A multi-level entity classification method and system based on a large language model and a rule engine

Abstract

The application provides a kind of multi-level entity classification method and system based on large language model and rule engine, belong to data processing technical field.The method includes: the first filtering step, based on the exclusion rule set of dynamic update, the basic attribute data of entity is filtered;Second filtering step, after the weight distribution of the multi-dimensional feature data of entity, hierarchical matching analysis is carried out with feature library;Third filtering step, the integrated feature data and target classification definition information are input into pre-training large language model to carry out semantic reasoning determination, and output confidence;Classification mark and feedback step, according to the comparison result of confidence and preset threshold, determine the entity attribution and mark, and at the same time, the determination result is fed back to the feature library for self-optimizing update.The application realizes efficient and accurate entity classification through multi-layer progressive filtering architecture, has the advantages of high processing efficiency, high recognition accuracy, high intelligence, strong expansibility and the like.

Description

Technical Field

[0001] This invention relates to the field of big data processing technology, and in particular to a multi-level entity classification method and system based on a large language model and rule engine. Background Technology

[0002] With the development of information technology, various industries have accumulated massive amounts of entity data, such as enterprise information, research institution data, and product information databases. How to efficiently and accurately identify entities that meet specific target classifications (e.g., "high-tech enterprises," "specialized and innovative enterprises," or enterprises in a specific technology field) from this large-scale, multi-source, and heterogeneous data, and to implement precise data labeling for them, has become a major challenge in the field of big data processing.

[0003] In existing technologies, the following methods are commonly used to classify entity data:

[0004] Keyword-based matching methods predefine a set of keywords related to the target category and then perform string matching in an entity database. This method is simple to implement and fast, but has low accuracy and is prone to generating a large number of false positives and false negatives. For example, entities whose names contain the word "technology" may have actual business operations completely unrelated to the target category; while entities using emerging terms may be missed due to outdated keyword databases. Essentially, this method is a simple string pattern matching approach and cannot understand the deeper semantics of entity data.

[0005] Classification methods based on a single data source: These methods typically rely solely on basic entity attributes, such as name and description, for judgment. Due to the limited information dimension, they cannot comprehensively assess the entity's true attributes. For example, based solely on a company's business scope, it's difficult to accurately determine whether it possesses technological R&D capabilities and market activities in a specific field, leading to unreliable classification results.

[0006] The manual annotation method involves business experts manually reading entity information and making judgments and annotations based on experience. This method has a high accuracy rate, but its processing efficiency is extremely low, and the labor cost is high, making it completely unsuitable for handling the real-time or near-real-time processing needs of massive amounts of data.

[0007] In summary, existing technical solutions have room for optimization in areas such as balancing processing efficiency and recognition accuracy, utilizing multi-source heterogeneous data, and intelligent processing capabilities when handling large-scale entity data classification tasks. Designing a computer-based method that can improve entity classification accuracy while ensuring high processing efficiency is a pressing technical problem to be solved in this field. Summary of the Invention

[0008] To address the aforementioned technical problems, the technical solution adopted by this invention is as follows:

[0009] According to a first aspect of the present invention, a multi-level entity classification method based on a large language model and a rule engine is provided, the method comprising the following steps:

[0010] The first filtering step is to obtain the basic attribute data of the entities to be processed, filter the basic attribute data based on the dynamically updated exclusion rule set, and remove entities that do not meet the basic attribute thresholds corresponding to the target classification to obtain the first candidate entity set.

[0011] The second filtering step is to obtain multi-dimensional feature data for each entity in the first candidate entity set, assign weights to the multi-dimensional feature data, perform hierarchical matching analysis with the feature library pre-configured for the target classification, and filter out the second candidate entity set based on the matching results; the multi-dimensional feature data includes at least semantic features and association features.

[0012] The third filtering step is to perform feature fusion on the multi-dimensional feature data of each entity in the second candidate entity set, input the fused feature data and the definition information of the target classification into the domain-adapted pre-trained large language model, use the pre-trained large language model to perform semantic reasoning and judgment, and output the confidence score.

[0013] Classification labeling and feedback steps: Based on the comparison result between the confidence level output by the third filtering step and the preset classification threshold, determine whether the entity belongs to the target category; label the entity that is determined to belong to the target category with the corresponding classification label, and feed back the feature data of the entity and the judgment result to the feature library for self-optimization and updating of the feature library.

[0014] In a second aspect, the present invention provides an entity classification system based on a multi-layer filtering mechanism, comprising:

[0015] The first filtering module is used to obtain the basic attribute data of the entities to be processed, filter the basic attribute data based on the dynamically updated exclusion rule set, and remove entities that do not meet the basic attribute threshold corresponding to the target classification to obtain the first candidate entity set.

[0016] The second filtering module is used to obtain multi-dimensional feature data of each entity in the first candidate entity set, assign weights to the multi-dimensional feature data, perform hierarchical matching analysis with the feature library pre-configured for the target classification, and filter out the second candidate entity set according to the matching results; the multi-dimensional feature data includes at least semantic features and association features.

[0017] The third filtering module is used to perform feature fusion on the multi-dimensional feature data of each entity in the second candidate entity set, input the fused feature data and the definition information of the target classification into the domain-adapted pre-trained large language model, use the model to perform semantic reasoning and judgment, and output the confidence score.

[0018] The classification labeling and feedback module determines whether an entity belongs to the target category based on the comparison result between the confidence level output by the third filtering module and the preset classification threshold; it labels the entity that is determined to belong to the target category with the corresponding classification label, and feeds back the feature data of the entity and the judgment result to the feature library for self-optimization and updating of the feature library.

[0019] The present invention has at least the following beneficial effects:

[0020] This invention achieves a dual improvement in entity classification efficiency and accuracy through a multi-layered progressive filtering architecture. The first filtering step quickly eliminates irrelevant entities, reducing the computational load on subsequent steps. The second filtering step further refines the classification through multi-dimensional feature fusion and hierarchical matching. The third filtering step introduces intelligent reasoning from a large model to handle complex semantics and boundary cases. Compared to existing technologies, this invention improves processing efficiency by tens of times and enhances classification accuracy. Furthermore, it achieves self-optimization and updating of the feature library through a feedback mechanism, exhibiting scalability and adaptability.

[0021] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of the present invention, nor is it intended to limit the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 A flowchart illustrating a multi-level entity classification method based on a large language model and a rule engine, provided for embodiments of the present invention. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of this invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0026] It should be noted that some exemplary embodiments are described as processes or methods depicted as flowcharts. Although the flowcharts describe the steps as sequential processes, many of these steps can be performed in parallel, concurrently, or simultaneously. Furthermore, the order of the steps can be rearranged. A process can be terminated when its operation is complete, but it may also have additional steps not included in the figures. A process can correspond to a method, function, procedure, subroutine, subroutine, etc.

[0027] To facilitate understanding of this invention, the key terms used herein are defined as follows:

[0028] Entity: refers to a digital data object stored in a computer system that has an independent identity, such as enterprise data object, organization data object, or product data object.

[0029] Target classification: refers to data category labels predefined in a computer system for classifying entities; each target classification contains definition information for semantic description and a pre-configured feature library for rule matching.

[0030] Core features: These are features that directly correspond to the core business, technology, or product of the target category, appear very infrequently in non-target category entities, and have high exclusivity. They are the key criteria for determining whether an entity belongs to the target category and are used for the first level of matching in the feature library.

[0031] Marginal features: These are features that are related to the target classification but have weak exclusivity. They may appear in upstream or downstream entities of the target classification or in supporting entities. They are used for the second level of matching in the feature library to help improve the accuracy of entity classification.

[0032] Consistency level characterization value: refers to the degree of consistency between the preliminary judgment results of multiple pre-trained large language models with different architectures for the same entity. The value range is [0,1]. The higher the value, the stronger the consistency of the judgment results between the models.

[0033] Feature library: refers to a collection of core features and marginal features that correspond one-to-one with the target classification. It includes feature words, feature weights, and feature levels and is the basis for multi-dimensional feature matching analysis of entities. It supports self-optimization and updating.

[0034] Domain adaptation: refers to adapting a general pre-trained large language model to the entity determination requirements of a specific target classification through structured prompt word engineering or domain data fine-tuning, without the need for large-scale reconstruction of the model itself.

[0035] This invention uses a multi-layered filtering mechanism to determine whether an entity (e.g., an enterprise data object) should belong to a predefined target category (e.g., the data category of "enterprise entities in the energy and power industry chain"), and automatically adds the label of the target category to the entities that are determined to belong, thereby achieving the orderly organization and management of massive entity data.

[0036] Based on this, embodiments of the present invention provide a multi-level entity classification method based on a large language model and a rule engine, such as... Figure 1 As shown, the method includes the following steps:

[0037] S100, First filtering step: Obtain the basic attribute data of the entities to be processed, filter the basic attribute data based on the dynamically updated exclusion rule set, remove entities that do not meet the basic attribute threshold corresponding to the target classification, and obtain the first candidate entity set.

[0038] The purpose of the first filtering step is to quickly eliminate entities from the massive entity dataset that clearly do not meet the basic requirements of the target classification, thereby significantly reducing the computational load of subsequent fine-grained analysis and improving overall processing efficiency. This step uses a dynamically updated set of exclusion rules to quickly filter the basic attribute data of the entities to be processed.

[0039] The basic attribute data includes at least objective fields such as the entity's name, type identifier, and existence status identifier. The exclusion rule set can be dynamically updated based on the characteristics of the target category and business feedback to ensure the timeliness and relevance of the rules. This rule set contains at least one of the following rules:

[0040] (1) General word removal rules based on entity names;

[0041] (2) Exclusion rules based on entity type identifier;

[0042] (3) Exclusion rules based on entity existence status identifiers;

[0043] (4) Exclusion rules based on foreign investment attributes.

