Enterprise information intelligent recommendation method and device, computer device and storage medium

Abstract

This invention discloses a method, apparatus, computer equipment, and storage medium for intelligent recommendation of enterprise information. The method includes: acquiring relevant enterprise data and creating an enterprise feature map containing a multi-level classification system; establishing a product coding table based on the enterprise feature map; acquiring enterprise information and user preferences from different sources, constructing a data model, and recording historical dialogues and current input queries; intelligently rewriting the current input query and identifying the query object and its business relationships; converting the rewriting intent into standardized query conditions; converting natural language descriptions into structured JSON format; generating specific query statements based on structured information and executing the query to obtain query results; evaluating and ranking the query results; and displaying detailed information about the recommended enterprises. By implementing the method of this invention, a precise conversion from natural language queries to structured enterprise information retrieval can be achieved, significantly improving the accuracy and intelligence level of enterprise recommendation systems.

Description

Technical Field

[0002] This invention relates to enterprise information retrieval and recommendation, and more specifically to intelligent enterprise information recommendation methods, devices, computer equipment, and storage media. Background Technology

[0004] Enterprise information retrieval and recommendation systems intelligently process and analyze massive amounts of enterprise data to accurately match user needs, thereby significantly improving work efficiency and decision-making quality, while also promoting information sharing and business cooperation among enterprises. Existing technologies such as Text2SQL and Elasticsearch-based enterprise information retrieval systems, as well as the open-source project Text2ESQuery, have achieved the conversion from natural language to structured query language. However, in the application scenarios of enterprise information recommendation systems, these technologies exhibit significant limitations, especially when handling complex business logic and multi-dimensional conditional filtering. Query results often deviate from the user's true intent, requiring manual intervention for correction. Text2ESQuery and Text2SQL perform well in general query scenarios, but lack in-depth optimization for specific business domains, making it difficult to meet the automation needs of enterprise-level applications.

[0005] While Elasticsearch is widely used as a distributed search and analytics engine and is beginning to integrate AI search capabilities to support semantic search and re-ranking models, it still falls short in areas such as anomaly handling, result optimization, and personalized recommendations, lacking a complete natural language to query language conversion mechanism. Overall, existing technologies face challenges in handling query complexity, domain adaptability, and specialization when dealing with enterprise information recommendation systems, limiting their effective application in commercial environments.

[0006] Therefore, it is necessary to design a new method to achieve accurate conversion from natural language queries to structured enterprise information retrieval, which will significantly improve the accuracy and intelligence of enterprise recommendation systems and solve the problems existing in current technologies. Summary of the Invention

[0008] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method, apparatus, computer equipment and storage medium for intelligent recommendation of enterprise information.

[0009] To achieve the above objectives, the present invention adopts the following technical solution: an intelligent recommendation method for enterprise information, comprising:

[0010] Acquire relevant enterprise data and create an enterprise feature map containing a multi-level classification system for enterprises;

[0011] A commodity coding table is established based on the enterprise feature map;

[0012] Acquire enterprise information and user preferences from different sources, build data models, and record historical conversations and current input queries;

[0013] The current input query is intelligently rewritten, and the query object and its business relationship are identified to obtain the rewriting intent;

[0014] Based on the enterprise feature map and the commodity coding table, the rewriting intent is transformed into standardized query conditions.

[0015] Based on the standardized query conditions, the natural language description is converted into a structured JSON format to obtain structured information;

[0016] Based on the structured information, a specific query statement is generated and executed to obtain the query results;

[0017] The query results are evaluated and sorted to obtain a ranking result based on scores.

[0018] The recommended companies are displayed based on their scores, ranked from highest to lowest.

[0019] The further technical solution is as follows: the acquisition of enterprise-related data and the creation of an enterprise feature map containing a multi-level classification system includes:

[0020] Acquire relevant enterprise data, establish a top-level classification of enterprise characteristics, and construct a systematic classification structure;

[0021] Based on each category in the classification structure, specific enterprise attribute dimensions are clearly defined, and semantic definitions and data types are set for each feature;

[0022] Create a thesaurus for feature terms;

[0023] A standard set of enumerated values ​​is developed for discrete features, and synonym processing is performed on the enumerated values ​​to identify and define the mutual exclusion relationships between feature values ​​in order to obtain the enterprise feature map.

[0024] The further technical solution is as follows: the establishment of a commodity coding table based on the enterprise feature map includes:

[0025] Based on the invoice data and expert experience in the enterprise feature map, a hierarchical commodity coding table containing multi-dimensional commodity information is constructed.

[0026] The further technical solution is as follows: The intelligent rewriting of the current input query and the identification of the query object and its business relationships to obtain the rewriting intent include:

[0027] Analyze the current input query and historical dialogue information to determine the query category of the current input query, so as to obtain the first analysis result;

[0028] Identify the core object of the current input query, and determine the business relationship between the user and the core object based on user information to obtain the second analysis result;

[0029] The relative and vague time expressions in the current input query are converted into absolute and specific time ranges or specific dates to obtain the third analysis result;

[0030] The first analysis result, the second analysis result, and the third analysis result are integrated to generate the rewriting intent.

[0031] The further technical solution is as follows: the step of converting the rewriting intent into standardized query conditions based on the enterprise feature map and the commodity coding table includes:

[0032] Based on the enterprise feature map and product coding table, the product and enterprise features involved in the rewriting intent are identified, matched with the corresponding entries in the database, and a corresponding query condition structure is generated.

[0033] The further technical solution is as follows: based on the enterprise feature map and product coding table, the product and enterprise features involved in the rewriting intent are identified, matched with the corresponding entries in the database, and a corresponding query condition structure is generated, including:

[0034] Combining the rewriting intent with the enterprise feature map and commodity coding table, prompt words are generated; the enterprise characteristics involved in the rewriting intent are identified and verified and supplemented through the enterprise feature map; the specific numerical values, text, or time values ​​in the current input query are obtained, and unit conversion and expression processing are performed; the comparison operators and special operation types in the current input query are analyzed and determined; the logical relationships and nested structures in the current input query are understood and constructed to obtain parsed information;

[0035] The parsed information is organized into a tree structure to ensure that the logical and conditional nodes correctly express the query intent, thus obtaining the query condition structure.

[0036] The further technical solution is as follows: generating a specific query statement based on the structured information and executing the query to obtain the query results includes:

[0037] The structured information is converted into an Elasticsearch query DSL according to predefined rules;

[0038] Based on the operator type of the node, the operator type is mapped to the corresponding content in the Elasticsearch query DSL. For non-equality comparison conditions, the value is excluded by nesting terms in bool.must_not. The match or wildcard query type is used to support flexible fuzzy search of text. The sort and size parameters are used to control the sorting method of the results and the number of returned results to meet the TopN / BottomN requirements. The logic tree is traversed and the ES query DSL is generated through a Python recursive function. At the same time, the requirements on the root node are processed to obtain the specific query statement.

[0039] The generated query statement is executed, and a list of companies that meet the criteria is obtained from the company profile table to obtain the query results.

[0040] This invention also provides an intelligent recommendation device for enterprise information, comprising:

[0041] The graph creation unit is used to acquire enterprise-related data and create an enterprise feature graph that includes a multi-level classification system for enterprises.

[0042] The coding table establishment unit is used to establish a commodity coding table based on the enterprise feature map.

[0043] The recording unit is used to acquire enterprise information and user preferences from different sources, build data models, and record historical dialogues and current input queries;

[0044] The rewriting unit is used to intelligently rewrite the current input query and identify the query object and its business relationship to obtain the rewriting intent;

[0045] The condition conversion unit is used to convert the rewriting intent into standardized query conditions based on the enterprise feature map and the commodity code table.

[0046] The format conversion unit is used to convert the natural language description into a structured JSON format based on the standardized query conditions to obtain structured information.

[0047] The query unit is used to generate a specific query statement based on the structured information and execute the query to obtain the query result;

[0048] A sorting unit is used to evaluate and sort the query results to obtain a ranking result based on scores.

[0049] The display unit is used to display detailed information about the recommended companies based on the ranking results of the scores.

[0050] The present invention also provides a computer device, the computer device including a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the above-described method.

[0051] The present invention also provides a storage medium storing a computer program that, when executed by a processor, implements the above-described method.

