An intellectual property big data intelligent retrieval method and system based on semantic understanding

Abstract

The application discloses a kind of intellectual property big data intelligent retrieval method and system based on semantic understanding, it is related to big data retrieval technical field, first, collect intellectual property original data, store into cloud original distributed database after uniform format etc.Processing;Then, construct quality annotation dataset and semantic verification model to generate quality label and store into cloud library;Again, parse user search request to generate semantic vector;After hierarchical retrieval matching, low-quality data is corrected and matching score is calculated to generate initial result;Finally, collect user behavior data to update algorithm parameters, re-verify and write back quality information, the present application builds cloud original distributed database and stores intellectual property original data and standardizes processing, constructs quality annotation dataset and semantic verification model to realize quality accurate determination;At the same time, the field analysis of search request is resolved, hierarchical retrieval matching is carried out, the retrieval algorithm is corrected and optimized through closed loop system, the accuracy and efficiency are improved, and user demand is adapted.

Description

Technical Field

[0001] This invention relates to the field of big data retrieval technology, specifically to an intelligent retrieval method and system for intellectual property big data based on semantic understanding. Background Technology

[0002] In the current context of rapid development in the intellectual property industry, intellectual property data is characterized by massive growth and diverse types. Data such as patents, trademarks, copyrights, and integrated circuit layout designs are scattered across various public channels, becoming core resources supporting intellectual property research, planning, and rights protection. With the integration and application of big data and artificial intelligence technologies, the field of intellectual property retrieval is gradually upgrading from traditional keyword retrieval to semantic understanding retrieval. The industry's requirements for the accuracy, efficiency, and standardization of data processing are constantly increasing. At the same time, intellectual property data is characterized by high professionalism, complex field information, and rigorous classification systems, which increases the difficulty of its storage management and quality verification. The maturity of technologies such as cloud-native distributed storage and natural language processing provides a technical foundation for the intelligent processing and retrieval of intellectual property big data. Building a semantic understanding intelligent retrieval method adapted to the intellectual property field has become an inevitable requirement for industry development.

[0003] Traditional intellectual property big data retrieval technologies have many shortcomings in practical applications. First, in the data processing stage, there is a lack of systematic collection and standardized processing procedures. Raw data from different sources has varying formats and chaotic fields, and there are common problems such as data duplication and missing information. Moreover, the data storage architecture is difficult to achieve elastic expansion and dynamic scheduling, and cannot adapt to the storage needs of massive intellectual property data. At the same time, there is a lack of a professional quality verification system, making it difficult to accurately identify various quality problems in the data, such as abnormal classification number matching and incorrect terminology labeling. Low-quality data directly affects the retrieval results. Second, in the retrieval stage, the parsing of user requests is limited to the keyword level, without combining the professional characteristics of the intellectual property field to identify technical entities and normalize terms. It is difficult to extract deep technical elements and retrieval constraints. The retrieval matching is not processed hierarchically, and there is a lack of a scientific correction mechanism for low-quality data. Furthermore, the retrieval algorithm lacks closed-loop optimization capabilities and cannot iteratively update based on user operation behavior, resulting in low retrieval accuracy and large deviations in results. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a semantic understanding-based intelligent retrieval method and system for intellectual property big data. It innovates on the entire process of intellectual property big data retrieval by first standardizing and distributing the collection of various types of original intellectual property data, then constructing a semantic verification model to accurately determine data quality, subsequently performing domain-specific semantic analysis on user search requests, and using hierarchical retrieval matching combined with proprietary algorithms to correct low-quality data and calculate matching scores. Finally, it relies on user behavior data to form a feedback loop optimization between algorithms and data quality. This invention forms a complete technical system from data source to retrieval implementation and iterative optimization, significantly improving the accuracy and efficiency of intellectual property big data retrieval, adapting to diverse retrieval needs in the intellectual property field, and providing reliable technical support for intellectual property-related work for various entities.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution: On one hand, a method for intelligent retrieval of intellectual property big data based on semantic understanding, the specific steps of which are as follows: S1, Data Collection and Storage: Collect original intellectual property data, unify the format, standardize fields, remove duplicates and complete information, extract unique identifiers, full text, IPC classification numbers, terminology annotations and bibliographic fields, and store them in a cloud-native distributed database. The cloud-native distributed database adopts a distributed sharded storage architecture. The database supports elastic expansion of data storage capacity and dynamic scheduling of data storage nodes. The database constructs a distributed retrieval index based on the unique identifier of the stored intellectual property data. The database uses a multi-replica storage method to persistently store intellectual property data. The database sets independent storage partitions for different types of intellectual property data. The database supports real-time writing and batch writing operations of intellectual property data. S2, Quality Model Construction: Based on standardized data, a quality labeling dataset is constructed, a semantic verification model is built, and the data is pre-verified through the semantic verification model to generate a list of quality labels and quality problem dimensions, which are then bound with unique identifiers and stored in the cloud database; The quality annotation dataset consists of annotated samples of classification number mismatch, annotated samples of terminology mislabeling, annotated samples of missing field information, and annotated samples of text semantic conflict. All samples in the quality annotation dataset are taken from standardized intellectual property data. The sample annotation content includes unique data identifiers, full-text annotation information, IPC classification number annotation information, terminology annotation information, bibliographic field annotation information, and quality problem type annotation information. S3, Request Semantic Parsing: Receives user text retrieval requests, performs word segmentation, stop word filtering, technical entity recognition and terminology normalization, extracts technical elements and retrieval constraints, and generates standardized retrieval semantic vectors; S4, Hierarchical retrieval and matching: Data is filtered based on semantic vectors and limiting conditions, hierarchical retrieval is performed according to quality labels, causal contribution algorithm is run on low-quality data to obtain causal contribution, correction priority is generated and corrected step by step based on causal contribution, matching score is calculated through closed-loop optimization retrieval and matching algorithm, hierarchical retrieval results are merged, deduplicated and sorted to generate initial retrieval results. S5, Feedback Closed-Loop Optimization: Collect user operation behavior data, incrementally update the parameters of data quality confidence algorithm, causal contribution algorithm and closed-loop optimization retrieval matching algorithm, re-verify the associated data through the updated semantic verification model, and write back the updated quality tags and quality problem dimension list to the cloud database. User action behavior data includes search result click data, favorite data, rejection data, erroneous labeling data, and correct result supplementary data. Incremental updates adjust the coefficients of the data quality confidence algorithm, the factors of the causal contribution algorithm, and the weights of the closed-loop optimization search matching algorithm. The updated semantic verification model performs full-field quality checks on the associated data.

[0006] Furthermore, in S1, during data collection and storage, the collected original intellectual property data includes patent data, trademark data, copyright data, and integrated circuit layout design data. Unified format processing performs format conversion operations on text data, tabular data, and layout data from different public sources. Field standardization processing performs operations to unify field names, field types, and field lengths for all data. Deduplication processing identifies and removes duplicate data based on unique data identifiers. Information completion processing fills in the corresponding basic information for missing fields in the bibliographic entries. The processed data is classified according to unique data identifiers, and the classified data is stored in a cloud-native distributed database.

[0007] Furthermore, in S2, the construction of the quality model includes building a semantic verification model, which involves selecting a pre-trained language model based on the Transformer architecture as the initial base model, dividing the quality-labeled dataset into a training subset, a validation subset, and a test subset according to a preset ratio, inputting the sample data from the training subset into the initial base model, performing multiple rounds of iterative adjustments on the hidden layer parameters of the encoder, the weight parameters of the attention mechanism, and the connection parameters of the feedforward network in the initial base model, using the validation subset to perform data feature matching verification on the initial base model after parameter adjustment, using the test subset to perform data classification verification on the initial base model after data feature matching verification, determining the initial base model that passes the data feature matching verification and data classification verification as the semantic verification model, and embedding a data quality confidence algorithm in the internal operation layer of the semantic verification model.

[0008] Furthermore, in S2, during the quality model construction, the semantic verification model consists of a data input layer, a feature refinement and extraction layer, a multi-layer feature encoding layer, an algorithm integration and operation layer, and a classification result output layer. The data input layer receives standardized intellectual property data that has been formatted, field-regulated, deduplicated, and information-completed. The feature refinement and extraction layer extracts unique identifier character features, full-text semantic features, IPC classification number encoding features, terminology annotation text features, and bibliographic field attribute features from the intellectual property data. The multi-layer feature encoding layer performs hierarchical feature fusion and standardized vector transformation on the extracted features. The algorithm integration and operation layer loads and runs the data quality confidence algorithm to complete the data quality score calculation. The classification result output layer outputs the corresponding quality labels and quality problem dimension list for the intellectual property data.

