Enterprise-level recommendation method, system and equipment based on large language model and medium

Abstract

The invention discloses an enterprise-level recommendation method, system and device based on a large language model and a medium, and relates to the technical field of information recommendation. The method comprises the steps of obtaining a login ID and a query request of a user, and obtaining an enterprise identity of the user and historical behavior characteristics of the user based on the login ID; determining a user portrait of the user based on the historical behavior characteristics and the enterprise identity; determining a keyword corresponding to the user based on the user portrait; converting the keyword and the query request into a target vector; and searching in a preset vector database based on the target vector to obtain recommendation information, and converting the recommendation information into a visual recommendation page for display. By implementing the technical scheme provided by the invention, the deviation between the pushed information and the actual demand of an enterprise can be reduced, and the service effectiveness of enterprise-level information recommendation is improved.

Description

Technical Field

[0001] The present application relates to the field of information recommendation technology, and in particular to an enterprise-level recommendation method, system, device and medium based on a large language model. Background Art

[0002] As the pace of scientific and technological innovation accelerates and industrial transformation deepens, companies are increasingly demanding information on scientific and technological intelligence, market trends, and other areas. In particular, timely and accurate access to technological development trends, industry chain changes, and market opportunities in related fields is crucial for companies' strategic decision-making and innovative development.

[0003] Currently, the primary way for enterprises to obtain intelligence information is through various information retrieval and recommendation systems. These systems typically use keyword matching to retrieve information from various data sources and filter and sort the results based on pre-set rules. However, with the development of the internet and big data technologies, the amount of accessible information has grown exponentially. Enterprise users often receive a large amount of redundant or irrelevant information. Existing systems provide superficial analysis of enterprise users, resulting in a significant mismatch between the information pushed and the actual needs of enterprises, leading to low effectiveness of enterprise-level information recommendation services. Summary of the Invention

[0004] This application provides an enterprise-level recommendation method, system, device and medium based on a large language model, which can reduce the deviation between pushed information and the actual needs of the enterprise and improve the service effectiveness of enterprise-level information recommendation.

[0005] In a first aspect, the present application provides an enterprise-level recommendation method based on a large language model, the method comprising: Obtaining the user's login ID and query request, and obtaining the user's corporate identity and the user's historical behavior characteristics based on the login ID; Determine a user profile of the user based on the historical behavior characteristics and the corporate identity; Determining keywords corresponding to the user based on the user portrait; Converting the keyword and the query request into a target vector; A search is performed in a preset vector database based on the target vector to obtain recommendation information, and the recommendation information is converted into a visual recommendation page for display.

[0006] By adopting the above technical solution, the user's corporate identity and historical behavioral characteristics can be obtained based on the user's login ID, and then the information from these two dimensions can be combined to construct a user profile, allowing the system to simultaneously grasp the user's personal behavioral characteristics and the business characteristics of the company to which they belong. On this basis, the user's corresponding keywords are determined by analyzing the user profile, and the keywords and query requests are converted into target vectors, achieving a precise mapping from user needs to vector space. Finally, based on the target vector, the vector database is searched and a visual recommendation page is generated, which not only ensures the relevance of the recommendation results to user needs, but also provides a good information display experience. This recommendation method, which combines user personal characteristics with corporate attributes, effectively reduces the deviation between pushed information and the actual needs of the enterprise, significantly improving the service effectiveness of enterprise-level information recommendation.

[0007] In a second aspect of the present application, an enterprise-level recommendation system based on a large language model is provided, the system comprising: A behavior feature acquisition module, configured to acquire a user's login ID and query request, and acquire the user's corporate identity and the user's historical behavior features based on the login ID; A user portrait generation module, configured to determine a user portrait of the user based on the historical behavior characteristics and the corporate identity; A keyword determination module, configured to determine keywords corresponding to the user based on the user portrait; a target vector conversion module, configured to convert the keyword and the query request into a target vector; The recommendation information display module is used to search a preset vector database based on the target vector to obtain recommendation information, and convert the recommendation information into a visual recommendation page for display.

[0008] In a third aspect of the present application, a computer storage medium is provided. The computer storage medium stores a plurality of instructions, and the instructions are suitable for being loaded by a processor and executing the above method steps.

[0009] In a fourth aspect of the present application, an electronic device is provided, comprising: a processor and a memory; wherein the memory stores a computer program, and the computer program is suitable for being loaded by the processor and executing the above-mentioned method steps.

[0010] In summary, one or more technical solutions provided in the embodiments of the present application have at least the following technical effects or advantages: This application can obtain the user's corporate identity and historical behavior characteristics based on the user's login ID, and then combine the information of these two dimensions to build a user portrait, so that the system can simultaneously grasp the user's personal behavior characteristics and the business characteristics of the company to which he belongs; on this basis, by analyzing the user portrait, the corresponding keywords of the user are determined, and the keywords and query requests are converted into target vectors, realizing the precise mapping from user needs to vector space; finally, based on the target vector, the vector database is searched and a visual recommendation page is generated, which not only ensures the relevance of the recommendation results to user needs, but also provides a good information display experience. This recommendation method that combines user personal characteristics with corporate attributes effectively reduces the deviation between the pushed information and the actual needs of the enterprise, and significantly improves the service effectiveness of enterprise-level information recommendation. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 This is a flowchart of an enterprise-level recommendation method based on a large language model provided in an embodiment of the present application; Figure 2 This is a module diagram of an enterprise-level recommendation system based on a large language model provided by an embodiment of the present application; Figure 3 This is a structural diagram of an electronic device provided in an embodiment of the present application.

[0012] Description of reference numerals: 300, electronic device; 301, processor; 302, communication bus; 303, user interface; 304, network interface; 305, memory. DETAILED DESCRIPTION

[0013] In order to enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below in conjunction with the drawings in the embodiments of this specification. Obviously, the described embodiments are only part of the embodiments of this application, not all of the embodiments.

[0014] In the description of the embodiments of this application, words such as "for example" or "for instance" are used to indicate examples, illustrations, or explanations. Any embodiment or design described as "for example" or "for instance" in the embodiments of this application should not be construed as being preferred or advantageous over other embodiments or designs. Rather, the use of words such as "for example" or "for instance" is intended to present the relevant concepts in a concrete manner.

