A method and device for quickly grouping electronic files based on metadata intelligent extraction

Abstract

The application relates to the technical field of electronic file management and intelligent data processing, and particularly discloses a method and device for quickly grouping electronic files based on metadata intelligent extraction, wherein the method comprises the following steps: generating an initial metadata set; performing weighted normalization processing; outputting a conflict set and a consistent metadata set; generating a case file abstract set and an element anchor mapping table; constructing a case file directory index tree; and performing high-concurrency and quick output of multi-modal evidence files. Compared with the prior art, the method can solve the technical problem that the efficiency of manual indexing and grouping is limited when multi-modal evidence files are processed, especially under the conditions of cross-source and multi-thread concurrency, and it is difficult to realize high-consistency and low-delay quick grouping of case files. By introducing a BERT semantic normalization mechanism, cross-file consistency scoring and multi-thread consistency scheduling algorithms, the application realizes efficient organization and real-time quick grouping of evidence files under the condition of multi-modal high concurrency, and significantly improves the efficiency of case file generation.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of electronic file management and intelligent data processing, and particularly relates to a method and device for quickly grouping electronic files based on intelligent metadata extraction. BACKGROUND

[0002] At present, with the rapid development of judicial informatization, administrative management digitization and enterprise compliance supervision automation, the role of electronic files in case handling, examination and supervision and document management is becoming increasingly important. However, the existing electronic file grouping technology still has obvious deficiencies in multi-modal evidence processing, cross-file consistency determination, abstract generation and directory construction, etc. For example, the traditional method relies on keyword search or fixed rule template extraction, and can only process structured text, and the processing effect of unstructured evidence such as scanned copies, pictures, audio and video is poor, which often needs manual transcription and annotation, and is low in efficiency and easy to make mistakes. In the evidence fusion scene of cross-source, multi-format and multi-language, due to the lack of unified metadata extraction and normalization mechanism, there are redundancies and conflicts in the case file elements in different files, resulting in incomplete evidence chain and poor consistency. At the same time, the existing method is based on static rules or shallow semantic matching in cross-file consistency determination, and cannot comprehensively model the context semantics and evidence credibility, resulting in insufficient determination accuracy in complex cases involving multi-modal evidence, witness testimony and audio and video records, and chaotic case file logic.

[0003] In addition, with the expansion of case size and the popularity of electronic files, the number of evidence files involved in a single case can reach hundreds to thousands, and the formats are various. The existing system usually adopts single-level cache and simple index structure, and in the multi-thread high-concurrency access scene, it is easy to have problems such as cache jitter, access delay and lock conflict, and it is difficult to realize real-time reading and rapid assembly of evidence. In the case of file arrangement and output, the existing method usually relies on manual abstract writing and directory, although some systems introduce automatic abstract algorithm, but due to the lack of context-based conflict highlighting mechanism and anchor mapping method based on semantic alignment, the accuracy and explainability of the generated abstract are insufficient, and it is difficult to meet the high standard requirements of judicial case handling and enterprise audit.

[0004] Therefore, there is an urgent need for an electronic file grouping method that can realize high-precision metadata intelligent extraction, cross-file consistency determination and efficient automatic grouping under the conditions of multi-modal, multi-source and high concurrency, so as to improve the intelligent level, grouping efficiency and logical consistency of case file processing. SUMMARY

[0005] In view of the above technical deficiencies, the purpose of the present application is to provide a metadata intelligent extraction-based electronic file fast grouping method, aiming to solve the technical problems in the prior art that manual indexing is relied on when processing multi-modal evidence files, and the efficiency is limited, especially under the conditions of cross-source and multi-thread concurrent, it is difficult to realize high consistency, low delay and fast grouping of case files.

[0006] To solve the above technical problems, the present application adopts the following technical solutions: the present application provides a metadata intelligent extraction-based electronic file fast grouping method,

[0007] The metadata intelligent extraction-based electronic file fast grouping method comprises:

[0008] Step S10: acquiring a multi-modal evidence file set, converting the multi-modal evidence file set into an initial metadata set M by using an OCR and ASR fusion extraction mechanism, and performing weighted normalization processing on the initial metadata set M based on a BERT semantic coding mode, and outputting a normalized metadata set ;

[0009] Step S20: constructing a preliminary case file time axis , according to the preliminary case file time axis and the normalized metadata set , a cross-file element consistency score S is calculated by using a consistency mechanism based on BERT semantic similarity and weight addition; based on the cross-file element consistency score S, it is determined whether there is a cross-source element conflict, and a conflict set and a consistency metadata set V are outputted;

[0010] Step S30: a candidate case file abstract generation model is constructed by using a preset BERT model combined with a self-attention mechanism, the conflict set and the consistency metadata set V are inputted into the candidate case file abstract generation model, and a case file abstract set is outputted; and based on the outputted case file abstract set , a semantic alignment and evidence backtracking association method is used to generate an element anchor mapping table ;

[0011] Step S40: based on the case file abstract set and the element anchor mapping table , a case file directory index tree is constructed, and a multi-thread transaction consistency parallel scheduling algorithm is used to organize the logical sequence based on the case file directory index tree, so as to obtain a multi-thread consistency index tree;

[0012] Step S50: Based on the multi-thread consistency index tree, the preset multi-level cache layering strategy is adopted to perform access frequency statistics and file importance evaluation, and the file cache priority is obtained; based on the file cache priority, an adaptive prefetch loading mechanism is adopted to perform high-concurrency stable output on the multi-modal evidence file set.

[0013] Preferably, in step S10, the multi-modal evidence file set is obtained, the multi-modal evidence file set is converted into an initial metadata set M by adopting an OCR and ASR fusion extraction mechanism, and a weighted normalization processing is performed on the initial metadata set M based on a BERT semantic encoding mode, and a normalized metadata set V is output. The step S10 specifically comprises the following steps.