[0009] Furthermore, the mathematical expression for the data quality confidence algorithm used in the quality model construction in S2 is: in The confidence score for intellectual property data quality. For multidimensional quality characteristic values ​​of intellectual property data, The weighted coefficients for intellectual property features are calculated statistically based on the annotation frequency of different quality features in the intellectual property field, with values ​​ranging from 0 to 1. β is the semantic anomaly penalty factor, calculated statistically based on the distribution characteristics of semantic deviation samples, with values ​​ranging from 0.5 to 2. This represents the semantic deviation value for a single piece of intellectual property data. This represents the global semantic mean value of the entire intellectual property data.

[0010] Furthermore, in S3, the semantic parsing request uses a dedicated dictionary for intellectual property to perform text splitting, technical entity identification identifies technical names, structural names, and method names, terminology normalization converts different professional terms into unified expressions, technical elements include technical themes, technical means, and application scenarios, search limiting conditions include IPC classification number range, application time period, and applicant range, and the standardized search semantic vector is a dense vector with fixed dimensions.

[0011] Furthermore, the mathematical expression for the causal contribution algorithm used in the hierarchical retrieval and matching process in S4 is: in The numerical value of causal contribution to intellectual property data quality issues. The deviation value of the search results is obtained by calculating the difference between the search result vector and the standard result vector. The change in quality confidence caused by a quality issue is calculated by subtracting the quality confidence score before and after the quality issue was corrected. γ is the confusion factor in the intellectual property field, calculated by statistically analyzing the distribution of easily confused features in the intellectual property field, with a value ranging from 1 to 3. The weighted index value for the retrieval scenario is calculated by statistically analyzing the sample proportions of different retrieval scenarios, with a value range of 0 to 2. q is the quality problem type identifier or unique identifier of the intellectual property data.

[0012] Furthermore, in S4, the hierarchical retrieval and matching process divides the data into a high-quality data set, a data set to be corrected, and a low-quality data set based on the quality tags corresponding to the intellectual property data. The intellectual property data in the low-quality data set is assigned a corresponding data correction priority based on the order of causal contribution values ​​from largest to smallest. The step-by-step correction process performs correction operations on a single type of quality problem in sequence according to the determined data correction priority. After the correction operation of a single type of quality problem is completed, the data quality score value of the corresponding intellectual property data is calculated. The deduplication process of the hierarchical retrieval results is performed to remove duplicate retrieval results based on the unique identifier of the intellectual property data. The sorting process of the hierarchical retrieval results is performed to sort the intellectual property data in descending order of matching score values.

[0013] Furthermore, the mathematical expression for the closed-loop optimized retrieval matching algorithm used in S4, hierarchical retrieval matching, is: Where S is the intellectual property data matching score, and V is the semantic relevance between the retrieval semantics and the data text, obtained by calculating the cosine distance between the retrieval semantic vector and the data text vector. ω represents the causal contribution value of the quality problem, ω represents the feedback iterative optimization weight value, which is obtained through iterative statistical calculation of user operation behavior data and ranges from 0.1 to 1. Q(θ) represents the confidence score of intellectual property data quality, which is obtained through data quality confidence algorithm.

[0014] On the other hand, a semantic understanding-based intelligent retrieval system for intellectual property big data includes: a cloud-native distributed database and a system communicatively connected to the cloud-native distributed database. The data collection and storage module is used to collect raw intellectual property data, perform operations such as format standardization, field regularization, deduplication and completion, extract feature fields, and store the processed data into a cloud-native distributed database. The quality model building module is used to build a quality labeling dataset and a semantic verification model. The semantic verification model verifies the standardized data, generates quality labels and a list of quality problem dimensions, and binds them with unique identifiers to be stored back to the cloud-native distributed database. The request semantic parsing module is used to receive user text search requests, perform word segmentation, stop word filtering, technical entity recognition, terminology normalization, extract technical elements and search constraints, and generate standardized search semantic vectors. The hierarchical retrieval and matching module is used to filter data based on semantic vectors and limiting conditions, retrieve data hierarchically according to quality labels, determine the priority of low-quality data correction through the causal contribution algorithm and correct it step by step, calculate the matching score through the closed-loop optimization retrieval and matching algorithm, and generate the initial retrieval results after deduplication and sorting of the retrieval results. The feedback closed-loop optimization module is used to collect user operation behavior data, incrementally update the parameters of each algorithm, re-verify the associated data through the updated semantic verification model, and write back the updated quality labels and quality problem dimension list to the cloud-native distributed data.

[0015] Compared with existing technologies, this semantic understanding-based intelligent retrieval method and system for intellectual property big data has the following advantages: I. This invention achieves the aggregation and storage of various types of original intellectual property data by building a cloud-native distributed database. Relying on a distributed sharding architecture, multi-replica storage, and independent partition design, it realizes elastic expansion, dynamic scheduling, and persistent storage of data. At the same time, it completes standardized processing of data format unification, field regularization, and deduplication and completion, laying a solid data foundation for retrieval. Furthermore, based on the standardized data, it constructs multi-type quality-labeled datasets and builds a multi-level semantic verification model. Through the model, it completes full-dimensional feature extraction and quality calculation of data, generates quality labels and problem lists, and achieves accurate judgment of intellectual property data quality. It avoids retrieval interference caused by quality problems from the data source, significantly improving the standardization and effectiveness of the retrieval data base, and adapting to the storage and quality verification needs of intellectual property big data.

[0016] Second, this invention performs domain-specific semantic analysis on user search requests, completing professional word segmentation, technical entity recognition, and terminology normalization. It accurately extracts technical elements and search constraints and generates standardized semantic vectors, making the analysis of search requests more aligned with the professional characteristics of the intellectual property field. Simultaneously, it performs tiered search matching based on quality tags, uses a causal contribution algorithm to prioritize and correct low-quality data step-by-step, combines a closed-loop optimization search matching algorithm with multi-dimensional indicators to calculate matching scores, and then uses user operation behavior data to incrementally update relevant algorithm parameters. It re-verifies associated data and writes back quality information, forming a closed-loop system of retrieval, correction, and optimization. This not only improves the accuracy and efficiency of search matching but also allows the search algorithm to continuously iterate and optimize, adapting to the actual search needs of users and significantly reducing the deviation of search results.

[0017] Other advantages, objectives and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination or study, or may be learned from the practice of the invention. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0019] Figure 1 This is a flowchart of a big data intelligent retrieval method for intellectual property based on semantic understanding. Figure 2 This is a framework diagram of an intellectual property big data intelligent retrieval system based on semantic understanding. Figure 3 This diagram illustrates the data transmission between the steps of a semantic understanding-based intelligent big data retrieval method for intellectual property. Detailed Implementation

[0020] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0021] The present invention provides an intelligent intellectual property big data retrieval system based on semantic understanding, comprising a data collection and storage module, a quality model construction module, a request semantic parsing module, a hierarchical retrieval and matching module, and a feedback closed-loop optimization module, such as... Figure 2As shown, the data collection and storage module and the quality model construction module have a one-way data connection, with the former outputting standardized intellectual property data. The latter interacts with the former, binding the generated quality tags, quality issue dimension lists, and unique data identifiers back to the cloud-native distributed database of the former. The request semantic parsing module and the hierarchical retrieval and matching module have a one-way data connection, with the former outputting standardized retrieval semantic vectors. The hierarchical retrieval and matching module and the feedback loop optimization module have a two-way data connection, with the former pushing initial retrieval results and related data information to the latter while receiving incrementally updated algorithm parameters. The feedback loop optimization module also establishes a real-time data interaction relationship with the quality model construction module, synchronizing the updated algorithm parameters to this module to complete model iteration optimization. All modules operate collaboratively according to predetermined logic, achieving standardized collection of intellectual property big data, full-dimensional quality detection, precise semantic parsing, hierarchical retrieval and matching, and dynamic iterative optimization of the algorithm model, ultimately achieving intelligent and high-precision retrieval of intellectual property big data.