[0015] In the description of the embodiments of the present application, the term "multiple" means two or more. For example, multiple systems refer to two or more systems, and multiple screen terminals refer to two or more screen terminals. In addition, the terms "first" and "second" are used for descriptive purposes only and are not to be understood as indicating or implying relative importance or implicitly indicating the indicated technical features. Thus, the features defined as "first" and "second" may explicitly or implicitly include one or more of the features. The terms "including", "comprising", "having" and their variations all mean "including but not limited to", unless otherwise specifically emphasized.

[0016] The following will provide a clear and complete description of the technical solutions in the embodiments of the present application in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments.

[0017] Please refer to Figure 1 , a flowchart of an enterprise-level recommendation method based on a large language model is proposed. The method can be implemented by a computer program, a single-chip microcomputer, or run on an enterprise-level recommendation system based on a large language model. The computer program can be integrated into a computer device or run as an independent tool application. Specifically, the method includes steps 10 to 50, which are as follows: Step 10: Obtain the user's login ID and query request, and obtain the user's corporate identity and historical behavior characteristics based on the login ID.

[0018] In the embodiment of this application, the login ID refers to a unique identity identifier assigned in the enterprise-level information service platform, which is used to associate with the enterprise user management system to obtain basic information such as the user's organizational affiliation and responsibilities and roles.

[0019] In the embodiment of the present application, a query request refers to the information demand expression input or selected by the user through the system interface, which can be in the form of keywords, phrases or complete questions, and is used to convey the user's specific information retrieval intention to the system.

[0020] In the embodiment of the present application, historical behavior characteristics refer to the user's search history, content interaction, evaluation feedback and other operation trajectory data recorded by the system over the past period of time, reflecting the user's information consumption habits and professional interest preferences.

[0021] Specifically, the system receives the user's login credentials through a unified authentication interface and returns a login ID upon successful authentication. Simultaneously, the system retrieves the user's input or selected query request through the system's search input box or pre-set query criteria selection interface. After obtaining the login ID, the system accesses the enterprise user management database to extract the corresponding enterprise organizational information, including department affiliation, job responsibilities, and other information, thereby determining the user's corporate identity. For example, if the user with login ID "E2023001" belongs to the R&D department, the system will identify them as a R&D employee. While obtaining the corporate identity, the system also queries the user behavior log database to extract the login ID's activity history from the past three months. This data includes the user's search terms, browsed content, and duration of stay on the platform. The system processes this raw behavior data using a behavioral feature extraction model to extract the user's historical behavioral characteristics. For example, by analyzing a user's frequent searches for "semiconductor materials" and their high frequency of access to technical patent literature, the system can identify the user as exhibiting a sustained interest in the field of semiconductor technology.

[0022] Based on the above embodiment, as an optional embodiment, the step of obtaining the user's corporate identity and the user's historical behavior characteristics based on the login ID may further include the following steps: Step 101: query the preset user database based on the login ID to obtain basic information of the company to which the user belongs.

[0023] Specifically, it is first necessary to obtain the basic information of the enterprise to which the user belongs, because the basic information of the enterprise is the basic data for building the user's corporate identity. The system uses the login ID as the query key to access the preset user database, which adopts a relational database architecture and stores the related information between users and enterprises. Through SQL query statement retrieval, you can obtain the basic information of the enterprise, including enterprise registration information, organizational structure information, business areas, etc. For example, when a user with a login ID of "E2023001" is queried, basic information such as the registered capital, establishment time, and main business direction of the "Technology Co., Ltd." to which he belongs can be obtained. This database-based rapid retrieval method provides complete data support for subsequent corporate identity extraction Step 102: Extract the enterprise identity from the basic information, and obtain the corresponding enterprise knowledge graph based on the enterprise identity. The enterprise knowledge graph includes the enterprise's main product information, technical patent information, upstream and downstream industry chain information and enterprise customer information.

[0024] Specifically, after obtaining the basic information of the enterprise, the system needs to further extract the enterprise identity and obtain the corresponding enterprise knowledge graph. This is because the enterprise knowledge graph contains a complete portrait of the enterprise in the industry. In specific implementation, the system first analyzes the basic information of the enterprise through the rule engine, extracts key elements such as the industry attributes, technical fields, and market positioning of the enterprise, and forms an enterprise identity. Then, based on the enterprise identity, the system accesses the pre-built enterprise knowledge graph database, which adopts a graph database architecture and extracts relevant node information through a graph query language. For example, the system can obtain information such as the core patent layout of the enterprise in the field of semiconductor materials, the cooperative relationship with upstream raw material suppliers, and the application scenarios of downstream end customers. This information association method based on the knowledge graph enables the system to fully grasp the position and characteristics of the enterprise in the industrial ecosystem.

[0025] Step 103: Obtain the user's historical behavior data within a preset time period based on the login ID, and determine the user's historical behavior characteristics based on the enterprise knowledge graph and the historical behavior data.

[0026] Specifically, the system obtains historical user behavior data within a preset time period based on the login ID and segments this data into time segments. Using a sliding window approach, the system divides the past three months' behavior data into multiple time windows, each spanning one week, with adjacent windows overlapping by three days. This segmentation approach maintains the temporal continuity of the behavior sequence while capturing the evolution of user behavior patterns. For example, for a user with the login ID "E2023001," their 12-week historical data will be divided into approximately 20 overlapping time windows, each containing the user's complete behavior record for a week.

[0027] From the segmented behavior sequences, the system needs to extract the user's query keyword set and click content set. Specifically, the system uses a natural language processing model to segment and tag search records within each time window, extracting keywords with actual search intent. Simultaneously, topic modeling is performed on the document content clicked by the user to extract the document's core subject terms and key concepts. This yields a query keyword set that reflects the user's information needs and a click content set that reflects their interests.

[0028] We extract various attribute information about enterprise nodes from the knowledge graph, including the technical parameters and performance indicators of core products, the innovative features and application areas of technical patents, the cooperation methods and business models of upstream and downstream industry chains, and the needs and application scenarios of enterprise customers. This information is then converted into structured feature vectors through feature engineering, forming a complete enterprise feature set that includes the core product feature set, the technical patent feature set, the industry chain feature set, and the enterprise customer feature set.