[0052] The beneficial effects of this invention compared to existing technologies are as follows: This invention creates an enterprise feature map containing a multi-level classification system of enterprises and a product coding table based on this map. Combined with enterprise information and user preference data from different sources, it constructs a data model and records historical dialogues and current query inputs. Then, it intelligently rewrites the current query, identifying the query object and its business relationships to clarify the rewriting intent, and transforms this intent into standardized query conditions based on the enterprise feature map and product coding table. Subsequently, these conditions are converted from natural language descriptions into structured JSON format information, generating specific query statements for execution. Finally, by evaluating and ranking the query results, it displays detailed information about the recommended enterprises, thereby achieving a precise conversion from natural language queries to structured enterprise information retrieval. This significantly improves the accuracy and intelligence level of the enterprise recommendation system, effectively solving the problems of low accuracy and the need for manual intervention in existing technologies under complex query conditions.

[0053] The present invention will be further described below with reference to the accompanying drawings and specific embodiments. Attached Figure Description

[0055] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0056] Figure 1 This is a schematic diagram illustrating an application scenario of the intelligent enterprise information recommendation method provided in this embodiment of the invention.

[0057] Figure 2 A flowchart illustrating the intelligent recommendation method for enterprise information provided in an embodiment of the present invention;

[0058] Figure 3 A schematic diagram of the enterprise feature classification system provided in an embodiment of the present invention;

[0059] Figure 4 A schematic block diagram of an intelligent enterprise information recommendation device provided in an embodiment of the present invention;

[0060] Figure 5 A schematic block diagram of a computer device provided for an embodiment of the present invention. Detailed Implementation

[0062] 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, not all, of the embodiments of the present invention. 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.

[0063] It should be understood that, when used in this specification and the appended claims, the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0064] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0065] It should also be further understood that the term "and / or" as used in this specification and the appended claims refers to any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0066] Please see Figure 1 and Figure 2 , Figure 1 This is a schematic diagram illustrating an application scenario of the intelligent enterprise information recommendation method provided in an embodiment of the present invention. Figure 2This is a schematic flowchart illustrating the intelligent enterprise information recommendation method provided in this embodiment of the invention. The method is applied to a server. The server interacts with the terminal, creating an enterprise feature map containing a multi-level classification system and a product coding table built upon it. Combining user preferences and historical dialogue data, it intelligently parses and rewrites natural language queries, transforming them into standardized query conditions. Subsequently, based on these conditions, the natural language description is converted into structured JSON format information, generating specific query statements for retrieval. Finally, detailed information about recommended enterprises is displayed according to the evaluation ranking. This method achieves accurate conversion from natural language queries to structured enterprise information retrieval, significantly improving the accuracy and intelligence of the enterprise recommendation system and effectively solving problems existing in the prior art. The entire process encompasses multiple steps, from raw data acquisition, feature definition, query intent recognition, query condition standardization, structured information generation, to specific query execution, ensuring the relevance and accuracy of the query results.

[0067] Figure 2 This is a flowchart illustrating the intelligent recommendation method for enterprise information provided in an embodiment of the present invention. Figure 2 As shown, the method includes the following steps S110 to S190.

[0068] S110. Obtain relevant enterprise data and create an enterprise feature map containing a multi-level classification system for enterprises.

[0069] In this embodiment, enterprise-related data refers to various information about an enterprise collected from multiple sources (such as publicly available government data, third-party business information providers, and the enterprise's official website). This information includes, but is not limited to, data on the company's basic information (such as name and registered address), operating status (such as operating revenue and profit), certification status (such as ISO certification and industry licenses), risk assessment (credit rating and legal proceedings), industry characteristics, and marketing models.

[0070] A corporate feature graph is a structured approach used to organize and represent the aforementioned corporate information. By constructing a multi-layered, multi-dimensional descriptive system, it enables the system to more accurately understand the corporate feature requirements in user queries. This graph not only helps improve the ability of natural language processing models to identify corporate characteristics but also provides users with a more intuitive and systematic perspective to understand and explore corporate information.

[0071] In one embodiment, step S110 described above may include steps S111 to S114.

[0072] S111. Obtain relevant enterprise data, establish a top-level classification of enterprise characteristics, and construct a systematic classification structure.

[0073] In this embodiment, this step mainly involves data collection and preliminary classification of the data. First, comprehensive information about the enterprise needs to be obtained from multiple channels; then, this information is divided into different categories based on business attributes, such as basic information, operating status, and qualifications / certifications. This process aims to establish a clear and systematic classification framework as the basis for subsequent detailed work.

[0074] S112. Based on each category in the classification structure, clarify the specific enterprise attribute dimensions, and set semantic definitions and data types for each feature.

[0075] In this embodiment, after determining the top-level classification, the next step is to delve into each classification to define the specific attribute dimensions in detail. For example, the "Basic Information" category might include registered capital, establishment date, etc. For each such feature item, an accurate semantic interpretation and corresponding data type restrictions need to be provided to facilitate subsequent data processing and query optimization.

[0076] S113. Create a thesaurus of feature terms.

[0077] In this embodiment, to enhance the system's flexibility and adaptability, this step requires building a thesaurus for each feature. This means that if a user uses terms different from standard terminology to describe a feature, the system can still correctly understand the user's intent. For example, "registered capital" can correspond to multiple terms such as "registered funds" or "capital."

[0078] S114. Develop a standard set of enumerated values ​​for discrete features, perform synonym processing on the enumerated values, identify and define the mutual exclusion relationships between feature values, and obtain the enterprise feature map.

[0079] The final step is to develop a standardized list of enumerated values ​​for certain features with specific value ranges (such as company size), and to consider different ways of representing these enumerated values. Furthermore, it is necessary to analyze and record any potential exclusionary relationships between different feature values ​​to ensure that logical contradictions do not occur when executing queries. For example, "large enterprise" and "registered capital less than or equal to 1 million" cannot be satisfied simultaneously.

[0080] Through the above steps, the resulting enterprise feature map is not only a powerful tool to support high-level data analysis and decision-making, but also a key factor in improving user experience.

[0081] In this embodiment, the knowledge graph aims to provide a structured foundation for feature recognition and mapping for natural language understanding models. Its construction employs a multi-layered, multi-dimensional description system to ensure the system can accurately understand the specific expressions of enterprise characteristics made by users in natural language queries. For example... Figure 3 As shown.

[0082] First, the classification hierarchy serves as the top-level architecture of the entire feature system, categorizing enterprise features according to business attributes. These categories primarily include, but are not limited to, basic information, operational status, qualifications and certifications, risk assessment, industry characteristics, and marketing models. Each major category contains multiple specific feature items, thus forming a systematic organizational framework for enterprise features.

[0083] Next, at the feature level, specific enterprise attribute dimensions are defined, such as key features like registered capital, establishment date, business scope, employee size, and main business. Each feature item has a clear semantic definition and data type constraints, providing a standardized foundation for subsequent feature extraction and query transformation. Furthermore, by establishing a feature thesaurus, it's possible to achieve multi-faceted mapping relationships between different terms for the same enterprise feature concepts. For example, the feature "registered capital" can correspond to multiple expressions such as "registered funds," "capital," or "registered amount," thus improving the system's ability to understand user natural language input.

[0084] For discrete features, a standardized set of values ​​is established for the enumeration hierarchy. For example, company size can be enumerated as standard values ​​such as micro-enterprise, small enterprise, medium-sized enterprise, and large enterprise. The design of these enumeration values ​​strictly adheres to industry standards and regulatory requirements to ensure the standardization and comparability of feature values. Simultaneously, to address potential differences between user expression habits and standard enumeration values, diverse expressions are created for the enumeration values; for example, "large enterprise" can correspond to "large company," "relatively large enterprise," or "large corporation," among other terms. This approach resolves the comprehension barriers caused by inconsistent expression.

[0085] Finally, the enumerated value mutual exclusion hierarchy defines the exclusion relationships between feature values ​​to avoid logical conflicts in query conditions. For example, the feature "large enterprise" is mutually exclusive with the condition "registered capital less than or equal to 1 million". When the system detects a query containing such mutual exclusion conditions, it can perform intelligent error correction or provide corresponding prompts to help the user revise the query conditions.