[0014] Step S101: Obtain the multi-modal evidence file set, the multi-modal evidence file set comprising scanned PDF files, photographed PNG files and voice transcription TXT files; adopt an OCR and ASR fusion extraction mechanism to uniformly convert the multi-modal evidence file set into an intermediate JSON structure, and obtain an initial metadata set M;

[0015] Step S102: Input the initial metadata set M into a preset BERT semantic encoding network, and extract an initial semantic vector; based on the initial semantic vector, a case core element is extracted by adopting a conditional random field (CRF) method, the case core element comprising a time element, a party element, an amount element and a location element;

[0016] Step S103: Based on the extracted case core element, a multi-dimensional dynamic semantic similarity method with fusion credibility is adopted to calculate an element similarity score, and a multi-dimensional weighted normalization processing is performed on the initial metadata set M according to the element similarity score, and a normalized metadata set V is output. .

[0017] Preferably, in step S20, a preliminary case file time axis is constructed , according to the preliminary case file time axis and the normalized metadata set V , a cross-file element consistency score S is calculated based on a BERT semantic similarity and weight addition consistency mechanism; based on the cross-file element consistency score S, it is determined whether there is a cross-source element conflict, and a conflict set C and a consistency metadata set V are output. The step S20 specifically comprises the following steps.

[0018] Step S201: Based on the normalized metadata set V The core elements of the case related to time are extracted. Based on the extracted core elements, the element similarity score is calculated using a multi-dimensional dynamic semantic similarity method that integrates credibility. The time elements of different source files are uniformly converted using a preset time format standardization function. The core elements of the case with element similarity scores greater than the preset element similarity score threshold are aggregated into the same event node, thereby constructing a preliminary case file time axis T. The preliminary case file time axis T is used to provide temporal constraints for the subsequent calculation of consistency scores.

[0019] Step S202: For the normalized metadata set The core elements of the case are extracted using a pre-defined BERT semantic encoding network to obtain the corresponding semantic vector h. Based on the semantic vector h, the cosine similarity calculation method is used to calculate the element semantic similarity sim of the core elements of the case between different files. Then, the semantic similarity sim and the pre-defined similarity threshold τ are used to determine whether the core elements of the case are semantically equivalent: elements with a semantic similarity sim higher than the similarity threshold τ are marked as semantically equivalent, and those with a similarity sim lower than the threshold τ are marked as non-equivalent. The set of equivalence relations is output.

[0020] Step S203: Obtain the credibility of the source of element evidence. Based on the set of equivalence relations and the credibility of the source of evidence as weighting factors, perform multi-dimensional dynamic consistency scoring on the core elements of the case and output the cross-file element consistency score S. The cross-file element consistency score S consists of the credibility of the source of element evidence, the semantic similarity of elements sim, and the contextual consistency score. The contextual consistency score is obtained by analyzing the paragraph semantic alignment relationship of the core elements of the case in the original multimodal evidence file set.

[0021] Step S204: Determine whether there are cross-source element conflicts based on the cross-document element consistency score S, and output the conflict set. And a consistent metadata set V.

[0022] Preferably, in step S30, the candidate case file summary generation model includes an input encoding layer, a semantic feature extraction layer, a multi-head self-attention layer, a conflict element highlighting layer, a summary generation layer, and a semantic alignment and output layer. Specifically, the input encoding layer performs multimodal feature encoding on the output conflict set C and the consistency metadata set V; the semantic feature extraction layer uses a pre-defined BERT semantic encoding network to extract contextual features from the input semantic vector; the multi-head self-attention layer models the correlation between cross-document elements and marks conflicting elements using a multi-head self-attention mechanism; the conflict element highlighting layer introduces explicit conflict weights for the marked conflicting elements; the summary generation layer generates a case file summary set based on a sequence-to-sequence structure Seq2Seq; and the semantic alignment and output layer generates and outputs the case file summary set. .

[0023] Preferably, in step S30, and based on the output case summary set The element anchor mapping table is generated by using the semantic alignment and evidence backtracking association method The step specifically includes: performing a semantic alignment operation on each case core element in the case summary set , and associating it with the original multi-modal evidence file set to generate an element anchor mapping table Map; the element anchor mapping table Map is used to record the source information of each case core element in the original multi-modal evidence file set, and the source information includes evidence source identification, original text offset, extraction confidence and backtracking path identification, to ensure that the case summary set and the multi-modal evidence file set realize a one-to-one corresponding backtracking association relationship.

[0024] Preferably, in step S40, based on the case summary set and the element anchor mapping table , a case directory index tree is constructed, and a multi-threaded consistent index tree is obtained by logically sequentially organizing the case directory index tree based on a multi-threaded transaction consistency parallel scheduling algorithm, which specifically includes:

[0025] Step S401: based on the case summary set and the element anchor mapping table , a case directory index tree with a bidirectional hierarchical index structure is constructed ;

[0026] Step S402: topological dependency analysis is performed on the i th node and the j th node in the case directory index tree , to obtain a task dependency matrix ; Based on the task dependency matrix , the task priority of the i th node in the case directory index tree is defined

[0027] Step S403: based on the task priority , a multi-threaded parallel scheduling queue Q is constructed, an improved two-phase commit protocol 2PC is introduced, and a three-phase commit task is executed based on the improved two-phase commit protocol 2PC, the three-phase commit task including a pre-commit phase, a consistency verification and a commit phase

[0028] Among them, for the pre-commit phase: for the task priority If the task priority is greater than the preset task priority threshold, the multi-thread preloading and locking operation are performed to avoid cross-thread write conflict;

[0029] For the consistency verification phase: perform a consistency check function on all nodes that have performed the locking operation :

[0030] ;

[0031] For the commit phase: if the consistency check function of all nodes is broadcasted by the global coordinator ; otherwise, a preset rollback function is executed ;

[0032] Step S404: After the completion of the commit phase, a multi-thread consistency index tree is generated.

