learning, and relates to the technical field of information retrieval processing. Comprising the steps that an archive feature data set after feature coding is obtained, and the data set comprises text feature data, meta feature data and knowledge feature data; inputting the data set into a preset weight gating model for weighted fusion, then performing classification through a neural network based on deep learning, and establishing topic index data corresponding to archive topic aggregation; after receiving a retrieval request, analyzing the retrieval request into query intention data and adjusting weight parameters of the weight gating model; generating candidate fragment feature data from the candidate archive data and the corresponding feature data set according to the adjusted model; and outputting retrieval result data according to the matching degree of the candidate fragment feature data and the query intention data. According to the method, high-precision classification and retrieval of the archives are realized, and the matching degree of the retrieval result and the query intention of the user is improved.

Description

Technical Field

[0001] This invention relates to the field of information retrieval and processing technology, specifically to an intelligent archive classification and retrieval system and method based on deep learning. Background Technology

[0002] In the field of archival information management and retrieval, as the scale of archival data continues to expand, archival content exhibits diversified and unstructured characteristics. Traditional retrieval methods based on keywords or simple feature matching are unable to accurately capture the complex relationships between text, metadata, and knowledge semantics.

[0003] While existing technologies have incorporated deep learning models for document content classification and retrieval, most of them simply concatenate different types of features and input them into the model uniformly. They lack the ability to dynamically adjust the weights of various features under different retrieval intentions. This results in irrelevant or low-relevance document content in the search results when the user's query intent emphasizes a specific information dimension, reducing the accuracy and practicality of the retrieval and failing to meet the user's specific information needs. Existing technologies generally lack effective processing mechanisms for multi-scale semantic matching.

[0004] Therefore, there is an urgent need for a method for classifying and retrieving archives that can improve the overall performance of classification and retrieval. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an intelligent archive classification and retrieval system and method based on deep learning.

[0006] To achieve the above objectives, the technical solution of the present invention is as follows:

[0007] In a first aspect, this invention discloses an intelligent archive classification and retrieval method based on deep learning, comprising the following steps:

[0008] Obtain a feature-encoded archive feature dataset, which includes text feature data, meta-feature data, and knowledge feature data;

[0009] The archive feature dataset is input into a preset weighted gating model for weighted fusion, then classified using a deep learning-based neural network, and topic index data corresponding to the archive topic aggregation is established.

[0010] Receive retrieval requests, parse them into query intent data, and adjust the weight parameters of the weight gating model;

[0011] The topic index data is input into the adjusted weight gating model to obtain candidate file data;

[0012] The candidate archive data and the corresponding archive feature dataset are input again into the feature encoding network to generate candidate fragment feature data;

[0013] Multi-scale matching is performed on the candidate fragment feature data. The matching degree between the candidate fragment feature data and the query intent data is calculated at different semantic scales of different granularities. Based on the weighted result of the matching degree, the search result data that meets the preset threshold is selected.

[0014] Secondly, this invention discloses a deep learning-based intelligent document classification and retrieval system, which uses the aforementioned deep learning-based intelligent document classification and retrieval method, including:

[0015] The feature encoding module is used to obtain the feature-encoded archive feature dataset, which includes text feature data, meta-feature data, and knowledge feature data.

[0016] The classification index module is used to input the archive feature dataset into a preset weighted gating model for weighted fusion, then classify it through a deep learning-based neural network, and establish topic index data corresponding to the archive topic aggregation.

[0017] The weight gating module is used to receive retrieval requests, parse them into query intent data, and adjust the weight parameters of the weight gating model.

[0018] The recall module is used to input the topic index data into the adjusted weight gating model to obtain candidate file data;

[0019] The feature regeneration module is used to input the candidate archive data and the corresponding archive feature dataset back into the feature encoding network to generate candidate fragment feature data;

[0020] The matching output module is used to perform multi-scale matching on the candidate fragment feature data, calculate the matching degree between the candidate fragment feature data and the query intent data at different semantic scales, and filter out the search result data that meets the preset threshold based on the weighted result of the matching degree.

[0021] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0022] 1. This invention dynamically weights and fuses text features, metadata features, and knowledge features in a weighted gating model, which can adaptively adjust the importance of various features under different search intentions. This allows the classification and retrieval process to highlight the most relevant feature dimensions, thereby significantly improving the matching degree between archive content and user query intent.

[0023] 2. In the candidate file retrieval stage, this invention introduces processing steps of re-encoding and candidate fragment feature generation, so that the retrieval not only stays at the topic aggregation level, but also performs feature reconstruction of information at the fragment granularity, ensuring that the retrieval results can accurately reflect the content in the file that best matches the query intent, and effectively reduce the interference of redundant information and low-relevance content.

[0024] 3. This invention calculates the matching degree between candidate fragments and query intent, and filters target fragments based on weighted results, so that the retrieval results can maintain overall semantic consistency while capturing key information at the detail level, thereby improving the adaptability and accuracy of the retrieval system in complex query tasks. Attached Figure Description

[0025] The disclosure of this invention is illustrated with reference to the accompanying drawings. It should be understood that the drawings are for illustrative purposes only and are not intended to limit the scope of protection of this invention. In the drawings, the same reference numerals are used to refer to the same parts. Wherein:

[0026] Figure 1 This is a flowchart of the steps of the present invention;

[0027] Figure 2 This is a flowchart illustrating the working principle of the present invention;

[0028] Figure 3 This is a flowchart of the feature encoding process of the present invention;

[0029] Figure 4 This is a system module connection diagram of the present invention;

[0030] Figure 5 This is a topology diagram of the system hardware deployment of the present invention. Detailed Implementation

[0031] It is readily understood that, based on the technical solution of this invention, those skilled in the art can propose various interchangeable structural methods and implementations without altering the essential spirit of the invention. Therefore, the following detailed embodiments and accompanying drawings are merely illustrative examples of the technical solution of this invention and should not be considered as the entirety of the invention or as limitations or restrictions on the technical solution of this invention.

[0032] In existing technologies, the field of archival information management and retrieval faces the challenges of diversification and unstructured data brought about by the expansion of archival data scale. Traditional retrieval methods based on keywords or simple feature matching are difficult to effectively capture the complex relationships between text, metadata, and semantic knowledge. Although existing deep learning models can handle document classification and retrieval tasks, they generally use feature concatenation as input to the model and lack dynamic weight adjustment mechanisms for different retrieval intentions. When users' query intentions focus on specific information dimensions, fixed weight fusion can easily lead to retrieval results containing irrelevant or low-relevance archival content, reducing retrieval accuracy. The lack of a multi-scale semantic matching mechanism further limits the system's ability to handle complex semantic relationships.

