Structured data matching method and device based on natural language processing

Abstract

The embodiment of the application provides a kind of structured data matching method and device based on natural language processing, by the semantic analysis module of natural language query sentence is executed word segmentation and entity recognition and syntax dependency analysis obtains semantic feature vector and with preset intent template library is executed vector similarity calculation obtains intent matching result after generation structured query condition buffer to session context, associated historical query record executes condition fusion screening candidate enterprise set after input potential assessment model executes multidimensional weighted calculation and obtains potential financing score and descending order arrangement obtains ranking enterprise list, according to user self-defined field configuration executes list rendering and pushes enterprise recommendation list page, effectively solve the deficiency of traditional technology in semantic analysis intent identification, multi-dimension fusion search potential assessment and list rendering score attribution etc., for the technical support that structured enterprise data intelligent matching based on natural language processing is provided.

Description

Technical Field

[0001] This application relates to the field of data processing, specifically to a structured data matching method and apparatus based on natural language processing. Background Technology

[0002] Existing structured data matching technologies have significant shortcomings in natural language semantic parsing and intent recognition, multi-dimensional conditional fusion retrieval and potential assessment, personalized list rendering and scoring attribution interpretation, etc. They are unable to support the fully automated processing from natural language query to structured enterprise recommendation, resulting in low query understanding accuracy and a lack of personalized ranking criteria for search results, which seriously restricts the practicality and user experience of intelligent enterprise data matching systems. Summary of the Invention

[0003] To address the problems in existing technologies, this application provides a structured data matching method and apparatus based on natural language processing, which can effectively solve the shortcomings of traditional technologies in semantic parsing and intent recognition, multi-dimensional fusion retrieval potential assessment, and list rendering scoring attribution, providing technical support for intelligent matching of structured enterprise data based on natural language processing.

[0004] To solve at least one of the above problems, this application provides the following technical solution: Firstly, this application provides a structured data matching method based on natural language processing, including: The system receives natural language query statements input by users through the search interface and inputs the natural language query statements into the semantic parsing module to perform word segmentation, entity recognition, and syntactic dependency analysis to obtain semantic feature vectors. The semantic feature vectors are then compared with a preset intent template library to calculate vector similarity and obtain intent matching results. Based on the intent matching results, structured query conditions are generated and cached in the session context. The structured query conditions are input as retrieval parameters into a multidimensional filtering engine and associated with historical query records in the session context to perform condition fusion to obtain fused retrieval conditions. Based on the fused retrieval conditions, enterprise records that meet the conditions are filtered from the enterprise database to obtain a candidate enterprise set. The candidate enterprise set is input into a potential assessment model to perform multidimensional weighted calculation to obtain the potential financing score corresponding to each enterprise. The candidate enterprise set is sorted in descending order according to the potential financing score to obtain a sorted enterprise list. The sorted company list is rendered according to the user-defined field configuration to generate a company recommendation list page, which is then pushed to the user interface. In response to user clicks, the detailed data of the corresponding company is read from the sorted company list. The potential assessment model is then invoked to perform attribution analysis on the detailed data to generate a score explanation text, which is then embedded in the company details page.

[0005] Furthermore, it also includes: listening to user input events through the search interface and capturing the input content when a query submission action is detected to obtain a natural language query statement; inputting the natural language query statement into a preprocessing unit to perform text cleaning and stop word filtering to obtain normalized text; and inputting the normalized text into a word segmenter to perform segmentation processing to obtain a word sequence. The word sequence is input into the semantic parsing module and named entity recognition is performed based on a preset entity dictionary to obtain an entity annotation sequence. Syntactic dependency analysis is performed on the entity annotation sequence to obtain a dependency tree. The dependency tree and the entity annotation sequence are input into a vector encoder to perform joint encoding to obtain a semantic feature vector.

[0006] Furthermore, it also includes: reading the template vector set corresponding to each intent template from the preset intent template library, performing cosine similarity calculation on the semantic feature vector and the template vector set one by one to obtain a similarity score sequence, and filtering intent identifiers that meet the conditions from the similarity score sequence according to preset threshold conditions to obtain intent matching results; Based on the intent matching result, the corresponding query parameter mapping rule is read, and the entity information in the semantic feature vector is converted into a structured query condition according to the query parameter mapping rule. The structured query condition is then associated with the current user identifier and written into the session context.

[0007] Furthermore, it also includes: inputting the structured query conditions into a multidimensional filtering engine and reading historical query records from the session context, and merging and resolving conflicts between the structured query conditions and the historical query records according to a preset fusion strategy to obtain fused retrieval conditions; Based on the fusion search conditions, a database search statement is constructed and sent to the enterprise database. The enterprise records returned by the enterprise database are received, and secondary filtering is performed according to the filtering dimensions in the fusion search conditions to obtain a set of candidate enterprises.

[0008] Furthermore, it also includes: reading the records of each enterprise from the candidate enterprise set one by one and extracting the financing window period, the number of institutional endorsements, risk indicators, and industry popularity index to obtain enterprise feature vectors; inputting the enterprise feature vectors into the potential assessment model and performing weighted summation calculation according to the preset dimension weight configuration to obtain the potential financing score corresponding to each enterprise; The potential financing score is associated with the enterprise records in the candidate enterprise set to obtain a set of enterprises with scores. The set of enterprises with scores is then sorted in descending order according to the potential financing score to obtain a sorted enterprise list.

[0009] Furthermore, it also includes: reading user-defined field configuration and extracting field data that matches the user-defined field configuration from the sorted enterprise list to obtain a dataset to be rendered, and performing column mapping on the dataset to be rendered according to the field order in the user-defined field configuration to obtain structured list data; The structured list data is input into the list rendering engine and a preset list page template is loaded. Data population and style rendering are performed to generate a company recommendation list page, which is then pushed to the user interface.

[0010] Furthermore, it also includes: listening to click events on the user interface and parsing the target enterprise identifier from the click events, locating the corresponding enterprise record from the sorted enterprise list based on the target enterprise identifier, and reading the detailed data; The detailed data is input into the potential assessment model, and the contribution of each scoring dimension is calculated to obtain the dimension attribution results. Based on the dimension attribution results, a scoring explanation text is generated according to a preset text template. The enterprise details page template is loaded, and the detailed data and the scoring explanation text are filled into the corresponding areas to generate the enterprise details page.