[0029] The system uses a semantic similarity calculation model to calculate the similarity score between query keywords and clicked content, and company features. For each feature pair, if the similarity score exceeds a preset threshold, it is considered a valid match. For example, if users frequently search for "gallium nitride substrates" and this term is highly correlated with semiconductor materials products in the company's core product feature set, it will be marked as a matching feature.

[0030] Finally, the system uses a graph traversal algorithm to retrieve the associated paths for each matching feature in the enterprise knowledge graph and calculate the depth and strength of the paths. Only features that can find a clear associated path in the knowledge graph are retained as historical behavioral features of the user. This knowledge graph-based feature screening method ensures that the extracted behavioral features not only reflect the user's personal interests but are also closely linked to the company's business scenarios. For example, if a user's continued attention to a certain technical patent happens to be on the company's key technology route, this feature has higher business relevance.

[0031] Step 20: Determine the user profile based on historical behavioral characteristics and corporate identity.

[0032] Specifically, in an optional embodiment, a two-layer feature fusion network can be constructed. In the first layer of feature processing, the system classifies historical behavioral features according to the main product dimension, technical patent dimension, industrial chain dimension, and corporate customer dimension, and calculates the importance weight of the features of each dimension through the attention mechanism. For example, when it is found that the user's behavioral characteristics in the technical patent dimension have high time persistence and content relevance, the system will give this dimension a higher weight. At the same time, the system converts the corporate identity into a vector representation, which contains information such as the user's department, scope of responsibilities, and technical field.

[0033] In the second layer of feature fusion, the system uses a deep neural network model to combine weighted historical behavior feature vectors with the company identity vector. Using a multi-layer perceptron, the system performs feature interaction and nonlinear transformations to generate a fused feature vector. This fusion process considers the mutual influence between features. For example, the user's R&D department identity information will strengthen the weight of their behavioral features in the technology patent dimension.

[0034] Based on the fused feature vectors, the system further constructs a user profile tagging system. This system consists of three layers: the first layer is the professional role tag, which reflects the user's functional position within the enterprise, such as "Technical R&D Engineer"; the second layer is the professional field tag, which describes the user's technical expertise and focus, such as "Semiconductor Materials R&D" and "Process Optimization"; and the third layer is the information need tag, which describes the user's specific information consumption preferences, such as "Frontier Technology Trends" and "Patent Analysis Reports."

[0035] The system uses a probabilistic graph model to calculate the probability of matching a user with each layer of tags, selecting the tag combination with the highest probability to form the final user profile. For example, for a user who frequently views patent literature related to semiconductor materials and works in the R&D department, their user profile might include a tag combination such as "Technical R&D Engineer - Semiconductor Materials R&D - Patent Analysis." This multi-layered tagging system can characterize users from different dimensions, making the user profile more three-dimensional and complete.

[0036] Based on the above embodiment, as another optional embodiment, the step of determining the user profile of the user based on historical behavior characteristics and corporate identity may further include the following steps: Step 201: Classify historical behavior features to obtain search interest features, browsing habit features, and interaction behavior features.

[0037] Specifically, before constructing a user profile, historical behavior characteristics must first be categorized, as different types of behavior characteristics reflect the user's different information needs. The system employs a feature classification model to categorize historical behavior characteristics into three categories: Search interest characteristics reflect the user's active search focus, derived by analyzing the semantic content and frequency of search keywords; Browsing habit characteristics reflect the user's content preferences, derived by counting document types, reading duration, and access frequency; and Interaction behavior characteristics reflect the user's engagement, extracted by recording actions such as collections, comments, and downloads. For example, the system may detect that a user frequently searches for content related to "gallium nitride," prefers to browse technical analysis reports, and frequently downloads related patent documents.

[0038] Step 202: Combine the search interest features, browsing habit features, and interaction behavior features to generate a user behavior profile.

[0039] Specifically, after obtaining the classified behavioral features, these features need to be organically combined to form a user behavior profile, because a single type of feature cannot fully describe the user's behavior pattern. The system uses a feature fusion network to first calculate the weight of each type of feature. The weight value is determined based on the timeliness, frequency, and user feedback intensity of the feature. Then, the weighted feature vectors are nonlinearly combined through a deep neural network to generate a unified behavioral feature representation. For example, when it is found that the user's search keywords, browsing content, and interactive behaviors are all concentrated in the field of semiconductor materials, the system will strengthen the feature weight of this field in the user behavior profile. This feature combination method can fully reflect the user's behavioral preferences on the platform.

[0040] Step 203: Generate an enterprise attribute portrait based on the enterprise feature set corresponding to the enterprise identity.

[0041] Specifically, based on the user's corporate identity, the system needs to generate a corporate attribute portrait that reflects their professional background. This is because the corporate feature set contains information about the user's business environment. In specific implementation, the system first extracts attribute information related to the user's corporate identity from the corporate feature set, including the business scope of the department to which they belong, job responsibilities, technical field divisions, etc. Then, through the attribute mapping model, these discrete attribute information are converted into structured feature vectors to form a corporate attribute portrait. For example, for technical personnel in the R&D department, their corporate attribute portrait will highlight professional characteristics such as technological innovation and patent analysis. This corporate identity-based portrait construction method can accurately locate the user's professional background and job responsibilities.

[0042] Step 204: Integrate the user behavior profile and the enterprise attribute profile to obtain a user profile of the user.

[0043] Specifically, because a user's information needs are influenced by both their personal behavior and their corporate role, the system employs a multimodal feature fusion algorithm. First, the correlation matrix between the two profile features is calculated to identify mutually supportive and complementary feature combinations. Then, an attention mechanism dynamically adjusts the feature weights, prioritizing combinations that demonstrate significant characteristics across both the behavioral and attribute dimensions. Finally, the fused feature vectors are mapped to a pre-set labeling system to generate a multidimensional user profile. For example, if a user's behavioral profile indicates a strong interest in semiconductor materials, while their corporate profile indicates they are responsible for materials innovation in R&D, the system will highlight the "Semiconductor Materials R&D Expert" feature label in the user profile. This approach to constructing a profile that integrates behavioral and corporate attributes more accurately depicts a user's expertise and information needs within a specific corporate context, providing more precise user profile support for subsequent personalized recommendations.