[0033] To address the aforementioned issues, the inventors discovered that existing methods suffer from technical bottlenecks in both feature fusion and semantic matching. First, fixed-weight feature fusion methods cannot adapt to the differentiated needs of various query intents regarding text, metadata, and knowledge features. Second, single-scale semantic matching struggles to cover multi-level semantic relationships such as sentences, paragraphs, and documents. Based on this, the inventors proposed constructing a dynamic weight gating model that adjusts feature weights in real-time by parsing query intent, while simultaneously introducing a multi-scale matching mechanism. Through the synergistic effect of feature encoding and weight adjustment, the retrieval process achieves accurate responses to user intent.

[0034] After introducing the basic concept of the present invention, the embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0035] Example 1:

[0036] like Figure 1 As shown, a deep learning-based intelligent document classification and retrieval method includes the following steps:

[0037] Obtain the archive feature dataset after feature encoding. The archive feature dataset includes text feature data, meta feature data, and knowledge feature data.

[0038] After the archival feature dataset is input into a pre-defined weighted gating model for weighted fusion, it is classified using a deep learning-based neural network, and topic index data corresponding to the archival topic aggregation is established.

[0039] Receive retrieval requests, parse them into query intent data, and adjust the weight parameters of the weight gating model;

[0040] The topic index data is input into a candidate file data obtained based on an adjusted weighted gating model;

[0041] The candidate archive data and the corresponding archive feature dataset are input again into the feature encoding network to generate candidate fragment feature data;

[0042] Multi-scale matching is performed on the candidate fragment feature data. The matching degree between the candidate fragment feature data and the query intent data is calculated at different semantic scales of different granularities. Based on the weighted result of the matching degree, the search result data that meets the preset threshold is selected.

[0043] like Figure 2 As shown, the working principle of this application is to obtain a feature dataset of archives after feature encoding. The archive feature dataset includes at least text feature data, meta-feature data, and knowledge feature data. Among them, text feature data is used to represent the semantic and syntactic characteristics of the archive text, meta-feature data is used to represent the structured information of the archives, such as attributes like time, author, and type, and knowledge feature data is used to introduce concepts, entities, and their relationships from a domain knowledge graph to improve the comprehensiveness of feature expression.

[0044] The archival feature dataset is input into a pre-defined weighted gating model for weighted fusion. The weighted gating model assigns weights based on the importance of each feature channel, allowing different features to dynamically adjust their contribution during the fusion process, thus improving the targeting and expressive power of feature fusion. The fused feature data is then input into a deep learning-based neural network for classification. Through the neural network's multi-layered feature abstraction and non-linear mapping capabilities, the archives are divided into different thematic categories, and thematic index data corresponding to the archive thematic aggregations is established to support subsequent retrieval.

[0045] During the retrieval phase, retrieval requests are received and parsed into query intent data. This query intent data can be used by a natural language understanding model to identify keywords, entities, and their semantic relationships, thereby more accurately reflecting the user's true retrieval needs. The parsed results are used to dynamically adjust the weight parameters of the weight gating model, enabling the feature fusion process to optimize the importance distribution of feature channels according to retrieval requirements.

[0046] Candidate profile data is obtained by inputting topic index data into a weighted gating model. This process uses optimized feature weights to match the topic index, achieving the selection of a candidate set highly relevant to the query intent.

[0047] After candidate screening, the candidate archive data and their corresponding archive feature datasets are input again into the feature encoding network to generate candidate fragment feature data. The candidate fragment feature data integrates multimodal features, including fragment-level semantic information, contextual features, and metadata features, providing high-dimensional representation support for subsequent multi-scale matching.

[0048] Finally, multi-scale matching is performed on the candidate fragment feature data, calculating the matching degree between the candidate fragment feature data and the query intent data at different semantic scales. Multi-scale matching can simultaneously consider fine-grained semantic precision matching and coarse-grained contextual consistency, thereby improving retrieval relevance at different semantic levels. The matching results are weighted and fused to filter out retrieval result data that meets the preset threshold, achieving an accurate response to the user's query intent.

[0049] This invention achieves a balance between high-precision matching and high recall in the process of archival classification and retrieval by introducing deep learning models and dynamic weight adjustment mechanisms in multiple stages, including feature encoding, feature fusion, semantic parsing, and multi-scale matching. This method not only improves the accuracy and robustness of retrieval but also adaptively adjusts the feature fusion strategy according to different retrieval needs, thereby significantly optimizing the efficiency of archival management and information acquisition.

[0050] like Figure 3 As shown, this application further proposes that the specific steps for obtaining the feature-encoded archival feature dataset include:

[0051] Acquire the archival text data, the archival metadata corresponding to the archival text data, and the knowledge anchor data mapped from the archival text data, and encode them by feature respectively;

[0052] Extract text feature data from archival text data; obtain the original archival text data, perform structured segmentation according to title / chapter / section / paragraph / sentence level, and generate text fragment data with hierarchical tags; encode the text fragment data respectively, and output text feature data;

[0053] Extract meta-feature data from archival metadata;

[0054] A knowledge graph is constructed based on knowledge anchor data, and knowledge feature data is mapped from the knowledge graph. Specifically, knowledge node data is obtained from the knowledge graph, which includes terms, legal clauses, and related timeliness status. The archival text data is aligned with the knowledge node data to generate knowledge anchor association data.

[0055] Based on knowledge anchor points, data and time status are associated with each other to obtain knowledge feature data;

[0056] Sentence segmentation is performed on the archival text data to extract fragment-level text features.

[0057] The acquisition of archival text data may include reading text fields from archival databases, full-text indexes, or external data sources. After acquiring the original archival data, structured segmentation is performed based on the hierarchical relationship of titles / chapters / sections / paragraphs / sentences to generate hierarchically labeled text fragments. Hierarchical labels can use tree indexes or nested identifiers to preserve the original document structure information. The text fragments are then input into a text feature encoding network to extract text feature data, including semantic vectors, positional encodings, and context window features, and output to the archival feature dataset.

[0058] Meta-feature data is extracted from the archive metadata, which may include, but is not limited to, structured attributes such as creation date, author, version number, file type, and business tags. This metadata is then mapped into a vectorized representation through the meta-feature encoding module so that it can be integrated with other features in a unified feature space.

[0059] A knowledge graph is constructed based on knowledge anchor data. The rules for constructing the knowledge graph include extracting 'debt-to-asset ratio' and 'cash flow coverage ratio' as key factors from the financial regulatory database. The knowledge graph may include knowledge node data and their edge relationships. The knowledge node data includes at least domain terminology, regulatory clauses, and their associated timeliness status.

[0060] In addition, sentence segmentation is performed on the archival text data to extract fragment-level text features at the sentence level.

[0061] To further enhance the robustness and generalization ability of feature representation, this application may optionally perform normalization processing on text feature data, meta-feature data and knowledge feature data, and introduce a feature masking mechanism to reduce interference from invalid information.