[0044] The following is a detailed explanation of each rule:

[0045] (1) General word removal rules based on entity names

[0046] This rule uses a predefined, general exclusion thesaurus to quickly match entity names and filter out entities whose names contain keywords that explicitly point to non-target category categories. The construction and maintenance of the general exclusion thesaurus is the core of this rule; each entry in the thesaurus typically contains the keyword itself and its corresponding category identifier. The vocabulary in the thesaurus can include the following categories:

[0047] Non-target industry keywords: words that explicitly point to other industry sectors, such as "catering", "food", "clothing", "real estate", "finance", "education", "medical care", etc. (specific words can be dynamically adjusted according to the target category).

[0048] Non-productive or intermediary keywords: words that refer to non-productive entities or intermediary organizations, such as "trade", "agency", "consulting", "service", "logistics", etc. (whether to exclude them needs to be determined based on the specific application scenario).

[0049] Abnormal or invalid entity keywords: words that point to entities that may be abnormal or invalid, such as "branch office", "office", "representative office", "training center", etc.

[0050] The specific implementation steps of the general word removal rule are as follows:

[0051] Step 1, Lexicon Building and Maintenance: Based on the business scope of the target category, build an initial general lexicon for elimination and set up a dynamic update mechanism. The lexicon can be continuously optimized based on human feedback or machine learning methods.

[0052] Step 2, Name Extraction: Obtain the name data of the entity to be processed.

[0053] Step 3, Matching Detection: Perform string matching or semantic matching between the entity name and each term in the general exclusion thesaurus to determine whether the name contains any general exclusion term.

[0054] Step 4, Filtering Decision: If the entity name contains any of the common exclusion words, the entity is directly determined to be unqualified and removed from the candidate set, and will not proceed to the next filtering step; if it does not contain any common exclusion words, the entity is retained and enters the next round of filtering.

[0055] This rule is extremely efficient (in milliseconds), quickly filtering out a large number of obviously irrelevant entities, thus reducing the burden on subsequent detailed analysis. Practice shows that the general word removal rule can typically filter out 40% to 60% of entities.

[0056] (2) Exclusion rules based on entity type identifier

[0057] This rule filters entities that do not meet the basic type requirements of the target classification based on the type identifier field stored in the data system. Entity type identifiers are usually derived from authoritative data sources such as business registration and organization registration, for example, enterprise type codes and organization category codes.

[0058] The specific implementation steps are as follows:

[0059] Step 1: Obtain the entity type identifier, that is, read the type identifier field of each entity from the entity database, such as "enterprise type" (domestic enterprise, foreign-invested enterprise, individual business, etc.) or "organization type" (public institution, social organization, etc.).

[0060] Step 2: Pre-set the exclusion type set. Based on the needs of the target classification, pre-configure the list of type identifiers to be excluded. For example, if the target classification only focuses on domestic enterprises, then identifiers such as "foreign-invested enterprises," "Sino-foreign joint ventures," and "resident representative offices of foreign enterprises" can be included in the exclusion set; if the target classification focuses on manufacturing enterprises, entities with type identifiers such as "trading companies" and "consulting companies" can be excluded.

[0061] Step 3: Matching and Exclusion. The type identifier of the entity is compared with the preset exclusion type set. If the type identifier of the entity matches the exclusion type set, the entity is directly eliminated; otherwise, the entity is retained and proceeds to the next step.

[0062] The aforementioned set of preset exclusion types is not static but supports dynamic updates to adapt to the needs of different application scenarios. The update mechanism includes:

[0063] Adjustments based on changes in target classification: When the definition or scope of the target classification changes, the composition of the exclusion type set is adjusted accordingly;

[0064] Optimization based on data distribution characteristics: By analyzing the type distribution patterns of confirmed target entities and non-target entities in the entity dataset, new types that should be excluded are automatically discovered and included, or types that are no longer applicable are removed;

[0065] Correction based on feedback data: When certain types are found to have been incorrectly excluded or missed based on subsequent filtering steps or manual review, the set is corrected. This rule utilizes authoritative structured data, ensuring high accuracy and fast processing speed, and can effectively filter out a large number of entities that do not match the data type.

[0066] (3) Exclusion rules based on entity existence status identifiers

[0067] This rule filters out invalid or abnormal entities such as those that are deregistered, revoked, or suspended, based on their legally recognized operating status. The operating status identifier also originates from authoritative data sources, such as the "Operating Status" field in business registration (existing, deregistered, revoked, relocated, etc.). The specific implementation steps are as follows:

[0068] Step 1, Obtain the existence status identifier: Read the existence status field of the entity.

[0069] Step 2: Preset a set of valid states. Define a list of valid existing states based on the requirements of the target classification, such as "existing", "operating", "open for business", etc.

[0070] Step 3: Preset the exclusion state set, which includes all states that are not in the valid state set, such as "cancellation", "revocation", "relocation", "closure", etc.

[0071] Step 4: Matching and Exclusion. Determine whether the existence status of an entity belongs to the exclusion status set. If yes, the entity is directly removed; otherwise, the entity is retained for subsequent steps.

[0072] The aforementioned set of valid states and set of excluded states are not static and immutable, but rather support dynamic updates to adapt to the needs of different application scenarios. The update mechanisms include:

[0073] Adjustments based on business rules: When the definition of target classification changes, the criteria for defining valid and excluded states can be adjusted accordingly.

[0074] Adaptation based on data source changes: When the status identifier field or its enumeration value provided by the data source changes, the status set is updated synchronously;

[0075] Optimization based on feedback data: When certain state definitions are found to be unreasonable based on subsequent filtering steps or manual review results, the set is corrected.

[0076] This rule ensures that all entities entering subsequent analysis are currently valid and in normal operating condition, avoiding invalid calculations for invalid entities and further improving the screening quality.

[0077] (4) Exclusion rules based on foreign investment attributes

[0078] This rule excludes foreign-invested enterprises or enterprises with foreign majority ownership, and applies to scenarios where the target classification focuses on domestic entities or entities with a specific country of origin. Determining foreign ownership requires a comprehensive assessment of multiple dimensions, including the entity's name, the enterprise type identifier in business registration, and the equity structure.

[0079] The specific criteria for judgment include:

[0080] Company name characteristics: If an entity name contains keywords such as "foreign-invested", "foreign-owned enterprise", "joint venture", or "cooperation", or contains the name of a foreign country or region (such as "United States", "Japan", "Germany", etc.), it may be judged as having foreign investment attributes.

[0081] Enterprise type identifier: Based on the enterprise type field in the business registration information, if the entity type is "foreign-invested enterprise", "Sino-foreign joint venture", "Sino-foreign cooperative enterprise", "resident representative office of foreign enterprise", etc., it is directly identified as a foreign-invested enterprise.

[0082] Equity structure analysis: By analyzing the shareholder information of an entity, if the major shareholder is a foreign enterprise or a foreign individual and the shareholding ratio exceeds a preset threshold (such as 50%), it is determined to be a foreign-invested enterprise.

[0083] The specific implementation steps of the foreign investment exclusion rule are as follows:

[0084] Step 1: Obtain basic information about the entity to be processed, including entity name, enterprise type identifier, shareholder information and shareholding ratio, etc.

[0085] Step Two: Preliminary Name Feature Assessment: Check if the entity name contains keywords related to foreign investment or foreign region names. If a match is found, it is marked as suspected foreign investment and requires further verification; if no match is found, continue to the next steps.

[0086] Step 3: Enterprise Type Identifier Confirmation: Compare the entity type identifier with the preset set of foreign-invested enterprise types (such as "foreign-invested enterprise"). If a match is found, the entity is directly identified as a foreign-invested entity and removed from the list.

[0087] Step 4: Equity Structure Analysis: If Steps 2 and 3 cannot provide a clear determination, further analysis of shareholder information is conducted to calculate the foreign shareholding ratio (total shareholding by foreign shareholders). If the foreign shareholding ratio exceeds a preset threshold (e.g., 50%), it is determined to be a foreign-invested entity; otherwise, it is considered a domestic entity.

[0088] Step 5: Screening Decision: If an entity is determined to be foreign-invested, it will be directly eliminated and will not proceed to the next filtering step; if it is domestic-invested, it will be retained and enter the next round of screening.

[0089] The rules also support a dynamic update mechanism to adapt to policy adjustments, data source changes, and business feedback. Updates include:

[0090] Based on policy and regulatory adjustments: When the definition of foreign investment or access policies change, the set of foreign investment types, keyword database, or shareholding ratio thresholds will be adjusted accordingly.

[0091] Data source optimization: When more accurate shareholder information or new data fields are obtained, optimize the logic of equity structure analysis.

[0092] Feedback-based correction: Based on the review results of subsequent stages, correct any misjudgments or omissions, such as adjusting keywords or shareholding ratio thresholds.

[0093] The aforementioned exclusion rule set is not static but supports dynamic updates to adapt to the needs of different target classifications and application scenarios. The dynamic update mechanism includes:

[0094] Updates based on business feedback: If misjudgments or omissions are found in certain rules based on subsequent filtering steps or manual review, rule updates can be triggered. For example, if it is found that some entities mistakenly excluded by general exclusion rules actually belong to the target category, the relevant keywords can be removed from the exclusion keyword library; if it is found that some type identifiers should be excluded but were not, the exclusion type set can be updated.

[0095] Data distribution-based updates: By periodically analyzing the distribution characteristics of entity data, new high-frequency invalid keywords or type identifiers are automatically discovered and incorporated into the rule set.

[0096] Adaptation based on target classification: Different exclusion rule sets can be configured for different target classifications, enabling pluggable and fast switching of rules.

