Commodity pushing method and device
By combining a large language model and an expert knowledge base, the product recommendation method solves the problems of low accuracy and high maintenance difficulty in existing systems with fixed rules. It enables precise product recommendations based on user needs, improving system flexibility and user experience.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- BEIJING DIANJIEZHI TECH CO LTD
- Filing Date
- 2026-01-12
- Publication Date
- 2026-05-01
AI Technical Summary
Existing product recommendation systems rely on fixed rules, making it difficult to flexibly recommend products based on users' actual needs. This results in low recommendation accuracy, and the system becomes increasingly difficult and complex to maintain as merchants expand their business and product range.
A product recommendation method based on a large language model is adopted. By generating query instructions through intent recognition and entity extraction, and combining an expert knowledge base, the recommendation conditions are adjusted in multi-turn dialogues to achieve accurate understanding of user needs and product recommendation.
It improves the accuracy of product recommendations and the flexibility of the system, reduces the difficulty of system maintenance, and enhances the user shopping experience.
Smart Images
Figure CN121961693A_ABST
Abstract
Description
A product delivery method and apparatus Technical Field
[0001] This invention relates to the fields of artificial intelligence and e-commerce technology, and in particular to a product recommendation method and apparatus. Background Technology
[0002] Currently, in the e-commerce and industrial product sectors, to provide users with a better shopping experience, merchants often use shopping guide systems and push notification systems to offer intelligent product recommendation processes, enabling users to quickly obtain the products they need. These systems primarily rely on pre-defined conditional judgment rules when recommending products. Based on these rules and product data, they perform keyword filtering and product matching to determine the final products to be pushed to the user.
[0003] This method of product recommendation relies on manual maintenance of the rules. As merchants' businesses grow and product types increase, the business scenarios become more complex, making rule creation more difficult and increasing the number of rules required, ultimately leading to system maintenance difficulties. Furthermore, fixed rules make it difficult to flexibly recommend products based on users' actual needs, resulting in low accuracy and negatively impacting the user shopping experience. Summary of the Invention
[0004] In view of this, embodiments of the present invention provide a product recommendation method and apparatus, which can understand user intent based on user input information and generate an accurate product recommendation scheme based on the entities contained in the information and an expert knowledge base, thereby improving the accuracy of product recommendation.
[0005] To achieve the above objectives, according to one aspect of the present invention, a product push method is provided, comprising: step S101, performing intent recognition and entity extraction based on user input information, and generating a query instruction based on the intent recognition result and the entity extraction result; step S102, determining whether the push conditions corresponding to the intent recognition result are met according to the query instruction, and if the push conditions are not met, generating a guiding question based on the query instruction, the push conditions, and an expert knowledge base to guide the user to make the next round of input; step S103, performing entity extraction on the user input information of the next round, and updating the query instruction according to the entity extraction result of the user input information of the next round; step S104, repeating steps S102 and S103 until the query instruction meets the push conditions, and generating a push scheme based on the query instruction to push products.
[0006] Optionally, before determining whether the push conditions corresponding to the intent recognition result are met according to the query instruction, the method further includes: calculating a confidence score for the intent recognition result; determining the agent unit corresponding to the intent recognition result in response to the confidence score being greater than a preset threshold; the step of determining whether the push conditions corresponding to the intent recognition result are met according to the query instruction includes: obtaining the push conditions corresponding to the intent recognition result based on the agent unit corresponding to the intent recognition result; and determining whether the push conditions are met according to the query instruction.
[0007] Optionally, the method further includes: generating a first inquiry request based on the intent recognition result in response to the confidence score not being greater than the preset threshold; and updating the intent recognition result based on the first supplementary information input by the user based on the first inquiry request in response to receiving the first supplementary information input by the user based on the first inquiry request.
[0008] Optionally, before generating the query instruction and updating the query instruction, the method further includes: performing conflict detection on the entity recognition result based on the expert knowledge base; generating a second query request based on the conflict point in response to the detection result indicating a conflict point; and modifying the entity recognition result based on the second supplementary information input by the user based on the second query request in response to receiving second supplementary information input by the user based on the second supplementary information.