[0062] This application enables the simultaneous fusion of document semantic content, structured meta-information, and domain knowledge information in a feature dataset, forming a multi-dimensional, high-information-density feature representation. This multimodal feature set provides rich feature support for subsequent weighted gating fusion and deep learning classification, not only improving classification accuracy but also significantly enhancing the multi-scale matching effect in the retrieval stage. This allows the system to maintain a balance between high precision and high recall when handling file retrieval requests of different types and contexts.

[0063] This application further proposes that, in receiving a retrieval request, parsing it into query intent data, and adjusting the weight parameters of the weight gating model, the specific steps for dynamically adjusting the weight parameters include:

[0064] Extract time constraints, exact match indicators, and concept expansion indicators from query intent data;

[0065] Acquire user behavior log data and generate feedback intensity data based on the user behavior log data. The user behavior log data includes click location, dwell time and secondary search behavior.

[0066] Based on time constraints, exact matching flags, concept expansion flags, and feedback intensity data, update the weight quadruples data of the weight gating model;

[0067] The weighted gating model includes text channels, metadata channels, knowledge channels, and fragment channels.

[0068] Specifically, first, the system receives user-submitted search requests and converts them into structured query intent data using a natural language parsing module. This query intent data includes at least time constraint information, exact match flags, and concept expansion flags.

[0069] Time constraint information is used to limit the time range of search results and can be parsed from explicit time expressions (such as "2023") or implicit time intentions (such as "recent policies").

[0070] The exact match flag is used to indicate whether the search process should prioritize a strict keyword or entity matching strategy.

[0071] The concept expansion flag is used to indicate whether it is necessary to introduce synonyms, hyponyms, or domain-related concepts to broaden the search coverage.

[0072] Subsequently, the system acquires user behavior log data related to the user's search behavior, including but not limited to click location, click order, page dwell time, and records of secondary search behaviors during the search process (such as modifying search terms, adding filter conditions, etc.). To improve the accuracy of the adjustment strategy, the system can perform noise reduction and normalization processing on the above behavior data, eliminating irrelevant clicks or invalid dwell times.

[0073] Based on the aforementioned behavioral log data, feedback intensity data is generated. This data is used to quantify the user's satisfaction with the current search results and their interaction intent. For example, positive or negative feedback signals from the user regarding the current search results can be derived by calculating the clicked document's ranking position in the results list, the ratio of dwell time to the global average dwell time, and the frequency of secondary searches.

[0074] Then, by combining time constraint information, exact matching flags, concept expansion flags, and feedback strength data, the weight quadruples of the weight-gated model are updated. The weight quadruples correspond to:

[0075] Text channel weights (used to adjust the contribution ratio of text features to search results);

[0076] Metadata channel weights (used to adjust the impact of meta-features on search ranking and filtering);

[0077] Knowledge channel weight (used to adjust the weight of knowledge features in semantic expansion and professional relevance matching);

[0078] Fragment channel weights (used to adjust the importance of fragment-level semantic matching in the final result score).

[0079] During the update process, the system can adopt an online learning strategy based on gradient updates or a weighted regression strategy based on historical feedback data to ensure that the weight adjustment can not only respond to the user's personalized needs in a timely manner, but also gradually optimize the global performance in long-term interaction.

[0080] This application enables real-time personalized optimization of search requests, allowing the search process to dynamically balance text matching accuracy, metadata filtering strength, knowledge semantic expansion capability, and fragment-level matching fine granularity under different users and task scenarios. This significantly improves the relevance of search results and user satisfaction, while also taking into account the system's search speed and scalability.

[0081] Through the above technical solutions, this application effectively solves the problem that static feature weights cannot adapt to diverse search intentions. It achieves time-sensitive retrieval through time constraint recognition, ensures accurate keyword matching through precise matching marker detection, and activates knowledge-related retrieval capabilities using concept expansion markers. Combined with a continuous optimization mechanism based on user behavior feedback, it significantly improves the rationality of feature weight allocation in complex query scenarios, making search results more closely match users' actual needs.

[0082] This application further proposes that the generation of weighted quadruples for the weighted gating model includes:

[0083] Extract intent element vectors from query intent data ,in, Indicates time constraints, used to reflect the degree of restriction on the time range in the query request, such as the length of the time window in units of days, months, or years, or the time weight value calculated based on the time decay function;

[0084] The concept expansion flag indicates whether synonyms, hyponyms, or domain-related concepts need to be introduced. Its value can be a binary flag (0 indicates expansion is off, 1 indicates expansion is on) or a normalization strength value.

[0085] This indicates an exact match flag, used to reflect whether a strict keyword matching strategy should be used for the search. The value can be binary or a confidence value in the range [0,1].

[0086] Extracting feedback vectors from feedback intensity data ,in This indicates the offset value of the click position. This is used to measure the relative sorting position of a document clicked by a user in the results list. The value can be negative (clicked before the expected ranking), zero (clicked at the expected ranking position), or positive (clicked after the expected ranking), and can be normalized.

[0087] The dwell time ratio measures the ratio of the time a user spends on a search result page to the global average dwell time. A value greater than 1 indicates a dwell time longer than the average, reflecting a higher relevance, while a value less than 1 indicates a potentially lower relevance.

[0088] Based on the intent element vector and feedback vector, the weighted quadruple data is calculated using a preset function, as shown in the following formula:

[0089]

[0090] in, These correspond to the weight values ​​for the text channel, metadata channel, knowledge channel, and fragment channel, respectively; when there is no user behavior data, the default weights are used. ;

[0091] σ is the sigmoid activation function, which is used to map the calculation result of the linear combination to the (0,1) interval to ensure that the weights of each channel fluctuate within a reasonable range;

[0092] and Here is a preset coefficient vector for channel j, used to adjust the sensitivity of intent elements and feedback signals to the channel weights. This vector can be obtained through offline training using historical retrieval data. Specifically, gradient descent is used (based on 100,000 training experiments, a learning rate of 0.01 and a batch size of 128 are determined), and the loss function is... ;

[0093] and The following example values ​​can be used, for example:

[0094] Text channel (j=1), intent coefficient vector (correspond ): =[0.6,0.2,0.8];

[0095] High weight is assigned to the exact match flag ( (Weight 0.8), prioritizing the accuracy of text literal matching;

[0096] Time constraints With a weight of 0.6, it is suitable for timely text retrieval needs.

[0097] Feedback coefficient vector (correspond ): = [0.3, 0.7];

[0098] Dwell time ratio ( The weight is 0.7, which strengthens the impact of users' in-depth feedback on the text content.

[0099] In practical applications, gradient descent (the document mentions a learning rate of 0.01 and a batch size of 128) should be used to iteratively optimize the coefficient values ​​using the mean squared error between the "click position offset" and the "prediction offset" as the loss function.

[0100] When user behavior or business scenarios change (such as an increase in knowledge-intensive search demands), incremental updates can be used for fine-tuning. and There is no need to retrain the entire model.