[0033] Preferably, in step S50, based on the multi-thread consistency index tree, a preset multi-level cache layering strategy is used to perform access frequency statistics and file importance evaluation to obtain a file cache priority; and based on the file cache priority, an adaptive prefetch loading mechanism is used to perform high-concurrency and stable output of the multi-modal evidence file set, which specifically includes:

[0034] Step S501: First, a three-layer cache layering strategy is established based on the multi-thread consistency index tree, including a cache L1, a shared cache L2, and a long tail cache L3; a preset multi-level cache layering strategy is used to perform access frequency statistics, and based on the results of the access frequency statistics, the files in the multi-modal evidence file set are preliminarily allocated to the cache L1, the shared cache L2, or the long tail cache L3, and a first file cache priority is output;

[0035] Step S502: According to the element anchor mapping table , the coverage degree score of the core elements of the case with respect to the files in the multi-modal evidence file set is evaluated, the first file cache priority is secondarily weighted and adjusted according to the coverage degree score, and a second file cache priority is output as the final file cache priority;

[0036] Step S503: After obtaining the file cache priority, an adaptive prefetch loading mechanism is used to perform high-concurrency and stable output of the multi-modal evidence file set based on the file cache priority, which specifically includes: based on the multi-thread consistency index tree, for files with high cache priority, more thread resources are allocated to improve the loading speed; for low-priority files, a delayed loading strategy is adopted to avoid resource competition and bottlenecks under high concurrency conditions.

[0037] The application also provides an electronic file fast grouping device based on metadata intelligent extraction, which comprises:

[0038] A metadata extraction and normalization module is configured to obtain a multi-modal evidence file set, convert the multi-modal evidence file set into an initial metadata set M by using an OCR and ASR fusion extraction mechanism, perform weighted normalization processing on the initial metadata set M based on a BERT semantic encoding mode, and output a normalized metadata set .

[0039] A cross-file consistency analysis module is configured to construct a preliminary case file timeline , calculate a cross-file element consistency score S based on the preliminary case file timeline and the normalized metadata set by using a consistency mechanism based on BERT semantic similarity and weight addition, determine whether there is a cross-source element conflict based on the cross-file element consistency score S, and output a conflict set and a consistent metadata set V.

[0040] A candidate case file abstract generation and anchor mapping module is configured to construct a candidate case file abstract generation model by using a preset BERT model combined with a self-attention mechanism, input the conflict set and the consistent metadata set V into the candidate case file abstract generation model, output a case file abstract set , and generate an element anchor mapping table based on the output case file abstract set by using a semantic alignment and evidence backtracking association method .

[0041] A case file directory construction and index scheduling module is configured to construct a case file directory index tree based on the case file abstract set and the element anchor mapping table , and perform logical sequential organization based on the case file directory index tree by using a multi-thread transaction consistency parallel scheduling algorithm to obtain a multi-thread consistency index tree.

[0042] A high-concurrency cache and adaptive prefetching module is configured to obtain a file cache priority by using a preset multi-level cache hierarchical strategy to perform access frequency statistics and file importance evaluation based on the multi-thread consistency index tree, and perform high-concurrency stable output on the multi-modal evidence file set by using an adaptive prefetching loading mechanism based on the file cache priority.

[0043] The application also provides an electronic case file rapid grouping device based on metadata intelligent extraction, which comprises a memory, a processor, and an electronic case file rapid grouping program based on metadata intelligent extraction stored in the memory and capable of running on the processor.

[0044] The application also provides a computer program product comprising an electronic file fast grouping program based on metadata intelligent extraction, which realizes the electronic file fast grouping method based on metadata intelligent extraction when executed by a processor.

[0045] The application has the advantages that: the application realizes high-precision semantic fusion and conflict determination among multi-source and multi-modal evidence files by introducing multi-dimensional weighted normalization processing based on BERT semantic coding and cross-file consistency scoring mechanism S; meanwhile, high consistency association and semantic alignment of core elements of the file are realized by using file time axis T and element anchor mapping table, thereby significantly reducing the file logical confusion and element conflict problems caused by heterogeneous file sources in traditional methods.

[0046] The application realizes adaptive prefetching and stable loading of multi-modal evidence files in a high-concurrency access scenario by using multi-thread transaction consistency scheduling algorithm T_con combined with multi-level cache priority control strategy P; compared with the traditional serial grouping scheme, the application improves the file organization efficiency to a second-level response while maintaining the integrity and accuracy of the file, significantly improves the cache hit rate, and greatly optimizes the overall grouping efficiency and system throughput capacity. BRIEF DESCRIPTION OF DRAWINGS

[0047] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0048] Figure 1 The flowchart of the first embodiment of the electronic file fast grouping method based on metadata intelligent extraction of the present application.

[0049] Figure 2 The device schematic diagram of the electronic file fast grouping method based on metadata intelligent extraction of the present application. DETAILED DESCRIPTION

[0050] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0051] Embodiment one: as Figure 1As shown, it is the flow diagram of the first embodiment of the electronic file fast grouping method based on metadata intelligent extraction of the application, and the first embodiment of the electronic file fast grouping method based on metadata intelligent extraction of the application is proposed.

[0052] In the first embodiment, the electronic file fast grouping method based on metadata intelligent extraction comprises:

[0053] Step S10: Obtain a multi-modal evidence file set, convert the multi-modal evidence file set into an initial metadata set M by using an OCR and ASR fusion extraction mechanism, perform weighted normalization processing on the initial metadata set M based on a BERT semantic coding mode, and output a normalized metadata set ;

[0054] It should be noted that the multi-modal evidence file set refers to various types of digital evidence involved in case handling, electronic file sorting or judicial archive management. These evidences not only include common scanned documents, bills and contract texts, but also include audio calls, video surveillance, photographed pictures, structured reports, database export files and other multi-source heterogeneous data. Due to the significant differences in content expression and storage format of different evidence sources, the traditional single-modal processing method often cannot simultaneously consider the integrity and semantic consistency of data. The application proposes an OCR and ASR fusion extraction mechanism to solve this problem. By using OCR technology to extract text information in images, video frames and scanned documents, and combining ASR technology to analyze the voice content in audio and video, these scattered information is uniformly converted into an initial metadata set M. Each metadata item in the initial metadata set M not only records the original content, but also contains core elements such as data source identifier, timestamp, belonging case, field category and context position.