[0011] Secondly, this application provides a structured data matching device based on natural language processing, comprising: The semantic analysis module is used to receive natural language query statements entered by users through the search interface and input the natural language query statements into the semantic parsing module to perform word segmentation, entity recognition, and syntactic dependency analysis to obtain semantic feature vectors. The semantic feature vectors are then compared with a preset intent template library to calculate the vector similarity and obtain intent matching results. Based on the intent matching results, structured query conditions are generated and cached in the session context. The data matching module is used to input the structured query conditions as retrieval parameters into the multidimensional filtering engine and associate them with the historical query records in the session context to perform condition fusion to obtain fused retrieval conditions. Based on the fused retrieval conditions, it filters enterprise records that meet the conditions from the enterprise database to obtain a candidate enterprise set. It inputs the candidate enterprise set into the potential assessment model to perform multidimensional weighted calculation to obtain the potential financing score corresponding to each enterprise. Based on the potential financing score, it sorts the candidate enterprise set in descending order to obtain a sorted enterprise list. The results feedback module is used to render the sorted enterprise list according to the user-defined field configuration to generate an enterprise recommendation list page and push it to the user interaction interface. In response to user click operations, it reads the detailed data of the corresponding enterprise from the sorted enterprise list, calls the potential assessment model to perform attribution analysis on the detailed data, generates a score explanation text, and embeds it into the enterprise details page.

[0012] Thirdly, this application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the structured data matching method based on natural language processing.

[0013] Fourthly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the structured data matching method based on natural language processing.

[0014] Fifthly, this application provides a computer program product, including a computer program / instructions that, when executed by a processor, implement the steps of the structured data matching method based on natural language processing.

[0015] As can be seen from the above technical solution, this application provides a structured data matching method and apparatus based on natural language processing. It generates semantic feature vectors by performing word segmentation, entity recognition, and syntactic dependency analysis on natural language query statements through a semantic parsing module. These vectors are then compared with a preset intent template library to calculate the intent matching result, generating structured query conditions that are cached in the session context. Historical query records are associated with these conditions to perform condition fusion and filter a set of candidate companies. This data is then input into a potential assessment model to perform multi-dimensional weighted calculations to obtain a potential financing score, which is then sorted in descending order to obtain a ranked list of companies. The list is then rendered and pushed to a company recommendation list page according to user-defined field configurations. This effectively solves the shortcomings of traditional technologies in semantic parsing intent recognition, multi-dimensional fusion retrieval potential assessment, and list rendering score attribution, providing technical support for intelligent matching of structured enterprise data based on natural language processing. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0017] Figure 1 This is a flowchart illustrating the structured data matching method based on natural language processing in the embodiments of this application; Figure 2 This is a flowchart illustrating a specific embodiment of the structured data matching method based on natural language processing in this application. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some embodiments of this application, but not all embodiments.

[0019] In view of the problems existing in the prior art, this application provides a structured data matching method and apparatus based on natural language processing. The method generates semantic feature vectors by performing word segmentation, entity recognition, and syntactic dependency analysis on natural language query statements through a semantic parsing module. These vectors are then compared with a preset intent template library to calculate the intent matching result, generating structured query conditions which are cached in the session context. Historical query records are associated with these conditions to perform condition fusion and filter a set of candidate companies. These are then input into a potential assessment model to perform multi-dimensional weighted calculations to obtain a potential financing score, which is then sorted in descending order to obtain a ranked list of companies. The list is then rendered and pushed to a recommended company list page according to user-defined field configurations. This method effectively solves the shortcomings of traditional technologies in semantic parsing intent recognition, multi-dimensional fusion retrieval potential assessment, and list rendering score attribution, providing technical support for intelligent matching of structured enterprise data based on natural language processing.

[0020] To effectively address the shortcomings of traditional technologies in semantic parsing and intent recognition, multi-dimensional fusion retrieval potential assessment, and list rendering scoring attribution, and to provide technical support for intelligent matching of structured enterprise data based on natural language processing, this application provides an embodiment of a structured data matching method based on natural language processing. See [link to embodiment]. Figure 1 and Figure 2 The structured data matching method based on natural language processing specifically includes the following: Step S101: Receive the natural language query statement entered by the user through the search interface and input the natural language query statement into the semantic parsing module to perform word segmentation, entity recognition, and syntactic dependency analysis to obtain a semantic feature vector. Perform vector similarity calculation between the semantic feature vector and the preset intent template library to obtain the intent matching result. Generate structured query conditions based on the intent matching result and cache the structured query conditions in the session context. In this embodiment, a natural language query statement is obtained from the user interface as the starting input for the semantic parsing process. The user input event is monitored through the search interface, and the input content is captured when a query submission action is detected, resulting in a natural language query statement. This natural language query statement, along with a user session identifier and submission timestamp, is written to the input buffer. If the input content is empty or exceeds a preset length limit, an abnormal state and session identifier are recorded in the input abnormality table, and a formatting prompt is returned to the user interface. The natural language query statement is then read by the preprocessing unit.

[0021] First, the natural language query statement is input into the preprocessing unit for text normalization. The preprocessing unit performs text cleaning on the natural language query statement, removing redundant whitespace and special characters, and performs stop word filtering based on a preset stop word list to obtain normalized text. The normalized text is then input into the word segmenter for segmentation to obtain a word sequence. The word sequence includes word content, word position index, and initial part-of-speech tagging fields, which are written to the word buffer for the semantic parsing module to read.

[0022] Next, the word sequence is input into the semantic parsing module for entity recognition and syntactic analysis. The semantic parsing module loads a preset entity dictionary and performs named entity recognition on the word sequence. During the recognition process, entity types such as company name, industry category, geographical scope, and financing stage are labeled to obtain an entity-labeled sequence. Syntactic dependency analysis is performed on the entity-labeled sequence to construct a dependency tree, which describes the modification, dominance, and coordination relationships between words. When recognition fails, the failed word and its context are written into a recognition exception table, and a fuzzy matching strategy is used for fallback processing.

[0023] Then, the dependency tree and the entity annotation sequence are input to a vector encoder for joint encoding to obtain a semantic feature vector. The vector encoder performs fusion encoding on the structural features of the dependency tree and the semantic features of the entity annotation sequence based on a pre-trained language model, with the encoding dimension set according to the model configuration. The semantic feature vector is written to a feature buffer for the intent matching step to read.

[0024] Subsequently, a set of template vectors is read from a preset intent template library and similarity calculation is performed with the semantic feature vector. Cosine similarity calculation is then performed on each of the semantic feature vectors and the set of template vectors to obtain a similarity score sequence. Intent identifiers that meet the conditions are selected from the similarity score sequence based on a preset threshold to obtain the intent matching result. The threshold is determined based on the balance between precision and recall on the validation set. If a match fails, the semantic feature vector and its nearest neighbor template are recorded in the matching log, and a manual review process is triggered.