[0101] This is a preset bias term for channel j, used to maintain the basic weight level of the channel in the absence of user behavior data or intent signals; the value sources for the corresponding text, metadata, knowledge, and fragment channels include:

[0102] Based on prior knowledge of the document retrieval scenario, a basic weight offset is set for different channels (e.g., the text channel needs to maintain a higher basic weight to ensure the stability of literal matching). Optimization is then performed using historical retrieval log data (e.g., retrieval accuracy when there is no user feedback), with the goal of "minimizing the basic weight deviation". Adjustments to search characteristics for specific fields (such as financial archives and legal documents), for example, improving knowledge channels in knowledge-intensive scenarios. The initial value.

[0103] Typically, the values ​​range from [−0.2, 0.8]. After mapping with the sigmoid function, the basic weights of each channel stabilize in the range of [0.45, 0.69], avoiding model failure caused by extreme values. When user intent signals (such as time constraints) are present... ) or feedback data (such as click position offset) When ) is missing, Ensure that the channel weight is not lower than the basic level (e.g., the minimum weight of the knowledge channel is σ(−0.2)≈0.45).

[0104] The weighted quadruples are used to control the proportion of each channel in the weighted fusion of the weighted gating model. After generation, the weighted quadruples are directly input into the weighted gating model to control the proportion of text feature data, meta feature data, knowledge feature data, and fragment feature data in the weighted fusion process, so that the contribution value of different channels can be adaptively adjusted according to the user's real-time needs and historical interaction feedback.

[0105] Specifically, when a user submits a search request that includes time constraints, the time sensitivity parameter... This will be assigned a higher value, increasing the weighting coefficient of time-related features in the metadata channel. Significant improvement. For example, in historical archive retrieval scenarios, if a user specifies to retrieve documents from 1990 to 2000, the time constraint parameter... It will be quantized to 0.8, at which point the weight calculation process of the metadata channel will be performed. This item will enhance the contribution of time features. Additionally, if a user's click position on a document category continuously deviates from the current sorting result, the click position offset value will be... This will trigger the feedback vector on the knowledge channel weights. The dynamic adjustment allows the semantic association features implicit in the knowledge graph to be given priority in subsequent retrieval.

[0106] Through the weight generation mechanism driven by both intent elements and user feedback, this invention can dynamically and accurately balance the influence of multiple feature channels during the retrieval stage. This enables comprehensive optimization of the relevance, coverage, and accuracy of retrieval results when facing diverse user needs, significantly improving user experience and system intelligence.

[0107] Through the above technical solution, this application solves the problem of the lack of a dynamic weight adjustment mechanism in the existing technology when fusing multiple features, enabling text features, metadata features, knowledge features, and fragment features to be proportionally optimized according to real-time retrieval intent and user feedback. For example, in scenarios where users need concept expansion, the knowledge channel weight... Automatically boosted to enhance semantic association retrieval capabilities; in scenarios requiring precise matching, text channel weighting. Enhanced to improve literal matching accuracy. This application effectively reduces the deviation of search results caused by fixed weight allocation, and improves the search accuracy and user satisfaction under complex query intent.

[0108] This application further proposes that, in establishing subject index data corresponding to the subject aggregation of archives, the construction of the subject index data includes:

[0109] The archive feature dataset (including text feature data, meta feature data, and knowledge feature data) is input into a pre-defined weighted gating model for weighted fusion to generate fused feature data. The weighted fusion process can adaptively adjust the contribution ratio of the text channel, metadata channel, knowledge channel, and fragment channel, thereby forming the optimal feature representation for the current retrieval target in the feature space.

[0110] The fused feature data is classified using a deep learning-based neural network to generate classification result data; the neural network classification model can be a multi-layer convolutional neural network, a bidirectional long short-term memory network (Bi-LSTM), or an encoder network based on the Transformer structure; the classification result data includes the category label of each file record or fragment and the corresponding classification confidence score;

[0111] The target text fragment is determined based on the fused feature data and classification results. Topic clustering is then performed on the target text fragment to generate topic cluster centroid vector data. and member list data;

[0112] Based on the fusion of feature data and classification results, the system identifies a set of target text fragments and performs topic clustering on these fragments. During clustering, algorithms based on K-means++, hierarchical clustering, or density clustering (DBSCAN) can be used to group semantically similar fragments into the same topic cluster, outputting the topic cluster centroid vector data for each cluster. and its member list data (i.e., the text fragment identifiers and their basic characteristics contained in the cluster).

[0113] For topic cluster centroid vector data For each member of the topic cluster, a confidence score for the target text fragment is calculated using the following formula:

[0114]

[0115] in, Segment-level semantic vector data refers to the result of mapping the semantic information of several consecutive text segments into vector representations when semantically encoding archives. It not only contains the word meaning and grammatical information of the segment, but also retains the contextual features of the segment in the context of the whole text.

[0116] sim is the cosine similarity function, used to measure the similarity of the angle between the semantic vector of a segment and the centroid vector of a cluster in the vector space. Its value ranges from [-1, 1], and the closer the value is to 1, the more similar the semantics are.

[0117] To extract knowledge consistency scores related to candidate fragment feature data from knowledge feature data, which reflect the degree of matching between fragment content and related concepts, rules or clauses in the domain knowledge graph, the score can be calculated based on triple matching rate, rule verification results, etc., and the value range can be [0,1].

[0118] α and β are preset weight coefficients that satisfy α+β=1, used to balance the contribution ratio of semantic similarity and knowledge consistency in confidence assessment, and can be set through offline training or expert experience.

[0119] After obtaining the confidence score of each topic cluster member, the system can filter based on a preset threshold, including only members with a confidence score higher than the threshold in the final topic index data, thereby ensuring the accuracy and representativeness of the index results. Ultimately, the constructed topic index data not only includes the centroid vector and member list of each topic cluster, but also records the semantic and knowledge-based dual matching information of each member, facilitating rapid location of relevant document fragments during the subsequent retrieval and matching stage.

[0120] Specifically, when generating topic index data, multi-source features are first dynamically fused using a weighted gating model. For example, when a user's query intent involves professional knowledge, the weight of the knowledge channel is automatically increased. The fused feature data is then processed by a classification network to determine topic categories, such as classifying legal files into contract and litigation categories. Based on the classification results, text fragments related to the target topic are selected, such as extracting sentences containing keywords like "clause" and "liability for breach of contract" under the contract category. These fragments are then subjected to cluster analysis, such as using a hierarchical clustering algorithm to generate topic clusters with a hierarchical structure, where the centroid vector of each cluster represents the core semantics of the topic. Finally, the semantic similarity between each fragment and the cluster centroid is calculated, and combined with entity association scores in the knowledge graph, such as determining whether the "income verification information" in the loan application material fragment has a graph association with the "debt repayment ability model key factors" in the topic cluster "high-credit customers," a comprehensive confidence score is derived. When the preset coefficients α and β are set to 0.6 and β to 0.4, the system can balance the influence of semantic expression similarity and financial attribute association, effectively identifying high-risk customers.