[0097] By combining the above rules, the first filtering step can eliminate a large number of entities that clearly do not meet the basic requirements in a very short time (milliseconds for a single rule, seconds for the whole process). Practice shows that this step typically filters out 60% to 80% of the entities, with the remaining 20% ​​to 40% proceeding to the second filtering step for more refined analysis. This efficient coarse-screening mechanism significantly reduces the computational load of subsequent steps, laying a solid foundation for processing massive amounts of entity data with limited resources. Simultaneously, because the rules are based on objective attribute data, the false positive rate is low, ensuring the reliability of the screening.

[0098] S200, Second filtering step: Obtain multi-dimensional feature data of each entity in the first candidate entity set, assign weights to the multi-dimensional feature data, perform hierarchical matching analysis with the feature library pre-configured for the target classification, and filter out the second candidate entity set based on the matching results.

[0099] The purpose of the second filtering step is to conduct a more refined multi-dimensional feature analysis on the remaining entities, building upon the initial screening in the first filtering step. This further filters out entities highly relevant to the target classification, providing high-quality input for subsequent intelligent judgment by the large-scale model. The core of this step is to construct a feature library for the target classification and perform hierarchical matching analysis on the multi-dimensional feature data of entities based on this feature library.

[0100] 1. Construction of the target classification feature library

[0101] The feature library is fundamental for determining whether an entity belongs to a target category, and its quality directly affects the accuracy of subsequent screening. The feature library corresponds one-to-one with the target category and contains core and marginal features used to characterize that category. The feature library construction process is as follows:

[0102] Step 1: Define the scope of the target classification. First, clarify the specific connotation and boundaries of the target classification to be identified. Each target classification has its specific business scope and technical field, such as "energy and power industry chain enterprises", "high-end equipment manufacturing enterprises", "high-tech enterprises", etc.

[0103] Step 2: Collect feature words. For each target category, collect relevant feature words through multiple channels:

[0104] Expert consultation: Invite experts in relevant fields to provide professional advice;

[0105] Public data analysis: Analyzing publicly available information such as industry reports, standards and specifications, and policy documents;

[0106] Sample entity analysis: Study the business descriptions of typical entities that have been identified as belonging to the target category;

[0107] Patent data analysis: Extracting patent technology terms relevant to the target classification;

[0108] Bidding data analysis: Extract business keywords from bidding announcements that are relevant to the target category.

[0109] Step 3: Feature Classification and Organization. The collected feature words are classified according to type to form a feature word library. Feature words can be divided into the following categories:

[0110] Product category characteristics: vocabulary describing specific products or equipment, such as "photovoltaic modules", "wind turbines", "transformers", etc.

[0111] Technical characteristics: Terms describing technologies or processes, such as "photovoltaic power generation", "wind power generation", "smart grid", etc.

[0112] Application-related features: words describing the application field or scenario, such as "new energy", "clean energy", "distributed energy", etc.

[0113] Material characteristics: words describing raw materials or auxiliary materials, such as "silicon wafers", "steel", "composite materials", etc.

[0114] Business-related characteristics: words describing business activities or services, such as "equipment installation", "operation and maintenance services", "project general contracting", etc.

[0115] Step 4: Feature importance classification and weight assignment.

[0116] Based on the functional role of features in the matching process, this step divides the feature words in the feature library into two levels: core features and marginal features, and assigns a weight to each feature word to reflect its importance within that level.

[0117] (1) Feature grading

[0118] Core Features: Features used for the first level of filtering. The criterion for determining core features is: when an entity's multi-dimensional feature data contains the feature, the statistical probability of that entity belonging to the target category is significantly higher than the preset baseline probability. Core features typically correspond directly to the core business activities of the target category and appear very infrequently in other categories, exhibiting high exclusivity. In the matching process, only when an entity's matching degree on core features reaches the first preset threshold is it eligible to enter the second level of matching; otherwise, it is directly determined not to pass the second filtering step.

[0119] Marginal features: Features used for second-level matching. The criterion for determining marginal features is that when an entity's multi-dimensional feature data contains the feature, the statistical probability of that entity belonging to the target category is higher than random, but significantly lower than that of core features. Marginal features may be related to the target category, but may also appear in the descriptions of non-core entities such as upstream and downstream enterprises and supporting service providers related to the target category, exhibiting weak exclusivity. After passing the first-level screening, the overall matching score of entities is further refined by matching marginal features, used to distinguish between "entities that have just reached the core threshold" and "entities that highly match the target category".

[0120] (2) Assigning feature weights

[0121] Assign a weight to each feature word in the feature library to quantify its contribution to the matching score calculation. Weights can be assigned using one of the following methods or a combination thereof:

[0122] Expert experience method: Domain experts manually assign values ​​to features based on their business importance;

[0123] Machine learning methods: Based on labeled training data, feature weights are automatically learned through feature selection algorithms (such as TF-IDF, information gain, etc.).

[0124] Hybrid approach: Based on expert experience, combined with machine learning results for calibration and optimization.

[0125] (3) Examples of hierarchical and weighted classification

[0126] Taking the target category "new energy vehicle manufacturers" as an example, the feature hierarchy and weights can be set as follows:

[0127]

[0128] (4) Basis for determining the classification and weight

[0129] The classification and weighting of features are not based on subjective assumptions, but rather on statistical analysis of confirmed target entity samples and non-target entity samples:

[0130] Identification of core features: Statistical analysis of the frequency P of each feature word in the target entity sample. target (f) and the frequency P of occurrence in non-target entity samples non The ratio of (f). When this ratio is greater than a preset first threshold (e.g., P). target (f) / P non When (f) > 5), the feature can be identified as a core feature candidate.

[0131] Edge feature identification: When the above ratio is greater than a preset second threshold (e.g., 2 < P) target (f) / P non When (f) < 5), the feature can be identified as a candidate edge feature.

[0132] Determining the weights: The weights of features can be the normalized values ​​of the ratios mentioned above, or the feature importance scores output by the machine learning model.

[0133] Step 5: Forming the feature library. The set of feature words that have been categorized, graded, and weighted is defined as the feature library F = {F_{i}} for the target classification. core F edge}, where F core F is the core feature set. edge This is a set of edge features. The feature library can be continuously optimized and updated based on a dynamic feedback mechanism.

[0134] 2. Composition of Multidimensional Feature Data

[0135] This step obtains multi-dimensional feature data for each entity in the first candidate entity set. This multi-dimensional feature data includes at least semantic features and relational features, and may also include entity attribute features. These features characterize the entity's attributes from different perspectives, providing a rich information foundation for comprehensive judgment.

[0136] Semantic features: Semantic information extracted from the textual description data of entities, including entity name data, brief description data, and business scope description data. This textual data contains the core business information of the entity and is an important basis for determining the entity's affiliation. In specific analysis, preprocessing such as word segmentation and stop word removal is required to extract keywords.

[0137] Relationship characteristics: Relationship data reflecting an entity's business activities and technological capabilities, including at least one of intellectual property-related data and market activity-related data. Among them:

[0138] Intellectual property-related data includes patent data and trademark data. Patent data reflects an entity's technological R&D direction and innovation capabilities, from which technical keywords (such as patent names, technical terms in abstracts, IPC classification numbers, etc.) can be extracted; trademark data reflects an entity's brand layout and business areas. Market activity-related data includes bidding data and project application data. Bidding data reflects an entity's market performance and competitiveness, from which business keywords (such as business terms in bidding project names, types of winning products / services, etc.) can be extracted; project application data reflects an entity's participation in government or industry projects. Entity attribute characteristics: These are characteristics in the entity's basic attribute data that may be relevant to the target classification, including entity size identification data (such as staff size, registered capital), geographic identification data (such as registered address, place of business), and establishment duration data (such as years of establishment). These attribute characteristics can serve as auxiliary judgment criteria.

[0139] 3. Multi-dimensional feature matching analysis

[0140] After acquiring the multi-dimensional feature data of the entity, this step performs hierarchical matching analysis against a feature library pre-configured for the target classification. The core concept of hierarchical matching is to prioritize matching core features with higher importance. Only when the matching degree of core features reaches a certain level will edge features be further matched, thereby optimizing the allocation of computing resources and improving the screening accuracy.

[0141] The specific matching analysis steps are as follows:

[0142] Step 1: Semantic Feature Matching

[0143] Text analysis is performed on the semantic feature data of the entity (i.e., the entity's name, introduction, and business scope text), keywords are extracted, and they are matched with feature words in the feature library. The specific algorithm for semantic feature matching is as follows: (1) Preprocess the entity's name, introduction, and business scope text, including word segmentation, stop word removal, and part-of-speech tagging;

[0144] (2) Extract the keyword set K from the preprocessed semantic feature data;

[0145] (3) Traverse each feature word in the feature library and determine whether the feature word appears in the keyword set K;

[0146] (4) Statistically analyze the matched feature words and their weights, and calculate the preliminary matching results of the semantic features. Calculate the preliminary matching score of the semantic features according to the following formula: .

[0147] Among them, f j1 Let w be the j1-th feature word in the intersection (K∩F), where j1 ranges from 1 to m1, and m1 is the number of feature words in the intersection (K∩F); j1 For the characteristic word f j1 The preset weights are used to determine the semantic feature matching score. This formula means: take the intersection K∩F of the keyword set K and the feature library F, which represents all feature words that appear simultaneously in both the entity text and the feature library; then sum the preset weights of these feature words to obtain the semantic feature matching score. This score reflects the number and importance of the target classification feature words contained in the entity text.

[0148] Step 2: Feature Matching

[0149] Targeted analysis of entity association feature data is performed to extract key information and match it with a feature database. Association feature matching is divided into two sub-steps: patent feature matching and bidding feature matching. 1. Patent Feature Matching

[0150] Obtain the entity's patent data, including patent name, abstract, classification number, etc. Construct a patent feature database (which may include technical keywords and typical applicants) for the target classification. Match the entity's patents against the patent feature database and calculate the patent feature matching score. The matching rules include: if the patent name or abstract contains keywords from the feature database, add the corresponding weight; if the patent applicant is the same as or similar to a typical entity within the target classification, add appropriate points. The specific algorithm for patent feature matching is as follows:

[0151] (1) Extract the set of technical keywords K from each patent data of the entity. patent (Mainly derived from patent titles and abstracts);

[0152] (2) Traverse each feature word in the feature library and determine whether the feature word appears in K. patent middle;

[0153] (3) Statistically analyze the matched feature words and their weights, and calculate the patent keyword matching score S. patent-kw : .