[0009] Optionally, generating a second query request based on the conflict point includes: determining the conflicting entities and their priorities based on the conflict point; adjusting the non-highest priority entities based on the expert knowledge base; and generating the second query request based on the highest priority entity, the non-highest priority entities, and the adjusted non-highest priority entities.
[0010] Optionally, before generating a query instruction based on the intent recognition result and the entity extraction result, the method further includes: validating the parameter range and parameter format of the entity extraction result, and correcting the entity extraction result based on the validation result.
[0011] Optionally, before generating a push plan to push products based on the query instruction, the method further includes: extracting user preferences from context information, the context information including user input information and guidance questions in each round; generating a push plan to push products based on the query instruction includes: generating a push plan to push products based on the query instruction and the user preferences.
[0012] Optionally, the step of generating a push plan based on the query instruction and the user preferences to push products includes: determining the product query range in the product database based on the user preferences; calculating the similarity between the parameters in the query instruction and the parameters of the products in the product query range; determining the push products that meet the preset similarity conditions based on the similarity calculation results; and generating a push plan based on the push products to push products, wherein the push plan includes the push products and the association information of the push products.
[0013] According to another aspect of the present invention, a product push device is provided, comprising: a generation module, configured to perform intent recognition and entity extraction based on user input information, and generate a query instruction based on the intent recognition result and the entity extraction result; a judgment module, configured to determine whether the push condition corresponding to the intent recognition result is met according to the query instruction, and if the push condition is not met, generate a guiding question according to the query instruction, the push condition, and an expert knowledge base to guide the user to make the next round of input; an update module, configured to perform entity extraction on the user input information of the next round, and update the query instruction according to the entity extraction result of the user input information of the next round; and a generation module, configured to repeatedly execute the judgment module and the update module until the judgment module determines that the query instruction meets the push condition, and generate a push scheme according to the query instruction to push products.
[0014] According to another aspect of the present invention, an electronic device is provided, comprising: one or more processors; and a storage device for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the product push method provided in the embodiments of the present invention.
[0015] According to another aspect of the present invention, a computer-readable medium is provided having a computer program stored thereon, which, when executed by a processor, implements the product push method provided in the embodiments of the present invention.
[0016] According to another aspect of the present invention, a computer program product is provided, including a computer program that, when executed by a processor, implements the product push method provided in the embodiments of the present invention.
[0017] One embodiment of the above invention has the following advantages or beneficial effects: it can perform intent recognition and entity extraction based on user input information and generate query instructions, achieving accurate structured understanding of natural language data from users in vertical domains. Furthermore, it determines whether push conditions are met based on the query instructions; if not, it updates the query instructions through expert knowledge-driven updates, improving the reliability of the query instructions. Simultaneously, it manages multi-turn dialogues through adjustable push conditions, adjusting the timing of product pushes and improving the accuracy of product pushes.
[0018] The further effects of the aforementioned unconventional alternative methods will be explained below in conjunction with specific implementation methods. Attached Figure Description
[0019] The accompanying drawings are provided to better understand the present invention and do not constitute an undue limitation thereof. Specifically: Figure 1 is a schematic diagram of the main steps of a product push method according to an embodiment of the present invention; Figure 2 is a schematic diagram of a product push process according to an embodiment of the present invention; Figure 3 is a schematic diagram of the main modules of a product push device according to an embodiment of the present invention; Figure 4 is an exemplary system architecture diagram in which embodiments of the present invention can be applied; Figure 5 is a schematic diagram of the structure of a computer system suitable for implementing terminal devices or servers according to embodiments of the present invention. Detailed Implementation
[0020] The following description, in conjunction with the accompanying drawings, illustrates exemplary embodiments of the present invention, including various details to aid understanding. These details should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of the invention. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.
[0021] It should be noted that the technical solutions disclosed in this invention, regarding the collection, updating, analysis, processing, use, transmission, and storage of user personal information, all comply with relevant laws and regulations, are used for legitimate purposes, and do not violate public order and good morals. Necessary measures are taken to prevent unauthorized access to user personal information data and to safeguard user personal information security, network security, and national security.