[0121] This application introduces a fusion-feature-driven topic clustering and a dual-index confidence calculation mechanism. This invention effectively improves the semantic purity and knowledge consistency of topic clusters during the topic index construction stage, avoiding potential misjudgments that may arise from relying solely on a single similarity index. Consequently, it significantly improves the accuracy and interpretability of results in subsequent retrieval and matching processes. This method retains the abstraction capabilities of deep learning models for semantic features while incorporating rule constraints based on knowledge features, making the topic index data more robust and usable across multiple application scenarios.

[0122] This application further proposes that the step of inputting the candidate archival data and the corresponding archival feature dataset back into the feature encoding network to generate candidate fragment feature data includes:

[0123] Retrieve relevant topic cluster data from the topic index data based on the query intent data;

[0124] Perform document-level semantic matching within the relevant topic cluster data to output a preliminary candidate document set.

[0125] Extract fragment-level semantic vector data corresponding to the primary candidate document set. , as candidate fragment feature data.

[0126] Based on the query intent data (obtained from the retrieval request and dynamically weighted), a topic-level recall is performed in the pre-built topic index data to determine target topic clusters that are semantically relevant to the query intent. This recall process can employ an approximate nearest neighbor (ANN) retrieval algorithm based on cosine similarity, vector inner product, or semantic hashing, which improves the semantic relevance of the recall results while ensuring recall speed.

[0127] Within the acquired relevant topic cluster data, document-level semantic matching is performed on the included archival records. Specifically, the fused feature vector of each archival record is calculated with the query intent data vector to generate a document matching score; the fused feature vector is obtained by fusing text feature data, meta-feature data, knowledge feature data, and fragment feature data according to a weighted gating model. In this step, semantic similarity and knowledge consistency scores can be combined to comprehensively score the documents, and several documents can be selected as the initial candidate document set data according to the scores from high to low.

[0128] For each record in the initial candidate document set, segment-level feature extraction is performed: the document is segmented into several continuous segments based on the document structure information (such as title / chapter / section / paragraph / sentence level segmentation markers); these segments are input into the same feature encoding network as the aforementioned feature encoding stage, which may include a text encoding sub-network (such as BERT, RoBERTa), a metadata encoding sub-network, and a knowledge feature encoding sub-network; the segment-level semantic vector data corresponding to each segment is extracted from the encoding output, and this vector data is used as candidate segment feature data for subsequent multi-scale matching.

[0129] In one embodiment, the extraction of fragment-level semantic vector data not only considers the semantic and syntactic features within the fragment, but also integrates contextual information from the surrounding fragments, thereby ensuring that the deep semantic relationship between the query intent and the fragment can be captured more accurately in the subsequent matching stage.

[0130] This application introduces a two-stage screening mechanism—document-level semantic matching and fragment-level feature extraction—after topic-level recall. This significantly improves the semantic purity and relevance of candidate fragment feature data while ensuring broad recall coverage. This two-stage filtering + re-encoding strategy effectively reduces the probability of irrelevant fragments entering the multi-scale matching stage, thereby reducing computational redundancy and improving retrieval efficiency and result accuracy.

[0131] This application further proposes that, in the multi-scale semantic matching of candidate fragment feature data to output retrieval results data, the specific steps of multi-scale semantic matching include:

[0132] Within the context of the document to which it belongs, the candidate fragment feature data is used to calculate the matching degree between the candidate fragment feature data and the query intent data at the sentence granularity, paragraph granularity, and document granularity, respectively, to generate fragment-level matching score data; based on the experimental optimization results of 100,000 retrieval samples, the multi-scale matching weight allocation is set to 0.4 for sentence granularity, 0.4 for paragraph granularity, and 0.2 for document granularity;

[0133] By integrating fragment-level matching score data, fragment-level semantic vector data, and knowledge consistency score data, a fused matching score data is generated.

[0134] Based on the fusion matching score data, fragments that satisfy the confidence score conf≥θ are selected, and a high-confidence fragment index is constructed, where θ is a preset threshold that can be optimized through backtesting with historical data according to different application scenarios. These high-confidence fragments are then used to construct the high-confidence fragment index and returned to the user as the final search results.

[0135] Sentence-level matching is used to capture precise semantic correspondences within a short text, and is suitable for handling scenarios such as keyword matching and specific fact-based question answering;

[0136] Segment-level matching can identify semantic relevance within a larger context, making it suitable for retrieving content that is thematic or analytical.

[0137] Document-level matching compares the fused feature vector of the entire document with the query intent to ensure consistency at the global semantic level, making it suitable for judging the topic relevance of long documents.

[0138] At each granularity, a cosine similarity function, a dual-tower deep matching network, or an attention-based interactive coding network can be used to calculate the matching score and output the corresponding fragment-level matching score data.

[0139] Next, the fragment-level matching scores at different granularities are fused with the corresponding fragment-level semantic vector data and the knowledge consistency score data generated from the knowledge feature data. The fusion process can employ weighted linear combination, MLP (Multilayer Perceptron) fusion, or gating mechanism fusion to generate fused matching score data, which simultaneously reflects a comprehensive evaluation of fine semantic matching degree, contextual semantic relevance, and knowledge consistency.

[0140] By introducing a multi-scale semantic matching mechanism, this invention achieves a balance between precise matching of short texts and semantic relevance judgment of long texts; combined with knowledge consistency scoring, it significantly reduces the risk of false detections caused by pure semantic similarity; the construction of a high-confidence fragment index enables subsequent retrieval responses to quickly and directly locate the most relevant content, thereby improving retrieval accuracy and response speed.

[0141] This application further proposes that, after outputting the retrieval results data, the intelligent document classification and retrieval method also includes:

[0142] The weight parameters of the weight gating model are updated based on user behavior log data, and weight adjustment instruction data is generated. This instruction data includes incremental adjustment coefficients for four channels (text channel, metadata channel, knowledge channel, and fragment channel).

[0143] Based on the weight adjustment instruction data, the weight quadruple data is incrementally updated;

[0144] When the knowledge graph triggers a predefined invalidation rule by updating data, it generates knowledge invalidation marker data and associates it with the corresponding archive data.

[0145] Among them, the weighted quadruple data Perform incremental updates. Update methods may include:

[0146]

[0147] in: The original weight value of channel j;

[0148] The adjustment value for channel j is calculated from the weight adjustment instruction data;

[0149] γ is the learning rate coefficient, used to control the magnitude of a single update and prevent model parameter oscillations, γ∈(0,0.1];

[0150] These correspond to the text, metadata, knowledge, and fragment channels, respectively.