[0022] Data collection and storage module: The data collection and storage module is the data source support unit of the entire system. It is responsible for the multi-channel collection, standardized processing, and distributed secure storage of original intellectual property data. It realizes unified management and efficient retrieval of various types of intellectual property data, and provides a standardized data source with unified format, complete information, and unique identification for the entire subsequent retrieval process. The specific implementation method is as follows: Intellectual Property Original Data Collection: The collected original intellectual property data includes four main categories: patent data, trademark data, copyright data, and integrated circuit layout design data. Collection channels cover official public channels such as the State Intellectual Property Office website, China Trademark Office website, copyright registration and protection platform, and integrated circuit layout design registration platform. It also accommodates existing data from professional intellectual property databases such as CNKI, Wanfang, and VIP, as well as private data imported from enterprise internal intellectual property management systems. The collection methods support both real-time incremental collection and historical batch collection. Real-time incremental collection targets newly published intellectual property data from various channels, with a collection frequency of hourly to ensure data timeliness. Historical batch collection targets existing publicly available data from various channels, collecting data in batches and by field, with each batch not exceeding 50,000 records to avoid excessive data processing pressure and ensure the comprehensiveness and completeness of the collected data.

[0023] Raw data standardization processing: For raw data collected from multiple channels that has different formats, different fields, and incomplete information, a four-step standardization process is performed sequentially: format unification, field normalization, deduplication, and information completion. Each step has its own specific processing rules and verification standards to ensure the standardization and accuracy of the data. Specific implementation details are as follows: Unified Formatting: Dedicated format conversion operations are performed on text data (TXT, DOC), tabular data (EXCEL, CSV), and layout data (PDF, CAJ) from different public sources. Text data retains its original content and is uniformly encoded in UTF-8. Tabular data is converted to a standardized CSV format while fully preserving the relationships between fields. Layout data undergoes text-image separation using OCR optical character recognition technology, extracting the plain text content and then uniformly converting it to TXT format. After conversion, an original format identifier field is added to each data entry to record the original data format and conversion operation information, enabling format traceability.

[0024] Field standardization: All types of intellectual property data undergo a unified process in terms of field names, field types, and field lengths. Specifically, patent data is standardized with 20 core bibliographic fields, including application number, applicant, application date, publication date, IPC classification number, invention name, technical abstract, and claims. Trademark data is standardized with 16 core fields, including application number, trademark name, applicant, application date, trademark category, and trademark design description. Copyright and integrated circuit layout design data are standardized with unified field systems according to their respective official standards. Field types are precisely defined as numeric, character, and date types. For example, application date and publication date are standardized as YYYY-MM-DD date fields, and IPC classification number and trademark category are standardized as character fields. Field lengths are set according to the actual maximum effective content length of each field to avoid truncation. Excessive length is truncated and marked accordingly.

[0025] Deduplication: A globally unique identifier is generated for each piece of intellectual property data. Patent data uses the official application number as the unique identifier, while trademark, copyright, and integrated circuit layout design data use their respective official registration / application numbers as unique identifiers. Based on this unique identifier, the Spark distributed deduplication algorithm is used to perform real-time deduplication on the data during the standardization process, identifying and removing duplicate data stored in the database. After the deduplication operation is completed, a deduplication log is generated, recording the unique identifier, source channel, removal time, and operator of the duplicate data, so as to achieve traceability of the deduplication operation.

[0026] Information Completion: Based on the constructed knowledge graph of the intellectual property field, automated completion operations are performed on missing fields in the bibliographic entries. For example, if the patent data is missing an IPC classification number, the corresponding accurate IPC classification number is matched through semantic analysis of the invention name and technical abstract. If the trademark data is missing an applicant's address, the valid address is completed by associating the applicant's name with the company's business registration information. For missing fields that cannot be automatically completed by the knowledge graph, information missing markers are marked and the names of the missing fields are recorded to form a list of missing information, which will be updated after manual supplementation. After the completion operation, an information completeness index is added to each data entry to quantify the degree of information completeness of the data.

[0027] After completing the above four standardization steps, the data is classified according to the types of patents, trademarks, copyrights, and integrated circuit layout designs. At the same time, five major information sets are extracted for each data entry: unique identifier, full text, IPC classification number, terminology annotation, and bibliographic field, to prepare the data for subsequent distributed storage.

[0028] Cloud-native distributed database storage: The categorized, standardized intellectual property data and the extracted five key information sets are stored in a cloud-native distributed database. This database features a dedicated architecture designed for the storage and retrieval needs of intellectual property big data. It employs a distributed sharded storage architecture, is deployed and managed using Kubernetes containerization technology, and possesses multi-dimensional storage and management characteristics. Specific implementation details are as follows: Elastic expansion and dynamic scheduling: The database adopts a horizontal sharding expansion method with a preset data storage capacity threshold of 85%. When the threshold is reached, new storage nodes are automatically added and the data shards are migrated seamlessly without manual intervention. The preset node load rate threshold is 90%. When the load rate of a node exceeds the threshold, some data tasks of that node are scheduled to low-load nodes to ensure the overall operating efficiency and stability of the database.

[0029] Distributed retrieval index construction: For the stored intellectual property data, a distributed retrieval master index is built based on the unique data identifier, while auxiliary indexes are built for frequently searched fields such as full-text text, IPC classification number, terminology annotation, and applicant; the index adopts an incremental update mechanism, and the index is synchronized and updated within 10 seconds after new data is stored, ensuring the consistency between the index and the data during retrieval; the index is stored separately on a dedicated index node, physically separated from the data storage node, which greatly improves the response speed of data retrieval.

[0030] Multi-replica persistent storage: A 3-replica storage mechanism is adopted, with each piece of data stored once on the primary storage node and once on each of two different secondary storage nodes. The replicas on the secondary nodes are synchronized within 1 second after the primary node updates the data. When the primary storage node fails, the secondary storage nodes can immediately switch to become the primary node within 5 seconds, ensuring data storage security and disaster recovery capabilities, and effectively avoiding data loss.

[0031] Independent storage partition settings: Independent storage partitions are set up in the database for four types of data: patents, trademarks, copyrights, and integrated circuit layout designs. Each partition adopts a dedicated storage strategy and retrieval rules. At the same time, a dedicated cross-partition retrieval channel is built to support joint retrieval and precise filtering of multiple types of intellectual property data.

[0032] Multi-mode data writing: Supports both real-time writing and batch writing. Real-time writing is for incrementally collected standardized data with a response time in milliseconds, meeting the needs of real-time data updates. Batch writing is for historically collected standardized data, with a maximum of 5,000 data entries per batch. A data verification mechanism is included during the writing process to verify the unique identifier and field integrity of each data entry, avoiding data errors during batch writing. Both writing modes generate detailed write logs, recording the amount of data written, the writing time, verification results, and abnormal data information, achieving full traceability of the data writing process.

[0033] Quality model building module: The quality model building module, based on the standardized data output by the data collection and storage module, completes the construction of the quality labeling dataset, the construction and training of the semantic verification model, and realizes full-dimensional quality detection and labeling management of intellectual property data. This provides accurate data quality basis for subsequent hierarchical retrieval and matching. The specific implementation method is as follows: Setting up a quality-labeled dataset: The quality-annotated dataset is the foundation for training, validating, and testing the semantic verification model. It consists of four types of samples: anomaly-annotated samples with mismatched classification numbers, incorrectly annotated samples with mislabeled terms, samples with missing field information, and samples with semantic conflicts in the text. The dataset was constructed following the principles of authenticity, representativeness, and completeness. Specific implementation details are as follows: Sample selection: All samples were taken from standardized intellectual property data. The four types of samples were randomly selected in a ratio of 1:1:1:1, with no less than 5,000 samples in each type and a total sample size of no less than 20,000. The samples cover four major categories of intellectual property data: patents, trademarks, copyrights, and integrated circuit layout designs. The sample proportion of each type of data is consistent with the data distribution in the actual database. At the same time, the samples cover intellectual property data from different technical fields and different application times to ensure the representativeness of the samples.

[0034] Sample labeling: The labeling work is completed by 5 or more professionals with more than 5 years of experience in the field of intellectual property, using the LabelStudio professional labeling platform. The labeling content includes unique data identifiers, full-text labeling information, IPC classification number labeling information, terminology labeling information, bibliographic field labeling information, and quality issue type labeling information. Among them, the quality issue type labeling must accurately correspond to the four sample types. Samples with multiple quality issues are labeled with multiple labels. Samples without quality issues are labeled as having no quality issues. During the labeling process, the labeler and labeling time are added to each labeling content.