[0025] Therefore, structured query conditions are generated based on the intent matching result and written into the session context. The query parameter mapping rule corresponding to the intent matching result is read, and the entity information in the semantic feature vector is converted into structured query conditions according to the query parameter mapping rule. The structured query conditions include a query condition identifier field, an intent type field, an entity parameter field, and a generation timestamp field. The user session identifier and query sequence number are used as primary keys and written into the session context for step S102 to read at the multi-dimensional filtering engine input interface for constructing fusion search conditions.

[0026] Step S102: Input the structured query conditions as retrieval parameters into the multidimensional filtering engine and perform condition fusion with the historical query records in the session context to obtain fused retrieval conditions. Based on the fused retrieval conditions, filter enterprise records that meet the conditions from the enterprise database to obtain a candidate enterprise set. Input the candidate enterprise set into the potential assessment model to perform multidimensional weighted calculation to obtain the potential financing score corresponding to each enterprise. Sort the candidate enterprise set in descending order according to the potential financing score to obtain a sorted enterprise list. In this embodiment, structured query conditions are read from the session context output in step S101 as the starting input for the multidimensional filtering process. These structured query conditions are then input as retrieval parameters into the multidimensional filtering engine. The multidimensional filtering engine loads the current user session identifier and reads historical query records from the session context. If the read fails, the session identifier and read status are recorded in the context exception table, and execution continues using the structured query conditions as independent retrieval conditions.

[0027] The fused retrieval conditions are obtained by fusing the structured query conditions with the historical query record execution conditions. The fusion process merges query parameters of the same dimension according to a preset fusion strategy, and resolves conflicting parameters based on a timestamp priority principle. The fused retrieval conditions include a fusion condition identifier, filtering parameters for each dimension, a fusion strategy flag, and a generated timestamp field, which are written to a condition buffer for the database retrieval steps to read.

[0028] To this end, a database retrieval statement is constructed based on the fused retrieval conditions and a retrieval request is sent to the enterprise database. The multidimensional filtering engine converts the fused retrieval conditions into a structured query statement, sends the database retrieval statement to the enterprise database, and receives the returned enterprise records. A secondary filtering is performed on the returned records according to the filtering dimensions in the fused retrieval conditions to obtain a candidate enterprise set. When a retrieval timeout occurs, the timeout status and query conditions are recorded in the retrieval log, and the retrieved partial records are returned.

[0029] Therefore, each enterprise record is read from the candidate enterprise set one by one, and enterprise feature vectors are extracted. The extraction process reads the financing window period, number of institutional endorsements, risk indicators, and industry popularity index fields from each enterprise record to obtain the enterprise feature vector. When a field is missing, it is filled with a default value or a mean imputation strategy according to the field type, and a filling mark is registered in the feature status table. The enterprise feature vectors are written to the feature buffer for the potential assessment model to read.

[0030] Accordingly, the enterprise feature vector is input into a potential assessment model to perform multi-dimensional weighted calculations to obtain a potential financing score. The potential assessment model loads a preset dimension weight configuration and performs a weighted summation calculation on each dimension feature in the enterprise feature vector. The optimal configuration of the dimension weights is determined on a validation set using a grid search based on historical successful financing cases. When anomalies occur, the abnormal features and model state are recorded in an evaluation anomaly table, and a conservative weight backoff calculation is used. The potential financing score is associated with and bound to the enterprise identifier to obtain a set of enterprises with scores.

[0031] Finally, the set of rated companies is sorted in descending order based on their potential financing scores to obtain a sorted list of companies. The sorted list includes a company identifier field, a company basic information field, a potential financing score field, a score dimension detail field, and a sorting sequence number field. The user session identifier and sorting batch number are used as primary keys for this information, which is read by step S103 at the list rendering engine input interface for generating the company recommendation list page.

[0032] Step S103: Render the sorted company list according to the user-defined field configuration to generate a company recommendation list page and push it to the user interface. Respond to the user's click operation, read the detailed data of the corresponding company from the sorted company list, call the potential assessment model to perform attribution analysis on the detailed data, generate a score explanation text, and embed it into the company details page.

[0033] In this embodiment, the sorted enterprise list output from step S102 is read from the sorting buffer as the starting input for the list rendering process. User-defined field configurations are read, and field data matching the user-defined field configurations is extracted from the sorted enterprise list to obtain the dataset to be rendered. The user-defined field configurations define the field types, field order, and column width ratios displayed in the list. If configuration reading fails, the user identifier and configuration status are recorded in the configuration exception table, and the system default field configuration is loaded to continue execution.

[0034] Based on the foregoing, the dataset to be rendered is subjected to column mapping according to the field order in the user-defined field configuration to obtain structured list data. The column mapping process traverses each enterprise record in the dataset to be rendered, reorganizes the field arrangement according to the field order configuration, and adds display format markers. The structured list data is written to the rendering buffer and associated with the enterprise identifier index of the sorted enterprise list for the list rendering engine to read.

[0035] Further, the structured list data is input into the list rendering engine to generate the page. The list rendering engine loads a preset list page template and performs data population and style rendering on the structured list data to generate a company recommendation list page. During the rendering process, pagination is performed for excess records according to the pagination configuration. If rendering fails, the failure record and template status are recorded in the rendering exception table and a fallback rendering process is triggered. The company recommendation list page is then pushed to the user interface.

[0036] Simultaneously, click events on the user interface are monitored and the user's intent is analyzed. When a click on a list item is detected, the target company identifier is parsed from the click event. Based on the target company identifier, the corresponding company record is located from the sorted company list and its details are read. If locating fails, the target company identifier and a list snapshot are recorded in a location error table, and a data unavailable message is returned to the user interface. The details are then written to a details buffer.

[0037] Therefore, the detailed data is input into the potential assessment model to perform attribution analysis and generate dimensional attribution results. The potential assessment model performs contribution calculations on each scoring dimension in the detailed data, identifying the influence weights of each dimension—financing window period, number of institutional endorsements, risk indicators, and industry popularity index—on the potential financing score to obtain dimensional attribution results. In case of attribution calculation anomalies, the abnormal dimensions and model status are recorded in the attribution log and rolled back using equal weights.