[0151] Furthermore, to ensure the reliability of knowledge feature data during long-term use, the system generates knowledge expiration marker data when knowledge graph updates trigger predefined expiration rules (such as revisions to legal provisions, deconstruction of knowledge node relationships, or expiration of associated entities). Expiration rules may include mechanisms such as time validity checks, version number verification, and content conflict detection. The generated knowledge expiration marker data will be directly associated with the corresponding archival data to automatically reduce or eliminate the impact of the expired knowledge point during subsequent retrieval and matching stages. For example, knowledge expiration marker data is triggered when the weight of an entity association edge in the knowledge graph is below 0.2, or when the update time is more than 5 years earlier than the current time.

[0152] This application introduces a user behavior-driven adaptive weight update mechanism, which enables continuous optimization of the weight gating model, allowing retrieval results to dynamically align with users' actual usage habits and feedback trends. Simultaneously, the knowledge expiration marking mechanism ensures the accuracy and timeliness of knowledge feature data in long-term operation, effectively reducing retrieval errors caused by knowledge expiration, thereby achieving continuous evolution of the retrieval model and guaranteeing knowledge credibility.

[0153] This application further proposes that the intelligent archive classification and retrieval method also includes an asynchronous maintenance process, used to maintain the validity and aggregation quality of the subject index data and knowledge feature data in the background without affecting the online retrieval performance. The steps are as follows:

[0154] Real-time monitoring of fusion matching score data and knowledge expiration marker data; fusion matching score data integrates fragment-level matching score, fragment-level semantic vector similarity, and knowledge consistency score; knowledge expiration marker data is used to identify knowledge nodes that have expired or become invalid.

[0155] When the fusion matching score data is lower than a preset threshold, a topic dispersion signal is generated; in the statistical window Within (e.g., the last 7 days), if the mean of the fused matching score data is consistently lower than the preset threshold (e.g., 0.65), and below the proportion exceeding If a retrieval batch (e.g., 50%) is processed, a topic dispersion signal is generated. This signal reflects phenomena such as decreased similarity within topic clusters and shift in cluster centers, indicating a need to re-optimize the topic index structure.

[0156] When the total amount of knowledge invalidation markers exceeds a preset quantity, a knowledge obsolescence signal is generated; this is displayed in the statistics window. If the total amount N of knowledge failure marker data detected by the system exceeds a preset threshold, (For example, 500 entries) will generate a knowledge obsolescence signal. This signal is used to reflect situations where a large number of nodes in the knowledge graph fail or are frequently updated;

[0157] The "match-precision" curve is determined based on historical retrieval data, and is usually taken as the critical value that maximizes the harmonic average of precision and recall. =0.65 means that in 100,000 labeled samples, when the fusion matching score is ≥0.65, the accuracy of the retrieval results is consistently higher than 85%, and below this value, the accuracy drops sharply to below 60%.

[0158] By statistically analyzing the score distribution within a "normal retrieval cycle," setting it to 50% means that optimization is triggered when "more than half of the retrieval batches fail to meet quality standards," thus avoiding interference from a single outlier.

[0159] Combining the "node density" and "update frequency" of a knowledge graph, the formula is as follows:

[0160]

[0161] in, The statistical window (the document example shows 7 days) is used to calculate the average monthly failure rate using historical knowledge update logs (e.g., the average monthly failure rate for policy-related knowledge in the financial field is about 2%).

[0162] In financial records, =500 entries (total knowledge nodes: 100,000; average monthly failure rate: 2%; expected failure count within a 7-day window: ≈467, rounded up to 500). Through online learning, when the system detects "poor index reconstruction effect after threshold triggering" (e.g., matching score improvement <5% after reconstruction), it automatically fine-tunes. (±0.05) or (±10%).

[0163] In response to topic dispersion signals or knowledge obsolescence signals, a local index reconstruction process is triggered, which includes the removal of invalid data, recalculation of topic cluster centroid vectors, and updating of the member list.

[0164] When any signal is triggered, the system enters a local index reconstruction process, which includes:

[0165] (1) Remove entries corresponding to knowledge failure marker data from the topic index data and related fragment index structures.

[0166] (2) Based on the remaining valid fragment-level semantic vector data, recalculate the topic cluster centroid vector. This ensures that cluster centers accurately reflect the latest data distribution.

[0167] (3) Adjust the list of topic cluster members, remove invalid members and add new candidate fragments to ensure the integrity and representativeness of the topic cluster.

[0168] This application introduces an asynchronous maintenance mechanism driven by quantifiable parameters. This method can dynamically discover the problems of topic aggregation decay and knowledge base obsolescence without affecting the online retrieval speed, and perform targeted local index reconstruction. This maintains the semantic aggregation accuracy and knowledge timeliness of the retrieval system in long-term operation, significantly reduces the risk of retrieval result quality decay over time, and achieves high availability and low maintenance cost of the archive retrieval platform.

[0169] The following is a specific example of a smart file classification and retrieval system for loan companies:

[0170] A loan company has deployed a deep learning-based intelligent document classification and retrieval system for the entire loan customer risk management process. The system accesses a massive loan application database, including textual materials submitted by customers (such as income statements and asset lists), metadata (loan amount, term, historical repayment records), and industry knowledge graphs (including financial risk factors such as "debt-to-equity ratio" and "cash flow coverage ratio").

[0171] The loan application form submitted by the customer is parsed into structured text data using OCR technology. The system segments the data according to the "title / paragraph / key field" hierarchy (e.g., the "income certificate" section). Metadata standardization: "Loan amount ≥ 5 million yuan" is mapped to a high-risk label, and missing values ​​in historical repayment records are filled in to match industry averages. Knowledge graph alignment: When "real estate mortgage" is mentioned in the file, the associated knowledge node "collateral valuation decay rate ≤ 15%" is linked, and a timeliness score is output (e.g., marked "expired" after local policy revision). The feature encoding module generates text feature vectors (BERT-768-dimensional semantic encoding), meta-feature vectors (numerically normalized), and knowledge consistency score data, which are then fused and input into a weighted gating model.

[0172] When a risk auditor inquires about "Short-term loan risks for manufacturing enterprises in the Yangtze River Delta region in 2023," the system analyzes the intent as follows:

[0173] Time constraints =0.9 (high timeliness); Conceptual expansion marker =1 (requires industry knowledge); exact match flag =0 (allows semantic generalization);

[0174] Based on user behavior logs (70% of clicks were on the "high-risk customer" tag, and the average dwell time was higher than...), =1.5), generating quadruplets using a dynamic weighting formula:

[0175]

[0176] Output weights: Knowledge channels =0.85 (Enhanced Industry Risk Analysis), Metadata Channel =0.75 (emphasis on time filtering). After weighted fusion, the neural network classifier clusters the files into themes such as "highly indebted enterprises" and "stable cash flow enterprises," and calculates the confidence score of the theme clusters:

[0177]

[0178] Filter segments with a confidence level > 0.8 (such as the segment “Quarterly cash flow decreased by 30% year-on-year”) to build a topic index.