[0035] Labeling Quality Control and Review: A three-tiered quality control mechanism is established, consisting of two labelers and one reviewer. The same sample is independently labeled by two labelers. Labels with consistent results are directly included in the dataset candidate pool, while those with inconsistent results are reviewed and judged by the reviewer. The reviewer conducts a 100% full review of all samples in the candidate pool, focusing on the accuracy of labeling quality issue types and the completeness of labeling information. A labeling accuracy rate of 95% or higher is required for formal inclusion in the dataset. After review, a labeling quality control report is generated, recording indicators such as labeling consistency rate, review correction rate, and the number of abnormal samples. Samples that fail the review are returned to the labelers for relabeling.

[0036] Dataset Management and Updates: The completed quality-labeled dataset is stored in a dedicated sample partition of the database. Version identifiers are added to the dataset to support version updates and backtracking. Subsequently, as new data is collected and standardized, the dataset is expanded once a quarter, adding no less than 1,000 samples of each type to ensure the timeliness and coverage of the dataset.

[0037] Construction and training of semantic validation models: The semantic verification model is the core carrier for realizing intellectual property data quality detection. The model selects a multi-encoder pre-trained language model based on the Transformer architecture as the initial base model. After sample training, parameter iterative adjustment, and double verification, the final model is determined. At the same time, a data quality confidence algorithm is embedded in the internal operation layer of the model to realize the synchronous output of data quality score and quality issues. The specific implementation details are as follows: Initial base model selection: The BERT-base multi-encoder pre-trained language model based on the Transformer architecture was selected as the initial base model. This model contains 12 encoder layers, 12 attention mechanisms, and a hidden layer dimension of 768, which is suitable for the text semantic analysis needs in the field of intellectual property. At the same time, the input layer of the model was lightweighted to support the simultaneous input of multiple types of features such as unique identifiers, IPC classification numbers, and term annotations, thereby improving the model's feature reception capability.

[0038] Dataset partitioning: The completed quality labeling dataset is randomly divided into training subset, validation subset and test subset according to the preset ratio of 7:2:1. Stratified sampling is used in the partitioning process to ensure that the distribution ratio of the four types of quality problem samples in the three subsets is consistent with the original dataset. The training subset is used for iterative adjustment of model parameters, the validation subset is used for numerical feature matching verification of the model, and the test subset is used for data classification verification of the model.

[0039] Model parameter iterative adjustment: The sample data from the training subset is input into the initial base model in batches, with a batch size of 32. The PyTorch deep learning framework is used for model training. The cross-entropy loss function is selected, the AdamW optimizer is selected, and the learning rate is set to 0.0001. The hidden layer parameters of the encoder, the weight parameters of the attention mechanism, and the connection parameters of the feedforward network are adjusted iteratively for no less than 100 rounds. After every 10 rounds of iteration, the model is initially validated using the validation subset. If the accuracy does not improve after 20 consecutive rounds of validation, the iteration is stopped to avoid model overfitting.

[0040] The model undergoes dual validation: First, a validation subset is used to perform data feature matching validation on the initial basic model after parameter adjustments. The model's extraction and matching accuracy for intellectual property data quality issues is calculated, and the matching accuracy must reach above 90% before proceeding to the next validation step. Then, a test subset is used to perform data classification validation on the model after feature matching validation. The model's classification accuracy for the four types of quality issues is calculated, and the classification accuracy must reach above 88%. After both validations pass, the model is determined as the final semantic validation model. If it fails, the model's learning rate and number of iterations are readjusted, and training and validation are performed again.

[0041] Algorithm embedding: A data quality confidence algorithm is embedded in the algorithm integration operation layer of the semantic verification model. The calculation parameters of the algorithm are associated and bound with the training parameters of the model, so that the model can simultaneously calculate the data quality confidence score while identifying data quality problems. The various parameters of the algorithm can be dynamically adjusted with the iterative optimization of the model to ensure the consistency between the algorithm calculation results and the model detection results.

[0042] Hierarchical operation of the semantic validation model: The semantic verification model consists of a five-layer architecture: a data input layer, a feature refinement and extraction layer, a multi-layer feature encoding layer, an algorithm integration and operation layer, and a classification result output layer. Each layer operates in a unidirectional manner according to a predetermined logic, enabling full-process, multi-dimensional quality inspection of standardized intellectual property data. The specific operational details of each layer are as follows: Data Input Layer: As the entry point of the model, it receives standardized intellectual property data output from the data collection and storage module, which has been formatted, fields are regularized, deduplicated, and information is completed. The data input format is structured tensor. At the same time, it performs legality verification on the input data, eliminating invalid data with format errors, missing unique identifiers, or severely incomplete fields to ensure the validity of the input data. Invalid data that fails verification generates an exception log, recording the exception data identifier and the reason for the exception, and returns it to the data collection and storage module for processing.

[0043] Feature Refinement Extraction Layer: For different dimensions of the input data, specific feature extraction methods are used to extract features separately, achieving a refined decomposition of data features. Unique identifier character features are extracted using character embedding, converting characters into fixed-dimensional vectors. Full-text semantic features are extracted using BERT embedding, capturing deep semantic information and contextual relationships within the text. IPC classification number encoding features are extracted using one-hot encoding, quantifying the encoding characteristics of the classification number. Terminology annotation text features are extracted using TF-IDF, reflecting the importance of terms in the text. Bibliographic field attribute features are extracted differently based on field type: numerical fields are extracted using normalization, character fields using the bag-of-words model, and date fields are converted to timestamps and then extracted using numerical methods.

[0044] Multi-layer feature encoding layer: Performs hierarchical feature fusion and standardized vector transformation on the five types of features output by the feature refinement extraction layer; first, each type of feature is encoded separately to generate a single feature vector; then, the five single feature vectors are concatenated in a fixed order to generate a fused feature vector; finally, the fused feature vector is deeply encoded by the Transformer encoder to complete the dimensional unification and semantic fusion of the features, generating a standardized feature vector with a fixed dimension of 768, providing an accurate feature foundation for subsequent algorithm operations.

[0045] Algorithm Integration and Operation Layer: Loads and runs the embedded data quality confidence algorithm, calculates the intellectual property data quality confidence score based on the standardized feature vectors output by the multi-layer feature coding layer, and combines the feature recognition results of the model to make a preliminary judgment on the type of data quality problems, providing data support for subsequent output results.

[0046] Classification Result Output Layer: Based on the quality confidence score and preliminary quality problem judgment results output by the algorithm integration and operation layer, the layer outputs the corresponding intellectual property data quality labels and quality problem dimension list. The quality labels are divided into three categories according to the quality confidence score: high quality, needing correction, and low quality. The quality problem dimension list accurately lists the type of quality problem, the field where the problem is located, and the severity of the problem. At the same time, the quality labels, quality problem dimension list, and data unique identifiers are bound and written back to the corresponding data entries in the cloud-native distributed database to complete the quality label management of the data.

[0047] Data quality confidence algorithm calculation: The data quality confidence algorithm is used to quantify the quality level of intellectual property data. The mathematical expression of the algorithm is: in, The confidence score for intellectual property data quality ranges from 0 to 1, with a higher score indicating better data quality. The numerical values ​​of multidimensional quality features of intellectual property data are obtained by parsing the standardized feature vectors output by multiple feature coding layers, covering multidimensional feature values ​​such as classification number matching, term annotation, field completeness, and text semantics. The weighted coefficient values ​​for features in the intellectual property field are obtained by statistically calculating the annotation frequency of different quality features in the intellectual property field. The values ​​range from 0 to 1. For example, the annotation frequency of the IPC classification number matching feature is the highest. Take 0.8, for the feature of missing field information. β is set to 0.6; β is the semantic anomaly penalty factor, which is obtained by statistical analysis of the distribution characteristics of semantic deviation samples, and its value ranges from 0.5 to 2. In this embodiment, β is set to 1.2. The semantic deviation value for a single piece of intellectual property data is calculated from the difference between the textual semantic features of the data and the domain standard semantic features. The global semantic mean value of the full set of intellectual property data is calculated from the textual semantic features of the full set of standardized data in the database.