[0154] Among them, f j2 For the intersection (K) patent The j2th feature word in ∩F), where j2 ranges from 1 to m2, and m2 is the intersection (K). patent The number of feature words in ∩F); w j2 For the characteristic word f j2 The preset weights.

[0155] (4) Obtain the patent applicant information A of the entity and determine whether A belongs to the preset typical applicant set AT (i.e., typical entities that have been confirmed to belong to the target classification). If A∈AT, then add the patent applicant bonus item S. patent-a The bonus value is the preset bonus coefficient β. patent (e.g., β) patent =0.2);

[0156] (5) Calculate the total patent feature matching score S patent =S patent-kw +S patent-a .

[0157] 2. Matching of bidding and tendering features

[0158] Acquire bidding data for entities, including bidding announcements, winning bid announcements, and other information. Construct a bidding feature library (which may include project keywords, purchasing parties, and types of winning products) for the target category. Match entity bidding records with the feature library and calculate the bidding feature matching score. The matching rules include: if the bidding project name or description contains keywords from the feature library, add the corresponding weight; if the purchasing party is a typical unit related to the target category, add appropriate points; if the winning product / service belongs to a category related to the target category, add points based on relevance. The specific algorithm for bidding feature matching is as follows:

[0159] (1) Extract the set of business keywords K from each bidding record of the entity. bid (Mainly derived from project name and project description);

[0160] (2) Traverse each feature word in the feature library and determine whether the feature word appears in K. bid middle;

[0161] (3) Statistically analyze the matched feature words and their weights, and calculate the keyword matching score S for bidding. bid-kw : .

[0162] Among them, f j3 For the intersection (K) bid The j3rd feature word in ∩F), where j3 ranges from 1 to m3, and m3 is the intersection (K). bid The number of feature words in ∩F); w j3 For the characteristic word f j3 The preset weights.

[0163] (4) Obtain the purchaser information P from the bidding records and determine whether P belongs to the preset typical purchaser set PT (i.e., typical units related to the target classification). If P∈PT, add the purchaser bonus item S. bid-b The bonus value is the preset bonus coefficient β. bid(e.g., β) bid =0.1);

[0164] (5) Obtain the winning product / service type T and calculate its relevance score to the target category. Pre-set the relevance weight ρ for each product / service category. c The value ranges from [0,1], reflecting the degree of association between the category and the target category. If the winning product / service type T belongs to the preset category set, then the relevance score S of the winning product is increased. bid_p =ρ T , where ρ T The relevance weights corresponding to category T;

[0165] (6) Calculate the total score S for matching bidding features. bid :S bid =S bid-kw +S bid-b +S bid_p .

[0166] Step 3: Stratified Matching and Overall Score Calculation

[0167] Based on the above matching results, the overall matching score of the entity is calculated according to the hierarchical matching strategy:

[0168] First-level matching: Matching the entity's multi-dimensional feature data with the core feature set in the feature library, and calculating the first matching degree S. core The first matching degree reflects the degree of consistency between the entity and the core features of the target classification. Among them, , where f r1 For the intersection (K) all ∩F core The r1th feature word in the set, where r1 ranges from 1 to n1, and n1 is the intersection of (K) all ∩F core The number of characteristic words in ); w r1 f r1 The preset weights, K all This is the union of keywords extracted from all feature data of an entity (semantic features, patent features, bidding features).

[0169] First-level judgment: Assign the first matching degree S... core With the first preset threshold θ core Compare them. If S core <θ core If the entity does not match the core features of the target classification sufficiently, the entity is directly determined not to pass the second filtering step, and the process ends; if S core ≥θ core If so, then proceed to the second level of matching.

[0170] Second-level matching: Matching the multi-dimensional feature data of the entity with the edge features in the feature library, and calculating the edge feature matching degree S. edge Edge feature matching degree reflects the degree of consistency between entity and target classification auxiliary features. Among them, Among them, f r2 For the intersection (K) all ∩F edge The r2th feature word in the set, where r2 ranges from 1 to n2, and n2 is the intersection of (K) all ∩F edge The number of feature words in ); w r2 f r2 The preset weights.

[0171] Overall matching degree calculation: Based on the first matching degree and the edge feature matching degree, the overall matching degree S is calculated according to the following formula:

[0172] S=α×S core +β×S edge .

[0173] Here, α and β are the preset weight coefficients for the core feature and the peripheral feature, respectively, and satisfy α + β = 1. The setting of the weight coefficients reflects the dominant role of the core feature in the judgment.

[0174] Comprehensive judgment: The comprehensive matching degree S is compared with the preset matching similarity threshold S0. If S≥S0, the entity passes the second filtering step and is included in the second candidate entity set; if S<S0, the entity fails the screening and the process ends.

[0175] The weights and thresholds in the second filtering step are key parameters and can be adjusted and optimized according to actual business needs and data characteristics.

[0176] Core feature weights and marginal feature weights: By default, α = 0.7 and β = 0.3 can be set to reflect the dominant role of core features. For target classifications where core features are particularly prominent, the α value can be appropriately increased; for classifications where marginal features are equally important, the β value can be appropriately increased.

[0177] First preset threshold θ core Used to determine whether an entity has sufficient basic matching degree. It is usually set to a high value (such as 0.5-0.7) to ensure that entities entering the second level have a certain core feature matching foundation.

[0178] Match similarity threshold S0: The threshold used for final filtering, which can be dynamically adjusted according to the strictness of target classification and business needs. In scenarios requiring high precision, a higher threshold (e.g., 0.8) can be set; in scenarios requiring high recall, the threshold can be appropriately lowered.

[0179] The second filtering step, through multi-dimensional feature fusion analysis and a hierarchical matching strategy, can further refine the selection of entities highly relevant to the target category, building upon the first filtering step. Practice shows that after the second filtering step, the entity-to-target-category matching rate can be increased to 70%-85%, a significant improvement compared to simple keyword matching methods. Simultaneously, the hierarchical matching mechanism avoids performing full feature matching for all entities, optimizing computational resource allocation, and keeping the processing time for a single entity within seconds, meeting the needs of large-scale data processing.

[0180] S300, Third filtering step: Perform feature fusion on the multi-dimensional feature data of each entity in the second candidate entity set, input the fused feature data and the definition information of the target classification into the domain-adapted pre-trained large language model, use the pre-trained large language model to perform semantic reasoning and judgment, and output the confidence score.

[0181] The purpose of the third filtering step (also known as the AI ​​analysis stage) is to perform intelligent semantic reasoning on entities that have passed the rule-based filtering, building upon the second filtering step. This step inputs entities that are still difficult to determine after the previous rule filtering (i.e., entities with ambiguous boundaries and complex semantics) into a domain-adapted pre-trained large language model. The model's semantic understanding capabilities are then used to make a final determination on whether the entity belongs to the target category. The domain-adapted pre-trained large language model can be a single model or a combination of multiple models with different architectures.

[0182] This step is one of the core innovative aspects of this invention. Its core idea is to combine the entity information filtered through the rules of the first two stages with the definition information of the target classification to construct a structured input, which is then input into a large language model. Utilizing the powerful natural language understanding and knowledge reasoning capabilities of the large model, it intelligently determines whether an entity belongs to the target classification.

[0183] Compared with traditional rule-based methods, large language models have the following advantages:

[0184] Semantic understanding capability: It can understand complex business descriptions, industry terms, and contextual semantics, rather than simple keyword matching;

[0185] Boundary handling capability: It can handle situations where information is ambiguous or boundaries are unclear, such as when the description of a company's business is too general or contains emerging terms.

[0186] Knowledge transfer capability: Possesses extensive knowledge reserves and is able to transfer analytical methods from one field to related fields;

[0187] Reasoning ability: Ability to perform logical reasoning based on given information, rather than relying solely on statistical matching.

[0188] This step, by introducing a large language model, solves the complex semantic problems that traditional rule-based methods struggle to handle, significantly improving the accuracy of the judgment.

[0189] Before proceeding with AI analysis, feature fusion is first required to analyze the multi-dimensional feature data of entities. The purpose of feature fusion is to integrate feature data from different sources and in different formats into a unified input format suitable for large-scale models. The fused feature data should include at least:

[0190] Key information extracted from semantic features (entity name, brief introduction, core content of business scope);

[0191] Key information extracted from associated features (core points of patented technology, main business content of bidding projects);

[0192] Auxiliary information extracted from entity attribute features (size, region, length of establishment, etc., optional).

[0193] Feature fusion can be achieved through simple text concatenation or structured summary generation. The fused feature data, along with the definition information of the target classification, should be constructed into structured prompt information according to a preset template.

[0194] The design of prompt words is a crucial step in the AI ​​analysis stage, directly impacting the accuracy of large-scale models. This invention, through structured prompt word engineering, organizes the definition information of the target classification and the multi-dimensional feature data of the entities into a unified input format, guiding a general-purpose pre-trained large-scale model to understand classification tasks in a specific domain, thereby achieving domain adaptation without model fine-tuning. The prompt word template contains the following five essential components:

[0195]

[0196] Through the structured design described above, prompt words can provide comprehensive domain context information for general pre-trained large models, enabling them to accurately understand the definition and boundaries of target classifications. This allows for accurate classification of specific domains without requiring fine-tuning of the model itself. This prompt word-based domain adaptation method has advantages such as low cost, rapid iteration, and ease of adjustment, and can flexibly adapt to different target classifications and application scenarios.