[0055] It can be understood that after the initial metadata set M is formed, the application introduces a BERT-based semantic coding network, which maps each metadata item to a unified high-dimensional semantic space by bidirectional modeling of the context environment, and assigns weights according to multi-dimensional factors such as data source, confidence, file type, etc. Then, the metadata of different sources are weighted and normalized to ensure that the same semantic concept can be expressed consistently in different files and different modalities.

[0056] For example, when the same transaction event is represented as "transfer amount: 12000 yuan" in a bank statement scanned document, and is extracted as "one million two" in the voice recognition result of a telephone recording, the semantic equivalence of the two can be recognized, and the two are merged into the same transaction event through a unified metadata standard. This mechanism effectively avoids the common semantic fragmentation problem in the traditional keyword matching method, so that the metadata remains highly consistent at the semantic level.

[0057] It should be understood that the present application has significant innovation and advantage in multi-modal fusion, semantic understanding and metadata normalization compared with traditional technology: multi-modal fusion capability, traditional methods often only support a single data source, such as only being able to process scanned copies or structured data, resulting in a lack of analysis capability for evidence content such as recordings, videos, etc. The OCR and ASR fusion mechanism of the present application can process multiple modal data at the same time, ensuring the integrity of the metadata set. Contextual semantic understanding, traditional keyword extraction methods rely on fixed templates, which are difficult to handle expression differences. The present application realizes context modeling through BERT semantic encoding, which can identify synonymous expressions, sequence transformations and cross-file references, improving cross-modal consistency. Cross-file consistency optimization, in traditional solutions, data from different sources are often stored and processed in isolation, resulting in inconsistencies in the same case in different files. The present application significantly reduces the cross-file conflict rate through weighted normalization and semantic unified modeling, providing a high-quality basis for subsequent case file grouping and index construction.

[0058] Step S20: constructing a preliminary case file timeline According to the preliminary case file timeline and the normalized metadata set A consistency mechanism based on BERT semantic similarity and weight addition is used to calculate the cross-file element consistency score S; based on the cross-file element consistency score S, it is determined whether there is a cross-source element conflict, and the conflict set and the consistency metadata set V are output;

[0059] It should be noted that the cross-file element consistency score S refers to a comprehensive score index defined in the electronic case file rapid grouping process to measure the semantic similarity and information credibility between the same case elements in different source evidence files. The preliminary case file timeline is used to preliminarily sort the event elements in all normalized metadata sets in chronological order, including case acceptance time, evidence collection time, event occurrence time, key witness statement time, etc. multi-dimensional time information.

[0060] It can be understood that this step uses a consistency mechanism based on BERT semantic similarity and weight addition, which maps the case file elements in different evidence files into a unified semantic space through semantic vector representation. The basic similarity value is calculated by calculating the semantic similarity of each pair of elements, and then combined with the source credibility, time proximity, context relevance, etc. of the case elements. dimensional weight addition correction, so as to obtain the cross-file element consistency score S. When the score S exceeds the preset consistency threshold, it is determined that the elements in different evidence files are semantically consistent; otherwise, the elements are marked as possibly conflicting. In this way, this step realizes the automatic calculation of the consistency of cross-source evidence elements, and provides a basis for generating the consistency metadata set V and the conflict set.

[0061] It should be understood that, compared with the prior art based on keyword matching or templated verification, the present application effectively breaks through the recognition bottleneck caused by inconsistent expression forms of case elements in different files by introducing BERT semantic modeling and weight addition mechanism. The traditional method is prone to semantic mismatch and logical breakage when facing the same case facts with multiple modalities, multiple sources and multiple expressions, resulting in insufficient conflict detection accuracy. The method of the present application can realize equivalent recognition through context semantic understanding, and combine the preliminary case file timeline to constrain the temporal consistency of evidence, thereby significantly improving the accuracy and integrity of cross-source evidence fusion, avoiding the fragmentation of case elements caused by expression differences, and ensuring the logical coherence of subsequent automatic case compilation.

[0062] Step S30: a candidate case summary generation model is constructed by using a preset BERT model combined with a self-attention mechanism, the conflict set C and the consistency metadata set V are input into the candidate case summary generation model, and a case summary set S is output. ; and based on the output case summary set S An element anchor mapping table is generated by using a semantic alignment and evidence backtracking association method ;

[0063] It should be noted that the candidate case summary generation model refers to a generative model based on a preset BERT deep semantic understanding model, and combined with a self-attention mechanism to perform semantic modeling and content compression on the input conflict set C and consistency metadata set V. The core function of the model is to extract a case summary set S with high expressiveness and high generalization for the core elements of the case from a large number of scattered, redundant or even conflicting evidence information. The self-attention mechanism is used to dynamically model the semantic dependency and logical correlation between different evidences within the input sequence, so as to automatically determine which evidence fragments need to be preferentially retained in the summary, and which redundant or low-relevance information can be weakened or discarded.

[0064] ​It can be understood that the technical effect of this step is to automatically realize the high integration and information condensation of multi-source evidence through the semantic-driven evidence summary and backtracking association mechanism. The BERT model obtains the semantic vector representation of each evidence element through context bidirectional encoding, and then accurately captures the long-distance dependence and cross constraint between elements by means of self-attention mechanism, thereby establishing a semantic co-occurrence network between different evidences. Based on this semantic network, the candidate case file summary generation model can automatically select high-confidence elements with representative for the conflicts and redundancies in multi-modal evidence data, so that the output case file summary set has high coverage and high consistency. At the same time, the semantic alignment and evidence backtracking association method maps the summary elements and the original evidence one by one, ensuring the traceability and verifiability of the summary content, and avoiding semantic distortion or loss of legal effectiveness caused by information compression.

[0065] It should be understood that, compared with the prior art which relies on manual screening or fixed summary method based on templates, the present application introduces BERT semantic modeling and self-attention mechanism, which significantly improves the intelligent degree and context awareness of case file summary generation. Traditional technologies often cannot effectively process complex evidence from different agencies, different formats and different modalities, especially when encountering inconsistent expressions or conflicting information, the summary generation result is easy to miss, repeat or even have logical errors. The method of the present application not only understands the deep semantic relationship between case elements, but also realizes the explainable linking of cross-file multi-modal evidence through evidence backtracking, thereby ensuring the accuracy and reliability of the case file summary content in legal application scenarios. This technology breaks through the bottleneck of low efficiency and high error rate of traditional manual methods, and realizes the automation, intelligence and traceability of electronic files in large-scale case processing.