[0048] During the algorithm's operation, the calculation results for each parameter are recorded, generating an algorithm operation log to ensure traceability of score calculation; simultaneously, based on different data types, [the following is applied]... Make differentiated adjustments to adapt to the quality inspection needs of different types of intellectual property data.

[0049] Request the semantic parsing module: The request semantic parsing module is the interaction unit between the user and the system. It is responsible for receiving various forms of text retrieval requests from users, extracting key retrieval information through a series of professional text processing operations, and converting it into standardized retrieval semantic vectors. This eliminates semantic ambiguity and inconsistent terminology in user retrieval requests, providing accurate and standardized semantic retrieval basis for subsequent hierarchical retrieval and matching. The specific implementation method is as follows: Search request reception: Supports three forms of receiving user text search requests: web page input, mobile APP input, and API interface call. The request text format supports various forms such as plain text, short sentences, long sentences, and technical description paragraphs. Each received search request is given a unique request identifier and a request timestamp, which are stored in the request log database to record information such as user ID, request content, and receiving time, so as to realize the traceability of search requests.

[0050] Word segmentation: A specialized dictionary for the intellectual property field is used to perform accurate word segmentation on the user's search request text. This dictionary contains more than 50,000 professional terms, technical terms, and legal terms in the field of intellectual property, and is updated quarterly to add new terms in the field. The word segmentation algorithm uses a customized version of Jieba word segmentation, which is optimized for the field-specific dictionary to avoid the incorrect segmentation of professional terms by general word segmentation algorithms and ensure the accuracy of word segmentation results. For example, integrated circuit layout design is segmented into a complete term rather than a single character.

[0051] Stop word filtering: Based on a custom stop word list for the intellectual property field, meaningless function words, auxiliary words, conjunctions, adverbs, and other stop words in the search request text are removed. This stop word list contains more than 800 words that have no search significance in the field, and users can also add their own stop words. The filtering operation only removes meaningless words, while retaining all professional terms, technical keywords, and search limiting words, simplifying the text content and improving the efficiency of subsequent processing.

[0052] Technical Entity Recognition: A Named Entity Recognition (NER) model based on the BERT+BiLSTM+CRF architecture is adopted to accurately identify three types of technical entity information from the segmented text: technical name, structural name, and method name. At the same time, structured information such as applicant, application time, and IPC classification number are also identified. During the recognition process, a knowledge graph in the field of intellectual property is used for auxiliary judgment to accurately match ambiguous technical entity information. The recognition accuracy rate needs to reach more than 92% to lock in the core technology and information dimensions of the user's search.

[0053] Terminology normalization: A terminology mapping table for the intellectual property field is constructed, which includes more than 3,000 sets of professional terms with different expressions. For example, AI, artificial intelligence, and artificial intelligence control are unified into artificial intelligence, and CNC machining and numerical control machining are unified into numerical control machining. Through this mapping table, the same professional term with different expressions is converted into a standard and unified expression in the field of intellectual property, eliminating the search bias caused by differences in terminology and ensuring the consistency of search semantics.

[0054] Technical elements and search constraints extraction: A combination of rule extraction and deep learning extraction is used to extract technical elements and search constraints from the processed text. Rule extraction targets structured search constraints such as IPC classification number range, application time period, and applicant range, using pre-defined regular expressions for accurate extraction. Deep learning extraction targets unstructured technical elements such as technical topics, technical methods, and application scenarios, using semantic analysis models for extraction. After extraction, the technical elements and search constraints are structured and organized to generate a set of key search information.

[0055] Standardized retrieval semantic vector generation: The extracted set of key retrieval information is input into the pre-trained BERT-base model, which is then converted into a dense vector with a fixed dimension of 768 through the model's encoding layer. This is the standardized retrieval semantic vector. This vector fully preserves the semantic information and retrieval constraints of the user's retrieval request. After the vector is generated, it is normalized to ensure that the numerical range of the vector is consistent with the feature vector of the intellectual property data, thus providing a foundation for subsequent semantic similarity calculation and retrieval matching.

[0056] Hierarchical search and matching module: The hierarchical retrieval and matching module is the execution unit for intelligent retrieval of intellectual property big data. Based on the standardized retrieval semantic vector output by the request semantic parsing module, it combines the quality tags of the data to perform hierarchical filtering, matching, correction, and integration of data in the database. Through the fusion of multiple algorithms to calculate the matching score, it finally generates accurate and ordered initial retrieval results. The specific implementation method is as follows: Preliminary data screening: Based on the search constraints and semantic features contained in the standardized search semantic vector, a two-layer preliminary screening is performed on the standardized intellectual property data in the cloud-native distributed database to quickly eliminate data that obviously does not meet the search requirements and narrow the search matching scope. Specifically: The first layer of conditional filtering: First, it accurately matches the search criteria, such as the range of IPC classification numbers, application time period, and applicant range, and directly eliminates data that does not meet the criteria; the conditional filtering is completed based on the index nodes of the database, and the search response time is in the millisecond level, which greatly improves the filtering efficiency.

[0057] The second layer of semantic screening: For the remaining data after conditional screening, calculate the cosine similarity between its text feature vector and the standardized retrieval semantic vector. Set the cosine similarity screening threshold to 0.5, remove data with similarity below the threshold, and retain data with similarity above the threshold to complete the preliminary data screening and obtain the retrieval candidate dataset.

[0058] Classified by quality label: Based on the quality labels generated from the intellectual property data in the quality model building module, and combined with the specific range of quality confidence scores, the initially screened candidate datasets are precisely divided into three categories. The specific classification criteria are as follows: High-quality datasets: Quality confidence scores Intellectual property data with a value of ≥0.8 indicates that the data set has no obvious quality issues, is complete in information, semantically accurate, and uses standard terminology, and can be directly searched and matched.

[0059] Data set to be corrected: Quality confidence score ≤ 0.5 Intellectual property data with a value <0.8 indicates minor quality issues, such as non-standard terminology, incomplete information in a single field, and small semantic deviations. These issues require simple corrections before retrieval and matching.

[0060] Low-quality datasets: Quality confidence scores Intellectual property data with a value <0.5 indicates serious quality issues, such as incorrect classification number matching, missing information in multiple fields, semantic conflicts in text, and numerous errors in terminology labeling. These issues need to be corrected step by step according to priority before retrieval and matching can proceed.

[0061] Causal contribution calculation and step-by-step correction for low-quality data: For intellectual property data in low-quality datasets, the causal contribution of quality issues is calculated using a causal contribution algorithm. Correction priorities are determined based on the contribution values, and precise corrections are performed on each type of quality issue sequentially according to priority, ensuring both effectiveness and efficiency. Specific implementation details are as follows: Causal contribution calculation: For each data point in the low-quality dataset, run the causal contribution algorithm to calculate its causal contribution to the quality problem. The mathematical expression of this algorithm is: in, The numerical value represents the causal contribution of intellectual property data quality issues. The larger the value, the greater the impact of the quality issue on the search results. The deviation value of the search results is obtained by calculating the difference between the search result vector of this data and the domain standard result vector; The change in confidence level caused by the quality problem is calculated by the difference in confidence level scores before and after the quality problem is corrected; γ is the confusion factor in the field of intellectual property, which is obtained by statistical calculation of the distribution of easily confused features in the field of intellectual property, and the value ranges from 1 to 3. In this embodiment, γ is 2; λ is the weighted index of the search scenario, which is obtained by statistical calculation of the sample proportion of different search scenarios, and the value ranges from 0 to 2. For example, λ is 1.5 for the patent technology search scenario and λ is 0.8 for the trademark name search scenario; q is the quality problem type identifier or data unique identifier of intellectual property data.

[0062] Correction priority determination: Based on the calculated causal contribution values, the data in the low-quality dataset are sorted in descending order to determine the correction priority of the data; the larger the causal contribution value, the higher the correction priority, and the data is corrected first to ensure that quality issues that have a greater impact on the search results are resolved first.