[0038] Accordingly, based on the dimensional attribution results, rating explanation text is generated according to a preset text template and assembled into a company details page. The company details page template is loaded, and the details data is filled into the basic information area. The rating explanation text is then filled into the rating interpretation area to generate the company details page. The company details page includes a page identifier field, a company identifier field, details data fields, rating explanation text fields, and a generation timestamp field. The user session identifier and page access sequence number are used as primary keys for the user interface to read at the details display entry point for presenting company details and supporting user decision-making.

[0039] As described above, the structured data matching method based on natural language processing provided in this application can generate structured query conditions and cache them in the session context by performing word segmentation, entity recognition, and syntactic dependency analysis on natural language query statements through a semantic parsing module to obtain semantic feature vectors, performing vector similarity calculation with a preset intent template library to obtain intent matching results, as well as inputting them into a potential assessment model to perform multi-dimensional weighted calculation to obtain potential financing scores and sorting them in descending order to obtain a sorted list of companies. The list is then rendered and pushed to a company recommendation list page according to user-defined field configurations. This effectively solves the shortcomings of traditional technologies in semantic parsing intent recognition, multi-dimensional fusion retrieval potential assessment, and list rendering score attribution, and provides technical support for intelligent matching of structured enterprise data based on natural language processing.

[0040] In one embodiment of the structured data matching method based on natural language processing in this application, the method may further include the following: Step S201: Listen for user input events through the search interface and capture the input content when a query submission action is detected to obtain a natural language query statement. Input the natural language query statement into the preprocessing unit to perform text cleaning and stop word filtering to obtain normalized text. Input the normalized text into the word segmenter to perform segmentation processing to obtain a word sequence. Step S202: Input the word sequence into the semantic parsing module and perform named entity recognition based on the preset entity dictionary to obtain the entity annotation sequence. Perform syntactic dependency analysis on the entity annotation sequence to obtain the dependency tree. Input the dependency tree and the entity annotation sequence into the vector encoder to perform joint encoding to obtain the semantic feature vector.

[0041] This embodiment uses the search interface to listen for user input events as the trigger entry point for the query processing flow. The search interface employs an event-driven mechanism to continuously listen for keyboard input and submit button click events. When a query submission action is detected, the current input box content is captured to obtain the natural language query statement. The natural language query statement, along with the user session identifier, client timestamp, and input method source marker, is written to the input buffer. If the input content is empty or contains illegal characters, the exception type and the original input are recorded in the input exception table, and a format verification prompt is returned to the user interface.

[0042] Preferably, the natural language query statement is input into the preprocessing unit for text normalization. The preprocessing unit performs text cleaning on the natural language query statement, removing leading and trailing whitespace, consecutive spaces, HTML tags, and special control characters. Based on a preset stop word list, the cleaned text is filtered to obtain normalized text. The stop word list includes common function words, modal particles, and words without semantic connections. When filtering anomalies occur, the anomaly location and the version of the stop word list are recorded in the preprocessing log, and the original vocabulary is retained for continued processing. The normalized text is written to the text buffer.

[0043] Optionally, the normalized text is input into a word segmenter for segmentation to obtain a word sequence. The word segmenter loads a domain dictionary and performs joint segmentation of the normalized text using forward maximum matching and a statistical model to identify domain-specific terms such as company names, industry terms, and financing stages. The word sequence includes a word content field, a word position index field, a part-of-speech tagging field, and a segmentation confidence field, and is written to a word buffer for reading at the semantic parsing module input interface in step S202.

[0044] In one implementation, the word sequence is input into a semantic parsing module for named entity recognition. The semantic parsing module loads a preset entity dictionary and performs entity type matching on each word sequence. The preset entity dictionary defines entity types such as company name, industry category, geographical scope, financing round, and amount expression, along with their recognition rules. The recognition process combines dictionary matching with a sequence labeling model. Successfully matched words are labeled with entity types and entity boundaries to obtain an entity-labeled sequence. When recognition fails, the failed words and their context windows are recorded in an entity anomaly table and marked as requiring manual review.

[0045] In another embodiment, syntactic dependency analysis is performed on the entity-labeled sequence to obtain a dependency tree. The syntactic analyzer constructs dependency relationships such as dominance, modification, and coordination between lexical units in the entity-labeled sequence, organizing a tree structure with the core predicate as the root node. The dependency tree includes fields for node identifier, dependency type, parent node pointer, and child node list. When analysis fails, the failed sequence that conflicts with the grammar rules is recorded in the syntactic log, and a rollback is performed using the shallow analysis results.

[0046] Therefore, the dependency tree and the entity annotation sequence are input into a vector encoder for joint encoding to obtain a semantic feature vector. The vector encoder extracts the semantic representation of the entity annotation sequence based on a pre-trained language model and performs joint encoding by fusing the structural features of the dependency tree. The semantic feature vector includes a vector identifier field, a vector dimension field, an encoding timestamp field, and a source lexical index field, using the user session identifier and query sequence number as primary keys, which are read by step S301 at the intent template library matching interface for vector similarity calculation and intent matching.

[0047] In one embodiment of the structured data matching method based on natural language processing in this application, it may further include the following: Step S301: Read the template vector set corresponding to each intent template from the preset intent template library, perform cosine similarity calculation on the semantic feature vector and the template vector set one by one to obtain a similarity score sequence, and filter the intent identifiers that meet the conditions from the similarity score sequence according to the preset threshold conditions to obtain the intent matching result; Step S302: Based on the intent matching result, read the corresponding query parameter mapping rule, perform field transformation on the entity information in the semantic feature vector according to the query parameter mapping rule to obtain structured query conditions, and write the structured query conditions, associated with the current user identifier, into the session context.

[0048] This embodiment obtains a set of template vectors from the intent template library management module as the basic dependency for the intent matching process. It reads the set of template vectors corresponding to each intent template from a preset intent template library, which defines intent types such as enterprise screening, financing inquiry, industry analysis, and risk assessment, and their standard vector representations. If the template library fails to load, the loading status and template library version are recorded in the template exception table, and a cache reconstruction process is triggered. The set of template vectors is then written to the template buffer.

[0049] In the initial stage, semantic feature vectors are read from the feature buffer output in step S202 and vector normalization is performed. L2 norm normalization is applied to the semantic feature vectors to eliminate the impact of vector magnitude differences on similarity calculation. If normalization fails, the abnormal vector and its numerical state are recorded in the vector anomaly table, and the original vector is used to continue execution. The normalized semantic feature vectors are then written to the calculation buffer for the similarity calculation step to read.