[0044] Step 30: Determine the keywords corresponding to the user based on the user portrait.

[0045] Specifically, the system first performs semantic analysis on the various dimensions of the user profile to construct a multi-layered keyword generation network. At the first layer, high-frequency search terms and key technical terms are extracted from the user's behavior profile, directly reflecting the user's active information needs. For example, if a user frequently searches for "gallium nitride power devices," this term will be prioritized for inclusion in the keyword set. At the second layer, topic modeling analyzes user browsing and interaction behavior, extracting key terms from frequently visited documents. At the third layer, based on the enterprise attribute profile, terms related to the user's professional identity are extracted from the enterprise knowledge graph.

[0046] Candidate keywords are scored for importance, taking into account factors such as word frequency, timeliness, and professional relevance. A word embedding model is used to calculate semantic similarity between keywords. Similar keywords are merged or filtered, and ultimately 10-20 core keywords with the highest scores are selected to form a keyword set. Each keyword is assigned a weight. For example, for semiconductor material R&D personnel, their keyword set might include: "gallium nitride material (weight 0.9)" and "epitaxial growth process (weight 0.8)." This keyword identification method provides reliable feature support for subsequent precise information recommendations.

[0047] Based on the above embodiment, as another optional embodiment, the step of determining the keywords corresponding to the user based on the user portrait may further include the following steps: Step 301: Extract core entity words from the user portrait and construct an entity relationship tree.

[0048] Specifically, the system uses an entity recognition model to extract core entity terms from user profiles. These entity terms are primarily derived from user behavioral characteristics and company attributes. For example, for an R&D engineer specializing in semiconductor materials, the system extracts core entity terms such as "gallium nitride," "power devices," and "epitaxial growth." The system then constructs an entity relationship tree based on these entity terms, establishing a hierarchical structure between entities using a depth-first search algorithm, and uses a graph database to store these relationships.

[0049] Step 302: Search for hypernyms and hyponyms in a preset professional knowledge base based on the entity relationship tree.

[0050] Specifically, based on the construction of the entity relationship tree, the system searches for hypernyms and hyponyms within a pre-set professional knowledge base. This knowledge base contains structured knowledge such as standard technical vocabulary and patent classification systems. Using a semantic similarity calculation model, the system performs upward and downward hierarchical searches for each node in the entity relationship tree. For example, for "gallium nitride," the hypernyms retrieved are "compound semiconductors" and "wide bandgap semiconductors," and hyponyms such as "vertical gallium nitride" and "lateral gallium nitride." This hierarchical search expands the technical dimension coverage of keywords.

[0051] Step 303: semantically expand the retrieved phrase to obtain an expanded phrase, which includes a synonym set and a related word set.

[0052] Specifically, the system uses a word vector model to calculate word semantic similarity and classifies phrases whose similarity exceeds a preset threshold into a synonym set, such as "GaN" and "gallium nitride", "power devices" and "power electronic devices", etc. At the same time, the system identifies technical terms related to the core words based on co-occurrence frequency and contextual relevance analysis to form a set of related words. For example, related words for "gallium nitride" include process-related terms such as "epitaxial", "substrate", and "doping". This two-way semantic expansion ensures that keywords have sufficient retrieval coverage while maintaining professional accuracy. For example, when a user searches for "gallium nitride" related technologies, the system can simultaneously retrieve related documents using different expressions, thereby improving the recall rate of the retrieval.

[0053] Step 304: Cross-validate the synonym set and the related word set with the scenario information in the user portrait, and combine the expanded word groups based on the cross-validation results to generate multi-dimensional keyword groups.

[0054] Specifically, contextual information is extracted from user profiles, including their technical responsibilities and professional background. For example, for a process engineer in a semiconductor R&D department, contextual information includes key scenarios such as "process development," "yield improvement," and "mass production introduction." A semantic matching model is used to calculate the relevance score between the expanded phrases and the contextual information. For each expanded phrase, the system compares its feature vector with the feature vector of the contextual information to determine a matching score. For example, the phrase "epitaxial growth" has a high correlation with the "process development" context, but a low correlation with the "market analysis" context. Based on these matching scores, the system reorganizes the expanded phrases, categorizing and combining highly relevant phrases by technical dimensions (such as materials, processes, devices, and applications) to form multi-dimensional keyword groups. For example, synonyms such as "gallium nitride," "GaN," and "III-nitride" are combined in the material dimension, while related terms such as "epitaxial," "MOCVD," and "growth process" are combined in the process dimension.

[0055] Step 305: extracting phrases that have a business relevance to the user from the multi-dimensional keyword group as keywords corresponding to the user.

[0056] Specifically, the system uses a deep learning model, taking the user's historical behavior data (such as search history and browsing history) and corporate identity information as input, to calculate the business relevance score of each phrase to the user. For example, if the user is primarily responsible for process development of power devices, phrases related to power device processes will receive a higher relevance score. The system then selects highly relevant phrases based on a preset threshold and uses them as the final user keywords. The keywords generated in this way maintain professional accuracy while being highly relevant to the user's actual business needs, effectively supporting subsequent information retrieval and recommendation services. For example, for a power device process engineer, the final keywords may include technical terms directly related to their work, such as "vertical gallium nitride," "trench gate process," and "source field plate structure."

[0057] Step 40: Convert keywords and query requests into target vectors.

[0058] Specifically, keywords and query requests are preprocessed. For keywords, the system retains their original weights, such as "gallium nitride material (weight 0.9)" and "epitaxial growth process (weight 0.8)." For queries, the system uses a natural language processing model to perform word segmentation and part-of-speech tagging, extracting core terms and semantic structures. For example, when a user enters a query like "Searching for the latest packaging solutions for gallium nitride power devices," the system identifies key terms such as "gallium nitride," "power devices," and "packaging solutions." The system then uses a pretrained domain-specific word embedding model to convert the keywords and query terms into high-dimensional vector representations. This word embedding model, trained on a large body of professional literature (such as patents and technical papers), accurately represents the semantic relationships between technical terms. For example, synonyms like "gallium nitride" and "GaN" are closely spaced in the vector space, while maintaining a moderate semantic distance from other semiconductor materials like "silicon."