[0063] Step-by-step correction execution: According to the determined correction priority, precise correction operations are performed on each type of quality issue in turn; if the classification number is mismatched, the correct IPC classification number is matched with the knowledge graph through semantic analysis; if the terminology label is incorrect, it is replaced with domain standard terms; if field information is missing, it is supplemented based on the knowledge graph and official data; if there is a semantic conflict in the text, the text content is corrected by combining the context and technical field specifications; after the correction operation of each type of quality issue is completed, the quality confidence score Q(θ) of the data is recalculated to verify the correction effect. If the score increases by ≥0.2 after correction, it is determined to be a valid correction, and the correction of the next type of quality issue continues until the quality confidence score of the data is ≥0.5 or there are no quality issues to be corrected.

[0064] Storage of Correction Results: Add correction identifiers to the corrected low-quality data, record the type of quality problem, correction content, correction personnel, correction time, and correction score, write back to the corresponding entry in the database, and include the corrected dataset in the dataset to be matched.

[0065] Closed-loop optimized retrieval matching algorithm operation: For high-quality datasets, corrected datasets to be corrected, and low-quality datasets, a closed-loop optimization retrieval matching algorithm is used to calculate the matching score for each data point. This algorithm integrates four factors: semantic relevance, causal contribution of quality issues, feedback iterative optimization weights, and quality confidence scores, to achieve accurate quantification of the matching score. The mathematical expression of the algorithm is as follows: Wherein, S is the intellectual property data matching score, ranging from 0 to 1. The higher the score, the higher the matching degree between the data and the user's search request; V is the semantic relevance between the search semantic and the data text, obtained by calculating the cosine distance between the search semantic vector and the data text feature vector, ranging from 0 to 1. ω represents the causal contribution of quality issues; this value is 0 for data without quality issues. ω is the feedback iteration optimization weight value, which is obtained through iterative statistical calculation of user operation behavior data. The value ranges from 0.1 to 1. In this embodiment, the initial value is 0.5, which is dynamically adjusted according to user feedback. The confidence score for intellectual property data quality is obtained by the data quality confidence algorithm.

[0066] The algorithm uses the Spark distributed computing framework to perform batch matching and score calculation on three types of datasets, with a batch size of 1000 records to improve computational efficiency. After the calculation is completed, a matching score field is added to each data record to record the calculation results and the values ​​of various parameters, so as to achieve traceability of score calculation.

[0067] Search results integration: The matching results of the three datasets are uniformly integrated, and the final initial search results are generated through merging, deduplication, and sorting operations, and then pushed to the user. The specific implementation details are as follows: Result merging: The matching results of the high-quality data set, the corrected data set to be corrected, and the low-quality data set are merged into a complete set of search results.

[0068] Deduplication: Based on the unique identifier of intellectual property data, duplicate data in the search result set is removed to ensure that each search result corresponds to a unique piece of intellectual property data.

[0069] Sorting operation: The deduplicated search results are arranged in descending order of matching score; if the matching scores are the same, they are sorted in descending order of quality confidence score; if the quality confidence scores are also the same, they are sorted in ascending order of data application / publication time, ensuring the orderliness and accuracy of the search results.

[0070] Output results: The sorted search results are truncated to a preset number. Users can customize the truncated number in the system. The system defaults to truncating the first 50 results and pushing them to the user's terminal as the initial search results. At the same time, each result is accompanied by information such as matching score, quality confidence score, quality tag, and core matching points, which makes it convenient for users to view, filter and refer to.

[0071] Feedback closed-loop optimization module: The feedback loop optimization module is a key unit for realizing the dynamic iteration of the system's algorithm model. By collecting user behavior data throughout the entire process of initial search results, it incrementally updates the core algorithm parameters within the system, iteratively optimizes the semantic verification model, and simultaneously re-verifies and updates the quality labels of associated data, forming a closed-loop system of retrieval-feedback-optimization-re-retrieval. This ensures that the system's retrieval capabilities continuously meet user needs. The specific implementation method is as follows: User behavior data collection: By combining front-end event tracking with back-end logging, we collect real-time data on various user actions related to the initial search results. The data collection covers the entire user process, ensuring the comprehensiveness and accuracy of the feedback data. Specific implementation details are as follows: Data collected includes five categories: search result click data, favorite data, rejection data, error annotation data, and correct result supplementary data. It also records related fields such as user ID, unique identifier of search request, unique identifier of result, operation type, operation time, and annotation content. For example, if a user annotates a result as a mismatch, the annotation content, the reason for the annotation, and the corresponding result identifier are recorded in detail.

[0072] Data collection method: Front-end tracking is implemented in user interfaces such as web pages and apps to accurately capture user actions such as clicks, favorites, rejections, annotations, and additions, and transmits them to the back-end server in real time; the back-end establishes a dedicated user behavior log database to store the collected behavior data in real time. The data storage format is standardized JSON format, which supports multi-dimensional filtering, statistics, and analysis.

[0073] Data cleaning: Clean the collected raw behavioral data to remove invalid operation data, such as accidental clicks, repeated operations within a short period of time, and meaningless annotations; after cleaning, add valid labels to valid data, record the reasons for invalid data and store them separately to ensure the quality of data sources for subsequent parameter updates and model optimization.

[0074] Incremental update of core algorithm parameters: Based on the cleaned and valid user behavior data, the parameters of the data quality confidence algorithm, causal contribution algorithm, and closed-loop optimization retrieval matching algorithm are incrementally updated to ensure that the algorithm's calculation results continuously match users' search habits and actual needs. Specific implementation details are as follows: Update method: The algorithm parameters are finely adjusted using mini-batch gradient descent, with the learning rate set to 0.001. Each update is based on the most recent 1000 valid user operation behavior data to ensure the timeliness and accuracy of parameter updates.

[0075] Algorithm parameter adjustment: Weighting coefficients for intellectual property features in the data quality confidence algorithm. The semantic anomaly penalty factor β is adjusted numerically. For example, if users pay high attention to IPC classification number matching and frequently label results with incorrect classification numbers, the penalty factor for the IPC classification number matching feature is appropriately increased. Values: Adjust the values ​​of the intellectual property confusion factor γ and the retrieval scenario weighting index λ in the causal contribution algorithm. For example, if users frequently label confusing results in a certain retrieval scenario, the value of λ in that scenario should be appropriately increased. Adjust the values ​​of the feedback iteration optimization weight ω in the closed-loop optimization retrieval matching algorithm. For example, if users have a high click-through rate or collection rate for a certain type of result, the value of ω for the corresponding match should be appropriately increased.

[0076] Update and Verification: After the parameters are updated, 50 test retrieval requests from different fields and of different types are selected for verification. The retrieval matching accuracy and user satisfaction of the updated algorithm are calculated. If the accuracy is improved by ≥3% and the user satisfaction is improved by ≥5%, the parameter update is deemed effective and can be officially put into use in the system. If the improvement standard is not met, the learning rate and parameter values ​​are readjusted, and the update and verification are performed again.

[0077] Semantic validation model iterative optimization: The incrementally updated algorithm parameters are embedded into the semantic verification model to complete the iterative optimization of the model, ensuring that the model's quality detection capability keeps pace with the algorithm's computational capability. Specific implementation details are as follows: Parameter association and binding: Deeply associate and bind the incrementally updated algorithm parameters with the internal parameters of the semantic verification model to achieve coordinated adjustment of algorithm parameters and model parameters, ensuring the consistency between model detection results and algorithm calculation results.

[0078] Model fine-tuning training: Newly added quality-labeled samples and user-labeled abnormal data samples were selected to fine-tune the semantic verification model for 20 rounds; the learning rate for fine-tuning was set to 0.00001, and a mini-batch training method was adopted with a batch size of 16 to avoid model overfitting.

[0079] Model performance verification: The fine-tuned semantic verification model is verified using a test subset. The accuracy of the model's quality problem classification and the precision of the quality confidence score calculation are calculated. If both the accuracy and precision are improved compared to before the fine-tuning, and the classification accuracy is ≥90%, the model is deemed to have been effectively optimized and can be put into use as a new version model.

[0080] Related data revalidation and information write-back: Using the iteratively optimized semantic verification model, a full-field quality check is performed on the related data in the database. A list of quality labels and quality issue dimensions is regenerated and written back to the database. Simultaneously, the distributed retrieval index is updated. Specific implementation details are as follows: Related data definition: Based on the semantics of the user's search request, calculate the semantic similarity between all intellectual property data in the database and the search request. Intellectual property data with a semantic similarity ≥ 0.7 are defined as related data. The scope of related data changes dynamically depending on the user's search request.