[0066] For example, when integrating bank statements, communication records, video surveillance, witness statements and other multi-modal evidence files, the traditional method needs to manually screen each evidence to extract core points. After using the method of the present application, the suspicious fund flow records in the conflict set and the cross-border communication records in the consistency metadata set V are uniformly input into the candidate case file summary generation model, the semantic importance of different evidence fragments is calculated through BERT and self-attention mechanism, and a case file summary set is automatically generated. The key transaction chain and cross-border contact relationship in the summary will be highlighted. At the same time, based on the evidence backtracking association method, the element anchor mapping table is generated, so that each high-value transaction node in the summary can be directly located to the original bank statement or call record. It has been verified that, in the processing of cases of the same size, the case file summary generation time is shortened, the accuracy of core element backtracking positioning is high, and the present application is significantly better than the traditional method.

[0067] Step S40: generating a case file summary set based on the case file summary set and the element anchor mapping table A case directory index tree is constructed, and logical sequential organization is performed based on the case directory index tree through a multi-thread transaction consistency parallel scheduling algorithm to obtain a multi-thread consistency index tree.

[0068] It should be noted that the case directory index tree refers to a multi-level tree data structure that is based on the case abstract set and the element anchor mapping table, and sequentially organizes the core elements, evidence fragments and their context logical relationships in the case through a hierarchical structure. Each index node corresponds to a case element or evidence abstract, and establishes a bidirectional reference relationship with the original evidence data through the anchor mapping table, thereby realizing efficient retrieval and traceability of the case content.

[0069] It can be understood that the technical effect of this step is to realize efficient logical organization and structured expression of large-scale case data. Through the case directory index tree, a multi-level index structure with semantic consistency can be constructed between multi-modal evidence, and high interpretability and low redundancy logical association can be established between evidence, abstracts and elements from different sources. At the same time, combined with the multi-thread transaction consistency parallel scheduling algorithm, the parallel computing capability can be fully utilized in a multi-core hardware environment, and the complex case organization task that needs to be processed in series is divided into multiple independent threads, which significantly reduces the processing delay and improves the throughput. In addition, through the transaction consistency mechanism, each update of the index tree has atomicity, isolation and recoverability, so that node disorder or data loss problems will not occur even in a high-concurrency scenario.

[0070] It should be understood that compared with the case directory generation method based on single-thread sequence in the prior art, the present application has made a significant breakthrough in two aspects. Firstly, the traditional method relies on manual or linear algorithm to sort and associate case abstracts one by one, and when faced with massive, multi-modal and cross-source evidence, the efficiency is low and logical breaks or reference failures are prone to occur. The present application realizes dynamic bidirectional binding of abstracts and original evidence through the element anchor mapping table, ensuring the integrity and verifiability of the logical relationship. Secondly, the prior art lacks high-concurrency processing capability, and once a case involves a large number of parallel evidence files, index update blocking and transaction inconsistency problems are prone to occur; while the present application introduces a multi-thread transaction consistency algorithm based on lock granularity optimization, which realizes real-time scheduling and parallel execution of high-concurrency tasks while ensuring data consistency, improving the speed and scalability of electronic case filing.

[0071] For example, in the test data processing of a commercial bribery case, the case evidence files exceeded 1.2TB, containing multimodal data sources such as scanned contracts, bank statements, chat logs, and video surveillance. Constructing the case file catalog using traditional sequential indexing methods took approximately 14 days, with an average response latency of up to 3.7 seconds when retrieving evidence. Applying the method of this invention, a case file catalog index tree is first generated based on the case file summary set and the element anchor mapping table. Then, a multi-threaded transactional consistency parallel scheduling algorithm is used to simultaneously create hundreds of index nodes on a 16-core CPU server and dynamically bind the original evidence files. In the test environment, the time to construct the complete case file catalog index tree was reduced to 26 hours, the average latency for evidence retrieval was reduced to 0.42 seconds, and the index tree data consistency and reference correctness were maintained even under high-concurrency scenarios (1000 concurrent requests), significantly outperforming the performance of traditional methods.

[0072] Step S50: Based on the multi-threaded consistent index tree, a preset multi-level caching strategy is used to perform access frequency statistics and file importance assessment to obtain file cache priority; based on the file cache priority, an adaptive prefetch loading mechanism is used to output the multimodal evidence file set with high concurrency and stability.

[0073] It's important to note that the multi-level caching strategy, based on a multi-threaded consistent index tree, divides the multimodal evidence file set into different levels of cache space according to access frequency, file importance, and latency sensitivity. This typically includes three types: high-speed cache (L1), shared cache (L2), and long-tail cache (L3). High-speed cache (L1) stores frequently accessed evidence files and supports low-latency access; shared cache (L2) stores files with medium access frequency and cross-thread sharing; and long-tail cache (L3) provides a capacity-optimized storage strategy for low-frequency long-tail files. This hierarchical design ensures fast response times for high-frequency files while preventing contention for high-performance resources by low-frequency files. File cache priority is a quantitative indicator calculated based on access frequency, importance weight, and latency modeling results, used to guide the subsequent adaptive prefetching mechanism to ensure the dynamic scheduling efficiency of multimodal evidence files in high-concurrency scenarios.

[0074] It can be understood that the technical effect of this step is to significantly improve the high concurrency performance and resource utilization of electronic files in large-scale, multi-modal evidence data processing. By combining multi-level cache layering strategy on the multi-threaded consistent index tree, the access characteristics of different evidence files can be accurately distinguished, thereby optimizing the cooperative processing efficiency between CPU and I / O. When combined with file cache priority for adaptive prefetch loading, the prefetch batch, cache level and loading order can be dynamically determined according to access heat prediction and cache delay model, realizing stable output under high concurrency. This design not only reduces the disk I / O bottleneck and cache jitter problem, but also significantly shortens the lock waiting time in the multi-threaded concurrent access process, improving the real-time and continuity of evidence retrieval and retrieval.