[0059] For queries containing multiple terms, the system uses an attention mechanism to weight the individual term vectors to generate an overall query vector. This weighting takes into account the importance of the terms in the query, such as giving higher weights to specialized terms and lower weights to more general terms. Furthermore, the system maintains the query vector and keyword vectors in the same vector space, ensuring that similarity calculations can be performed directly between them.

[0060] Ultimately, the system obtains a set of standardized vector representations, including a set of keyword vectors representing the user's professional characteristics and a query vector representing the current information need. These vectors are high-dimensional vectors of 768 or 1024 dimensions, with each dimension carrying specific semantic information. This vectorized representation provides the mathematical foundation for subsequent similarity calculations and information retrieval, enabling more accurate matching of user needs with relevant information. For example, the system can calculate the cosine similarity between the query vector and the document vector to quickly find the most relevant technical documents.

[0061] Based on the above embodiment, as another optional embodiment, the step of converting keywords and query requests into target vectors may further include the following steps: Step 401: Perform semantic analysis on keywords and query requests to obtain semantic units.

[0062] Specifically, a deep learning model is used to perform semantic analysis on the input text, breaking it down into its smallest semantic units. For example, for the query "Develop a new gate structure for high-efficiency gallium nitride power devices," the system uses specialized word segmentation and part-of-speech tagging models to identify semantic units such as "high efficiency" (an attribute term), "gallium nitride" (a material term), "power device" (a device term), and "gate structure" (a structure term). For each semantic unit, the system also annotates its part-of-speech information and semantic role information, providing a foundation for subsequent dependency analysis.

[0063] Step 402: Construct a semantic dependency tree based on the hierarchical relationship and dependency relationship between the semantic units.

[0064] Specifically, a semantic dependency tree is constructed based on the identified semantic units. The system uses a dependency syntax analysis model to analyze the hierarchical and dependency relationships between semantic units. During the construction process, the system first identifies the core semantic unit. For example, in this example, "gate structure" is the core node as the research object. Then, the system identifies the relationship between other semantic units and the core node. For example, "gallium nitride" and "power device" form a composite modification relationship, jointly describing the technical field, and "high efficiency" serves as an attribute modifier to modify "power device." The system uses a tree data structure to store these relationships. Each node contains the text content, part-of-speech tag, and semantic role of the semantic unit, and the edges between nodes represent the dependency relationship type.

[0065] Step 403: Reorganize the semantic units based on the semantic dependency tree to obtain a semantic structured expression.

[0066] Specifically, based on the constructed semantic dependency tree, the system reorganizes the semantic units to generate standardized semantic structured expressions. The reorganization process follows the predefined technical text semantic template to ensure that the reorganized expression conforms to the description specifications of the professional field. For example, the system reorganizes the above query into a structured form of "{Technical field: Gallium nitride power devices} {Technical features: high efficiency} {Technical solution: gate structure}". This reorganization method enables queries with different expressions but similar semantics to be mapped to similar structured expressions, improving the accuracy of subsequent retrieval. For example, a query such as "Study gate design to improve efficiency in gallium nitride devices" will obtain a similar structured expression after processing. Through this semantic analysis and structured processing, the system can more accurately understand the user's professional retrieval needs and provide a standardized semantic basis for subsequent vectorization conversion.

[0067] Step 404: Input the semantic structured expression into a preset text encoder to obtain a semantic vector.

[0068] Specifically, the semantically structured expression is input into a pre-set text encoder, a deep learning model pre-trained and fine-tuned with specialized domain data. For example, for the structured expression "{Technical field: GaN power devices} {Technical features: High efficiency} {Technical solution: Gate structure}," the system uses a modified BERT model as the encoder. This model has been trained on a large number of patents and technical documents and possesses specialized domain semantic understanding capabilities. During the encoding process, the system converts the structured expression into the model's input format, including adding special markers and positional encoding. It then encodes the expression through a multi-layer Transformer structure, ultimately obtaining a 768-dimensional base semantic vector from the model's [CLS] position.

[0069] Step 405: Perform semantic enhancement on the semantic vector to generate an enhanced vector, and use the enhanced vector as the target vector.

[0070] Specifically, the basic semantic vector undergoes multi-dimensional semantic enhancement. First, expertise enhancement: The system retrieves core concepts and relationships related to the current technical topic from a pre-set expertise database. For example, it retrieves technical parameters (such as breakdown voltage and on-resistance) and typical structural features (such as vertical and horizontal structures) related to "GaN power devices." The system converts this expertise into an enhanced knowledge vector using a knowledge encoder. Second, context enhancement: The system considers the user's professional background (e.g., device R&D engineer) and past query history to generate a context vector that reflects the user's professional perspective. The system uses an attention mechanism to fuse the basic semantic vector, the enhanced knowledge vector, and the context vector. The system calculates the relevance score of each vector to the current query topic and uses these scores as fusion weights. For example, for a query on gate structure, the system may assign a higher weight to knowledge vectors related to device structure. Through weighted summation and normalization, the system generates a final enhanced vector. This vector has the same dimensionality as the basic semantic vector but contains richer semantic information. This semantically enhanced vector representation has many advantages: first, it retains the core semantic information of the original query; second, through knowledge enhancement, the vector contains relevant professional and technical information, so that subsequent searches can find documents with higher technical relevance; finally, through context enhancement, the vector can reflect the user's professional perspective and demand characteristics. For example, when performing similarity retrieval, the system can not only match documents containing similar gate structures, but also consider whether the technical parameter indicators in the document meet the user's needs and whether the technical difficulty of the document is suitable for the user's professional level. This enhanced vector is ultimately used as the target vector for subsequent document retrieval and similarity calculation, which can significantly improve the professional relevance and practicality of the retrieval results.

[0071] Step 50: Search the preset vector database based on the target vector to obtain recommendation information, and convert the recommendation information into a visual recommendation page for display.