[0086] In summary, this multi-layered and multi-dimensional knowledge graph not only enhances the system's ability to understand enterprise characteristics, but also provides users with a more flexible and accurate query experience.

[0087] S120. Establish a commodity coding table based on the enterprise feature map.

[0088] In this embodiment, the product coding table is constructed based on invoice data from the enterprise feature map and expert experience. It aims to support accurate identification and querying of main product characteristics through a hierarchical structure containing multi-dimensional product information. The design of this product coding table considers the diversity of product name expressions and supports association and extended queries between products.

[0089] The components of a commodity code table include:

[0090] Product Name: The basic identifier for each product, used to describe the product's core attributes.

[0091] Code: A unique identifier assigned to each product to facilitate quick retrieval and processing within the system.

[0092] Definition: Provides a detailed description of the product, including but not limited to its functions and uses, to help distinguish similar products.

[0093] Hierarchy: Indicates the position of a product within its category, which helps to establish hierarchical relationships between products, such as the relationship between parent and child products.

[0094] Parent code: Indicates the direct parent product code of the current product in the hierarchy, and is used to reflect the subordinate relationship between products.

[0095] Synonymous product names: Considering that different users may use different terms to refer to the same product, all known product aliases or synonyms are recorded to improve search accuracy.

[0096] Alternative Product Names: List all other product names that can serve as substitutes for the current product. This is of great value for inventory management and sales recommendations.

[0097] Based on the invoice data and expert experience in the enterprise feature map, a hierarchical commodity coding table containing multi-dimensional commodity information is constructed.

[0098] Specifically, firstly, invoice data from the enterprise feature map is used as the primary data source. This invoice data contains rich information about commodity transactions, such as product name, quantity, and price. Next, this raw data needs to be cleaned and standardized to ensure the data quality used in subsequent steps.

[0099] Relying solely on automated analysis can lead to biases or inaccuracies, so the experience of industry experts is incorporated for manual correction. This step not only improves the accuracy of the product coding table but also ensures that it meets the needs of actual business scenarios.

[0100] The products are organized into a multi-layered structure based on their different characteristics. This design allows users to navigate flexibly between different levels according to their needs, thereby finding the products that best suit their requirements.

[0101] By recording synonyms and alternative product names, the system effectively solves the problem of information retrieval difficulties caused by differences in product naming. This way, even if the product name entered by the user is not the most frequently used one in the database, the system can still recognize it and return the correct result.

[0102] By leveraging the various dimensions of information mentioned above, the system can not only provide accurate product search services, but also automatically recommend relevant products based on the user's query criteria, thereby enhancing the user experience.

[0103] In summary, step S120 emphasizes how to leverage the rich resources within an enterprise's feature map and combine them with professional knowledge to create a comprehensive yet flexible product coding system. This achievement is significant for improving product management efficiency and optimizing customer service.

[0104] S130. Obtain enterprise information and user preferences from different sources, build a data model, and record historical dialogues and current input queries.

[0105] In this embodiment, historical dialogue refers to the system's recorded interactions with the user, particularly the dialogue records from the last five rounds. These records help understand the user's long-term needs, changes in preferences, and their interest trends in specific goods or services. By analyzing historical dialogue, the system can better capture the user's intent and provide more personalized recommendations and services.

[0106] The current input query refers to the specific question or request entered by the user in the latest round of interaction. This could be an information query about a particular product, a request for company details, or other types of business needs. This query will serve as a core input in the system's processing flow to determine the user's immediate needs.

[0107] A data model is a framework or system built by combining relevant enterprise data, user preferences, and historical dialogue information to support intelligent query rewriting, intent recognition, and transformation into standardized query conditions.

[0108] In this embodiment, heterogeneous enterprise data from multiple sources is integrated. After user authentication or authorization, a unified data model is established, which integrates the following user information:

[0109] Basic information: including industry, size, geographical location, etc.

[0110] Behavioral data: such as historical query behavior, browsing history, etc.

[0111] Preference tags: preferences for partner companies, preferences for traded goods, etc.

[0112] The integrated data model can be represented as: ;

[0113] in: The integrated unified data model;

[0114] : The authentication and authorization function for user U, ;

[0115] : No. One data source, ;

[0116] : No. The original datasets from each data source;

[0117] : No. Data transformation functions for each data source;

[0118] : Authentication condition constraint operator;

[0119] : Data union operator;

[0120] : The binding operator between the data source and the transformed data;

[0121] Total number of data sources.

[0122] In addition, the system records the historical dialogue within 5 rounds, collects the natural language input of the current user as the input for this round, and the current time when the user input was received, for the conversion calculation of time-related features. This integrated data will then serve as the input to the semantic understanding module.

[0123] S140. Intelligently rewrite the current input query and identify the query object and its business relationship to obtain the rewriting intention.

[0124] In this embodiment, rewriting intent refers to identifying and standardizing key elements such as objects, relationships, and timeliness by comprehensively analyzing the user's original input and its contextual information (including current time, user information, and historical intent). These elements are then integrated into a fluent, complete, and unambiguous natural language description to accurately reflect the user's query intent. This process ensures that even vague or incomplete user input can be accurately understood and efficiently processed. Ultimately, this single, coherent sentence serves as a standardized intent description, guiding the subsequent query and response generation of the enterprise-level intelligent recommendation system. For example, if a user's query involves finding Apple suppliers established within a specific timeframe, the "rewritten intent" will refine it into a clear statement like "find high-quality Apple suppliers established no more than 5 years ago."

[0125] In one embodiment, step S140 described above may include steps S141 to S144.

[0126] S141. Analyze the current input query and historical dialogue information to determine the query category of the current input query, so as to obtain the first analysis result.

[0127] In this embodiment, the first analysis result refers to the system's ability to identify the category (such as information query, product search, company details, etc.) of the current query by comprehensively analyzing the current input query and historical dialogue information. This process involves contextual analysis, that is, determining whether the new input is a continuation or modification of an existing topic or a completely new query direction.

[0128] S142. Identify the core object of the current input query, and determine the business relationship between the user and the core object based on the user information to obtain the second analysis result.

[0129] In this embodiment, the second analysis result refers to identifying the main entity (i.e., the core object) targeted by the current input query and inferring the specific business relationship between the user and that entity based on the information provided by the user. For example, if the user is inquiring about a certain raw material, it is necessary to clarify whether this material is a component of its product or part of its production tools, etc.

[0130] S143. Convert the relative and vague time expressions in the current input query into absolute and explicit time ranges or specific dates to obtain the third analysis result.

[0131] In this embodiment, the third analysis result means that any query containing time descriptions will be refined, that is, relative time expressions such as "recent years" or "established for more than X years" will be converted into specific years or date ranges. This is crucial for improving the accuracy and usability of queries.

[0132] S144. Integrate the first analysis result, the second analysis result, and the third analysis result to generate a rewriting intent.

[0133] In the final stage, the system integrates the results from the previous steps to form a clear, complete, and logically consistent intent description. This rewritten intent not only reflects the user's original needs but also considers all relevant background information, ensuring the effectiveness and relevance of the query.

[0134] The intent rewriting module is one of the core innovations of this system, implemented using a large language model based on the Transformer architecture. Specifically, it uses Qwen3-14B as the base model, and its technical features are as follows:

[0135] Infrastructure: Transformer Decoder-only architecture.

[0136] Parameter size: 14 billion parameters.

[0137] Context length: Supports 32K token context windows.

[0138] Attention mechanism: Multi-head self-attention mechanism, with 40 heads.

[0139] Hidden layer dimension: 5120.

[0140] Feedforward network dimension: 13824.

[0141] Number of floors: 40.

[0142] The main function of this module is to intelligently rewrite the user's original input by combining contextual elements such as historical dialogue information, current time information, and user information, and output a semantically complete, logically consistent, and clearly expressed intent description. The rewriting process follows these principles:

[0143] Eliminate semantic ambiguity: Ensure the uniqueness and accuracy of the query intent.

[0144] Optimize expression: Make the description of intent more standardized and formalized.

[0145] Maintain logical consistency: Ensure that the logical relationships between the preceding and following text are reasonable.

[0146] In particular, it is necessary to focus on object identification and relationship analysis, that is, to identify whether the input in this round is about "enterprises" or "products", and to analyze the business relationships between users and these objects.