[0022] It should be noted that the collection, use, storage, sharing and transfer of user personal information involved in the technical solution of the present invention all comply with the provisions of relevant laws and regulations, and require notification to users and obtaining their consent or authorization. When applicable, user personal information is subjected to de-identification and / or anonymization and / or encryption technical processing.
[0023] Currently, shopping guide and push systems used in e-commerce platforms and merchant business platforms can push products based on user needs, reducing the time users spend searching for products. Commonly used push systems include rule-based push systems, collaborative filtering push systems, and content-based push systems.
[0024] Rule-based push systems use pre-defined if-then conditional rules for product matching and leverage keyword filtering and a manually maintained decision tree structure for product guidance. The advantages of this approach are its clear principles, strong controllability, and ability to provide accurate matching in specific, simple scenarios. However, its drawbacks are also significant. First, these systems heavily rely on manual maintenance. As business grows and product types and business scenarios increase dramatically, the number of rules grows exponentially, ultimately making the system difficult to maintain. Second, this approach lacks dynamic adaptability, typically lacking a later re-ranking stage to introduce operational intervention strategies or real-time data, thus affecting the quality of the final push list. Furthermore, due to the rigidity of the rules, it struggles to handle vague or incomplete user expressions and cannot quickly respond to new products or new demand scenarios.
[0025] Collaborative filtering push algorithms analyze user behavior data (such as user-item rating matrices) to calculate the similarity between users or items, and then push recommendations accordingly. This technology is highly effective in scenarios with large amounts of historical data. However, it faces three major challenges: data sparsity, the cold start problem, and user interest drift. When processing large-scale data, the user-item rating matrix is high-dimensional and sparse, which significantly reduces the effectiveness of traditional similarity-based algorithms. For new users or new products, the lack of sufficient historical interaction data prevents the system from providing personalized recommendations, the so-called "cold start problem." Furthermore, this algorithm struggles to handle "gray sheep" users whose preferences differ from the majority, as the system cannot provide effective recommendations in the absence of similar users.
[0026] Content-based recommendation systems analyze product attribute features and use text similarity algorithms such as TF-IDF (term frequency–inverse document frequency, a common weighting technique used in information retrieval and data mining) to match the features of items a user has previously liked with the features of new items. The advantage of this method is its ability to effectively solve the cold start problem for new items, as it relies on the metadata features of the items themselves, not user interaction history. Furthermore, it enhances transparency by providing interpretable features. However, its core drawback lies in the problem of "over-specialization." The system tends to recommend items highly similar to the user's past preferences, lacking novelty and failing to guide users to discover new interests or explore other categories. The effectiveness of the recommendations is entirely limited by predefined item features; if certain elements favored by the user are not included in the feature description, the system will be unable to provide suitable recommendations. Simultaneously, traditional text processing methods are inefficient when handling large text fields, potentially leading to high query costs, especially when dealing with high-frequency terms, which may even result in execution costs approaching a "full table scan."
[0027] Considering the shortcomings of the aforementioned push systems, this invention provides a product push method based on LLM (Large Language Model). This method does not rely on fixed rules or require historical user purchase data to accurately push products based on user needs. Furthermore, by combining expert knowledge to accurately understand user needs and push products accordingly, it significantly improves the accuracy of the final product push and enhances the user's shopping experience.
[0028] Figure 1 is a schematic diagram of the main steps of the product push method according to an embodiment of the present invention. As shown in Figure 1, the product push method mainly includes steps S101 to S104.
[0029] Step S101: Perform intent recognition and entity extraction based on user input information, and generate query instructions based on intent recognition results and entity extraction results.