[0072] Specifically, in a preset vector database, an approximate nearest neighbor algorithm is used to calculate the similarity between the target vector and the document vector, and the document with the highest similarity is selected as the candidate recommendation set. The system then sorts the candidate set in multiple dimensions, including technical relevance sorting, timeliness sorting, and user matching sorting, to ensure the quality of the recommendation results. The system converts the filtered recommendation information into a structured visual page, using a multi-level display method: the top level displays the technical field knowledge map, intuitively showing the technical connections between documents; the middle level presents the document list, including core information such as title and abstract; the bottom level provides detailed technical analysis content. The page integrates interactive components such as technical feature filters, timelines, and performance indicator radar charts to facilitate users to filter documents and compare solutions.

[0073] The system also provides intelligent assistance functions, automatically extracting and highlighting key technical features, and adjusting recommendation strategies in real time based on user browsing feedback. This recommendation method accurately captures technical semantic associations, ensuring that the recommended results are both professionally relevant and practical. At the same time, through intuitive visual displays, it helps users quickly understand and compare different technical solutions, improving information acquisition efficiency. For example, when a user queries for gate structures, the system not only recommends relevant documents but also visually displays the performance comparison of different structures, assisting users in selecting the most suitable technical solution.

[0074] Based on the above embodiment, as another optional embodiment, the step of searching a preset vector database based on the target vector to obtain recommendation information and converting the recommendation information into a visual recommendation page for display may further include the following steps: Step 501: Search a preset vector database based on a target vector to obtain a candidate information set.

[0075] Specifically, the system searches a pre-set vector database based on the target vector, calculates vector similarity using an improved approximate nearest neighbor algorithm, and selects documents with a similarity exceeding a preset threshold (e.g., 0.7) as candidate information sets. For example, for a query related to gallium nitride devices, the system might retrieve hundreds of highly relevant technical documents.

[0076] Step 502: Determine the basic priority of each candidate information in the candidate information set according to the information type and timeliness.

[0077] Specifically, the candidate information is then evaluated for basic priority. This evaluation considers the importance of the information type (e.g., granted patents take precedence over patent applications) and timeliness (e.g., documents published within the past year receive a timeliness weighting). The system uses pre-set scoring rules, such as the product of the patent type weight (granted patents 1.2, patent applications 1.0) and the timeliness weight (within one year 1.2, one to two years 1.1, more than two years 1.0) to calculate the basic priority score.

[0078] Step 503: Calculate the click rate and conversion rate of each candidate information based on the historical behavior characteristics and generate an interaction priority.

[0079] Specifically, the system calculates interaction priority based on historical user behavior data. The system analyzes user groups' historical clickthrough and conversion data for different document types and uses a logistic regression model to predict the expected click-through rate and conversion rate for each candidate piece of information. For example, the system might find that the average click-through rate and conversion rate for a certain type of technical solution are 15% and 30%, respectively. The system then calculates an interaction priority score based on this. The calculation formula is: Interaction priority = 0.4 × expected click-through rate + 0.6 × expected conversion rate.

[0080] Step 504: Perform weighted fusion of the basic priority and the interactive priority to obtain a comprehensive priority score.

[0081] Specifically, a weighted fusion of basic and interactive priorities is performed to generate a composite priority score. This fusion employs a dynamic weighting mechanism, adjusting the weight ratio of the two priorities based on user expertise and usage scenarios. For example, R&D personnel are given a higher weight for basic priorities (0.7:0.3), while product managers are given a higher weight for interactive priorities (0.3:0.7).

[0082] Step 505: sort the candidate information in layers according to the comprehensive priority score, construct a hierarchical recommendation page, and display the recommendation page visually. The recommendation page includes a top display area, a key recommendation area, and a general display area.

[0083] Specifically, a hierarchical recommendation page is constructed based on the comprehensive priority score. The top 5% of information with the highest scores is placed in the top display area, highlighting its innovative features and core technical features. The next 15% of information is placed in the key recommendation area, presenting technical solutions with a combination of pictures and text. The remaining information is placed in the general display area, presenting basic information in a list format. Each display area is equipped with corresponding interactive components, such as a technical key point navigation in the top area, a solution comparison tool in the key area, and a quick filter in the general area. This layered display method ensures the eye-catching presentation of important information while maintaining the integrity of information acquisition, effectively improving users' information acquisition efficiency. For example, when browsing gallium nitride device technical solutions, users can quickly focus on the breakthrough technologies in the top area while simultaneously understanding the overall development of the technology through other display areas.

[0084] See Figure 2 , is a module diagram of an enterprise-level recommendation system based on a large language model provided in an embodiment of the present application, wherein the system includes: The behavior feature acquisition module is used to obtain the user's login ID and query request, and obtain the user's corporate identity and historical behavior features based on the login ID; User profile generation module, used to determine the user profile of a user based on historical behavioral characteristics and corporate identity; Keyword determination module, used to determine the keywords corresponding to the user based on the user portrait; A target vector conversion module, used to convert keywords and query requests into target vectors; The recommendation information display module is used to search the preset vector database based on the target vector to obtain recommendation information, and convert the recommendation information into a visual recommendation page for display.

[0085] Optionally, the behavior feature acquisition module is further configured to query a preset user database based on the login ID to obtain basic information of the enterprise to which the user belongs; Extracting the enterprise identity from the basic information, and obtaining a corresponding enterprise knowledge graph based on the enterprise identity, wherein the enterprise knowledge graph includes information on the enterprise's main products, technical patents, upstream and downstream industry chain information, and enterprise customer information; The historical behavior data of the user within a preset time period is obtained based on the login ID, and the historical behavior characteristics of the user are determined based on the enterprise knowledge graph and the historical behavior data.

[0086] Optionally, the behavior feature acquisition module is further configured to segment the historical behavior data according to a time dimension to obtain behavior sequences of multiple time windows; Extracting the user's query keyword set and click content set from the behavior sequence; Building an enterprise feature set based on the enterprise knowledge graph, wherein the enterprise feature set includes a main product feature set, a technology patent feature set, an industrial chain feature set, and an enterprise customer feature set; Matching the query keyword set and the click content set with the enterprise feature set to obtain matching features; Features associated with the enterprise knowledge graph are selected from the matching features as the historical behavior features of the user.