[0050] Subsequently, cosine similarity calculations are performed on the semantic feature vector and the template vector set one by one to obtain a similarity score sequence. The calculation process traverses each template vector in the template vector set, performing an inner product operation on the semantic feature vector and the current template vector to obtain a cosine similarity score. The similarity score sequence includes an intent identifier field, a similarity score field, and a calculation timestamp field. When a calculation overflows, the overflow position and numerical range are recorded in the calculation log, and the corresponding score is set to the minimum value.

[0051] In the parallel path, intent identifiers that meet the conditions are filtered from the similarity score sequence based on preset threshold conditions. The preset threshold conditions are determined based on the balance point of precision and recall on the validation set. The filtering process extracts intent identifiers with similarity scores higher than the threshold to form intent matching results. When all scores are lower than the threshold, the semantic feature vector and the highest score template are recorded in the low confidence table and a manual review process is triggered. The intent matching results are then written to the matching buffer.

[0052] When necessary, the corresponding query parameter mapping rules are read based on the intent matching result and field transformation is performed. The query parameter mapping rules corresponding to the intent matching result are read from the mapping rule configuration table. These rules define the correspondence between entity types and database fields, as well as the transformation functions. The entity information in the semantic feature vector is transformed according to the query parameter mapping rules to obtain structured query conditions. If the transformation fails, the failed entity and mapping rule are recorded in the transformation exception table and filled with default parameter values.

[0053] Finally, the structured query conditions are associated with the current user identifier and written into the session context to complete the condition caching. The structured query conditions include a condition identifier field, an intent type field, a filter parameter field, an entity source field, and a generation timestamp field, and are written into the session context using the user session identifier and query sequence number as primary keys. The session context maintains the current user's query history and condition evolution trajectory, which is read by step S401 at the multi-dimensional filtering engine input interface for constructing historical query record association and fusion retrieval conditions.

[0054] In one embodiment of the structured data matching method based on natural language processing in this application, it may further include the following: Step S401: Input the structured query conditions into the multidimensional filtering engine and read the historical query records from the session context. Merge the structured query conditions and the historical query records according to the preset fusion strategy and resolve conflicts to obtain the fused retrieval conditions. Step S402: Construct a database retrieval statement based on the fusion retrieval conditions and send the database retrieval statement to the enterprise database. Receive the enterprise records returned by the enterprise database and perform secondary filtering according to the filtering dimensions in the fusion retrieval conditions to obtain a candidate enterprise set.

[0055] In this embodiment, structured query conditions are read from the session context output in step S302 as the starting input for the multidimensional filtering process. These structured query conditions are then input into the multidimensional filtering engine, which parses the intent type and filtering parameter fields within the structured query conditions and establishes a parameter index. If input format validation fails, the validation status and exception fields are recorded in the input exception table, and a parameter reconstruction request is returned upstream. The structured query conditions are also written to the condition buffer.

[0056] Accordingly, the system reads the current user's historical query records from the session context and performs record filtering. The reading process locates the historical query sequence within the session context based on the user's session identifier and filters recently valid historical query records according to a time window configuration. The historical query records include historical condition identifiers, historical filtering parameters, query timestamps, and result feedback flag fields. If the read fails, the session identifier and read status are recorded in the context exception table, and execution continues with an empty historical record.

[0057] Subsequently, the structured query conditions and the historical query records are merged and conflict resolved according to a preset fusion strategy. The fusion process iterates through the same-dimensional parameters in the structured query conditions and the historical query records, merging conditions for consistent parameters to expand the search scope, and retaining the latest condition value for conflicting parameters based on a timestamp priority principle. The preset fusion strategy defines the merging weights and conflict resolution rules for each dimension. In case of fusion anomalies, conflicting parameters and fusion rules are recorded in the fusion log, and the current conditions overwrite the historical conditions. After fusion is complete, the fused search conditions are obtained and written to the search buffer.

[0058] Simultaneously, a database retrieval statement is constructed based on the fused retrieval conditions, and a database query is executed. The multi-dimensional filtering engine converts the fused retrieval conditions into a structured query statement, assembling query predicates and sorting clauses according to the filtering dimensions in the fused retrieval conditions. The database retrieval statement is sent to the enterprise database, and a query timeout threshold is set. If the timeout occurs, the query conditions and timeout duration are recorded in the retrieval log, and a portion of the retrieved records are returned.

[0059] Subsequently, the system receives enterprise records returned from the enterprise database and performs secondary filtering to obtain a candidate enterprise set. The secondary filtering process verifies each returned record according to the refined screening dimensions in the fusion search conditions, filtering out enterprise records that do not meet the composite conditions. The filtering process records the filtering status and failed dimensions for each enterprise. When filtering anomalies occur, the anomaly record and filtering rules are recorded in the anomaly table, and the anomaly record is moved to the pending review queue.

[0060] Finally, the filtered enterprise records are aggregated to form a candidate enterprise set and then encapsulated. The candidate enterprise set includes an enterprise identifier field, an enterprise basic information field, a filter hit dimension field, a data source identifier field, and an entry timestamp field. The user session identifier and the retrieval batch number are used as primary keys for step S501 to read at the potential assessment model input interface for enterprise feature vector extraction and multi-dimensional weighted scoring calculation.

[0061] In one embodiment of the structured data matching method based on natural language processing in this application, it may further include the following: Step S501: Read the records of each enterprise from the candidate enterprise set one by one and extract the financing window period, number of institutional endorsements, risk indicators, and industry popularity index to obtain the enterprise feature vector. Input the enterprise feature vector into the potential assessment model and perform weighted summation calculation according to the preset dimension weight configuration to obtain the potential financing score of each enterprise. Step S502: Associate the potential financing score with the enterprise records in the candidate enterprise set to obtain a set of enterprises with scores, and sort the set of enterprises with scores in descending order according to the potential financing score to obtain a sorted enterprise list.

[0062] In this embodiment, the candidate enterprise set output from step S402 is read from the candidate buffer as the starting input for the potential assessment process. Each enterprise record is read sequentially from the candidate enterprise set. These enterprise records contain fields such as basic enterprise information, operating data, financing history, and risk labeling. If a record reading fails, the failed enterprise identifier and reading status are recorded in the read exception table, and the current record is skipped to continue traversal. Enterprise records to be processed are written to the traversal buffer.

[0063] First, the financing window period field is extracted from each company's records and its value is normalized. The financing window period represents the expected time interval between a company's next round of financing, and its extraction is calculated based on the company's most recent financing date and the industry average financing cycle. When a field is missing, it is filled using the industry median based on the industry category, and a filling marker is registered in the feature status table. The financing window period value is written into the first dimension of the feature buffer.