[0030] Before pushing products, it's crucial to first obtain user needs information. This can be achieved through the user interaction layer, where user input is acquired. After receiving this input, an intelligent NLU (Natural Language Understanding) layer transforms the natural language input into structured data that the machine can understand, facilitating subsequent processes. This includes intent recognition and entity extraction based on the user input, yielding intent recognition and entity extraction results. The intent recognition result includes the intent category corresponding to the user input, such as "forklift inquiry," "other industrial product inquiry," or "casual conversation." The entity extraction result includes various parameters from the user input, such as "load capacity xx tons" or "climbing." In addition to product attribute parameters, extracted entities can also include application scenario parameters, product function, and other parameters. Once intent recognition and entity extraction are complete, a specific query command can be constructed based on the results, enabling product searches. For example, if a user inputs "I need a three-ton forklift for warehouse use," the identified intent category is "forklift selection," and the extracted entities are "three-ton," "warehouse utilization," and "forklift." The query command constructed based on the intent category and entities is {"intent": "forklift selection", "params": {"load capacity": "3000kg", "usage environment": "indoor / warehouse"}}. The query command mentioned above can be a pre-set query command with a specified format. It is generated by extracting parameters from the intent recognition results and entity extraction results and assembling them according to the specified format. For example, the specified format is: {"intent": "intent category", "params": {"parameter 1": "parameter content", "parameter 2": "parameter content"}}. Here, parameter 1, parameter 2, and their corresponding parameter content are determined based on the identified entities.
[0031] When performing intent recognition, the semantic understanding capabilities of large language models (such as LLM models) can be used to parse user input information and classify it into the corresponding intent category in a pre-defined intent system. The recognition process can achieve high-precision intent recognition using pre-defined prompts and a small amount of labeled instance data, ensuring the accuracy of the recognition results.
[0032] When extracting entities, a hybrid entity extraction method combining a large language model (such as an LLM model) with recognition rules can overcome the limitations of a single method and more flexibly and accurately identify all entities in the input information, such as "climbing" and "three tons". The obtained entities can then be further normalized and optimized. Specifically, a pre-defined entity mapping table and a corresponding rule engine are used to map all identified entities to entity parameters. For example, "three tons" is mapped to "load capacity 3000kg", and "climbing" is mapped to "applicable scenario: slope". Entity parameters help improve the accuracy and speed of subsequent product search results. The user input information can be used to generate the final query command through a three-layer natural language understanding architecture consisting of an intent recognition layer, an entity extraction layer, and a query rewriting layer. This three-layer structure effectively addresses the shortcomings of complex semantic recognition and accurately understands user input information. The query rewriting layer is the process layer that constructs the query command based on various entity parameters and intent categories.
[0033] The user interaction layer serves as both the entry and exit point of the push system of this invention. It receives natural language input from users and returns response information to them in a streaming manner. The response information includes replies to user-input questions, pushed product information, and guiding questions.
[0034] Step S102: Determine whether the push conditions corresponding to the intent recognition result are met based on the query instruction. If the push conditions are not met, generate guiding questions based on the query instruction, push conditions, and expert knowledge base to guide the user to the next round of input.
[0035] After receiving a query command, to ensure the accuracy of the product search, it's necessary to first determine if the current situation meets the product push conditions. The push conditions can differ depending on the intent category. For example, large products may involve more product parameters and applicable scenario parameters, requiring a larger number of parameters to be covered. Therefore, the judgment needs to determine if the number of parameters included in the query command reaches a certain threshold. Conversely, small or medium-sized products may involve fewer product parameters and applicable scenario parameters, significantly fewer than large products, resulting in a significantly lower threshold. Therefore, the judgment requires obtaining the push conditions corresponding to the intent recognition result of the current user input information. Push conditions include not only whether the number of parameters reaches the threshold but also whether the current dialogue round reaches a preset threshold, and whether the user input information contains a clear product push command. If the push conditions are met, the product database is directly queried according to the query command to obtain the pushed product. If the push conditions are not met, further adjustments are needed. Specifically, based on the query command, push conditions, and expert knowledge base, intelligent guidance questions are generated to guide the user to re-enter information to complete parameters or increase the dialogue round.