[0147] The relevant keywords are as follows:

[0148] You are a highly specialized query analysis and rewriting engine designed specifically for enterprise-level intelligent recommendation systems. Your core task is to receive the user's raw, potentially vague or incomplete, input and, combined with complete contextual information, transform it into a single, complete, explicit, and logically consistent standardized intent description. The contextual information required for processing this task includes the current time, user information, historical intents, and the current input. First, carefully read and understand the content of the current input. Then, perform correlation analysis between the current input and historical intents, determining, using the consideration tags, whether the current input is a "supplement," "correction," "replacement" of the historical intent, or a completely new query. This is the first step: Contextual Analysis.

[0149] The second step is Object & Relationship Analysis. In this step, the core target object (object) of the input in this round needs to be determined from the thought tags: is it a "company" or a "product"? If the input describes geographical location, years of establishment, company size, etc., then the object is usually a "company". If the input describes a specific product name, model, raw material, etc., then the object is a "product". Based on the object and user information, the business relationship between the user and the target company is further analyzed and determined. For example, if the user's business is canned apples and the query intent is to find apples, then the relationship should be precisely defined as a supplier of raw materials for canned apples. If a buyer-seller relationship is not explicitly stated, the default is to follow the relationship in the historical intent, or default to a supplier.

[0150] The third step involves temporal feature standardization. In this step, all time-related expressions need to be identified, and these relative times need to be converted into absolute formats based on the current time. For example, expressions such as "established for X years" or "established for more than X years" should be converted into specific establishment dates based on the current time. <yyyy-mm-dd>Format. Furthermore, vague time concepts such as "recent years" or "the last few years" need to be converted into specific time ranges, such as "years since establishment". "3 years" is replaced with "years since establishment". "1 year", while "established company" can be defined as "established for 10 years".

[0151] The final step is Intent Synthesis & Output. This step integrates all the analysis results above to generate the final rewritten intent. Newly identified features (such as geolocation, product, relationship, time, etc.) are merged with features from historical intents. If the current input is a modification of a feature (e.g., changing the product from banana to apple), that feature should be updated, not simply added. Ensure all valid features are combined into a fluent, complete, and unambiguous natural language description. The output should be a single, coherent sentence, with the final generated, single, standardized intent description string provided within the `rewritten_intent` tag. Through these four steps, user queries can be effectively analyzed and rewritten, thereby improving the accuracy and user experience of enterprise recommendation systems.

[0152] In summary, step S140, through systematic analysis of user input, historical interaction records, and relevant enterprise data, achieves a deep understanding and modeling of enterprise information and user preferences. This step is of great significance for improving the intelligence level of recommendation systems and user experience.

[0153] S150. Based on the enterprise feature map and the commodity code table, the rewriting intent is transformed into standardized query conditions.

[0154] In this embodiment, standardized query conditions refer to identifying the product and enterprise characteristics involved in the user's rewriting intent by combining enterprise feature maps and product coding tables, and matching these characteristics to corresponding entries in the database to generate structured query conditions that accurately reflect the user's query needs. This process not only ensures the accuracy of the query conditions but also guarantees their compatibility with existing data structures, providing clear and specific guidance for subsequent data retrieval.

[0155] Based on the enterprise feature map and product coding table, the product and enterprise features involved in the rewriting intent are identified, matched with the corresponding entries in the database, and a corresponding query condition structure is generated.

[0156] In one embodiment, step S150 described above may include steps S151 to S152.

[0157] S151. Combining the rewriting intent with the enterprise feature map and commodity coding table, generate prompt words; identify the enterprise characteristics involved in the rewriting intent, and verify and supplement them through the enterprise feature map; obtain the specific numerical value, text or time value in the current input query, and perform unit conversion and expression processing; analyze and determine the comparison operators and special operation types in the current input query; understand and construct the logical relationships and nested structures in the current input query to obtain parsed information.

[0158] In this embodiment, parsing information refers to analyzing the user's rewriting intent, combining enterprise feature maps and product coding tables, identifying and extracting the standard features, specific values, operators, and logical relationships involved in the query, and then constructing structured information that can accurately reflect the user's query needs.

[0159] First, in step S151, the system combines the user's rewriting intent, enterprise feature graph, and product coding table to generate prompt words for processing large models (such as qwen2.5-72b). The core tasks of this step include:

[0160] Feature Recognition and Verification Supplement: Identify the enterprise features mentioned in the rewriting intent, and use the enterprise feature map for verification and supplementation to ensure that the identified features are standard and consistent.

[0161] Value extraction and processing: Extract specific numerical, textual, or time values ​​from the current input query, and perform unit conversion and expression processing based on information in the knowledge graph to ensure that the value representation format meets the database requirements.

[0162] Operator analysis: Identify the comparison operators and special operation types (such as TopN / BottomN) in the query, as these operators will directly affect the way data is retrieved.

[0163] Logical relationship construction: Understanding the logical relationships and nested structures in a query involves basic logical operations such as AND, OR, and NOT, which are crucial for correctly interpreting user intent.

[0164] In this process, the parsed information contains all the necessary elements for the next step of organizing it into a tree structure, namely, explicit feature names, feature values, operators, and the logical relationships between them.

[0165] S152. Organize the parsed information into a tree structure to ensure that the logical and conditional nodes correctly express the query intent, so as to obtain the query condition structure.

[0166] Subsequently, in step S152, the parsed information is further organized into a tree-like logical structure. This structure aims to ensure that the query intent is expressed in the most intuitive and clear way, facilitating understanding and execution by downstream modules. Specifically, the design principles of the tree structure are as follows:

[0167] Condition block definition: Each node represents a query condition, which includes features, operators, and feature values, forming a condition block.

[0168] Logical relationship expression: Use logical nodes such as AND, OR, NOT to represent the relationship between condition blocks, allowing complex logical nesting.

[0169] Sorting operation restrictions: In particular, sorting and quantity restriction operations such as TopN / BottomN can only appear at the root node position to indicate the sorting basis and quantity requirements of the final result set.

[0170] Through these two steps, the system can transform the user's original fuzzy query into precise, structured query conditions, greatly improving the efficiency and accuracy of data retrieval. The final output tree structure not only reflects the user's query intent but also strictly adheres to the predefined JSON pattern, laying a solid foundation for intelligent and automated data retrieval.

[0171] In this embodiment, after initially rewriting the user's intent, the core task of this step is to accurately identify, standardize, and structure the query conditions involved in the intent description using a feature knowledge graph. This process does not directly retrieve enterprise data, but rather obtains authoritative and structured feature metadata to guide the generation of downstream logical structures. This process is specifically divided into two stages: The first stage is the extraction of candidate features and values. The system first analyzes the rewritten intent text and uses Entity Recognition (NER) technology to extract candidate text fragments that may represent query features and feature values. For example, when querying "technology companies established for more than 5 years," the system will extract the candidate feature "years established" and the candidate value "more than 5 years," as well as the candidate feature "company type" and the candidate value "technology." The second stage is feature graph alignment and information enhancement. The system aligns the candidate text fragments extracted in the previous stage with the constructed feature knowledge graph, completing this process through vector similarity calculation or keyword matching. The main objectives of this step include feature standardization, enumeration value normalization, and acquisition of feature metadata. The final output is a structured, feature knowledge graph-validated, and rich information set that clearly describes the standard features in the user query, the expected values ​​of these features, and their inherent attributes. This enhanced context is passed as the core input to the next step, the "tree-shaped logical structure generation" module, which greatly reduces the complexity and ambiguity of downstream model understanding and transformation tasks.

[0172] Next, a specific large-scale language model (such as qwen2.5-72b) is used to convert natural language into actionable structured query conditions. It primarily performs information extraction tasks such as feature extraction, feature value extraction, operator recognition, and logical relationship parsing, outputting the results in a tree structure. The basic idea of ​​the tree structure is to decompose the natural language query into multiple condition blocks, each containing features, operators, and feature values; the logical relationships between condition blocks are expressed using a tree structure, supporting nesting. Each node can be a logical node (AND / OR / NOT) or a condition node (leaf node). TopN / BottomN can only appear at the root node, indicating that all results satisfying the conditions are sorted and the top N / bottom N are selected. The entire process ensures that the output strictly follows the provided JSON pattern, paying attention to nested logical conditions and sorting requirements to accurately reflect the user's query needs. Finally, based on the user's intent and authoritative, structured feature metadata identified and aligned from the feature knowledge graph, a compliant JSON format output is provided for subsequent enterprise information retrieval.