[0087] Optionally, the user portrait generation module is further configured to classify the historical behavior characteristics to obtain search interest characteristics, browsing habit characteristics, and interaction behavior characteristics; Combining the search interest features, browsing habit features, and interaction behavior features to generate a user behavior profile; Generate an enterprise attribute portrait based on the enterprise feature set corresponding to the enterprise identity; The user behavior portrait and the enterprise attribute portrait are integrated to obtain a user portrait of the user.

[0088] Optionally, the keyword determination module is further configured to extract core entity words from the user portrait and construct an entity relationship tree; Performing hypernym and hyponym retrieval in a preset professional knowledge base based on the entity relationship tree; Performing semantic expansion on the retrieved phrase to obtain an expanded phrase, wherein the expanded phrase includes a synonym set and a related word set; Cross-validating the synonym set and the related word set with the scenario information in the user portrait, and combining the expanded word groups based on the cross-validation results to generate a multi-dimensional keyword group; A phrase having a business relevance to the user is extracted from the multi-dimensional keyword group as a keyword corresponding to the user.

[0089] Optionally, the target vector conversion module is further configured to perform semantic analysis on the keyword and the query request to obtain a semantic unit; Construct a semantic dependency tree based on the hierarchical relationship and dependency relationship between semantic units; Reorganizing the semantic units based on the semantic dependency tree to obtain a semantic structured expression; Inputting the semantic structured expression into a preset text encoder to obtain a semantic vector; Semantic enhancement is performed on the semantic vector to generate an enhanced vector, and the enhanced vector is used as the target vector.

[0090] Optionally, the recommended information display module is further configured to search a preset vector database based on the target vector to obtain a candidate information set; Determining the basic priority of each candidate information in the candidate information set according to the information type and timeliness; Calculate the click rate and conversion rate of each candidate information based on the historical behavior characteristics and generate an interaction priority; Performing weighted fusion of the basic priority and the interaction priority to obtain a comprehensive priority score; The candidate information is hierarchically sorted according to the comprehensive priority score, a hierarchical recommendation page is constructed, and the recommendation page is visually displayed. The recommendation page includes a top display area, a key recommendation area, and a general display area.

[0091] It should be noted that the above embodiments provide systems that implement their functions using only the division of the above functional modules as an example. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. In addition, the system and method embodiments provided in the above embodiments are based on the same concept. The specific implementation process is detailed in the method embodiment and will not be repeated here.

[0092] An embodiment of the present application also provides a computer storage medium, which can store multiple instructions. The instructions are suitable for being loaded by a processor and executing an enterprise-level recommendation method based on a large language model in the above embodiment. The specific execution process can be found in the specific description of the above embodiment and will not be repeated here.

[0093] Please refer to Figure 3 The present application also discloses an electronic device. Figure 3The electronic device 300 may include: at least one processor 301 , at least one network interface 304 , a user interface 303 , a memory 305 , and at least one communication bus 302 .

[0094] The communication bus 302 is used to implement the connection and communication between these components.

[0095] The user interface 303 may include a display screen (Display) and a camera (Camera). Optionally, the user interface 303 may also include a standard wired interface and a wireless interface.

[0096] The network interface 304 may optionally include a standard wired interface or a wireless interface (such as a WI-FI interface).

[0097] The processor 301 may include one or more processing cores. Using various interfaces and circuits, the processor 301 connects to various components within the server. It executes instructions, programs, code sets, or instruction sets stored in the memory 305, as well as accesses data stored in the memory 305, to perform various server functions and process data. Optionally, the processor 301 may be implemented using at least one of the following hardware forms: a digital signal processing (DSP), a field-programmable gate array (FPGA), or a programmable logic array (PLA). The processor 301 may integrate one or a combination of a central processing unit (CPU), a graphics processing unit (GPU), and a modem. The CPU primarily processes the operating system, user interface, and application programs; the GPU is responsible for rendering and drawing content displayed on the display screen; and the modem handles wireless communications. It is understood that the modem may not be integrated into the processor 301 but implemented as a separate chip.

[0098] Among them, the memory 305 may include a random access memory (RAM) or a read-only memory (Read-Only Memory). Optionally, the memory 305 includes a non-transitory computer-readable storage medium. The memory 305 can be used to store instructions, programs, codes, code sets or instruction sets. The memory 305 may include a program storage area and a data storage area, wherein the program storage area may store instructions for implementing an operating system, instructions for at least one function (such as a touch function, a sound playback function, an image playback function, etc.), instructions for implementing the above-mentioned various method embodiments, etc.; the data storage area may store data involved in the above-mentioned various method embodiments, etc. The memory 305 may also optionally be at least one storage device located away from the aforementioned processor 301. Refer to Figure 3 , as a computer storage medium, the memory 305 may include an operating system, a network communication module, a user interface module, and an application program of an enterprise-level recommendation method based on a large language model.

[0099] exist Figure 3 In the electronic device 300 shown, the user interface 303 is mainly used to provide an input interface for the user and obtain the data input by the user; and the processor 301 can be used to call the memory 305 to store an application program of an enterprise-level recommendation method based on a large language model. When executed by one or more processors 301, the electronic device 300 executes one or more methods in the above-mentioned embodiments. It should be noted that for the aforementioned method embodiments, for the sake of simplicity of description, they are all expressed as a series of action combinations, but those skilled in the art should know that this application is not limited to the order of the actions described, because according to this application, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also know that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily required for this application.

[0100] In the above embodiments, the description of each embodiment has its own focus. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0101] In the several embodiments provided in this application, it should be understood that the disclosed devices can be implemented in other ways. For example, the device embodiments described above are merely schematic, such as the division of units, which is only a logical function division. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some service interface, and the indirect coupling or communication connection of devices or units can be electrical or other forms.

[0102] Units described as separate components may or may not be physically separate, and components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment according to actual needs.

[0103] In addition, the functional units in the various embodiments of the present application may be integrated into a single processing unit, or each unit may exist physically separately, or two or more units may be integrated into a single unit. The aforementioned integrated units may be implemented in the form of hardware or software functional units.

[0104] 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 computer-readable memory. Based on this understanding, the technical solution of this application, or the portion 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 memory and includes several instructions for causing a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the various embodiments of the method of this application. The aforementioned memory includes various media that can store program code, such as USB flash drives, mobile hard drives, magnetic disks, or optical disks.