[0179] The query intent vector is used to recall the "manufacturing enterprises" topic cluster and perform multi-scale matching. At the sentence level, keywords such as "insufficient valuation of collateral" are matched; at the paragraph level, the context of the "impact of industry policies" chapter is analyzed; and at the document level, the similarity between the entire file and the topic "short-term loan default risk" is calculated. The matching score and knowledge consistency score are combined to output high-risk customer file fragments (such as "a certain company's debt-to-asset ratio exceeded 85% in Q2 2023").

[0180] The system continuously monitors users' "skip" behavior when viewing search results (feedback vector click position offset value). =−0.2), incrementally update weights:

[0181]

[0182] When the knowledge graph updates the policy on "tightening credit for manufacturing in the Yangtze River Delta," it triggers the reconstruction of the topic index. The predictive model shows an increased probability of default for this type of loan in the next three months, generating an adjustment instruction.

[0183] 1. Risk Control: Initiate secondary due diligence on clients with a "debt-to-asset ratio > 80%" in the matched segments;

[0184] 2. Alternative solution: Push the "collateral-enhanced loan" product library (match rate 92%);

[0185] 3. Process optimization: Shorten the approval cycle for high-risk customers to 48 hours.

[0186] In this embodiment, the system accurately captures "time-sensitive" query intents through dynamic weight gating. In the context of manufacturing loan risk assessment, topic indexing significantly improves the recall efficiency of high-risk customers. A multi-scale matching mechanism locates key risk segments (such as descriptions of abnormal cash flow), and knowledge consistency verification avoids misjudgments caused by outdated policies. User behavior-driven online optimization enables the model to continuously adapt to changes in risk preferences, ultimately achieving accurate risk stratification and proactive risk control in loan profiles.

[0187] Example 2:

[0188] like Figure 4 and Figure 5 As shown, a deep learning-based intelligent document classification and retrieval system, using the aforementioned deep learning-based intelligent document classification and retrieval method, includes:

[0189] The feature encoding module is used to obtain the feature-encoded archive feature dataset, which includes text feature data, meta-feature data, and knowledge feature data; it includes a high-performance central processing unit (CPU), a graphics processing unit (GPU), and high-speed memory (RAM), where the GPU is used to accelerate text vectorization, knowledge graph mapping, and multi-channel feature encoding operations in parallel.

[0190] The classification index module is used to input the archival feature dataset into a preset weighted gating model for weighted fusion, and then classify it through a deep learning-based neural network to build topic index data corresponding to the archival topic aggregations. It includes a multi-core CPU, GPU accelerator card, and solid-state drive (SSD) storage device for storing and quickly retrieving topic cluster centroid vector data and member list data; the classifier adopts a Transformer encoder (12 layers, 12 attention heads).

[0191] The weight gating module is used to receive retrieval requests, parse them into query intent data, and adjust the weight parameters of the weight gating model.

[0192] The recall module is used to input topic index data into a weighted gating model to obtain candidate file data;

[0193] The feature regeneration module is used to input the candidate archive data and the corresponding archive feature dataset back into the feature encoding network to generate candidate fragment feature data.

[0194] The matching output module is used to perform multi-scale matching on candidate fragment feature data. It calculates the matching degree between candidate fragment feature data and query intent data at different semantic scales, and filters out the retrieval result data that meets the preset threshold based on the weighted result of the matching degree.

[0195] The feature encoding module is a component that performs structured processing on the original archival data. Specifically, it uses a pre-trained language model to encode text data and extracts meta-features and knowledge features using a metadata parser and knowledge graph construction tools. Its role is to integrate multi-dimensional archival information to support subsequent classification and retrieval. The classification indexing module is a component that implements feature fusion and topic clustering. Specifically, it uses a gated attention mechanism to dynamically weight different feature channels and generates a topic index through a deep neural network. Its role is to establish a semantically related archival organizational structure. The weight gating module is a component that enables adaptive retrieval intent. Specifically, it dynamically adjusts the weight ratio of each feature channel by parsing elements such as time constraints and exact match flags in the query statement and combining user behavior logs. Its role is to improve the responsiveness of the retrieval process to user needs. The feature regeneration module is a component that implements deep representation of candidate data. Specifically, it extracts fragment-level semantic vectors through secondary encoding. Its role is to enhance the ability to capture fine-grained semantics during the retrieval process.

[0196] Specifically, the system constructs a multi-dimensional dataset containing text, metadata, and knowledge features through a feature encoding module, and a classification indexing module uses a weighted gating model to fuse features and generate a topic index. When a retrieval request is received, the weighted gating module parses the query intent and adjusts the weights of each feature channel, while the recall module filters candidate files based on the optimized weights. After the feature regeneration module extracts fragment-level features from the candidate data, the matching output module calculates semantic matching degrees at three granularities: sentence, paragraph, and document. Finally, highly relevant results that meet the threshold are selected through weighted fusion. This process achieves the synergistic effect of dynamic feature weighting and multi-scale matching, solving the problem of rigid feature utilization in traditional methods.

[0197] In a simplified embodiment, the system is deployed on a server equipped with two 32-core CPUs, four NVIDIA A100 GPUs, 512GB of memory, and 2TB of NVMe SSD storage, running a Linux operating system and the TensorFlow deep learning framework. The A100 GPUs improve feature encoding speed by 3 times. The feature encoding module receives raw text and metadata from the archive management database and performs batch BERT encoding and knowledge graph embedding mapping via the GPU. The classification indexing module performs topic clustering on the fused features and generates topic indexes. When a user issues a search request, the weight gating module adjusts the channel weights based on the query intent and the user's historical behavior data. The recall module performs preliminary candidate document screening based on the topic index, and the feature regeneration module performs secondary encoding on the candidate segments. The matching output module calculates the matching degree and merges the scores at the sentence, paragraph, and document granularities, and finally returns a list of high-confidence search results.

[0198] Through the aforementioned system architecture, this application achieves a closed-loop processing workflow from archival feature extraction, subject classification, dynamic weight adjustment, candidate recall, feature regeneration to multi-scale matching. Because the system hardware architecture supports high-parallelism computation and low-latency storage access, it significantly improves retrieval response speed and scalability while maintaining high accuracy. It can continuously adapt to the growth of archival data scale and changes in user retrieval patterns, demonstrating significant practical value and innovation.

[0199] The technical scope of this invention is not limited to the content described above. Those skilled in the art can make various modifications and variations to the above embodiments without departing from the technical concept of this invention, and all such modifications and variations should fall within the protection scope of this invention.