[0075] It should be understood that, unlike the prior art electronic file system which relies on single-level cache or fixed priority scheduling, the present application solves the problems of cache imbalance, response delay and unstable output under high concurrency access by introducing dynamic multi-level cache strategy and adaptive prefetch loading mechanism. Most of the prior art is based on static cache allocation or fixed access heat sorting, which cannot adjust the file priority in real time, resulting in inefficient behaviors such as cache overflow, frequent invalidation and repeated loading in multi-threaded scenarios. The present application not only realizes dynamic division driven by access characteristics in cache layering, but also through quantitative evaluation of file cache priority, combined with thread-safe prefetch loading algorithm, ensures stable output of evidence files under high concurrency conditions.

[0076] For example, in the processing of certain judicial case test data, a single case file contains more than 3TB of multi-modal evidence files, including scanned documents, audio recordings, video surveillance and chat records, with a peak concurrent access volume of 1200 threads / second. When using the traditional single-level cache strategy, the average evidence loading delay is 3.8 seconds under high concurrency, and the cache hit rate is only 41% at peak, with frequent access timeouts caused by cache jitter. After using the method of the present application, the file is dynamically prioritized by combining the multi-threaded consistent index tree with the three-layer cache strategy, and based on the adaptive prefetch loading mechanism driven by cache priority, the average evidence loading delay is reduced to 0.57 seconds, the cache hit rate is increased to 92%, and the stability and consistency of the output stream are maintained under high concurrency, improving the processing capacity of electronic files.

[0077] Embodiment two: In addition, the application provides a kind of electronic file fast group volume device based on metadata intelligent extraction, using the electronic file fast group volume method based on metadata intelligent extraction in above-mentioned embodiment, can solve the technical problem of a kind of electronic file fast group volume based on metadata intelligent extraction.Compared with prior art, the beneficial effects of the electronic file fast group volume device based on metadata intelligent extraction provided in the application are the same as the beneficial effects of the electronic file fast group volume method based on metadata intelligent extraction provided in the above-mentioned embodiment, and other technical features in the electronic file fast group volume device based on metadata intelligent extraction are the same as the features disclosed in the above-mentioned embodiment method, which will not be repeated here.

[0078] Embodiment three: the application provides a kind of electronic file fast group volume equipment based on metadata intelligent extraction, please refer to Figure 2An electronic file fast grouping device based on metadata intelligent extraction includes at least one processor, and a memory connected to the at least one processor in communication; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the electronic file fast grouping method based on metadata intelligent extraction in the above-mentioned embodiment one. The electronic file fast grouping device based on metadata intelligent extraction in the embodiment of the present application can include, but is not limited to, mobile terminals such as mobile phones, notebook computers, digital broadcast receivers, PDAs (Personal Digital Assistant), PADs (Portable Application Description), PMPs (Portable Media Player), vehicle-mounted terminals (such as vehicle-mounted navigation terminals), and the like, and fixed terminals such as digital TVs, desktop computers, and the like. The electronic file fast grouping device based on metadata intelligent extraction is only an example, and should not bring any limitation to the function and use range of the embodiment of the present application. The electronic file fast grouping device based on metadata intelligent extraction can include a processing device 1001 (such as a central processor, a graphics processor, etc.), which can perform various appropriate actions and processes according to programs stored in a read-only memory 1002 or loaded from a storage device 1003 to a random access memory 1004. In the random access memory 1004, various programs and data required for the operation of the electronic file fast grouping device based on metadata intelligent extraction are also stored. The processing device 1001, the read-only memory 1002, and the random access memory 1004 are connected to each other through a bus 1005. An I / O interface 1006 is also connected to the bus. Generally, the following devices can be connected to the I / O interface 1006: input devices 1007 including, for example, touch screens, touchpads, keyboards, mice, image sensors, microphones, accelerometers, gyroscopes, and the like; output devices 1008 including, for example, liquid crystal displays (LCDs), speakers, vibrators, and the like; storage devices 1003 including, for example, magnetic tapes, hard disks, and the like; and communication devices 1009. The communication devices 1009 can allow the electronic file fast grouping device based on metadata intelligent extraction to communicate with other devices wirelessly or by wire to exchange data. Although the electronic file fast grouping device based on metadata intelligent extraction with various devices is shown in the figure, it should be understood that all the shown devices are not required to be implemented or possessed. More or less devices can be alternatively implemented or possessed.

[0079] Embodiment four: the application also provides a computer program product, comprising a computer program, which, when executed by a processor, implements the steps of the electronic file fast grouping method based on metadata intelligent extraction as described above. The computer program product provided by the application can solve the technical problem of the electronic file fast grouping method based on metadata intelligent extraction. Compared with the prior art, the beneficial effects of the computer program product provided by the application are the same as those of the electronic file fast grouping method based on metadata intelligent extraction provided by the above-mentioned embodiment, and are not described here.

[0080] In particular, according to the embodiments disclosed by the application, the processes described above with reference to the flowcharts can be implemented as a computer software program. For example, the embodiments disclosed by the application include a computer program product comprising a computer program carried on a computer readable medium, the computer program comprising program code for executing the method shown in the flowchart. In such embodiments, the computer program can be downloaded and installed from a network through a communication device, or installed from a storage device 1003, or installed from a read-only memory 1002. When the computer program is executed by a processing device 1001, the above-mentioned functions defined in the method of the embodiments disclosed by the application are executed.

[0081] It should be understood that various parts of the application disclosed can be realized in hardware, software, firmware or a combination thereof. In the description of the above-mentioned embodiments, specific features, structures, materials or characteristics can be combined in any one or more embodiments or examples in a suitable manner.

[0082] Obviously, those skilled in the art can make various modifications and variations to the application without departing from the spirit and scope of the application. Thus, if these modifications and variations of the application fall within the scope of the claims of the application and their equivalent technologies, the application also intends to include these modifications and variations.