[0064] Secondly, the "Number of Institutional Endorsements" and "Risk Identification" fields are extracted from each company's records. The "Number of Institutional Endorsements" counts the number of influential institutions among the company's historical investors, while the "Risk Identification" indicates the company's current operational risk level and compliance status. The extraction process iterates through the company's financing history and risk assessment records, performing field aggregation. When a field is abnormal, the abnormal type and original value are recorded in the extraction log and replaced with a conservative default value.

[0065] Then, the industry popularity index is extracted from the records of each enterprise, and the enterprise feature vector is assembled. The industry popularity index is calculated based on the frequency of recent financing events, policy attention, and market growth rate of the industry to which the enterprise belongs. The financing window period, the number of institutional endorsements, the risk indicator, and the industry popularity index are assembled in dimensional order to obtain the enterprise feature vector, which is written into the feature buffer for the potential assessment model to read.

[0066] Subsequently, the enterprise feature vectors are input into the potential assessment model for weighted summation. The potential assessment model loads a preset dimension weight configuration, which determines the optimal weights for each dimension on a validation set using a grid search based on historical successful financing cases. The weighted summation of the values ​​for each dimension in the enterprise feature vectors according to their corresponding weights yields the potential financing score for each enterprise. If a calculation overflows, the overflowing enterprises and their numerical ranges are recorded in a calculation anomaly table and truncated using boundary value truncation.

[0067] Therefore, the potential financing score is associated with the enterprise records in the candidate enterprise set to obtain a set of enterprises with scores. The binding process appends the potential financing score to the corresponding enterprise record based on the enterprise identifier. If binding fails, the unmatched score and enterprise identifier are recorded in the binding exception table, and the exception record is moved to the pending verification queue. The set of enterprises with scores is sorted in descending order based on the potential financing score to obtain a sorted enterprise list. The sorted enterprise list includes an enterprise identifier field, an enterprise basic information field, a potential financing score field, a score dimension detail field, and a sorting sequence number field, using the user session identifier and sorting batch number as primary keys. This is read by step S601 at the list rendering engine input interface for generating the enterprise recommendation list page.

[0068] In one embodiment of the structured data matching method based on natural language processing in this application, it may further include the following: Step S601: Read the user-defined field configuration and extract the field data that matches the user-defined field configuration from the sorted enterprise list to obtain the dataset to be rendered. Perform column mapping on the dataset to be rendered according to the field order in the user-defined field configuration to obtain structured list data. Step S602: Input the structured list data into the list rendering engine and load the preset list page template to perform data filling and style rendering to generate the enterprise recommendation list page, and push the enterprise recommendation list page to the user interaction interface.

[0069] This embodiment obtains user-defined field configurations from the user configuration management module as a fundamental dependency for the list rendering process. The user-defined field configurations define the field types, field names, field order, column width ratios, and numerical formatting rules that the user expects to display on the list page. If configuration reading fails, the user identifier and configuration status are recorded in the configuration exception table, and the system default field configuration is loaded to continue execution. The user-defined field configurations are written to the configuration buffer.

[0070] Based on the user-defined field configuration, matching field data is extracted from the sorted enterprise list output in step S502. The extraction process iterates through each enterprise record in the sorted enterprise list, locates and reads the corresponding field value according to the field type definition in the user-defined field configuration. If a field does not exist, the missing field and enterprise identifier are recorded in the extraction log and placed with an empty value. After extraction is completed, the dataset to be rendered is obtained and written to the data buffer.

[0071] To this end, column mapping is performed on the dataset to be rendered according to the field order in the user-defined field configuration. The column mapping process rearranges the fields of each enterprise record in the dataset according to the field order configuration, calculates the display width of each column according to the column width ratio configuration, and performs format conversion on numeric fields according to the numeric formatting rules. In case of mapping errors, the abnormal fields and mapping rules are recorded in the mapping error table, and the original field order is preserved while execution continues. After mapping is completed, structured list data is obtained and written to the rendering buffer.

[0072] The structured list data is then input into the list rendering engine to assemble the page. The list rendering engine loads a preset list page template, which defines the layout structure and visual specifications for header styles, row styles, pagination controls, and sorting interaction components. Data population is performed on the structured list data, mapping each company's records to table rows. After population, the page is rendered with style settings, including font, color, and spacing configurations. If rendering fails, the failed components and template status are recorded in a rendering exception table, triggering a fallback rendering process.

[0073] Based on this, pagination processing and interactive component binding are performed on the rendered page. Excess records are paginated according to the pagination configuration, and a pagination navigation control is generated. A sorting click event is bound to the table header field to support interactive reordering. In case of pagination errors, the total number of records and pagination parameters are logged to the pagination log, and a single-page full-view rollback is implemented.

[0074] Finally, the enterprise recommendation list page is pushed to the user interface to complete the page presentation. The enterprise recommendation list page includes a page identifier field, a list data snapshot field, a pagination status field, a rendering timestamp field, and a user configuration version field, using the user session identifier and page number as primary keys. These are read by step S701 at the click event listener entry point in the user interface for triggering the enterprise details page and loading details data.

[0075] In one embodiment of the structured data matching method based on natural language processing in this application, it may further include the following: Step S701: Listen for click events on the user interface and parse the target enterprise identifier from the click events; locate the corresponding enterprise record from the sorted enterprise list based on the target enterprise identifier and read the details data. Step S702: Input the detailed data into the potential assessment model and perform contribution calculation on each scoring dimension to obtain the dimension attribution result. Based on the dimension attribution result, generate the scoring explanation text according to the preset text template, load the enterprise details page template, and fill the detailed data and the scoring explanation text into the corresponding area to generate the enterprise details page.

[0076] This embodiment establishes a click event listener through the user interface as the trigger entry point for the details display process. The listener monitors click events on the user interface, triggered by the user clicking the target company row or the details button on the company recommendation list page. The event listener captures the click event and extracts the interactive element identifier and click coordinates from the event payload. If an anomaly is detected, the anomaly type and interface state are recorded in the event anomaly table, and the operation is ignored.

[0077] Based on the foregoing, the target enterprise identifier is parsed from the click event and an identifier verification is performed. The parsing process reads the data attribute fields from the click event payload to extract the target enterprise identifier, and performs a validity check according to the identifier format rules. If parsing fails, the failure event and payload snapshot are recorded in the parsing exception table, and an invalid operation message is returned to the user interface. The target enterprise identifier is written to the identifier buffer.