[0036] To improve dialogue efficiency when generating guiding questions, a pre-set expert knowledge base can be used to generate targeted guiding questions, which helps to quickly meet the push conditions. For example, the number of parameters that need to be supplemented by the user can be determined based on the push conditions. Then, based on the parameters already included in the query command, the number of parameters that need to be supplemented, and the expert knowledge base, the key parameters that need to be supplemented are determined, and guiding questions are generated to guide the user to supplement the key parameters. When supplementing parameters, only key parameters, i.e., the most valuable parameters, are supplemented, avoiding the listing of all parameters including many unimportant ones. This reduces the amount of data that the user needs to supplement, improves the user experience, reduces the complexity of user operations, and ensures the completeness of parameters, thereby ensuring the accuracy of subsequent product push results.
[0037] Step S103: Extract entities from the user input information in the next round, and update the query instruction based on the entity extraction results of the user input information in the next round.
[0038] After obtaining the user's input information for the next round, entity extraction needs to be performed again on the input information to update the query command based on the extracted entities, resulting in a more accurate command. Specifically, the entities extracted from the newly obtained user input information need to undergo the same normalization and standardization operations as the initial user input information, and the processed entities are used to update the query command, including parameter addition and replacement operations.
[0039] Step S104: Repeat steps S102 and S103 until the query command meets the push conditions, and generate a push plan based on the query command to push the product.
[0040] After each step S103 operation, step S102 is used to determine whether the corresponding push conditions are met based on the updated query command. If not, steps S102 and S103 are repeated until the updated query command meets the push conditions, triggering the product push process. Specifically, the updated query command queries the product database to obtain the push products, and a push plan is generated based on these products to push them to the user. The generated push plan includes not only product information but also a reason for pushing the product. This reason is generated based on the product being pushed, the parameters in the query command, and the user's input information. For example, the push reason can specify which parameters are determined based on the user's input information, and the push information for each parameter (e.g., load capacity 3000kg: product a has a maximum load capacity of 4000kg, suitable for user needs, etc.). The push plan includes a purchase link or product details link for the pushed product, facilitating user browsing.
[0041] The product push method provided by this invention can perform intent recognition and entity extraction based on user input information, and generate query instructions. This achieves accurate and structured understanding of natural language data from users in a vertical domain. The method also determines whether push conditions are met based on the query instructions. If the conditions are not met, the query instructions are updated using expert knowledge to improve their reliability. Furthermore, adjustable push conditions manage multi-turn dialogues, adjusting the timing of product pushes and improving accuracy.
[0042] In one embodiment, before determining whether the push conditions corresponding to the intent recognition result are met according to the query instruction, the method further includes: calculating the confidence score of the intent recognition result; determining the agent unit corresponding to the intent recognition result in response to the confidence score being greater than a preset threshold; determining whether the push conditions corresponding to the intent recognition result are met according to the query instruction includes: obtaining the push conditions corresponding to the intent recognition result based on the agent unit corresponding to the intent recognition result; and determining whether the push conditions are met according to the query instruction.
[0043] To further ensure the accuracy of the recognition results, a confidence score can be calculated when the large language model generates the intent recognition results. This score characterizes the accuracy of the recognition results. When the confidence score is high (greater than a preset threshold), the intent category in the intent recognition results can be directly determined to be the intent category corresponding to the user input information. Based on this intent category, the corresponding agent unit is determined, and the push conditions for the corresponding agent unit are obtained. The query command is then judged based on these push conditions.
[0044] To improve the accuracy of product recommendations, a multi-agent unit collaborative architecture can be used to process query commands of different intent categories during the product recommendation process. Different agent units maintain expert knowledge related to different intent categories and apply different recommendation logics for each intent category. For example, the priority and range of parameters differ for different intent categories. That is, after a query command arrives at the multi-agent unit collaborative architecture, the corresponding agent unit is determined based on its intent category, and the recommendation criteria of the agent unit are then used for further evaluation.
[0045] In one embodiment, for a multi-agent unit collaborative architecture, request distribution is handled by a central agent router. This router intelligently routes requests to specific agent units based on the intent category and confidence level identified by the NLU layer. For example, if the intent is "forklift consultation," the request will be routed to the forklift agent unit; if the intent is general chat, it will be routed to the general agent unit. This multi-agent unit collaboration method ensures that each agent unit focuses on a specific domain, improving the accuracy of product recommendations.