Claims

1. A method for rapid electronic case file assembly based on intelligent metadata extraction, characterized in that, The methods include: Step S10: Obtain the multimodal evidence file set, use the OCR and ASR fusion extraction mechanism to convert the multimodal evidence file set into an initial metadata set M, and perform weighted normalization processing on the initial metadata set M based on the BERT semantic encoding method to output the normalized metadata set. ; Step S20: Constructing a preliminary case file timeline According to the preliminary case file timeline and normalized metadata set A consistency mechanism based on BERT semantic similarity and weighted summation is used to calculate the cross-file feature consistency score S; based on the cross-file feature consistency score S, it is determined whether there are cross-source feature conflicts, and a conflict set is output. and consistent metadata set V; Step S30: Construct a candidate case summary generation model using a pre-defined BERT model combined with a self-attention mechanism, and process the conflict set. The consistent metadata set V is input into the candidate case summary generation model, and the output is a case summary set. Based on the output set of case summary data. An anchor point mapping table is generated using semantic alignment and evidence backtracking association methods. ; Step S40: Based on the case file summary set and feature anchor mapping table Construct a case file directory index tree, and organize it logically in order using a multi-threaded transaction consistency parallel scheduling algorithm to obtain a multi-threaded consistent index tree; Step S50: Based on the multi-threaded consistent index tree, perform access frequency statistics and file importance assessment using a preset multi-level caching strategy to obtain file cache priority; based on the file cache priority, use an adaptive prefetch loading mechanism to output the multimodal evidence file set with high concurrency and stability.

2. The method for rapid electronic case file assembly based on intelligent metadata extraction as described in claim 1, characterized in that, In step S10, a multimodal evidence file set is obtained. An OCR and ASR fusion extraction mechanism is used to convert the multimodal evidence file set into an initial metadata set M. Then, based on BERT semantic encoding, a weighted normalization process is performed on the initial metadata set M to output a normalized metadata set. The steps specifically include: Step S101: Obtain a multimodal evidence file set, which includes scanned PDF files, photographed PNG files, and speech-to-text TXT files; use an OCR and ASR fusion extraction mechanism to uniformly convert the multimodal evidence file set into an intermediate JSON structure to obtain the initial metadata set M; Step S102: Input the initial metadata set M into the preset BERT semantic coding network to extract the initial semantic vector; based on the initial semantic vector, use the Conditional Random Field (CRF) method to extract the core elements of the case, which include time elements, party elements, amount elements, and location elements; Step S103: Based on the extracted core elements of the case, calculate the element similarity score using a multi-dimensional dynamic semantic similarity method that integrates credibility. Perform multi-dimensional weighted normalization processing on the initial metadata set M according to the element similarity score, and output the normalized metadata set. .

3. The method for rapid electronic case file assembly based on intelligent metadata extraction as described in claim 1, characterized in that, In step S20, a preliminary case file timeline is constructed. According to the preliminary case file timeline and normalized metadata set A consistency score S for cross-document elements is calculated using a consistency mechanism based on BERT semantic similarity and weighted addition. Based on the cross-document feature consistency score S, it is determined whether cross-source feature conflicts exist, and the conflict set is output. The steps for creating a consistent metadata set V specifically include: Step S201: Based on the normalized metadata set The core elements of the case related to time are extracted. Based on the extracted core elements, the element similarity score is calculated using a multi-dimensional dynamic semantic similarity method that integrates credibility. The time elements of different source files are uniformly converted using a preset time format standardization function. The core elements of the case with element similarity scores greater than the preset element similarity score threshold are aggregated into the same event node, thereby constructing a preliminary case file time axis T. The preliminary case file time axis T is used to provide temporal constraints for the subsequent calculation of consistency scores. Step S202: For the normalized metadata set The core elements of the case are extracted using a pre-defined BERT semantic encoding network to obtain the corresponding semantic vector h. Based on the semantic vector h, the cosine similarity calculation method is used to calculate the element semantic similarity sim of the core elements of the case between different files. Then, the semantic similarity sim and the pre-defined similarity threshold τ are used to determine whether the core elements of the case are semantically equivalent: elements with a semantic similarity sim higher than the similarity threshold τ are marked as semantically equivalent, and those with a similarity sim lower than the threshold τ are marked as non-equivalent. The set of equivalence relations is output. Step S203: Obtain the credibility of the source of element evidence. Based on the set of equivalence relations and the credibility of the source of evidence as weighting factors, perform multi-dimensional dynamic consistency scoring on the core elements of the case and output the cross-file element consistency score S. The cross-file element consistency score S consists of the credibility of the source of element evidence, the semantic similarity of elements sim, and the contextual consistency score. The contextual consistency score is obtained by analyzing the paragraph semantic alignment relationship of the core elements of the case in the original multimodal evidence file set. Step S204: Determine whether there are cross-source element conflicts based on the cross-document element consistency score S, and output the conflict set. And a consistent metadata set V.

4. The method for rapid electronic case file assembly based on intelligent metadata extraction as described in claim 1, characterized in that, In step S30, the candidate case file summary generation model includes an input encoding layer, a semantic feature extraction layer, a multi-head self-attention layer, a conflict element highlighting layer, a summary generation layer, and a semantic alignment and output layer. Specifically, the input encoding layer encodes multimodal features of the output conflict set C and the consistency metadata set V; the semantic feature extraction layer extracts contextual features from the input semantic vector using a pre-defined BERT semantic encoding network; the multi-head self-attention layer models the correlation between cross-document elements and marks conflicting elements using a multi-head self-attention mechanism; the conflict element highlighting layer introduces explicit conflict weights for the marked conflicting elements; the summary generation layer generates a case file summary set based on a sequence-to-sequence structure (Seq2Seq); and the semantic alignment and output layer generates and outputs the case file summary set. .