[0081] Full-field quality inspection: Using the updated semantic verification model, full-field quality inspection is performed on the associated data. The inspection scope covers 28 specific indicators in 5 major categories, including unique identifiers, full-text text, IPC classification numbers, terminology annotations, and bibliographical entries. During the inspection process, the updated data quality confidence algorithm is run simultaneously to recalculate the quality confidence score of each associated data.

[0082] Quality information regeneration: Based on the full-field quality inspection results and the recalculated quality confidence score, quality labels and a list of quality problem dimensions are regenerated for the associated data; for associated data with quality problems, the problem type, the field where the problem is located, and correction suggestions are accurately listed to provide a basis for subsequent retrieval matching and data correction.

[0083] Quality information write-back and index update: An incremental write-back method is adopted, only writing back the changed quality tags, quality issue dimension lists, and quality confidence scores to the corresponding data entries in the cloud-native distributed database. Unchanged information remains unchanged. The write-back response time is ≤1 second, avoiding the impact of large-scale data write-back on database operating efficiency. At the same time, based on the written-back quality information, the distributed retrieval index of the database is incrementally updated, focusing on updating the index content related to quality tags and quality confidence scores to ensure the consistency between the index and the data during retrieval. After the index update is completed, a validity check is performed to ensure that the retrieval function operates normally.

[0084] Optimize logging: Record all operations, including algorithm parameter updates, model iteration optimization, correlation data re-verification, quality information write-back, and index updates, in the system optimization log. The log will include update time, update content, amount of data involved, optimization effect, and verification results, enabling full traceability of system optimization.

[0085] Implementation process of semantic understanding-based intelligent big data retrieval method for intellectual property: The semantic understanding-based intelligent retrieval method for intellectual property big data of the present invention is implemented based on the aforementioned intelligent retrieval system. The method is achieved through the orderly execution of five steps, S1 to S5, as follows: Figure 1 As shown, this document describes the entire process of intellectual property big data collection, storage, and retrieval result output, while simultaneously achieving dynamic optimization of the algorithm model through a feedback loop. The specific implementation flow of each step is as follows: S1, Data Collection and Storage: The data collection and storage module gathers original intellectual property data (patents, trademarks, copyrights, and integrated circuit layout designs) from multiple channels, including official public channels such as the State Intellectual Property Office, professional intellectual property literature databases, and internal enterprise systems. The collected raw data undergoes standardized operations such as unified formatting, field normalization, deduplication, and information completion, generating a globally unique identifier for each data entry. Five key information sets are extracted: unique identifier, full-text, IPC classification number, terminology annotation, and bibliographical entries. These are then categorized by data type and stored in a cloud-native distributed database, completing the unified collection and standardized storage of data. Simultaneously, a distributed primary index is built based on the unique identifier, and auxiliary indexes are constructed for frequently searched fields to enable rapid data retrieval. Figure 3 As shown.

[0086] S2, Quality Model Construction: Through the quality model construction module, standardized intellectual property data is selected to build a quality-labeled dataset consisting of four types of samples: classification number mismatch, terminology labeling errors, missing field information, and text semantic conflicts. After professional labeling, three-level quality control, and review, the dataset is divided into training, validation, and test subsets in a 7:2:1 ratio. A pre-trained language model based on the Transformer architecture is selected as the initial base model. The training subset is input into the model and its parameters are iterated and adjusted for no less than 100 rounds. After feature matching verification of the validation subset and classification verification of the test subset, the final semantic verification model is determined, and a data quality confidence algorithm is embedded in the algorithm integration operation layer of the model. The semantic verification model is used to perform full-dimensional quality detection on the standardized data, extract multi-dimensional features of the data, calculate quality confidence scores, output corresponding data quality labels and a list of quality problem dimensions, bind unique data identifiers, and write back to the cloud-native distributed database.

[0087] S3, Request Semantic Parsing: The request semantic parsing module receives text search requests from users via web pages, mobile apps, APIs, etc., and adds a unique identifier and timestamp to the request; it segments the request text using a dictionary specific to the intellectual property field and removes meaningless words using a domain stop word list; it identifies technical entity information using a NER model with a BERT+BiLSTM+CRF architecture and normalizes terms using a domain terminology mapping table; it extracts technical elements and search constraints from the request using a combination of rule extraction and deep learning extraction; and it inputs the extracted key search information set into a pre-trained BERT-base model, converting it into a standardized search semantic vector with a fixed dimension of 768, providing accurate semantic basis for subsequent search matching.

[0088] S4, Hierarchical Retrieval and Matching: The hierarchical retrieval and matching module, based on the retrieval constraints and semantic features of the standardized retrieval semantic vector, performs conditional filtering and initial semantic screening on the data in the database to obtain candidate datasets. Based on quality confidence scores and quality labels, the candidate datasets are divided into three categories: high-quality, needing correction, and low-quality datasets. For the low-quality datasets, a causal contribution algorithm is run to calculate the causal contribution of quality issues and determine the correction priority based on the numerical values ​​from largest to smallest. Step-by-step corrections are then performed on each type of quality issue according to priority, and the correction effect is verified and written back to the database. For the three types of datasets, a closed-loop optimization retrieval and matching algorithm is run to calculate the matching score for each data point. The matching results of the three types of datasets are merged, duplicate data is removed based on unique data identifiers, and the results are sorted from largest to smallest by matching score. A user-preset number of results are selected as the initial retrieval results and pushed to the user.

[0089] S5, Feedback Loop Optimization: Through the feedback loop optimization module, front-end tracking and back-end logging are used to collect real-time user action data on initial search results, including clicks, favorites, rejections, error annotations, and correct result additions. After cleaning, effective feedback data is obtained. Based on this effective feedback data, the parameters of the three algorithms—data quality confidence, causal contribution, and loop-optimized search matching—are incrementally updated using mini-batch gradient descent. The update effect is verified through test requests. The updated algorithm parameters are embedded into a semantic verification model, and the model is fine-tuned using new samples to complete iterative optimization. Related data in the database with a semantic similarity ≥ 0.7 to the search request are calculated. The updated semantic verification model is used to perform full-field quality checks on these data, recalculating the quality confidence score and generating quality labels and a list of quality issue dimensions. The updated quality information is incrementally written back to the database, and the distributed search index is updated simultaneously, completing the feedback loop optimization process and providing more accurate algorithm and data support for subsequent intellectual property big data retrieval.

[0090] This invention presents a semantic understanding-based intelligent retrieval method and system for intellectual property big data. Through a cloud-native distributed database, it achieves multi-channel aggregation, standardized processing, and efficient storage of intellectual property big data, solving the problems of inconsistent data storage formats, low management efficiency, and slow retrieval response in traditional intellectual property data storage. Relying on a semantic verification model based on the Transformer architecture and a data quality confidence algorithm, it achieves full-dimensional detection and tag-based management of data quality, addressing the issue of inconsistent data quality affecting the accuracy of retrieval results in traditional retrieval. Combined with semantic parsing technology specific to the intellectual property field, it achieves accurate interpretation of user retrieval requests, eliminating retrieval biases caused by semantic ambiguity and inconsistent terminology. Through the fusion of multiple algorithms such as hierarchical retrieval matching, causal contribution, and closed-loop optimization retrieval matching, it achieves accurate hierarchical matching and score quantification of retrieval results, improving the accuracy and orderliness of retrieval results. Finally, relying on a feedback closed-loop optimization system, it achieves dynamic iteration of algorithm parameters and the semantic verification model, ensuring that the system's retrieval capabilities continuously meet user needs. This effectively solves the problems of low retrieval efficiency, poor matching accuracy, and low level of intelligence in traditional intellectual property retrieval, realizing intelligent, refined, and efficient retrieval of intellectual property big data.