[0046] According to the product push method provided in the embodiments of the present invention, the reliability of the intent recognition result can be judged. For highly reliable intent recognition results, they can be directly adopted. The corresponding agent unit is determined, and the professional agent unit judges the push conditions and pushes products based on the query command, thereby improving the accuracy of product push.
[0047] In one embodiment, the method further includes: generating a first inquiry request based on the intent recognition result in response to a confidence score not exceeding a preset threshold; and updating the intent recognition result based on the first supplementary information input by the user based on the first inquiry request in response to receiving first supplementary information input by the user based on the first inquiry request.
[0048] If the confidence score obtained is not greater than a preset threshold, it indicates that the accuracy of the identified intent category is low. In this case, to ensure the accuracy of the query instruction, a query request can be returned to the user to obtain more specific instructions or more parameter information to select the final intent category. A first query request is generated based on the intent recognition result and returned to the user. First supplementary information input by the user based on the first query request is obtained, and the final intent category is determined by combining the first supplementary information to update the intent recognition result. The generated first query request can be a request to the user to obtain more parameters. After obtaining the first supplementary information, intent recognition is performed again by combining the first supplementary information and the initial user input information, and the intent recognition result is updated based on the obtained result. Alternatively, the generated first query request can also be a request to the user to select which intent category, and the final intent category is determined based on the returned first supplementary information containing selection information.
[0049] The product push method provided by the present invention can send an inquiry request to the user when the confidence level of the intent recognition result is low, so as to further update the intent recognition result based on the information provided by the user, ensure the accuracy of the recognition result, and facilitate the subsequent distribution of the query instruction to the correct agent unit for product push.
[0050] In one embodiment, before generating and updating the query instruction, the method further includes: performing conflict detection on the entity recognition result based on an expert knowledge base; generating a second query request based on the conflict point in response to the detection result indicating a conflict; and modifying the entity recognition result based on the second supplementary information input by the user based on the second query request in response to receiving second supplementary information input by the user based on the second query request.
[0051] Since user input information is not absolutely accurate and conflicts may exist between entities, conflict detection is necessary to ensure the quality of entity detection results. During conflict detection, a maintained expert knowledge base is used to check for conflicts between entities. If a conflict point exists, multiple entities corresponding to the conflict point are identified, and a second query request is generated based on these entities to confirm with the user which entity to retain for subsequent query command generation. After receiving the second supplementary information from the user, the modification information for the entities corresponding to the conflict point is determined and modified to generate accurate query commands based on the modified entities.
[0052] During conflict detection, each entity is first retrieved from an expert knowledge base. If any two entities correspond to two pre-labeled conflicting knowledge fragments, then there is a conflict between the two entities; otherwise, the two entities do not conflict. For example, the detected entities include "fuel forklift" and "500W," where 500W corresponds to electrical power, clearly indicating a conflict between the two.
[0053] The expert knowledge base is constructed as follows: industry expert knowledge is stored in an easily maintainable format, such as tables, containing information including expert knowledge at four levels, clarification points, product parameters, and professional advice. The system incorporates a dynamic loading and parsing module that loads this formatted knowledge on demand during application and converts it into internally usable structured knowledge objects, enabling rapid knowledge updates and applications.
[0054] The product recommendation method provided by this invention can verify the accuracy of user input information. Specifically, it performs conflict detection on each extracted entity and modifies conflicting entities to eliminate entity conflicts. During modification, a query request can be sent to the user to obtain modification suggestions, and the conflicting entities are modified according to the user's suggestions to obtain accurate entity extraction results.
[0055] In one embodiment, generating a second query request based on a conflict point includes: determining the conflicting entities and their priorities based on the conflict point; adjusting the non-highest priority entities based on an expert knowledge base; and generating a second query request based on the highest priority entity, the non-highest priority entities, and the adjusted non-highest priority entities.