[0197] To further improve the robustness and reliability of the judgment, this invention preferably employs a multi-model fusion judgment mechanism. This mechanism inputs the same structured prompt information into multiple pre-trained large-scale language models with different architectures or different training data to obtain multiple preliminary judgment results. Then, the fusion model performs consistency analysis on these results to generate the final fused confidence score.

[0198] The specific steps for multi-model fusion determination are as follows:

[0199] S310: Construction of Structured Prompt Information

[0200] The fused entity feature data and target classification definition information are constructed into structured prompt information according to the preset template. This prompt information remains consistent across all language models involved in the judgment, ensuring the comparability of the inputs.

[0201] S311: Parallel Decision-Making for Multiple Models

[0202] The structured prompt information is input into multiple pre-trained large-scale language models with different architectures (such as GPT series, Claude series, Wenxin Yiyan, etc.) to obtain the preliminary judgment results and corresponding initial confidence scores of each model.

[0203] For the p-th model (p=1, 2, ..., Q, where Q represents the number of large language models participating in the decision), its output includes:

[0204] Preliminary judgment result r p : Usually a binary value, indicating whether the model considers an entity to belong to the target category (e.g., r). p =1 indicates that it belongs to, r p =0 indicates that it does not belong);

[0205] Initial confidence level c p The value range is [0,1], reflecting the model's confidence in its judgment results.

[0206] S312: Calculation of confidence score after fusion

[0207] The multiple preliminary judgment results are analyzed for consistency using a preset fusion model to generate a fused confidence level C, which is then used as the confidence level output by the third filtering step.

[0208] The design of the fusion model took the following factors into consideration:

[0209] Model authority: Different models may have varying accuracy and reliability, therefore a preset weight λ is set for each model. p This reflects its importance in the integration process;

[0210] Consistency of results: When the judgment results of multiple models are consistent, the confidence level of the fused model should be high; when the results are inconsistent, the confidence level of the fused model should be appropriately reduced to reflect uncertainty.

[0211] Based on the above considerations, this invention uses the following formula to calculate the confidence level C after fusion:

[0212] ;

[0213] Among them, c p λ represents the initial confidence level of the p-th model output. p This represents the preset weights of the p-th model, satisfying λ. p >0, can be set based on the model's historical performance or authority; r p This represents the preliminary judgment result of the p-th model, r. agg This represents the benchmark judgment result determined by an aggregation algorithm based on all preliminary judgment results, typically determined using a majority voting principle (i.e., taking the preliminary judgment result value that appears most frequently as the benchmark); δ(r) p r major ) is an indicator function, when r p With r agg The value is 1 if the match is consistent, otherwise it is 0.

[0214] The physical meaning of this formula is: the fused confidence score equals the weighted average of the initial confidence scores and weights of all models consistent with the benchmark judgment, divided by the sum of all weights. Outputs from inconsistent models are discarded and not included in the confidence score calculation. This ensures that the fused confidence score primarily reflects the judgments of the models that reached a consensus, while also reflecting the differences in importance between different models through weight adjustments.

[0215] The confidence score C obtained through the above calculation is the output of the third filtering step, which is used for subsequent classification labeling and feedback steps. The value of C ranges from [0,1], and the higher the value, the greater the probability that the entity belongs to the target category.

[0216] When using multi-model fusion for decision-making, the reliability of the fused confidence score is affected by the consistency between the models. If the decision results of multiple models differ significantly, even if the fused confidence score calculated by weighted averaging is high, its actual reliability should be appropriately reduced. Therefore, this invention introduces a consistency-based confidence score correction mechanism:

[0217] S320, calculate the consistency degree characterization value D of the multiple preliminary judgment results. The consistency degree characterization value D represents the degree of consistency between the multiple model judgment results. The value range is [0,1]. The higher the value, the stronger the consistency between the models.

[0218] The consistency level characterization value D can be calculated using one of the following methods:

[0219] Method 1 (based on majority ratio):

[0220] D = max(Np, Nn) / Q.

[0221] Where Np and Nn represent the number of models that determine that an entity belongs to the target category and the number that determine that it does not belong to the target category, respectively. The formula takes the proportion of models on the majority side as the consistency index.

[0222] Equation 2 (based on entropy):

[0223] D=1-[(-h×log2h-(1-h)×log2(1-h))] / log22=1+h×log2h+(1-h)×log2(1-h)).

[0224] Where h = Np / Q, is the proportion of models that belong to the target classification. This formula normalizes the entropy value of the result distribution and takes the complement. When the results are completely consistent, D = 1; when the results are completely divergent, D = 0.

[0225] Method 3 (Simplified Method):

[0226] D = 1 - |Np - Nn| / Q.

[0227] The formula measures the degree of disagreement by the absolute difference between the number of positive and negative determinations. The smaller the difference (i.e., the closer to equilibrium), the smaller D is, reflecting a lower degree of consistency.

[0228] Method 1 is preferred in this invention because it is simple to calculate and intuitive.

[0229] S330, when the consistency degree characterization value D is less than the preset consistency threshold D0, the fused confidence degree is penalized and corrected to obtain the corrected confidence degree.

[0230] In this invention, the corrected confidence level satisfies the following condition:

[0231] C adj =C×(1-γ×(1-D));

[0232] Among them, C adj C represents the corrected confidence level, C represents the uncorrected fusion confidence level, and γ represents the preset penalty coefficient, where 0 < γ < 1, used to control the intensity of the penalty. (1-D) represents the degree of inconsistency, serving as the basis for the penalty intensity.

[0233] If D≥D0, it indicates good consistency between models, and the confidence score after fusion does not need to be corrected. The original confidence score C after fusion can be used directly as the output. If D<D0, it indicates significant discrepancy between models, and the confidence score after fusion needs to be penalized.

[0234] The consistency threshold D0 ranges from (0,1), and is typically set between 0.6 and 0.8. It can be adjusted according to the reliability requirements of the business scenario. The higher the requirement, the closer D0 should be to 1.

[0235] The physical meaning of this formula is: the original confidence level C is reduced based on the degree of inconsistency (1-D) between models. The greater the degree of inconsistency, the greater the reduction; the penalty coefficient γ controls the sensitivity of the reduction, and the closer γ is to 1, the stronger the penalty.

[0236] S340, the corrected confidence level is used as the confidence level output by the third filtering step.

[0237] The corrected confidence level is used as the final output confidence level of the third filtering step, and is then used for subsequent classification labeling and feedback steps.

[0238] Through the aforementioned confidence correction mechanism, the technical solution of this invention can proactively reduce the confidence level when there are discrepancies between models, avoiding outputting overly optimistic judgment results when the results are inconsistent, thereby further improving the reliability and robustness of the overall classification system.

[0239] This invention achieves intelligent processing of complex semantics and boundary cases by introducing a large language model. Practice shows that after this step, the matching rate between entities and target classifications can be improved to over 90%, which is a significant improvement compared to the second filtering step based solely on rule matching (match rate 70%-85%).

[0240] Meanwhile, the cost control in this step is also quite reasonable. The cost of a single AI analysis (including multiple model calls) is usually between a few cents and a few dimes. Considering that this step is only performed on entities that have passed the first two rounds of rule screening (usually accounting for 20%-40% of the total), the overall computational cost is controllable, achieving a good balance between high accuracy and reasonable cost.

[0241] S400, Classification labeling and feedback step: Based on the comparison result between the confidence level output by the third filtering step and the preset classification threshold, determine whether the entity belongs to the target classification; label the entity that is determined to belong to the target classification with the corresponding classification label, and feed back the feature data of the entity and the judgment result to the feature library for self-optimization update of the feature library.

[0242] The classification labeling and feedback step is the final step of the entire method. It is responsible for determining whether an entity belongs to the target category based on the confidence level output by the third filtering step, classifying and labeling entities that are determined to belong to the category, and feeding the determination results back to the feature library for subsequent optimization and updates.

[0243] To achieve the orderly organization and management of entities, this invention pre-establishes a unified classification label system. Each label corresponds to a specific target category and can be designed as a hierarchical structure to support classification needs at different granularities. For example, a first-level label is "Energy and Electricity," a second-level label is "New Energy," and a third-level label is "Photovoltaic Power Generation."

[0244] Each tag contains the following attributes:

[0245] Tag encoding: A unique identifier for a tag;

[0246] Tag Name: The Chinese name of the tag;

[0247] Tag Description: The business definition and scope of the tag;

[0248] Parent tag: The encoding of the parent tag, used to establish hierarchical relationships;

[0249] This tagging system can be dynamically expanded and maintained according to business needs, and supports operations such as adding, modifying, and disabling tags.

[0250] This step first obtains the confidence score output from the third filtering step (such as the confidence score after multi-model fusion and correction), and compares it with a preset classification threshold. Based on the comparison result, the entity's classification is determined. To improve the reliability of the judgment result, this invention adopts a dual-threshold verification mechanism, dividing the confidence score into three intervals, each corresponding to a different processing method.

[0251] The specific implementation steps are as follows:

[0252] S410, compare the confidence level output by the third filtering step with the first preset classification threshold and the second preset classification threshold respectively.

[0253] The first preset classification threshold is less than the second preset classification threshold, and both thresholds range from 0 to 1. The first preset classification threshold is used to determine whether an entity may belong to the target category, and the second preset classification threshold is used to determine whether an entity can be definitively identified as belonging to the target category.

[0254] S411, when the confidence level is greater than or equal to the second preset classification threshold, it indicates that the model has a high degree of confidence in determining that the entity belongs to the target category. At this time, the entity is directly determined to belong to the target category, and a classification labeling operation is performed. The specific process of classification labeling includes:

[0255] Based on the definition of the target category, determine the corresponding label code and name;

[0256] Associate entity information with tag information to generate classification tag records;

[0257] Write the tagging results into the entity tag database to complete the tagging operation.