[0078] Further, based on the target enterprise identifier, the corresponding enterprise record is located from the sorted enterprise list output in step S502. The location process involves traversing the sorted enterprise list to establish an index match based on the target enterprise identifier. Upon a match, the complete fields of the corresponding enterprise record are read to obtain detailed data. If location fails, the target enterprise identifier and a list snapshot are recorded in the location exception table, and a data unavailable message is returned to the user interface. The detailed data includes basic enterprise information, financing history, operating indicators, risk labels, and detailed fields of scoring dimensions, and is written to the details buffer.

[0079] Simultaneously, the detailed data is input into the potential assessment model to perform dimensional attribution analysis. The potential assessment model reads the enterprise feature vector and potential financing score from the detailed data, and performs contribution calculations for each dimension, including financing window period, number of institutional endorsements, risk indicators, and industry popularity index. The contribution calculation uses the ratio of the weighted value of each dimension to the total score to measure the degree of influence of each dimension on the final score, thus obtaining the dimensional attribution result. When anomalies occur, the abnormal dimension and model status are recorded in the attribution log, and equal contribution rollback is performed.

[0080] Therefore, a rating explanation text is generated based on the dimensional attribution results according to a preset text template. The preset text template is read from the template configuration table; this template defines the natural language description format and numerical interpolation positions for the contribution of each dimension. The rating explanation text is generated by filling the preset text template with the contribution ranking and values ​​of each dimension in the dimensional attribution results. If the template filling is abnormal, the abnormal position and attribution data are recorded in the template abnormality table and output using a simplified text format.

[0081] Accordingly, a company details page template is loaded and data population is performed to generate the company details page. The company details page template defines the layout structure and style specifications for the basic information area, financing history area, operating indicator area, and rating interpretation area. The details data is populated into the basic information area and financing history area, and the rating explanation text is populated into the rating interpretation area to complete page assembly. The company details page includes a page identifier field, a company identifier field, a details data field, a rating explanation text field, and a generation timestamp field, using the user session identifier and page access sequence number as primary keys. These are read by the user interface at the details display entry point for presenting company details and supporting user investment decisions.

[0082] To effectively address the shortcomings of traditional technologies in semantic parsing intent recognition, multi-dimensional fusion retrieval potential assessment, and list rendering scoring attribution, and to provide technical support for intelligent matching of structured enterprise data based on natural language processing, this application provides an embodiment of a structured data matching device based on natural language processing for implementing all or part of the aforementioned structured data matching method. The structured data matching device based on natural language processing specifically includes the following components: The semantic analysis module is used to receive natural language query statements entered by users through the search interface and input the natural language query statements into the semantic parsing module to perform word segmentation, entity recognition, and syntactic dependency analysis to obtain semantic feature vectors. The semantic feature vectors are then compared with a preset intent template library to calculate the vector similarity and obtain intent matching results. Based on the intent matching results, structured query conditions are generated and cached in the session context. The data matching module is used to input the structured query conditions as retrieval parameters into the multidimensional filtering engine and associate them with the historical query records in the session context to perform condition fusion to obtain fused retrieval conditions. Based on the fused retrieval conditions, it filters enterprise records that meet the conditions from the enterprise database to obtain a candidate enterprise set. It inputs the candidate enterprise set into the potential assessment model to perform multidimensional weighted calculation to obtain the potential financing score corresponding to each enterprise. Based on the potential financing score, it sorts the candidate enterprise set in descending order to obtain a sorted enterprise list. The results feedback module is used to render the sorted enterprise list according to the user-defined field configuration to generate an enterprise recommendation list page and push it to the user interaction interface. In response to user click operations, it reads the detailed data of the corresponding enterprise from the sorted enterprise list, calls the potential assessment model to perform attribution analysis on the detailed data, generates a score explanation text, and embeds it into the enterprise details page.

[0083] As described above, the structured data matching device based on natural language processing provided in this application can generate structured query conditions and cache them in the session context by performing word segmentation, entity recognition, and syntactic dependency analysis on natural language query statements through a semantic parsing module to obtain semantic feature vectors, performing vector similarity calculation with a preset intent template library to obtain intent matching results, as well as associating historical query records to perform condition fusion and screening of candidate enterprise sets, inputting them into a potential assessment model to perform multi-dimensional weighted calculation to obtain potential financing scores, and sorting them in descending order to obtain a sorted enterprise list. The list is then rendered and pushed to an enterprise recommendation list page according to user-defined field configurations. This effectively solves the shortcomings of traditional technologies in semantic parsing intent recognition, multi-dimensional fusion retrieval potential assessment, and list rendering score attribution, and provides technical support for intelligent matching of structured enterprise data based on natural language processing.

[0084] This invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the structured data matching method based on natural language processing.

[0085] This invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described structured data matching method based on natural language processing.

[0086] This invention also provides a computer program product, which includes a computer program that, when executed by a processor, implements the above-described structured data matching method based on natural language processing.

[0087] In this embodiment of the invention, a semantic feature vector is obtained by performing word segmentation, entity recognition, and syntactic dependency analysis on a natural language query statement through a semantic parsing module. This vector feature vector is then used to calculate the intent matching result with a preset intent template library. The resulting structured query conditions are then cached in the session context. Historical query records are associated with these conditions to perform condition fusion and filter a set of candidate companies. This set is then input into a potential assessment model to perform multi-dimensional weighted calculations to obtain a potential financing score, which is then sorted in descending order to obtain a ranked list of companies. The list is then rendered and pushed to a company recommendation list page according to user-defined field configurations. This effectively solves the shortcomings of traditional technologies in semantic parsing intent recognition, multi-dimensional fusion retrieval potential assessment, and list rendering score attribution, providing technical support for intelligent matching of structured enterprise data based on natural language processing.

[0088] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A structured data matching method based on natural language processing, characterized in that, The method includes: The system receives natural language query statements input by users through the search interface and inputs the natural language query statements into the semantic parsing module to perform word segmentation, entity recognition, and syntactic dependency analysis to obtain semantic feature vectors. The semantic feature vectors are then compared with a preset intent template library to calculate vector similarity and obtain intent matching results. Based on the intent matching results, structured query conditions are generated and cached in the session context. The structured query conditions are input as retrieval parameters into a multidimensional filtering engine and associated with historical query records in the session context to perform condition fusion to obtain fused retrieval conditions. Based on the fused retrieval conditions, enterprise records that meet the conditions are filtered from the enterprise database to obtain a candidate enterprise set. The candidate enterprise set is input into a potential assessment model to perform multidimensional weighted calculation to obtain the potential financing score corresponding to each enterprise. The candidate enterprise set is sorted in descending order according to the potential financing score to obtain a sorted enterprise list. The sorted company list is rendered according to the user-defined field configuration to generate a company recommendation list page, which is then pushed to the user interface. In response to user clicks, the detailed data of the corresponding company is read from the sorted company list. The potential assessment model is then invoked to perform attribution analysis on the detailed data to generate a score explanation text, which is then embedded in the company details page.