[0173] S160. Based on the standardized query conditions, the natural language description is converted into a structured JSON format to obtain structured information.

[0174] Based on the standardized query conditions, the natural language description is converted into a structured JSON format to obtain structured information. This process is one of the key steps in the entire query processing flow, and its purpose is to capture and express the user's query intent through a precise structured representation, thereby providing clear and explicit guidance for subsequent data retrieval and analysis.

[0175] In this embodiment, structured information refers to one or more features and their corresponding values, operators, and logical relationships being systematically organized in a tree-structured JSON format. This structure not only clarifies the specific content of each query condition (i.e., what enterprise characteristics the user wants to find and what the expected feature values ​​are), but also details how these conditions are related to each other (e.g., some conditions may need to be met simultaneously, while others only need to be met one of them). Furthermore, this structured information can support more complex query requirements, such as sorting requirements (TopN / BottomN), allowing the final results to be sorted in ascending or descending order based on specific fields, and selecting the first N or last N items.

[0176] Specifically, structured information includes the following core components:

[0177] Feature extraction: Identify and extract enterprise features involved in the query, such as enterprise name, registered capital, establishment date, industry classification, etc. This step combines information from the knowledge graph for verification and supplementation to ensure that all mentioned features are accurate and standardized.

[0178] Feature extraction: Accurately identify specific numerical values, text values, time values, and other features in the query, and process various unit conversions and numerical expressions based on information from the knowledge graph. This ensures that the data format in the query conditions is consistent with the storage format in the database, facilitating efficient matching.

[0179] Operator recognition: Identifies comparison operators in queries, including equal to, greater than, less than, greater than or equal to, less than or equal to, not equal to, TopN, BottomN, contain, fuzzy matching, etc. This helps determine how to compare feature values ​​with actual values ​​in the database.

[0180] Logical Relationship Analysis: Accurately understand the logical relationships within a query, including basic logical operations such as AND, OR, and NOT, as well as their nesting and combination relationships. This approach allows for the construction of complex query logic to meet the diverse needs of different users.

[0181] Tree structure representation: All the above elements are integrated into a tree structure, where each node represents a condition or logical operation. Condition nodes contain specific features, operators, and feature values; logical nodes define the logical relationships between these conditions. Such a structure is not only easy to understand and execute, but also flexibly handles query requests of varying complexity.

[0182] The structured information generated through the above process can significantly improve the efficiency and accuracy of data retrieval, and also provide users with a more intuitive and personalized query experience. Therefore, it is a key bridge for realizing the transformation from natural language to actionable query conditions.

[0183] S170. Generate a specific query statement based on the structured information and execute the query to obtain the query result.

[0184] In this embodiment, the query result refers to the list of eligible companies obtained from the enterprise profile table by converting structured information into Elasticsearch Query DSL and executing the query.

[0185] In one embodiment, step S170 described above may include steps S171 to S173.

[0186] S171. Convert the structured information into an Elasticsearch query DSL according to predefined rules.

[0187] In this embodiment, this step mainly involves accurately translating the structured information tree-like logical structure generated in the previous step into an executable query DSL for Elasticsearch, according to predefined transformation rules. This process is similar to a text-to-SQL conversion task, but because the input is highly structured, its reliability is far higher than generating it directly from natural language. The specific transformation rules are as follows:

[0188] The AND logical operator is converted to bool.must.

[0189] The OR logical operator is converted to bool.should.

[0190] The NOT logical operator is converted to bool.must_not.

[0191] The equals operator (=) is converted to a term query.

[0192] The not-equals operator (!=) is implemented by nesting a term in bool.must_not.

[0193] Comparison operators such as greater than and less than are converted to range queries.

[0194] Other fuzzy matching conditions can use the match or wildcard query types.

[0195] S172. Based on the operator type of the node, map the operator type to the corresponding content in the Elasticsearch query DSL. For non-equality comparison conditions, exclude values ​​by nesting terms in bool.must_not. Use match or wildcard query types to support flexible fuzzy search of text. Use sort and size parameters to control the sorting method of the results and the number of returned results to meet the TopN / BottomN requirements. Use a Python recursive function to traverse the logic tree and generate the ES query DSL, while processing the requirements on the root node to obtain the generated specific query statement.

[0196] In this step, based on each node in the structured information and its associated operators, the query DSL is further refined and mapped to the corresponding ES query DSL elements. Specifically, for unequal comparison conditions, value exclusion is achieved by nesting terms within `bool.must_not`. Furthermore, flexible fuzzy searches on text are supported using the `match` or `wildcard` query types. Simultaneously, the `sort` and `size` parameters control the sorting of results and the number of results returned to meet TopN / BottomN requirements. The entire logical tree is traversed using a Python recursive function to generate the final ES query DSL. During this process, special attention is paid to handling special requirements at the root node, such as the TopN or BottomN settings.

[0197] S173. Execute the generated specific query statement and obtain a list of enterprises that meet the conditions from the enterprise profile table to obtain the query results.

[0198] The final step is to execute the specific query statements generated in the previous steps. By calling the Elasticsearch API, the constructed query DSL is executed to retrieve data from the enterprise profile table, thus obtaining a list of enterprises that meet the criteria as the query results. These enterprise lists will be written into the opportunity recommendation results table based on the scores of the "Opportunity Recommendation Enterprise Scoring Model," where the recommendation order is determined by the score, with higher scores receiving higher priority.

[0199] After the query is completed, the system will display company information in the recommended order. Each company is listed in card format, including but not limited to company name, main products, establishment date, legal representative, company address, business scope, and contact information. Users can interact with the recommendation results through the like, dislike, and favorite functions, enhancing the user experience while also providing data support for the optimization of subsequent recommendation algorithms.

[0200] S180. Evaluate and sort the query results to obtain a ranking result based on scores.

[0201] In this embodiment, the ranking result refers to evaluating and scoring each candidate company based on the list of candidate companies obtained from the company profile table using the "Business Opportunity Recommendation Company Scoring Model." This scoring model comprehensively evaluates companies based on multiple dimensions and calculates a specific score for each company. This process not only considers the initial screening criteria set by the user but also introduces additional factors such as quality, relevance, and user preferences, thereby ensuring that the final list of recommended companies displayed to the user meets basic requirements and has high business potential and matching degree.

[0202] First, the Elasticsearch (ES) query language generated in the previous step is used to query the corresponding enterprise profile table to obtain a preliminary list of candidate enterprises that meet the user's specific conditions. For each enterprise found, it is scored using the "Business Opportunity Recommendation Enterprise Scoring Model." This model considers factors such as, but is not limited to, the enterprise's operating status, market performance, credit rating, industry influence, and relevance to the user's needs. All candidate enterprises are then sorted from highest to lowest score, resulting in a ranking result. The higher the score, the higher the enterprise appears in the recommendation list, meaning it is given priority in being recommended to the user.

[0203] S190. Based on the ranking results of the scores, display the detailed information of the recommended companies.

[0204] Subsequently, in step S190, detailed information about the recommended companies is displayed based on the ranking results. This means the system will sequentially present detailed information about the recommended companies to the user according to the ranking results. The company information will be presented in an intuitive and easy-to-understand manner, such as in the form of company information cards, including but not limited to key information such as company name, main products, establishment date, legal representative, company address, business scope, and contact information. Furthermore, users can interact with recommended companies they are interested in by liking, disliking, or saving them, in order to further filter and manage their target companies.

[0205] This embodiment transforms natural language queries into a standardized tree-like logical structure, supporting complex nested logical relationships and automatically converting them to the Elasticsearch query language. Core technologies include condition block parsing (feature-operator-feature value triple extraction), logical relationship tree construction (supporting multi-level nesting of AND / OR / NOT), TopN / BottomN sorting constraint processing, and rule-based conversion from tree structure to Elasticsearch query statements. A feature knowledge graph system is constructed, encompassing classification levels, feature levels, feature synonyms, enumeration value levels, enumeration value synonyms, and enumeration value mutual exclusions, enabling standardized expression and intelligent mapping of enterprise features. It addresses the issue of diverse enterprise feature expressions in natural language by avoiding logical conflicts in query conditions through a mutual exclusion hierarchy, supporting unified mapping of diverse expression methods. A large-scale language model is used in conjunction with contextual elements such as historical dialogue information, current time information, and user basic information to intelligently rewrite the user's original input, particularly achieving object recognition and relationship analysis functions and time feature conversion. Semantic ambiguity is eliminated, ensuring the uniqueness and accuracy of the query intent.