[0258] S412, when the confidence level is greater than or equal to the first preset classification threshold and less than the second preset classification threshold, it indicates that the model's judgment result has a certain degree of credibility, but has not yet reached the level of direct certainty. At this time, the entity and its judgment result are marked as pending review, and a manual review task is generated. Business personnel conduct manual review based on the entity information and the model's judgment reasons, and ultimately determine whether the entity belongs to the target category based on the review results.

[0259] If the verification confirms that it belongs to the category, then perform the classification labeling operation in S411;

[0260] If the review confirms that it does not belong to the category, the process ends without any marking.

[0261] S413, when the confidence level is less than the first preset classification threshold, it indicates that the reliability of the model's judgment result is low, and the entity is very unlikely to belong to the target category. At this time, it is directly determined that the entity does not belong to the target category, the process ends, and no marking is performed.

[0262] Through the aforementioned dual-threshold verification mechanism, this invention achieves an organic combination of automated processing and manual review: high-confidence results are processed fully automatically, medium-confidence results are subject to manual review to ensure accuracy, and low-confidence results are quickly eliminated. This mechanism minimizes the workload of manual review while ensuring overall accuracy. Another important function of the classification labeling and feedback steps is to feed the judgment results back to the feature library for self-optimization and updating. The feedback data comes from the following channels:

[0263] The high confidence judgment result output by the third filtering step: When the confidence level output by the third filtering step is higher than the preset high confidence threshold (e.g., the second preset classification threshold), the multi-dimensional feature data of the entity and its judgment result are used as valid feedback.

[0264] Manual review results: For entities whose classification is confirmed to be correct or incorrect after manual review, their data is also included in the feedback;

[0265] Periodic sampling verification: Sampling verification is performed on the marked entities, and the verification results are used as feedback.

[0266] Based on the feedback data above, the self-optimization update of the feature library includes at least one of the following operations:

[0267] S420, based on the feedback judgment result, extract new feature words from the multi-dimensional feature data of the corresponding entity and add them to the feature library.

[0268] Specifically, based on the feedback results, natural language processing is performed on the multi-dimensional feature data of entities confirmed to belong to the target category to extract frequently occurring feature words that are not yet included in the feature library. After analysis and confirmation, these new feature words are added to the feature library and initial weights are assigned to them.

[0269] S421, Based on the feedback judgment result, calculate the relevance index of existing feature words in the feature library, and remove feature words whose relevance is lower than the preset threshold.

[0270] The relevance index reflects the degree of association between feature words and the target category, and can be calculated in the following way:

[0271] R(k)=(Nc(k) / Nc)-(Ni(k) / Ni).

[0272] in:

[0273] Nc(k) represents the number of entities containing feature word k among all entities correctly classified as the target category during the feedback period; Nc represents the total number of entities correctly classified as the target category during the feedback period.

[0274] Ni(k) represents the number of entities containing feature word k among all entities that are incorrectly classified as the target category (i.e., entities that are classified as belonging to the target category but do not actually belong to it) within the feedback period.

[0275] Ni represents the total number of entities that were incorrectly classified as the target category during the feedback period.

[0276] When the relevance index R(k) of a feature word is lower than the preset removal threshold, it indicates that the association between the feature word and the target classification has been significantly weakened, and the feature word is removed from the feature library.

[0277] S422, Based on the feedback judgment results, adjust the weights of existing feature words in the feature library.

[0278] The weights of existing feature words in the feature library are dynamically adjusted according to the following formula:

[0279] w(k) t+1 =w(k) t +η×△(k)

[0280] Where, w(k) t w(k) represents the weight of feature word k at time t. t+1η represents the updated weight of feature word k at the next time step t+1. η is a preset learning rate parameter used to control the step size of weight adjustment. The value range is usually 0 < η ≤ 1, and can be set according to actual needs (e.g., η = 0.1). △(k) is the change in relevance of feature word k calculated based on feedback data. It can be the relevance index R(k) in S421, or its normalized value.

[0281] The formula means that the weight of a feature word is iteratively adjusted based on its relevance changes in the feedback data. If the frequency of a feature word in correct cases is higher than its frequency in incorrect cases (i.e., Δ(k) > 0), its weight is increased; otherwise, its weight is decreased.

[0282] To prevent weights from growing or decaying indefinitely, constraints can be applied to the updated weights:

[0283] w(k) t+1 =min(max(w(k) t+1 w min ), w max )

[0284] Among them, w min and w max These are the preset minimum and maximum weights (e.g., w). min =0, w max =1), ensuring that the weight is always within a reasonable range.

[0285] The feature library can be updated periodically in batches or online in real time, depending on the system load and business requirements.

[0286] Periodic batch updates: At fixed time intervals (such as weekly or monthly), relevance indicators are calculated uniformly based on the accumulated feedback data within that period, and the feature library is adjusted in batches. This method is computationally efficient and suitable for scenarios with large amounts of data.

[0287] Online real-time updates: Whenever a high-confidence feedback data is received, an incremental update of the feature library is immediately triggered. This method is highly time-sensitive and suitable for scenarios with high real-time requirements.

[0288] Hybrid update strategy: Periodic batch updates are used daily, while real-time update channels are set up for key feedback data, balancing efficiency and timeliness.

[0289] The classification and feedback process automates the entire process from confidence level determination to label assignment. The processing time for a single labeling operation is typically in the millisecond range, meeting the performance requirements for large-scale entity labeling. Simultaneously, the feedback mechanism enables self-optimization and updating of the feature library, providing continuous learning capabilities. This allows it to adapt to changes in the target classification context and data distribution shifts, maintaining high classification accuracy over the long term.

[0290] To support the rapid migration and expansion of the method of this invention in different application scenarios, a rapid adaptation mechanism for new target classifications is further provided. When it is necessary to identify a new target classification that is not predefined, there is no need to build a feature library from scratch. Instead, the initial feature library is automatically generated through an entity-based capability knowledge graph.

[0291] The rapid adaptation mechanism for new target classifications includes the following steps:

[0292] S500 pre-builds a knowledge graph of the entity's basic capabilities.

[0293] The knowledge graph stores basic capability units abstracted from entity feature data and their relationships. The construction process of this knowledge graph is as follows:

[0294] Basic capability unit extraction: Deep analysis is performed on a large amount of multi-dimensional feature data (including semantic features, relational features, attribute features, etc.) of existing entities to extract basic capability units that can independently describe a specific aspect of an entity's business or technical capabilities. Basic capability units are more abstract and structured knowledge representations than feature words, such as: "high-power power electronic conversion technology," "precision machining capability," "energy storage battery management system development," "new energy bus manufacturing," and "charging facility operation services."

[0295] Capability Unit Vectorization: Each basic capability unit is converted into a vector representation through a text embedding model for similarity calculation and retrieval.

[0296] Capability Unit Relationship Construction: Analyze the semantic, technological, and upstream / downstream industry chain relationships between different basic capability units to establish a relationship network between them. For example, "energy storage battery management system development" and "battery pack integration" are technologically related, and "photovoltaic module manufacturing" and "power station operation and maintenance" are upstream / downstream industry chain related.

[0297] Knowledge graph storage: Basic capability units, their vector representations, and relationships are stored as knowledge graphs, supporting subsequent fast retrieval and combination.

[0298] S510, when a request to identify a new target category is received, the definition information of the new target category is obtained.

[0299] When a request to identify a new target category is received, the definition information of the new target category provided by the user or business system is obtained, including at least:

[0300] Category name (e.g., "Hydrogen Energy Storage and Transportation Equipment Manufacturer");

[0301] Category Description (detailing the business scope, core technologies, main products, etc. of this category);

[0302] Optional keywords or example entities.

[0303] S520, based on the defined information, retrieve and combine relevant basic capability units in the knowledge graph to automatically generate an initial feature library for the new target classification, specifically including:

[0304] S521, Parse the definition information and generate the retrieval vector.

[0305] Natural language processing is performed on the classification definition information to extract key descriptions, and a retrieval vector Vtarget is generated using the same text embedding model as in step 4.6.1.

[0306] S522, Retrieve relevant basic capability units from the knowledge graph.

[0307] Using the retrieval vector Vector as the query, a vector similarity search is performed in the entity basic capability knowledge graph to recall the M basic capability units most relevant to the target category (M is a preset number, such as 50). At the same time, combined with the association relationships in the knowledge graph, other capability units that are highly related to these basic capability units are recalled.

[0308] S523, combining basic capability units to generate an initial feature library.

[0309] The system analyzes multiple basic capability units retrieved from the data and extracts key feature words that constitute these capability units. Specific methods include:

[0310] Extract core keywords from the description text of each basic capability unit using methods such as TF-IDF or TextRank;

[0311] The extracted keywords are merged and deduplicated, and then sorted according to their frequency and importance in the recall capability unit;

[0312] The top-ranked keywords are selected as candidate feature words.

[0313] Based on the correlation between candidate feature words and the target classification, they are initially divided into core features and peripheral features:

[0314] Keywords extracted from the capability unit with the highest similarity to the retrieval vector, as well as keywords strongly related to the target category name, are given priority for inclusion in the core feature candidate list.

[0315] Other keywords are included as edge feature candidates.

[0316] To set initial weights for each feature word, one of the following methods can be used:

[0317] Based on the average weight of the feature word in the recall capability unit;

[0318] Based on the similarity between the feature word and the retrieval vector;

[0319] Set them to the default values ​​(e.g., core features 0.8, edge features 0.3).

[0320] This ultimately forms the initial feature library F for classifying the new target. new ={F new core ,F new edge} contains feature words and their initial weights.

[0321] S524, Manual Review and Optimization

[0322] The automatically generated initial feature library can be submitted to business experts for rapid review and fine-tuning to ensure its accuracy and applicability. Once approved, the feature library can be used for matching analysis in the second filtering step.