5. The method for rapid electronic case file assembly based on intelligent metadata extraction as described in claim 1, characterized in that, In step S30, based on the output case summary set... An anchor point mapping table is generated using semantic alignment and evidence backtracking association methods. The specific steps include: compiling a collection of case file summaries. Each core element of the case is semantically aligned and associated with the original multimodal evidence file set to generate an element anchor mapping table (Map). This Map records the source information of each core element in the original multimodal evidence file set, including evidence source identifier, original text offset, extraction confidence level, and backtracking path identifier, to ensure the accuracy of the case summary set. A one-to-one, traceable association is established between the multimodal evidence file set and the data.

6. The method for rapid electronic case file assembly based on intelligent metadata extraction as described in claim 1, characterized in that, In step S40, based on the case file summary set and feature anchor mapping table The steps of constructing a case file directory index tree and then logically organizing it using a multi-threaded transaction consistency parallel scheduling algorithm to obtain a multi-threaded consistent index tree include: Step S401: Based on the case file summary set and feature anchor mapping table Construct a case file catalog index tree with a two-way hierarchical index structure ; Step S402: Targeting the case file catalog index tree The i-th node in and the j-th node Perform topological dependency analysis to obtain the task dependency matrix. ; Based on the task dependency matrix Define the case file directory index tree The i-th node in Task priority ; Step S403: Based on task priority A multi-threaded parallel scheduling queue Q is constructed, and an improved two-phase commit protocol 2PC is introduced. Based on the improved two-phase commit protocol 2PC, a three-phase commit task is executed, which includes a pre-commit phase, a consistency verification phase, and a commit phase. Specifically, for the pre-submission phase: regarding task priority... Tasks with priorities greater than the preset threshold will undergo multi-threaded preloading and locking operations to avoid cross-thread write conflicts. During the consistency verification phase: a consistency check function is executed on all nodes that have undergone locking operations. : ； For the commit phase: if the consistency check function of all nodes... By the global coordinator Broadcast the commit command; otherwise, execute the preset rollback function. ; Step S404: After the commit phase is completed, generate a multi-threaded consistent index tree.

7. The method for rapid electronic case file assembly based on intelligent metadata extraction as described in claim 1, characterized in that, In step S50, based on the multi-threaded consistent index tree, a preset multi-level cache hierarchy strategy is used to perform access frequency statistics and file importance evaluation to obtain file cache priority. The steps for achieving high-concurrency and stable output of multimodal evidence file sets using an adaptive prefetching mechanism based on file cache priority include: Step S501: First, establish a three-level cache hierarchy strategy based on the multi-threaded consistent index tree, including high-speed cache L1, shared cache L2, and long-tail cache L3; perform access frequency statistics using the preset multi-level cache hierarchy strategy, and preliminarily allocate files in the multimodal evidence file set to high-speed cache L1, shared cache L2, or long-tail cache L3 based on the results of the access frequency statistics, and output the first file cache priority. Step S502: Based on the feature anchor point mapping table The coverage score of the core elements of the case relative to the files in the multimodal evidence file set is evaluated. The first file cache priority is then adjusted by a second weighting based on the coverage score, and the second file cache priority is output as the final file cache priority. Step S503: After obtaining the file cache priority, an adaptive prefetch loading mechanism is adopted based on the file cache priority to output the multimodal evidence file set in a high-concurrency and stable manner. Specifically, based on the multi-threaded consistent index tree, more thread resources are allocated to files with high cache priority to improve loading speed; for low priority files, a delayed loading strategy is adopted to avoid resource competition and bottlenecks under high concurrency conditions.

8. A device for rapid electronic case file assembly based on intelligent metadata extraction, applied to the method for rapid electronic case file assembly based on intelligent metadata extraction as described in any one of claims 1 to 7, characterized in that, The electronic case file rapid assembly device based on metadata intelligent extraction includes: The metadata extraction and normalization module is used to acquire a multimodal evidence file set. It employs a combined OCR and ASR extraction mechanism to convert the multimodal evidence file set into an initial metadata set M. Then, based on BERT semantic encoding, it performs weighted normalization processing on the initial metadata set M, outputting a normalized metadata set. ; The cross-file consistency analysis module is used to construct a preliminary case file timeline. According to the preliminary case file timeline and normalized metadata set A consistency mechanism based on BERT semantic similarity and weighted summation is used to calculate the cross-file feature consistency score S; based on the cross-file feature consistency score S, it is determined whether there are cross-source feature conflicts, and a conflict set is output. and consistent metadata set V; The candidate case file summary generation and anchor mapping module is used to construct a candidate case file summary generation model using a pre-defined BERT model combined with a self-attention mechanism, and to generate conflict sets. The consistent metadata set V is input into the candidate case summary generation model, and the output is a case summary set. Based on the output set of case summary data. An anchor point mapping table is generated using semantic alignment and evidence backtracking association methods. ; The case file catalog construction and index scheduling module is used to build and index the case file summary set. and feature anchor mapping table Construct a case file directory index tree, and organize it logically in order using a multi-threaded transaction consistency parallel scheduling algorithm to obtain a multi-threaded consistent index tree; The high-concurrency caching and adaptive prefetching module is used to perform access frequency statistics and file importance assessment based on a multi-threaded consistent index tree and a preset multi-level caching hierarchy strategy to obtain file cache priority. Based on the file cache priority, an adaptive prefetching loading mechanism is used to output the multimodal evidence file set with high concurrency and stability.

9. A rapid electronic case file assembly device based on intelligent metadata extraction, characterized in that, The electronic case file rapid assembly device based on metadata intelligent extraction includes: a memory, a processor, and an electronic case file rapid assembly program based on metadata intelligent extraction stored in the memory and executable on the processor. When the electronic case file rapid assembly program based on metadata intelligent extraction is executed by the processor, it implements the electronic case file rapid assembly method based on metadata intelligent extraction as described in any one of claims 1 to 7.

10. A computer program product, characterized in that, The computer program product includes a rapid electronic case file assembly program based on intelligent metadata extraction. When the rapid electronic case file assembly program based on intelligent metadata extraction is executed by the processor, it implements a rapid electronic case file assembly method based on intelligent metadata extraction as described in any one of claims 1 to 7.

Images