[0206] Unlike existing technologies, which often result in conversions that deviate from the user's true intent when handling multi-dimensional condition filtering and complex business logic judgments, leading to low accuracy and the need for manual intervention and correction, the method in this embodiment innovatively adopts a tree-like logical structure as an intermediate representation, supporting nested logical relationships of arbitrary complexity. Furthermore, it addresses logical conflicts in queries through pre-processing intent rewriting and ensures query performance through subsequent feature value standardization, enabling complex queries to reach a stable level required for industrial applications.

[0207] Existing technologies typically employ simple prompting engineering methods, resulting in limited depth of understanding of query intent and difficulty in accurately grasping the user's true business needs. The method in this embodiment adds object identification and relationship analysis, automatic processing of relative time representation conversion, and combines multi-dimensional context such as historical dialogue information, current time information, and user basic information to rewrite intent, thereby significantly improving the accuracy of user intent recognition in business scenarios.

[0208] Traditional Text2SQL technology is primarily designed for relational databases and has limited support for Elasticsearch query syntax; moreover, existing Elasticsearch systems lack a complete natural language conversion mechanism. This embodiment addresses this by designing conversion rules specifically for Elasticsearch DSL query syntax, achieving a complete conversion mapping from tree-like logical structures to ES query statements, supporting all core query types. Furthermore, compared to query languages ​​generated directly from large models in a single step, the method presented in this embodiment offers greater stability and accuracy.

[0209] Generalized technical solutions such as Text2ESQuery lack in-depth optimization for specific business domains, making it difficult to accurately understand the professional terminology, business rules, and query patterns of the enterprise information domain. The method in this embodiment constructs a multi-level feature knowledge graph specific to the enterprise information domain, including a complete enterprise feature system encompassing basic information, business status, qualification certification, and risk assessment categories. It also utilizes the company's internally accumulated product coding table to support reasoning about enterprise-specific business relationships and handle implicit mutual exclusion relationships within conditions.

[0210] Existing technologies require extensive code modifications when adding new query conditions or enterprise features, resulting in high maintenance costs and difficulties in expansion. The method in this embodiment is based on a loosely coupled knowledge graph architecture. Adding new features only requires modifying the knowledge graph configuration, supports a hot update mechanism, and takes effect without restarting the system. Furthermore, the large model suggestion words used do not need to iterate with each new feature, offering greater versatility.

[0211] The aforementioned intelligent enterprise information recommendation method creates an enterprise feature map containing a multi-level classification system and a product coding table based on this map. It then combines enterprise information from different sources with user preference data to construct a data model and record historical dialogues and current query inputs. Next, it intelligently rewrites the current query, identifying the query object and its business relationships to clarify the rewriting intent, and transforms this intent into standardized query conditions based on the enterprise feature map and product coding table. Subsequently, these conditions are converted from natural language descriptions into structured JSON format information, generating specific query statements for execution. Finally, by evaluating and ranking the query results, it displays detailed information about the recommended enterprises. This achieves a precise conversion from natural language queries to structured enterprise information retrieval, significantly improving the accuracy and intelligence of the enterprise recommendation system and effectively solving the problems of low accuracy and the need for manual intervention in complex query conditions under existing technologies.

[0212] Figure 4 This is a schematic block diagram of an intelligent enterprise information recommendation device 300 provided in an embodiment of the present invention. Figure 4 As shown, corresponding to the above-described intelligent enterprise information recommendation method, the present invention also provides an intelligent enterprise information recommendation device 300. This intelligent enterprise information recommendation device 300 includes a unit for executing the above-described intelligent enterprise information recommendation method, and the device can be configured in a server. Specifically, please refer to... Figure 4 The enterprise information intelligent recommendation device 300 includes a map creation unit 301, an encoding table establishment unit 302, a recording unit 303, a rewriting unit 304, a condition conversion unit 305, a format conversion unit 306, a query unit 307, a sorting unit 308, and a display unit 309.

[0213] The graph creation unit 301 is used to acquire enterprise-related data and create an enterprise feature graph containing a multi-level classification system of enterprises; the coding table establishment unit 302 is used to establish a commodity coding table based on the enterprise feature graph; the recording unit 303 is used to acquire enterprise information and user preferences from different sources, construct a data model, and record historical dialogues and current input queries; the rewriting unit 304 is used to intelligently rewrite the current input query and identify the query object and its business relationship to obtain the rewriting intent; the condition conversion unit 305 is used to convert the rewriting intent into standardized query conditions based on the enterprise feature graph and the commodity coding table; the format conversion unit 306 is used to convert natural language descriptions into structured JSON format based on the standardized query conditions to obtain structured information; the query unit 307 is used to generate specific query statements based on the structured information and execute the query to obtain query results; the sorting unit 308 is used to evaluate and sort the query results to obtain a high-low score ranking result; and the display unit 309 is used to display detailed information of recommended enterprises based on the high-low score ranking result.

[0214] In one embodiment, the map creation unit 301 includes:

[0215] The classification structure construction subunit is used to acquire enterprise-related data, establish the top-level classification of enterprise characteristics, and construct a systematic classification structure; the setting subunit is used to clarify the specific enterprise attribute dimensions based on each classification in the classification structure, and set semantic definitions and data types for each feature; the thesaurus creation subunit is used to create a thesaurus of feature items; the processing subunit is used to formulate a standard enumeration value set for discrete features, perform thesaurus processing on the enumeration values, identify and define the mutual exclusion relationships between feature values, so as to obtain the enterprise feature map.

[0216] In one embodiment, the coding table establishment unit 302 is used to construct a hierarchical commodity coding table containing multi-dimensional commodity information based on invoice data and expert experience in the enterprise feature map.

[0217] In one embodiment, the rewriting unit 304 includes:

[0218] The first analysis subunit is used to analyze the current input query and historical dialogue information to determine the query category of the current input query, so as to obtain a first analysis result; the second analysis subunit is used to identify the core object of the current input query and determine the business relationship between the user and the core object based on user information, so as to obtain a second analysis result; the third analysis subunit is used to convert the relative and vague time expressions in the current input query into absolute and specific time ranges or specific dates, so as to obtain a third analysis result; the integration subunit is used to integrate the first analysis result, the second analysis result, and the third analysis result to generate a rewriting intent.

[0219] In one embodiment, the condition conversion unit 305 is used to identify the product and enterprise characteristics involved in the rewriting intent based on the enterprise feature map and the product code table, match the corresponding entries in the database, and generate the corresponding query condition structure.

[0220] In one embodiment, the conditional transformation unit 305 includes:

[0221] The parsing subunit is used to combine the rewriting intent with the enterprise feature map and the product coding table to generate prompt words; identify the enterprise characteristics involved in the rewriting intent, and verify and supplement them through the enterprise feature map to obtain the specific numerical, text or time values ​​in the current input query, and perform unit conversion and expression processing; analyze and determine the comparison operators and special operation types in the current input query; understand and construct the logical relationships and their nested structures in the current input query to obtain parsed information; and organize the parsed information into a tree structure to ensure that the logical and conditional nodes correctly express the query intent to obtain the query condition structure.

[0222] In one embodiment, the query unit 307 includes:

[0223] The transformation subunit converts the structured information into an Elasticsearch query DSL according to predefined rules. The statement generation subunit maps the operator type of the node to the corresponding content in the Elasticsearch query DSL. For unequal comparison conditions, value exclusion is achieved by nesting terms in bool.must_not. The match or wildcard query type supports flexible fuzzy search of text. The sort and size parameters control the sorting method of the results and the number of returned results to meet the TopN / BottomN requirements. The logical tree is traversed using a Python recursive function to generate the ES query DSL, while processing the requirements on the root node to obtain the generated specific query statement. The execution subunit executes the generated specific query statement and retrieves a list of enterprises that meet the conditions from the enterprise profile table to obtain the query results.