[0105] The above are merely exemplary embodiments of the present disclosure and are not intended to limit the scope of the present disclosure. In other words, any equivalent variations and modifications made in accordance with the teachings of the present disclosure are still within the scope of the present disclosure. Those skilled in the art will readily conceive of other embodiments of the present disclosure after considering the disclosure and the practical implications thereof.

[0106] This application is intended to cover any variations, uses, or adaptations of the present disclosure that follow the general principles of the present disclosure and include common knowledge or customary techniques in the art not described herein. The description and examples are to be considered as exemplary only, and the scope and spirit of the present disclosure are to be defined by the claims.

Claims

1. An enterprise-level recommendation method based on a large language model, characterized in that: Methods include: Obtaining the user's login ID and query request, and obtaining the user's corporate identity and the user's historical behavior characteristics based on the login ID; Determine a user profile of the user based on the historical behavior characteristics and the corporate identity; Determining keywords corresponding to the user based on the user portrait; Converting the keyword and the query request into a target vector; A search is performed in a preset vector database based on the target vector to obtain recommendation information, and the recommendation information is converted into a visual recommendation page for display.

2. The enterprise-level recommendation method based on a large language model according to claim 1, characterized in that: The acquiring of the user's corporate identity and the user's historical behavior characteristics based on the login ID includes: Based on the login ID, a preset user database is searched to obtain basic information of the enterprise to which the user belongs; Extracting the enterprise identity from the basic information, and obtaining a corresponding enterprise knowledge graph based on the enterprise identity, wherein the enterprise knowledge graph includes information on the enterprise's main products, technical patents, upstream and downstream industry chain information, and enterprise customer information; The historical behavior data of the user within a preset time period is obtained based on the login ID, and the historical behavior characteristics of the user are determined based on the enterprise knowledge graph and the historical behavior data.

3. The enterprise-level recommendation method based on a large language model according to claim 2, characterized in that: The determining of the user's historical behavior characteristics based on the enterprise knowledge graph and the historical behavior data includes: Segmenting the historical behavior data according to the time dimension to obtain behavior sequences of multiple time windows; Extracting the user's query keyword set and click content set from the behavior sequence; Building an enterprise feature set based on the enterprise knowledge graph, wherein the enterprise feature set includes a main product feature set, a technology patent feature set, an industrial chain feature set, and an enterprise customer feature set; Matching the query keyword set and the click content set with the enterprise feature set to obtain matching features; Features associated with the enterprise knowledge graph are selected from the matching features as the historical behavior features of the user.

4. The enterprise-level recommendation method based on a large language model according to claim 3, characterized in that: The determining of the user profile of the user based on the historical behavior characteristics and the corporate identity includes: Classifying the historical behavior characteristics to obtain search interest characteristics, browsing habit characteristics, and interaction behavior characteristics; Combining the search interest features, browsing habit features, and interaction behavior features to generate a user behavior profile; Generate an enterprise attribute portrait based on the enterprise feature set corresponding to the enterprise identity; The user behavior portrait and the enterprise attribute portrait are integrated to obtain a user portrait of the user.

5. The enterprise-level recommendation method based on a large language model according to claim 1, characterized in that: The determining the keyword corresponding to the user based on the user portrait includes: Extract core entity words from the user portrait and construct an entity relationship tree; Performing hypernym and hyponym retrieval in a preset professional knowledge base based on the entity relationship tree; Performing semantic expansion on the retrieved phrase to obtain an expanded phrase, wherein the expanded phrase includes a synonym set and a related word set; Cross-validating the synonym set and the related word set with the scenario information in the user portrait, and combining the expanded word groups based on the cross-validation results to generate a multi-dimensional keyword group; A phrase having a business relevance to the user is extracted from the multi-dimensional keyword group as a keyword corresponding to the user.

6. The enterprise-level recommendation method based on a large language model according to claim 1, characterized in that: The converting the keyword and the query request into a target vector includes: Performing semantic analysis on the keywords and the query request to obtain semantic units; Construct a semantic dependency tree based on the hierarchical relationship and dependency relationship between semantic units; Reorganizing the semantic units based on the semantic dependency tree to obtain a semantic structured expression; Inputting the semantic structured expression into a preset text encoder to obtain a semantic vector; Semantic enhancement is performed on the semantic vector to generate an enhanced vector, and the enhanced vector is used as the target vector.

7. The enterprise-level recommendation method based on a large language model according to claim 1, characterized in that: The step of searching a preset vector database based on the target vector to obtain recommendation information and converting the recommendation information into a visual recommendation page for display includes: Searching a preset vector database based on the target vector to obtain a candidate information set; Determining the basic priority of each candidate information in the candidate information set according to the information type and timeliness; Calculate the click rate and conversion rate of each candidate information based on the historical behavior characteristics and generate an interaction priority; Performing weighted fusion of the basic priority and the interaction priority to obtain a comprehensive priority score; The candidate information is hierarchically sorted according to the comprehensive priority score, a hierarchical recommendation page is constructed, and the recommendation page is visually displayed. The recommendation page includes a top display area, a key recommendation area, and a general display area.

8. An enterprise-level recommendation system based on a large language model, characterized by: The system includes: A behavior feature acquisition module, configured to acquire a user's login ID and query request, and acquire the user's corporate identity and the user's historical behavior features based on the login ID; A user portrait generation module, configured to determine a user portrait of the user based on the historical behavior characteristics and the corporate identity; A keyword determination module, configured to determine keywords corresponding to the user based on the user portrait; a target vector conversion module, configured to convert the keyword and the query request into a target vector; The recommendation information display module is used to search a preset vector database based on the target vector to obtain recommendation information, and convert the recommendation information into a visual recommendation page for display.

9. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a plurality of instructions, and the instructions are suitable for being loaded by a processor and executing the method according to any one of claims 1 to 7.

10. An electronic device, characterized in that: The electronic device comprises a processor, a memory, a user interface and a network interface, wherein the memory is used to store instructions, the user interface and the network interface are used to communicate with other devices, and the processor is used to execute the instructions stored in the memory so that the electronic device executes the method according to any one of claims 1 to 7.

Images