2. The structured data matching method based on natural language processing according to claim 1, characterized in that, The process of receiving natural language queries input by users through the search interface and then inputting these queries into a semantic parsing module for word segmentation, entity recognition, and syntactic dependency analysis to obtain semantic feature vectors includes: By listening to user input events through the search interface and capturing the input content when a query submission action is detected, a natural language query statement is obtained. The natural language query statement is then input into a preprocessing unit to perform text cleaning and stop word filtering to obtain normalized text. The normalized text is then input into a word segmenter to perform segmentation processing to obtain a word sequence. The word sequence is input into the semantic parsing module and named entity recognition is performed based on a preset entity dictionary to obtain an entity annotation sequence. Syntactic dependency analysis is performed on the entity annotation sequence to obtain a dependency tree. The dependency tree and the entity annotation sequence are input into a vector encoder to perform joint encoding to obtain a semantic feature vector.

3. The structured data matching method based on natural language processing according to claim 1, characterized in that, The step of calculating the similarity between the semantic feature vector and a preset intent template library to obtain an intent matching result, generating structured query conditions based on the intent matching result, and caching the structured query conditions in the session context includes: Read the template vector set corresponding to each intent template from the preset intent template library, perform cosine similarity calculation on the semantic feature vector and the template vector set one by one to obtain a similarity score sequence, and filter the intent identifiers that meet the conditions from the similarity score sequence according to the preset threshold conditions to obtain the intent matching result; Based on the intent matching result, the corresponding query parameter mapping rule is read, and the entity information in the semantic feature vector is converted into a structured query condition according to the query parameter mapping rule. The structured query condition is then associated with the current user identifier and written into the session context.

4. The structured data matching method based on natural language processing according to claim 1, characterized in that, The process involves inputting the structured query conditions as retrieval parameters into a multidimensional filtering engine and associating them with historical query records in the session context to perform condition fusion and obtain fused retrieval conditions. Based on these fused retrieval conditions, candidate enterprise sets are obtained by filtering enterprise records that meet the conditions from the enterprise database, including: The structured query conditions are input into the multidimensional filtering engine and historical query records are read from the session context. The structured query conditions and historical query records are merged and conflict resolved according to the preset fusion strategy to obtain the fused retrieval conditions. Based on the fusion search conditions, a database search statement is constructed and sent to the enterprise database. The enterprise records returned by the enterprise database are received, and secondary filtering is performed according to the filtering dimensions in the fusion search conditions to obtain a set of candidate enterprises.

5. The structured data matching method based on natural language processing according to claim 1, characterized in that, The process involves inputting the candidate enterprise set into a potential assessment model to perform multi-dimensional weighted calculations to obtain the potential financing score for each enterprise, and then sorting the candidate enterprise set in descending order based on the potential financing scores to obtain a ranked enterprise list, including: The records of each enterprise in the candidate enterprise set are read one by one, and the financing window period, number of institutional endorsements, risk indicators and industry popularity index are extracted to obtain the enterprise feature vector. The enterprise feature vector is input into the potential assessment model and weighted summation is performed according to the preset dimension weight configuration to obtain the potential financing score of each enterprise. The potential financing score is associated with the enterprise records in the candidate enterprise set to obtain a set of enterprises with scores. The set of enterprises with scores is then sorted in descending order according to the potential financing score to obtain a sorted enterprise list.

6. The structured data matching method based on natural language processing according to claim 1, characterized in that, The step of rendering the sorted company list according to user-defined field configurations to generate a company recommendation list page and pushing it to the user interface includes: Read the user-defined field configuration and extract the field data that matches the user-defined field configuration from the sorted enterprise list to obtain the dataset to be rendered. Perform column mapping on the dataset to be rendered according to the field order in the user-defined field configuration to obtain structured list data. The structured list data is input into the list rendering engine and a preset list page template is loaded. Data population and style rendering are performed to generate a company recommendation list page, which is then pushed to the user interface.

7. The structured data matching method based on natural language processing according to claim 1, characterized in that, The response to a user click operation reads the detailed data of the corresponding company from the sorted company list, calls the potential assessment model to perform attribution analysis on the detailed data, generates a score explanation text, and embeds it into the company details page, including: Listen for click events on the user interface and parse the target company identifier from the click events. Based on the target company identifier, locate the corresponding company record from the sorted company list and read the detailed data. The detailed data is input into the potential assessment model, and the contribution of each scoring dimension is calculated to obtain the dimension attribution results. Based on the dimension attribution results, a scoring explanation text is generated according to a preset text template. The enterprise details page template is loaded, and the detailed data and the scoring explanation text are filled into the corresponding areas to generate the enterprise details page.

8. A structured data matching device based on natural language processing, characterized in that, The device includes: The semantic analysis module is used to receive natural language query statements entered by users through the search interface and input the natural language query statements into the semantic parsing module to perform word segmentation, entity recognition, and syntactic dependency analysis to obtain semantic feature vectors. The semantic feature vectors are then compared with a preset intent template library to calculate the vector similarity and obtain intent matching results. Based on the intent matching results, structured query conditions are generated and cached in the session context. The data matching module is used to input the structured query conditions as retrieval parameters into the multidimensional filtering engine and associate them with the historical query records in the session context to perform condition fusion to obtain fused retrieval conditions. Based on the fused retrieval conditions, it filters enterprise records that meet the conditions from the enterprise database to obtain a candidate enterprise set. It inputs the candidate enterprise set into the potential assessment model to perform multidimensional weighted calculation to obtain the potential financing score corresponding to each enterprise. Based on the potential financing score, it sorts the candidate enterprise set in descending order to obtain a sorted enterprise list. The results feedback module is used to render the sorted enterprise list according to the user-defined field configuration to generate an enterprise recommendation list page and push it to the user interaction interface. In response to user click operations, it reads the detailed data of the corresponding enterprise from the sorted enterprise list, calls the potential assessment model to perform attribution analysis on the detailed data, generates a score explanation text, and embeds it into the enterprise details page.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the structured data matching method based on natural language processing as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the steps of the structured data matching method based on natural language processing as described in any one of claims 1 to 7.

Images