[0224] It should be noted that those skilled in the art can clearly understand that the specific implementation process of the above-mentioned enterprise information intelligent recommendation device 300 and each unit can be referred to the corresponding description in the foregoing method embodiments. For the sake of convenience and brevity, it will not be repeated here.

[0225] The aforementioned intelligent recommendation device 300 for enterprise information can be implemented as a computer program, which can perform tasks such as... Figure 5 It runs on the computer device shown.

[0226] Please see Figure 5 , Figure 5 This is a schematic block diagram of a computer device provided in an embodiment of this application. The computer device 500 can be a server, wherein the server can be a standalone server or a server cluster composed of multiple servers.

[0227] See Figure 5 The computer device 500 includes a processor 502, a memory, and a network interface 505 connected via a system bus 501. The memory may include a non-volatile storage medium 503 and internal memory 504.

[0228] The non-volatile storage medium 503 may store an operating system 5031 and a computer program 5032. The computer program 5032 includes program instructions that, when executed, cause the processor 502 to perform an intelligent recommendation method for enterprise information.

[0229] The processor 502 provides computing and control capabilities to support the operation of the entire computer device 500.

[0230] The internal memory 504 provides an environment for the execution of the computer program 5032 in the non-volatile storage medium 503. When the computer program 5032 is executed by the processor 502, the processor 502 can execute an intelligent recommendation method for enterprise information.

[0231] This network interface 505 is used for network communication with other devices. Those skilled in the art will understand that... Figure 5 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device 500 to which the present application is applied. The specific computer device 500 may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0232] The processor 502 is used to run a computer program 5032 stored in the memory to perform the following steps:

[0233] The process involves: acquiring relevant enterprise data and creating an enterprise feature map containing a multi-level classification system; establishing a product coding table based on the enterprise feature map; acquiring enterprise information and user preferences from different sources, constructing a data model, and recording historical dialogues and current input queries; intelligently rewriting the current input query and identifying the query object and its business relationships to obtain the rewriting intent; transforming the rewriting intent into standardized query conditions based on the enterprise feature map and product coding table; converting natural language descriptions into structured JSON format based on the standardized query conditions to obtain structured information; generating specific query statements based on the structured information and executing the queries to obtain query results; evaluating and ranking the query results to obtain a ranking result based on scores; and displaying detailed information of recommended enterprises based on the ranking result.

[0234] In one embodiment, when the processor 502 implements the step of acquiring enterprise-related data and creating an enterprise feature map containing a multi-level classification system of enterprises, it specifically implements the following steps:

[0235] Acquire relevant enterprise data, establish a top-level classification of enterprise characteristics, and construct a systematic classification structure; based on each classification in the classification structure, clarify specific enterprise attribute dimensions, and set semantic definitions and data types for each feature; create a thesaurus of feature items; formulate a standard enumeration value set for discrete features, perform thesaurus processing on the enumeration values, identify and define the mutual exclusion relationships between feature values, so as to obtain an enterprise feature map.

[0236] In one embodiment, when implementing the step of establishing a commodity coding table based on the enterprise feature map, the processor 502 specifically implements the following steps:

[0237] Based on the invoice data and expert experience in the enterprise feature map, a hierarchical commodity coding table containing multi-dimensional commodity information is constructed.

[0238] In one embodiment, when the processor 502 implements the step of intelligently rewriting the current input query and identifying the query object and its business relationship to obtain the rewriting intent, it specifically implements the following steps:

[0239] Analyze the current input query and historical dialogue information to determine the query category of the current input query to obtain a first analysis result; identify the core object of the current input query and determine the business relationship between the user and the core object based on user information to obtain a second analysis result; convert the relative and vague time expressions in the current input query into absolute and specific time ranges or specific dates to obtain a third analysis result; integrate the first analysis result, the second analysis result, and the third analysis result to generate a rewriting intent.

[0240] In one embodiment, when the processor 502 implements the step of converting the rewriting intent into standardized query conditions based on the enterprise feature map and the commodity code table, it specifically implements the following steps:

[0241] Based on the enterprise feature map and product coding table, the product and enterprise features involved in the rewriting intent are identified, matched with the corresponding entries in the database, and a corresponding query condition structure is generated.

[0242] In one embodiment, when the processor 502 implements the steps of identifying the product and enterprise characteristics involved in the rewriting intent based on the enterprise feature map and product coding table, matching them to the corresponding entries in the database, and generating the corresponding query condition structure, the specific implementation is as follows:

[0243] Combining the rewriting intent with the enterprise feature map and product coding table, prompt words are generated; enterprise characteristics involved in the rewriting intent are identified and verified and supplemented through the enterprise feature map; specific numerical values, text, or time values ​​in the current input query are obtained, and unit conversion and expression processing are performed; comparison operators and special operation types in the current input query are analyzed and determined; logical relationships and their nested structures in the current input query are understood and constructed to obtain parsed information; the parsed information is organized into a tree structure to ensure that logical and conditional nodes correctly express the query intent, thereby obtaining the query condition structure.

[0244] In one embodiment, when the processor 502 implements the step of generating a specific query statement based on the structured information and executing the query to obtain the query result, it specifically implements the following steps:

[0245] The structured information is converted into an Elasticsearch query DSL according to predefined rules. Based on the operator type of each node, the operator type is mapped to the corresponding content in the Elasticsearch query DSL. For unequal comparison conditions, value exclusion is achieved by nesting terms in `bool.must_not`. The `match` or `wildcard` query types support flexible fuzzy searches of text. The `sort` and `size` parameters control the sorting method and the number of results returned to meet TopN / BottomN requirements. A Python recursive function traverses the logic tree and generates the Elasticsearch query DSL, while simultaneously processing the requirements at the root node to obtain a specific query statement. The generated query statement is executed, and a list of matching companies is retrieved from the enterprise profile table to obtain the query results.

[0246] It should be understood that in the embodiments of this application, the processor 502 may be a central processing unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor.

[0247] It will be understood by those skilled in the art that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program includes program instructions and can be stored in a storage medium, which is a computer-readable storage medium. The program instructions are executed by at least one processor in the computer system to implement the process steps of the embodiments of the above methods.

[0248] Therefore, the present invention also provides a storage medium. This storage medium can be a computer-readable storage medium. The storage medium stores a computer program, wherein when executed by a processor, the computer program causes the processor to perform the following steps:

[0249] The process involves: acquiring relevant enterprise data and creating an enterprise feature map containing a multi-level classification system; establishing a product coding table based on the enterprise feature map; acquiring enterprise information and user preferences from different sources, constructing a data model, and recording historical dialogues and current input queries; intelligently rewriting the current input query and identifying the query object and its business relationships to obtain the rewriting intent; transforming the rewriting intent into standardized query conditions based on the enterprise feature map and product coding table; converting natural language descriptions into structured JSON format based on the standardized query conditions to obtain structured information; generating specific query statements based on the structured information and executing the queries to obtain query results; evaluating and ranking the query results to obtain a ranking result based on scores; and displaying detailed information of recommended enterprises based on the ranking result.

[0250] In one embodiment, when the processor executes the computer program to acquire enterprise-related data and create an enterprise feature map containing a multi-level classification system, it specifically implements the following steps:

[0251] Acquire relevant enterprise data, establish a top-level classification of enterprise characteristics, and construct a systematic classification structure; based on each classification in the classification structure, clarify specific enterprise attribute dimensions, and set semantic definitions and data types for each feature; create a thesaurus of feature items; formulate a standard enumeration value set for discrete features, perform thesaurus processing on the enumeration values, identify and define the mutual exclusion relationships between feature values, so as to obtain an enterprise feature map.

[0252] In one embodiment, when the processor executes the computer program to implement the step of establishing a commodity coding table based on the enterprise feature map, it specifically implements the following steps:

[0253] Based on the invoice data and expert experience in the enterprise feature map, a hierarchical commodity coding table containing multi-dimensional commodity information is constructed.

[0254] In one embodiment, when the processor executes the computer program to intelligently rewrite the current input query and identify the query object and its business relationships to obtain the rewriting intent, it specifically implements the following steps:

[0255] Analyze the current input query and historical dialogue information to determine the query category of the current input query to obtain a first analysis result; identify the core object of the current input query and determine the business relationship between the user and the core object based on user information to obtain a second analysis result; convert the relative and vague time expressions in the current input query into absolute and specific time ranges or specific dates to obtain a third analysis result; integrate the first analysis result, the second analysis result, and the third analysis result to generate a rewriting intent.