[0323] Through the aforementioned rapid adaptation mechanism for new target classification, this invention achieves the following technical effects:

[0324] Significantly shortens the time to launch new categories: Traditional methods require days or even weeks of manual feature engineering, while this method can automatically generate an initial feature library in minutes, significantly improving response speed.

[0325] Make full use of existing knowledge accumulation: Based on the pre-built entity basic capability knowledge graph, cross-class knowledge transfer and reuse are realized, avoiding duplication of work.

[0326] Ensuring the quality of the initial feature library: Through semantic retrieval and association expansion of the knowledge graph, the generated initial feature library has high relevance and coverage, laying a good foundation for subsequent rule matching and large model judgment.

[0327] Support for continuous evolution: After the initial feature library is put into use, it can be further optimized through the aforementioned feedback mechanism to achieve a complete closed loop from rapid startup to continuous optimization.

[0328] Based on the same inventive concept, another embodiment of the present invention provides an entity classification system based on a multi-layer filtering mechanism, comprising:

[0329] The first filtering module is used to obtain the basic attribute data of the entities to be processed, filter the basic attribute data based on the dynamically updated exclusion rule set, and remove entities that do not meet the basic attribute threshold corresponding to the target classification to obtain the first candidate entity set.

[0330] The second filtering module is used to obtain multi-dimensional feature data of each entity in the first candidate entity set, assign weights to the multi-dimensional feature data, perform hierarchical matching analysis with the feature library pre-configured for the target classification, and filter out the second candidate entity set according to the matching results; the multi-dimensional feature data includes at least semantic features and association features.

[0331] The third filtering module is used to perform feature fusion on the multi-dimensional feature data of each entity in the second candidate entity set, input the fused feature data and the definition information of the target classification into the domain-adapted pre-trained large language model, use the model to perform semantic reasoning and judgment, and output the confidence score.

[0332] The classification labeling and feedback module determines whether an entity belongs to the target category based on the comparison result between the confidence level output by the third filtering module and the preset classification threshold; it labels the entity that is determined to belong to the target category with the corresponding classification label, and feeds back the feature data of the entity and the judgment result to the feature library for self-optimization and updating of the feature library.

[0333] It should be noted that the system embodiments and the corresponding method embodiments are based on the same inventive concept. Therefore, the technical features, implementation details and technical effects of the method embodiments are also applicable to the system embodiments, and will not be repeated here.

[0334] Those skilled in the art will understand that the above-described processing units can be implemented by hardware, software, firmware, or any combination thereof. When implemented in software, the system may further include a memory storing a computer program, which, when executed by the processing unit, performs the operations described above.

[0335] The multi-level entity classification method and system proposed in this invention, based on a large language model and rule engine, can be widely applied to various scenarios requiring accurate classification and annotation of massive amounts of entity data, demonstrating significant versatility and adaptability. For example:

[0336] Scenario 1: Industry Chain Analysis and Investment Promotion Target Identification

[0337] When conducting supply chain analysis, industry research institutions and investment promotion departments can use this invention to identify upstream and downstream enterprises in specific supply chain segments. For example, in the analysis of the new energy vehicle supply chain, "vehicle manufacturing," "power batteries," "electric drive systems," and "charging facilities" can be set as target categories. Through a multi-layered filtering mechanism, core enterprises belonging to each segment can be accurately identified from a massive number of companies and categorized accordingly. Investment promotion departments can then quickly locate potential investment targets for missing segments of the supply chain based on the identification results, improving the accuracy and efficiency of investment attraction.

[0338] Scenario 2: Corporate Risk Assessment by Financial Institutions

[0339] In scenarios where banks, investment institutions, and other financial institutions conduct credit assessments and due diligence on enterprises, this invention can be used to identify the industry category and technological field to which an enterprise belongs. For example, in green finance, "energy conservation and environmental protection," "clean energy," and "green building" can be set as target categories. By analyzing multi-dimensional data such as the enterprise's business scope, patented technologies, and bidding projects, enterprises that meet the conditions for green credit can be accurately identified. This provides financial institutions with an objective and efficient technical means for credit approval, differentiated pricing, and risk management.

[0340] Scenario 3: Talent Acquisition by Research Institutions

[0341] In research institutions and universities conducting talent acquisition and team evaluation, this invention can target technical fields such as "artificial intelligence," "quantum computing," and "biomedicine," and perform multi-dimensional analysis of researchers' papers, patents, project experience, and other data to identify core talents and teams in specific technical fields. This provides precise data support for research project applications, talent recruitment, and industry-academia-research collaborations.

[0342] Scenario 4: Product categorization and recommendations on e-commerce platforms

[0343] In e-commerce platforms that automatically classify and label massive amounts of goods, this invention sets each level of the product classification system as the target category. Based on multi-dimensional data such as product titles, descriptions, attributes, and user reviews, it achieves automatic product categorization through a multi-layered filtering mechanism. For products with vague descriptions or incomplete attributes, it uses intelligent reasoning based on a large model to supplement the judgment, significantly improving the accuracy and automation level of product classification, laying the foundation for accurate recommendations and search optimization.

[0344] Scenario 5: Patent Classification by Intellectual Property Service Agencies

[0345] In scenarios where intellectual property service agencies classify and manage patent data, this invention can set the International Patent Classification (IPC) or the Strategic Emerging Industries Classification as the target classification. Based on multi-dimensional data such as patent name, abstract, claims, and citation relationships, it can automatically classify patents. For patents with overlapping technologies or ambiguous classifications, intelligent reasoning through a large model assists in classification, improving the efficiency and consistency of patent classification.

[0346] In summary, this invention can be widely applied in various fields such as industrial research, financial risk control, scientific research management, e-commerce operations, and intellectual property services. It can effectively solve the problems of low efficiency, low accuracy, and insufficient utilization of multi-source data in the classification of massive entity data, and has broad application prospects and market value.

[0347] Taking a city's New Energy Vehicle Industry Promotion Center as an example, the center needs to conduct a comprehensive survey and classification management of enterprises engaged in new energy vehicle-related businesses throughout the city in order to accurately formulate industry support policies and optimize the industrial chain layout. The data to be processed includes approximately 500,000 enterprises, with data sources including business registration information, patent databases, and bidding platforms.

[0348] The first filtering step (coarse screening) involves obtaining business registration information for 500,000 companies and quickly screening them based on a dynamically updated set of exclusion rules. Approximately 250,000 obviously irrelevant companies are filtered out using common exclusion terms such as "catering," "real estate," and "trade." Simultaneously, foreign-invested enterprises and deregistered companies are excluded, leaving approximately 150,000 companies to proceed to the next stage.

[0349] The second filtering step (refined screening) involves analyzing the multi-dimensional feature data of 150,000 companies within a pre-configured feature library for target categories such as "new energy vehicle manufacturers," "power battery manufacturers," and "charging facility operators." For example, for the "power battery manufacturers" category, core features include "lithium battery," "cathode material," and "battery pack," while peripheral features include "energy density," "thermal management," and "BMS." Through hierarchical matching analysis, approximately 30,000 companies highly relevant to the target categories are selected from the 150,000 companies.

[0350] The third filtering step (AI analysis): For complex enterprises with ambiguous business descriptions and overlapping technical fields, the enterprise information and classification definitions are constructed into structured prompts and input into a large language model for intelligent reasoning. For example, for an enterprise whose business scope includes "new energy technology development", whose patents involve "battery management system", and whose bidding records are for "charging station construction projects", the model comprehensively judges that it belongs to "charging facility operation enterprise" and outputs a high confidence score of 0.92.

[0351] Classification and Feedback Steps: Companies with a confidence score higher than 0.9 were directly labeled with the corresponding classification tag; companies with a confidence score between 0.7 and 0.9 entered a manual review process; and companies with a confidence score lower than 0.7 were excluded. Ultimately, approximately 28,000 new energy vehicle-related companies were successfully identified, covering the entire industry chain, including vehicle manufacturing, power batteries, electric drive systems, and charging infrastructure. The identification results were fed back to the feature library for subsequent weight optimization and feature updates.

[0352] Compared to traditional manual review methods (which take months), this invention can classify and identify 500,000 companies within hours, with an accuracy rate exceeding 92%. Based on the identification results, the industry promotion center has precisely formulated support policies for different links in the industrial chain, effectively promoting the high-quality development of the local new energy vehicle industry.

[0353] In summary, the multi-level entity classification method and system based on a large language model and rule engine proposed in this invention has the following significant technical advantages compared with existing technologies:

[0354] 1. Improved screening efficiency

[0355] This invention improves the efficiency of entity screening through a multi-level progressive filtering architecture.

[0356] In terms of processing speed, traditional manual processing methods take an average of several minutes to tens of minutes to review a single entity; the automated processing method of this invention only takes a few seconds to a few minutes to process a single entity, improving efficiency by tens to hundreds of times.

[0357] In terms of large-scale data processing, this invention can process massive amounts of entity data, ranging from millions to tens of millions, with a daily processing capacity of hundreds of thousands of entities, meeting the business needs of large-scale data screening.

[0358] Regarding real-time response, the processing flow of this invention is highly automated, completing the process from input entity data to output classification and labeling results within seconds, supporting real-time classification and labeling business scenarios. 2. Improved recognition accuracy.

[0359] This invention improves the accuracy of entity recognition through multi-dimensional data fusion analysis and AI intelligent reasoning.

[0360] In terms of overall accuracy, according to test data, the target classification and recognition accuracy of this invention reaches over 90%, which is an improvement compared to the 60% to 80% accuracy of traditional keyword matching methods.

[0361] Regarding the false alarm rate, through a multi-level filtering mechanism, a large number of irrelevant entities are filtered out in advance, and the false alarm rate of the final output result is effectively controlled.

[0362] Regarding the false negative rate, the large-scale intelligent reasoning in the third filtering step can handle complex business semantics and boundary cases, reducing the risk of false negatives. Entities using emerging terms or special business descriptions can also be accurately identified. 3. Fully Leveraging the Value of Multi-Dimensional Information

[0363] This invention fully explores and utilizes the value of multi-dimensional information about entities.