Claims

1. A deep learning-based intelligent document classification and retrieval method, characterized in that: Includes the following steps: Obtain a feature-encoded archive feature dataset, which includes text feature data, meta-feature data, and knowledge feature data; The archive feature dataset is input into a preset weighted gating model for weighted fusion, then classified using a deep learning-based neural network, and topic index data corresponding to the archive topic aggregation is established. Receive retrieval requests, parse them into query intent data, and adjust the weight parameters of the weight gating model; The topic index data is input into the adjusted weight gating model to obtain candidate file data; The candidate archive data and the corresponding archive feature dataset are input again into the feature encoding network to generate candidate fragment feature data; Multi-scale matching is performed on the candidate fragment feature data. The matching degree between the candidate fragment feature data and the query intent data is calculated at different semantic scales of different granularities. Based on the weighted result of the matching degree, the search result data that meets the preset threshold is selected.

2. The intelligent archive classification and retrieval method based on deep learning according to claim 1, characterized in that: The specific steps for obtaining the feature-encoded archive feature dataset include: The document obtains the document text data, the document metadata corresponding to the document text data, and the knowledge anchor data mapped from the document text data, and performs feature encoding on each of them; Extract the text feature data from the archive text data; Extract the meta-feature data from the archive metadata; A knowledge graph is constructed based on the knowledge anchor data, and the knowledge feature data is obtained by mapping from the knowledge graph. Sentence segmentation is performed on the archive text data to extract fragment-level text features.

3. The intelligent archive classification and retrieval method based on deep learning according to claim 2, characterized in that: The specific steps for dynamically adjusting the weight parameters in receiving a retrieval request, parsing it into query intent data, and adjusting the weight parameters of the weight gating model include: Extract time constraints, exact match flags, and concept expansion flags from the query intent data; Acquire user behavior log data, and generate feedback intensity data based on the user behavior log data, wherein the user behavior log data includes click location, dwell time and secondary search behavior; Based on the time constraint, the exact matching flag, the concept expansion flag, and the feedback intensity data, update the weight quadruples data of the weight gating model; The weighted gating model includes a text channel, a metadata channel, a knowledge channel, and a fragment channel.

4. The intelligent archive classification and retrieval method based on deep learning according to claim 3, characterized in that: The generation of weighted quadruples for the weighted gating model includes: Extract intent element vectors from the query intent data. ,in This indicates the time constraint. Indicates the concept extension flag, This indicates the exact match flag; Extract the feedback vector from the feedback intensity data. ,in This indicates the offset value of the click position. Indicates the percentage of time spent in the residence area; Based on the intent element vector and feedback vector, the weighted quadruple data is calculated using a preset function, as shown in the following formula: in, These correspond to the weight values ​​of the text channel, the metadata channel, the knowledge channel, and the fragment channel, respectively; σ is the sigmoid activation function. and This is the preset coefficient vector for channel j; This is the preset bias term for channel j; The weighted quadruple data is used to control the proportion of each channel in the weighted fusion of the weighted gating model.

5. The intelligent archive classification and retrieval method based on deep learning according to claim 1, characterized in that: The construction of the topic index data corresponding to the archive topic aggregation includes: The archive feature dataset is input into a preset weighted gating model for weighted fusion to generate fused feature data; The fused feature data is then classified using a deep learning-based neural network to generate classification result data. Based on the fused feature data and the classification result data, a target text segment is determined, and topic clustering is performed on the target text segment to generate topic cluster centroid vector data. and member list data; For the centroid vector data of the topic cluster Each member of the topic cluster calculates the confidence score of the target text fragment, using the following formula: in, This represents fragment-level semantic vector data; sim is the cosine similarity function. To extract a knowledge consistency score related to the candidate fragment feature data from the knowledge feature data; α and β are preset weight coefficients that satisfy α+β=1.

6. The intelligent archive classification and retrieval method based on deep learning according to claim 5, characterized in that: The step of inputting the candidate archive data and the corresponding archive feature dataset back into the feature encoding network to generate candidate fragment feature data includes: Retrieve relevant topic cluster data from the topic index data based on the query intent data; Perform document-level semantic matching within the relevant topic cluster data to output a preliminary candidate document set data; Extract the fragment-level semantic vector data corresponding to the primary candidate document set. , as the feature data of the candidate fragment.

7. The intelligent archive classification and retrieval method based on deep learning according to claim 5, characterized in that: The specific steps of multi-scale semantic matching in outputting retrieval results data by performing multi-scale matching on the candidate fragment feature data include: Within the context of the document to which it belongs, the candidate fragment feature data is used to calculate the matching degree between the candidate fragment feature data and the query intent data at the sentence granularity, paragraph granularity and document granularity respectively, generating fragment-level matching score data; The fragment-level matching score data, the fragment-level semantic vector data, and the knowledge consistency score data are combined to generate fused matching score data. Based on the fusion matching score data, segments that satisfy the confidence score conf≥θ are selected, and high-confidence segment index data is constructed, where θ is a preset threshold.

8. The intelligent archive classification and retrieval method based on deep learning according to claim 3, characterized in that: The intelligent archive classification and retrieval method further includes, after outputting the retrieval result data: The weight parameters of the weight gating model are updated based on the user behavior log data, and weight adjustment instruction data is generated. Based on the weight adjustment instruction data, the weight quadruple data is incrementally updated; When the knowledge graph triggers a predefined failure rule by updating data, it generates knowledge failure marker data and associates it with the corresponding archive data.

9. The intelligent archive classification and retrieval method based on deep learning according to claim 8, characterized in that: The intelligent document classification and retrieval method also includes an asynchronous maintenance process: Real-time monitoring of the fusion matching score data and the knowledge failure marker data; When the fusion matching score data is lower than a preset threshold, a topic dispersion signal is generated; When the total amount of knowledge invalidation marker data exceeds a preset quantity, a knowledge obsolescence signal is generated; In response to the topic dispersion signal or knowledge obsolescence signal, a local index reconstruction process is triggered, which includes the removal of invalid data, recalculation of topic cluster centroid vectors, and updating of the member list.

10. A deep learning-based intelligent archive classification and retrieval system, characterized in that: The method for intelligent document classification and retrieval based on deep learning as described in any one of claims 1 to 9 includes: The feature encoding module is used to obtain the feature-encoded archive feature dataset, which includes text feature data, meta-feature data, and knowledge feature data. The classification index module is used to input the archive feature dataset into a preset weighted gating model for weighted fusion, then classify it through a deep learning-based neural network, and establish topic index data corresponding to the archive topic aggregation. The weight gating module is used to receive retrieval requests, parse them into query intent data, and adjust the weight parameters of the weight gating model. The recall module is used to input the topic index data into the adjusted weight gating model to obtain candidate file data; The feature regeneration module is used to input the candidate archive data and the corresponding archive feature dataset back into the feature encoding network to generate candidate fragment feature data; The matching output module is used to perform multi-scale matching on the candidate fragment feature data, calculate the matching degree between the candidate fragment feature data and the query intent data at different semantic scales, and filter out the search result data that meets the preset threshold based on the weighted result of the matching degree.

Images