[0256] In one embodiment, when the processor executes the computer program to implement the step of converting the rewriting intent into standardized query conditions based on the enterprise feature map and the commodity code table, it specifically implements the following steps:

[0257] Based on the enterprise feature map and product coding table, the product and enterprise features involved in the rewriting intent are identified, matched with the corresponding entries in the database, and a corresponding query condition structure is generated.

[0258] In one embodiment, when the processor executes the computer program to implement the steps of identifying the product and enterprise characteristics involved in the rewriting intent based on the enterprise feature map and product code table, matching the corresponding entries in the database, and generating the corresponding query condition structure, the specific implementation is as follows:

[0259] Combining the rewriting intent with the enterprise feature map and product coding table, prompt words are generated; enterprise characteristics involved in the rewriting intent are identified and verified and supplemented through the enterprise feature map; specific numerical values, text, or time values ​​in the current input query are obtained, and unit conversion and expression processing are performed; comparison operators and special operation types in the current input query are analyzed and determined; logical relationships and their nested structures in the current input query are understood and constructed to obtain parsed information; the parsed information is organized into a tree structure to ensure that logical and conditional nodes correctly express the query intent, thereby obtaining the query condition structure.

[0260] In one embodiment, when the processor executes the computer program to generate a specific query statement based on the structured information and executes the query to obtain query results, it specifically implements the following steps:

[0261] The structured information is converted into an Elasticsearch query DSL according to predefined rules. Based on the operator type of each node, the operator type is mapped to the corresponding content in the Elasticsearch query DSL. For unequal comparison conditions, value exclusion is achieved by nesting terms in `bool.must_not`. The `match` or `wildcard` query types support flexible fuzzy searches of text. The `sort` and `size` parameters control the sorting method and the number of results returned to meet TopN / BottomN requirements. A Python recursive function traverses the logic tree and generates the Elasticsearch query DSL, while simultaneously processing the requirements at the root node to obtain a specific query statement. The generated query statement is executed, and a list of matching companies is retrieved from the enterprise profile table to obtain the query results.

[0262] The storage medium can be any computer-readable storage medium capable of storing program code, such as a USB flash drive, portable hard drive, read-only memory (ROM), magnetic disk, or optical disk.

[0263] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0264] In the several embodiments provided by this invention, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative. For example, the division of each unit is merely a logical functional division, and there may be other division methods in actual implementation. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed.

[0265] The steps in the method of this invention can be adjusted, merged, or reduced in order according to actual needs. The units in the device of this invention can be merged, divided, or reduced according to actual needs. Furthermore, the functional units in the various embodiments of this invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0266] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a terminal, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention.

[0267] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. An intelligent recommendation method for enterprise information, characterized in that, include: Acquire relevant enterprise data and create an enterprise feature map containing a multi-level classification system for enterprises; A commodity coding table is established based on the enterprise feature map; Acquire enterprise information and user preferences from different sources, build data models, and record historical conversations and current input queries; The current input query is intelligently rewritten, and the query object and its business relationship are identified to obtain the rewriting intent; Based on the enterprise feature map and the commodity coding table, the rewriting intent is transformed into standardized query conditions. Based on the standardized query conditions, the natural language description is converted into a structured JSON format to obtain structured information; Based on the structured information, a specific query statement is generated and executed to obtain the query results; The query results are evaluated and sorted to obtain a ranking result based on scores. The recommended companies are displayed based on their scores, ranked from highest to lowest.

2. The intelligent recommendation method for enterprise information according to claim 1, characterized in that, The process of acquiring enterprise-related data and creating an enterprise feature map containing a multi-level classification system includes: Acquire relevant enterprise data, establish a top-level classification of enterprise characteristics, and construct a systematic classification structure; Based on each category in the classification structure, specific enterprise attribute dimensions are clearly defined, and semantic definitions and data types are set for each feature; Create a thesaurus for feature terms; A standard set of enumerated values ​​is developed for discrete features, and synonym processing is performed on the enumerated values ​​to identify and define the mutual exclusion relationships between feature values ​​in order to obtain the enterprise feature map.

3. The intelligent recommendation method for enterprise information according to claim 1, characterized in that, The step of establishing a commodity coding table based on the enterprise feature map includes: Based on the invoice data and expert experience in the enterprise feature map, a hierarchical commodity coding table containing multi-dimensional commodity information is constructed.

4. The intelligent recommendation method for enterprise information according to claim 1, characterized in that, The intelligent rewriting of the current input query, and the identification of the query object and its business relationships to obtain the rewriting intent, includes: Analyze the current input query and historical dialogue information to determine the query category of the current input query, so as to obtain the first analysis result; Identify the core object of the current input query, and determine the business relationship between the user and the core object based on user information to obtain the second analysis result; The relative and vague time expressions in the current input query are converted into absolute and specific time ranges or specific dates to obtain the third analysis result; The first analysis result, the second analysis result, and the third analysis result are integrated to generate the rewriting intent.

5. The intelligent recommendation method for enterprise information according to claim 1, characterized in that, The process of converting the rewriting intent into standardized query conditions based on the enterprise feature map and the product coding table includes: Based on the enterprise feature map and product coding table, the product and enterprise features involved in the rewriting intent are identified, matched with the corresponding entries in the database, and a corresponding query condition structure is generated.

6. The intelligent recommendation method for enterprise information according to claim 5, characterized in that, Based on the enterprise feature map and product coding table, the process identifies the product and enterprise features involved in the rewriting intent, matches them to corresponding entries in the database, and generates a corresponding query condition structure, including: Combining the rewriting intent with the enterprise feature map and commodity coding table, prompt words are generated; the enterprise characteristics involved in the rewriting intent are identified and verified and supplemented through the enterprise feature map; the specific numerical values, text, or time values ​​in the current input query are obtained, and unit conversion and expression processing are performed; the comparison operators and special operation types in the current input query are analyzed and determined; the logical relationships and nested structures in the current input query are understood and constructed to obtain parsed information; The parsed information is organized into a tree structure to ensure that the logical and conditional nodes correctly express the query intent, thus obtaining the query condition structure.

7. The intelligent recommendation method for enterprise information according to claim 1, characterized in that, The process of generating a specific query statement based on the structured information and executing the query to obtain the query results includes: The structured information is converted into an Elasticsearch query DSL according to predefined rules; Based on the operator type of the node, the operator type is mapped to the corresponding content in the Elasticsearch query DSL. For non-equality comparison conditions, the value is excluded by nesting terms in bool.must_not. The match or wildcard query type is used to support flexible fuzzy search of text. The sort and size parameters are used to control the sorting method of the results and the number of returned results to meet the TopN / BottomN requirements. The logic tree is traversed and the ES query DSL is generated through a Python recursive function. At the same time, the requirements on the root node are processed to obtain the specific query statement. The generated query statement is executed, and a list of companies that meet the criteria is obtained from the company profile table to obtain the query results.

8. An intelligent recommendation device for enterprise information, characterized in that, include: The graph creation unit is used to acquire enterprise-related data and create an enterprise feature graph that includes a multi-level classification system for enterprises. The coding table establishment unit is used to establish a commodity coding table based on the enterprise feature map. The recording unit is used to acquire enterprise information and user preferences from different sources, build data models, and record historical dialogues and current input queries; The rewriting unit is used to intelligently rewrite the current input query and identify the query object and its business relationship to obtain the rewriting intent; The condition conversion unit is used to convert the rewriting intent into standardized query conditions based on the enterprise feature map and the commodity code table. The format conversion unit is used to convert the natural language description into a structured JSON format based on the standardized query conditions to obtain structured information. The query unit is used to generate a specific query statement based on the structured information and execute the query to obtain the query result; A sorting unit is used to evaluate and sort the query results to obtain a ranking result based on scores. The display unit is used to display detailed information about the recommended companies based on the ranking results of the scores.

9. A computer device, characterized in that, The computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the method as described in any one of claims 1 to 7.

10. A storage medium, characterized in that, The storage medium stores a computer program that, when executed by a processor, implements the method as described in any one of claims 1 to 7.

Images