[0364] Comprehensive information coverage: It comprehensively analyzes information from multiple dimensions such as semantic features, association features, and attribute features of entities, covering aspects such as entity name features, business description, technical capabilities, and market performance.

[0365] Information complementarity enhancement: Information from different dimensions complements and verifies each other. For example, patent data verifies the entity's technical direction, and bidding data verifies the entity's business performance. Comprehensive analysis of multi-dimensional information is more accurate and reliable than analysis of a single dimension.

[0366] Maximizing information value: Through feature matching and intelligent analysis of large models, scattered raw data is transformed into valuable classification judgments, maximizing the utilization value of the data.

[0367] 4. High level of intelligence

[0368] This invention introduces a large language model into the field of entity classification, thereby improving the level of intelligence.

[0369] In terms of semantic understanding capabilities, large models can handle complex business semantics and fuzzy business descriptions, which is difficult to achieve with traditional rule matching.

[0370] In terms of knowledge transfer capability, large models have extensive knowledge reserves, enabling them to transfer analytical methods to new target classifications, thus possessing a certain degree of versatility.

[0371] In terms of continuous learning ability, the system can gradually improve the accuracy of judgment by continuously optimizing the design of prompt words, multi-model fusion mechanism and feature library feedback updates, and has the ability to continuously improve.

[0372] 5. Scalability and flexibility

[0373] This invention has good scalability and flexibility.

[0374] Rapid adaptation to new target categories: When adding a new target category, an initial feature library can be automatically generated through the entity basic capability knowledge graph, eliminating the need to build from scratch and enabling rapid deployment.

[0375] Flexible rule adjustment: The thresholds, weights and other parameters of each stage can be adjusted through configuration to adapt to different business needs and data characteristics.

[0376] Modular design: Each filtering step is relatively independent, and a certain stage can be enabled or skipped as needed, or the implementation method of a certain stage can be replaced.

[0377] 6. Cost-effectiveness

[0378] This invention achieves high accuracy while effectively controlling costs.

[0379] Reduced labor costs: Compared to purely manual review, the automated processing of this invention reduces manual input and lowers labor costs.

[0380] Computational costs are controllable: By using a multi-level filtering architecture, AI inference is concentrated in the final stage, and large-scale model analysis is only performed on entities that have passed the first two rounds of screening, thus controlling computational costs.

[0381] Maintenance costs: Based on configuration-driven design and a self-optimizing update mechanism for the feature library, the system does not require a lot of development work for maintenance and optimization, and maintenance costs are controlled.

[0382] This invention improves the efficiency and accuracy of entity classification through a four-stage filtering process: coarse screening, fine screening, AI analysis, and classification labeling. This method combines the efficiency of a rule engine with the intelligence of a large model; it utilizes multi-dimensional information about entities through multi-dimensional data fusion analysis; and through a scalable architecture design and a self-optimizing feedback mechanism, it supports adaptation to different target classifications and continuously optimizes classification accuracy.

[0383] Compared with existing technologies, this invention improves screening efficiency, recognition accuracy, intelligence level, expansion flexibility, and cost-effectiveness. It can improve the problems of existing technologies, such as difficulty in balancing efficiency and accuracy, single information dimension, insufficient intelligence, and poor scalability, and has good application prospects.

[0384] This invention also provides an electronic device, including: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being configured to perform the method described in this invention.

[0385] This invention also provides a computer-readable storage medium storing computer-executable instructions for performing the methods described in this invention.

[0386] It should be understood that the various forms of processes shown above can be used to reorder, add, or delete steps. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this invention can be achieved, and this is not limited herein.

[0387] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A multi-level entity classification method based on a large language model and a rule engine, characterized in that, The method includes the following steps: First filtering step: Obtain the basic attribute data of the entities to be processed, filter the basic attribute data based on the dynamically updated exclusion rule set, remove entities that do not meet the basic attribute threshold corresponding to the target classification, and obtain the first candidate entity set; The second filtering step is to obtain multi-dimensional feature data for each entity in the first candidate entity set, assign weights to the multi-dimensional feature data, perform hierarchical matching analysis with the feature library pre-configured for the target classification, and filter out the second candidate entity set based on the matching results; the multi-dimensional feature data includes at least semantic features and association features. The third filtering step is to perform feature fusion on the multi-dimensional feature data of each entity in the second candidate entity set, input the fused feature data and the definition information of the target classification into the domain-adapted pre-trained large language model, use the pre-trained large language model to perform semantic reasoning and judgment, and output the confidence score. Classification labeling and feedback steps: Based on the comparison result between the confidence level output by the third filtering step and the preset classification threshold, determine whether the entity belongs to the target classification; Entities identified as belonging to the target category are labeled with corresponding classification tags, and the feature data of the entity and the judgment result are fed back to the feature library for self-optimization and updating of the feature library.

2. The method of claim 1, wherein, The dynamically updated exclusion rule set includes at least one of the following rules: A general keyword filtering rule based on entity name is used to filter entities whose names contain preset non-target category keywords; Exclusion rules based on entity type identifiers are used to filter entities that do not meet the basic type requirements of the preset target classification based on the entity's legal type identifier. Exclusion rules based on entity survival status identifiers are used to filter invalid or abnormal entities according to their legal survival status identifiers.

3. The method of claim 1, wherein, The multi-dimensional feature data also includes entity attribute features; The semantic features include textual semantic information extracted from entity name data, brief description data, and business scope description data; The associated features include at least one of the entity's intellectual property association data and market activity association data. The intellectual property association data includes patent association data and trademark association data, and the market activity association data includes bidding association data and project application association data. The entity attribute features include at least one of the entity's size identification data, geographical identification data, and establishment duration data.

4. The method of claim 3, wherein, The matching analysis in the second filtering step is performed according to the following steps: The multi-dimensional feature data is matched with the core features in the feature library at the first level, and the first matching degree is calculated. If the first matching degree is greater than the first preset threshold, then the multi-dimensional feature data is further matched with the edge features in the feature library at the second level to calculate the edge feature matching degree. Calculate the comprehensive matching degree based on the first matching degree and the edge feature matching degree; The overall matching degree is compared with a preset matching similarity threshold, and a second candidate entity set is selected based on the comparison result.

5. The method of claim 4, wherein, The overall matching degree is calculated according to the following formula: S = a x S core + b x S edge ; wherein S core represents the first matching degree, S edge represents the edge feature matching degree, and α and β are preset weight coefficients of the core feature and the edge feature, respectively, and satisfy α+β=1.

6. The method of claim 1, wherein, The self-optimization update of the feature library includes at least one of the following operations: Based on the feedback judgment results, new feature words are extracted from the multi-dimensional feature data of the corresponding entity and added to the feature library; Based on the feedback judgment results, the relevance index of existing feature words in the feature library is calculated, and feature words with relevance below the preset threshold are removed. Based on the feedback results, the weights of existing feature words in the feature library are adjusted.

7. The method of claim 1, wherein, The semantic reasoning determination in the third filtering step specifically includes: The fused feature data and the definition information of the target classification are constructed into structured prompt information according to a preset template; The structured prompt information is input into multiple pre-trained large language models with different architectures to obtain multiple preliminary judgment results and corresponding initial confidence scores; The multiple preliminary judgment results are analyzed for consistency using a preset fusion model to generate a fused confidence score, which is then used as the confidence score output by the third filtering step.

8. The method of claim 7, wherein, The third filtering step also includes a confidence correction mechanism: Calculate the consistency degree characterization value of the multiple preliminary judgment results, whereby the consistency degree characterization value represents the degree of consistency between the multiple model judgment results; When the consistency level characterization value is less than the preset consistency threshold, the fused confidence level is penalized and corrected to obtain the corrected confidence level. The corrected confidence level is used as the confidence level output by the third filtering step.

9. The method according to claim 1, characterized in that, The step of determining whether an entity belongs to the target category based on the comparison result between the confidence level output by the third filtering step and the preset classification threshold specifically includes: The confidence level output by the third filtering step is compared with the first preset classification threshold and the second preset classification threshold, wherein the first preset classification threshold is less than the second preset classification threshold. When the confidence level is greater than or equal to the second preset classification threshold, the entity is directly determined to belong to the target classification, and classification labeling is performed. When the confidence level is greater than or equal to the first preset classification threshold and less than the second preset classification threshold, the entity and its judgment result are marked as pending review, a manual review task is generated, and it is determined whether it belongs to the target category based on the review result. When the confidence level is less than the first preset classification threshold, it is determined that the entity does not belong to the target classification.

10. An entity classification system based on multi-layer filtering mechanism, characterized in that, include: The first filtering module is used to obtain the basic attribute data of the entities to be processed, filter the basic attribute data based on the dynamically updated exclusion rule set, and remove entities that do not meet the basic attribute threshold corresponding to the target classification to obtain the first candidate entity set. The second filtering module is used to acquire multi-dimensional feature data of each entity in the first candidate entity set, assign weights to the multi-dimensional feature data, perform hierarchical matching analysis with the feature library pre-configured for the target classification, and filter out the second candidate entity set according to the matching results; the multi-dimensional feature data includes at least semantic features and association features. The third filtering module is used to perform feature fusion on the multi-dimensional feature data of each entity in the second candidate entity set, input the fused feature data and the definition information of the target classification into the domain-adapted pre-trained large language model, use the model to perform semantic reasoning and judgment, and output the confidence score. The classification labeling and feedback module determines whether an entity belongs to the target category based on the comparison result between the confidence level output by the third filtering module and the preset classification threshold. Entities identified as belonging to the target category are labeled with corresponding classification tags, and the feature data of the entity and the judgment result are fed back to the feature library for self-optimization and updating of the feature library.

Images