[0091] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for intelligent retrieval of intellectual property big data based on semantic understanding, characterized in that, The specific steps of this method are as follows: S1, Data Collection and Storage: Collect original intellectual property data, unify the format, standardize fields, remove duplicates and complete information, extract unique identifiers, full text, IPC classification numbers, terminology annotations and bibliographic fields, and store them in a cloud-native distributed database. S2, Quality Model Construction: Based on standardized data, a quality labeling dataset is constructed, a semantic verification model is built, and the data is pre-verified through the semantic verification model to generate a list of quality labels and quality problem dimensions, which are then bound with unique identifiers and stored in the cloud database; S3, Request Semantic Parsing: Receives user text retrieval requests, performs word segmentation, stop word filtering, technical entity recognition and terminology normalization, extracts technical elements and retrieval constraints, and generates standardized retrieval semantic vectors; S4, Hierarchical retrieval and matching: Data is filtered based on semantic vectors and limiting conditions, hierarchical retrieval is performed according to quality labels, causal contribution algorithm is run on low-quality data to obtain causal contribution, correction priority is generated and corrected step by step based on causal contribution, matching score is calculated through closed-loop optimization retrieval and matching algorithm, hierarchical retrieval results are merged, deduplicated and sorted to generate initial retrieval results. S5, Feedback Closed-Loop Optimization: Collect user operation behavior data, incrementally update the parameters of the data quality confidence algorithm, causal contribution algorithm and closed-loop optimization retrieval matching algorithm, re-verify the associated data through the updated semantic verification model, and write the updated quality tags and quality problem dimension list back to the cloud database.

2. The intelligent retrieval method for intellectual property big data based on semantic understanding according to claim 1, characterized in that, In S1, during data collection and storage, the collected original intellectual property data includes patent data, trademark data, copyright data, and integrated circuit layout design data. Unified format processing performs format conversion operations on text data, tabular data, and layout data from different public sources. Field standardization processing performs operations to unify field names, field types, and field lengths for all data. Deduplication processing identifies and removes duplicate data based on unique data identifiers. Information completion processing fills in the corresponding basic information for missing fields in the bibliographic entries. The processed data is then classified according to unique data identifiers, and the classified data is stored in a cloud-native distributed database.

3. The intelligent retrieval method for intellectual property big data based on semantic understanding according to claim 1, characterized in that, In step S2, the construction of the quality model includes building a semantic verification model. This involves selecting a pre-trained language model based on a Transformer architecture with multiple encoders as the initial base model, dividing the quality-labeled dataset into a training subset, a validation subset, and a test subset according to a preset ratio, inputting sample data from the training subset into the initial base model, performing multiple rounds of iterative adjustments on the hidden layer parameters of the encoder, the weight parameters of the attention mechanism, and the connection parameters of the feedforward network in the initial base model, using the validation subset to perform data feature matching verification on the initial base model after parameter adjustment, using the test subset to perform data classification verification on the initial base model after data feature matching verification, and determining the initial base model that passes both data feature matching verification and data classification verification as the semantic verification model. A data quality confidence algorithm is then embedded in the internal computational layer of the semantic verification model.

4. The intelligent retrieval method for intellectual property big data based on semantic understanding according to claim 1, characterized in that, In S2, during the quality model construction, the semantic verification model consists of a data input layer, a feature refinement and extraction layer, a multi-layer feature encoding layer, an algorithm integration and operation layer, and a classification result output layer. The data input layer receives standardized intellectual property data that has undergone format unification, field regularization, deduplication, and information completion. The feature refinement and extraction layer extracts unique identifier character features, full-text semantic features, IPC classification number encoding features, terminology annotation text features, and bibliographic field attribute features from the intellectual property data. The multi-layer feature encoding layer... The layer performs hierarchical feature fusion and standardized vector transformation on the extracted features. The algorithm integration and operation layer loads and runs the data quality confidence algorithm to complete the data quality score calculation. The classification result output layer outputs the corresponding intellectual property data quality labels and a list of quality problem dimensions.

5. The intelligent retrieval method for intellectual property big data based on semantic understanding according to claim 1, characterized in that, The mathematical expression for the data quality confidence algorithm used in the construction of the quality model in S2 is: in The confidence score for intellectual property data quality. For multidimensional quality characteristic values ​​of intellectual property data, The weighted coefficients for intellectual property features are calculated statistically based on the annotation frequency of different quality features in the intellectual property field, with values ​​ranging from 0 to 1. β is the semantic anomaly penalty factor, calculated statistically based on the distribution characteristics of semantic deviation samples, with values ​​ranging from 0.5 to 2. This represents the semantic deviation value for a single piece of intellectual property data. This represents the global semantic mean value of the entire intellectual property data.

6. The intelligent retrieval method for intellectual property big data based on semantic understanding according to claim 1, characterized in that, In S3, the semantic parsing request uses a dedicated dictionary for intellectual property to perform text splitting. Technical entity identification identifies technical names, structural names, and method names. Terminology normalization converts different professional terms into unified expressions. Technical elements include technical themes, technical means, and application scenarios. Search limiting conditions include the range of IPC classification numbers, application time periods, and applicant ranges. The standardized search semantic vector is a dense vector with fixed dimensions.

7. The intelligent retrieval method for intellectual property big data based on semantic understanding according to claim 1, characterized in that, The mathematical expression for the causal contribution algorithm used in the hierarchical retrieval and matching process in S4 is: in The numerical value of causal contribution to intellectual property data quality issues. The deviation value of the search results is obtained by calculating the difference between the search result vector and the standard result vector. The change in quality confidence caused by a quality issue is calculated by subtracting the quality confidence score before and after the quality issue was corrected. γ is the confusion factor in the intellectual property field, calculated by statistically analyzing the distribution of easily confused features in the intellectual property field, with a value ranging from 1 to 3. The weighted index value for the retrieval scenario is calculated by statistically analyzing the sample proportions of different retrieval scenarios, with a value range of 0 to 2. q is the quality problem type identifier or unique identifier of the intellectual property data.

8. The intelligent retrieval method for intellectual property big data based on semantic understanding according to claim 1, characterized in that, In step S4, the hierarchical retrieval and matching process divides the data into a high-quality data set, a data set to be corrected, and a low-quality data set based on the quality tags corresponding to the intellectual property data. The intellectual property data in the low-quality data set is assigned a data correction priority based on the order of causal contribution values ​​from largest to smallest. The step-by-step correction process performs correction operations on each type of quality problem in turn according to the determined data correction priority. After the correction operation of each type of quality problem is completed, the data quality score value of the corresponding intellectual property data is calculated. The deduplication process of the hierarchical retrieval results is performed to remove duplicate retrieval results based on the unique identifier of the intellectual property data. The sorting process of the hierarchical retrieval results is performed to sort the intellectual property data in descending order of the matching score values.

9. The intelligent retrieval method for intellectual property big data based on semantic understanding according to claim 1, characterized in that, The mathematical expression for the closed-loop optimized retrieval matching algorithm used in S4, hierarchical retrieval matching, is as follows: Where S is the intellectual property data matching score, and V is the semantic relevance between the retrieval semantics and the data text, obtained by calculating the cosine distance between the retrieval semantic vector and the data text vector. ω represents the causal contribution value of the quality problem, ω represents the feedback iterative optimization weight value, which is obtained through iterative statistical calculation of user operation behavior data and ranges from 0.1 to 1. Q(θ) represents the confidence score of intellectual property data quality, which is obtained through data quality confidence algorithm.

10. A semantic understanding-based intelligent retrieval system for intellectual property big data, the system being applicable to the semantic understanding-based intelligent retrieval method for intellectual property big data as described in any one of claims 1-9, characterized in that, The system includes: a cloud-native distributed database and a communication connection to the cloud-native distributed database. The data collection and storage module is used to collect raw intellectual property data, perform operations such as format standardization, field regularization, deduplication and completion, extract feature fields, and store the processed data into a cloud-native distributed database. The quality model building module is used to build a quality labeling dataset and a semantic verification model. The semantic verification model verifies the standardized data, generates quality labels and a list of quality problem dimensions, and binds them with unique identifiers to be stored back to the cloud-native distributed database. The request semantic parsing module is used to receive user text search requests, perform word segmentation, stop word filtering, technical entity recognition, terminology normalization, extract technical elements and search constraints, and generate standardized search semantic vectors. The hierarchical retrieval and matching module is used to filter data based on semantic vectors and limiting conditions, retrieve data hierarchically according to quality labels, determine the priority of low-quality data correction through the causal contribution algorithm and correct it step by step, calculate the matching score through the closed-loop optimization retrieval and matching algorithm, and generate the initial retrieval results after deduplication and sorting of the retrieval results. The feedback closed-loop optimization module is used to collect user operation behavior data, incrementally update the parameters of each algorithm, re-verify the associated data through the updated semantic verification model, and write back the updated quality labels and quality problem dimension list to the cloud-native distributed data.

Images