[0056] After detecting a conflict, a second query request containing modification suggestions for the conflicting entities can be returned to the user. First, when generating modification suggestions for conflicting entities, it's necessary to determine the entities corresponding to the conflict point and their priorities. This is because, in the event of a conflict, lower-priority entities are more likely to err; that is, the accuracy of important entities in the user's input is very high. Therefore, lower-priority entities are prioritized for generating adjustment suggestions. From the entities corresponding to the identified conflict points, entities other than the highest priority are selected, and adjustment suggestions for these entities are determined based on the highest-priority entity. The second query request is then generated by combining these adjustment suggestions with the entities corresponding to the conflict point. For example, if a user specifies "indoor" (priority 1), "3-ton load capacity" (priority 2), and "fuel-powered forklift" (priority 3), and the expert knowledge base states that fuel-powered forklifts with a capacity of 3 tons or more are generally not used indoors, the system will identify this conflict point. Subsequently, utilizing the LLM generation capabilities and combining the expert knowledge base, a second query request is sent to the user, such as, "The fuel-powered forklift you mentioned may not be suitable for indoor use; would you like to consider an electric forklift?" Then, adjust the entity based on the second supplementary information returned by the user. For example, if the second supplementary information is "Yes, an electric forklift is needed", then adjust "fuel forklift" in the entity extraction result to "electric forklift".
[0057] In one embodiment, before generating a query instruction based on the intent recognition result and the entity extraction result, the method further includes: validating the parameter range and parameter format of the entity extraction result, and correcting the entity extraction result based on the validation result.
[0058] Further validation of the entity extraction results, including parameter range and format verification, can ensure accuracy. An entity validation mechanism can be introduced, employing different validation strategies for different types of entities (e.g., numerical entities, scene entities, etc.). For numerical entities, validation can be performed based on the numerical range and format provided in the expert knowledge base. For scene entities, semantic analysis can be conducted using an LLM model combined with the expert knowledge base, and the results can be converted into structured parameters for validation. For entities that fail validation, error messages can be returned to the user, prompting correction, or modifications can be made based on the reference range and format provided in the expert knowledge base.
[0059] In one embodiment, before generating a push plan to push products based on the query instruction, the method further includes: extracting user preferences from context information, which includes user input information and guidance questions in each round; generating a push plan to push products based on the query instruction includes: generating a push plan to push products based on the query instruction and user preferences.
[0060] Before generating the final push notification scheme, contextual information from multiple rounds of dialogue can be obtained. User preferences are extracted based on this contextual information, and the scope of pushed products is determined according to these preferences. For example, if a user prefers a certain category of products, then when querying the product database according to layer query instructions, the system will directly query products of that category and generate a push notification scheme based on the query results. Specifically, when extracting user preferences, deep learning can be used to construct user profiles based on contextual information, and then user preferences can be determined based on these profiles.
[0061] The product recommendation method provided by the present invention can identify user preferences based on contextual information in multi-turn dialogues, so as to combine user preferences and query instructions for more accurate product recommendation.
[0062] In one embodiment, generating a push plan based on query instructions and user preferences for product push includes: determining the product query range in the product database based on user preferences; calculating the similarity between parameters in the query instructions and parameters of products within the product query range; determining push products that meet preset similarity conditions based on the similarity calculation results; and generating a push plan based on the push products for product push, wherein the push plan includes the push products and their associated information.
[0063] Finally, when determining the push notification strategy, similarity calculations are performed based on the parameters included in the query command and the parameters of each product in the product database. This yields a similarity value between the query command and each product. Products that meet preset similarity criteria are then selected for push notifications. For example, a target number of products can be selected based on their similarity values from highest to lowest, or products with similarity values greater than a first preset threshold can be chosen. Push notification suggestions are generated based on the selected products, and a push notification strategy is created by combining these suggestions with product links and then pushed to the user. The product parameters include the product's attribute parameters, applicable scenarios, and functional parameters.
[0064] In one embodiment, before performing similarity calculation, the product query scope can be determined based on user preferences to reduce the number of products to be queried, thereby improving the efficiency of product recommendation. Accordingly, after determining the product query scope, the aforementioned product similarity calculation process can be performed within this scope based on the parameters of the query command. The user preference information may include the user's preferred product types; based on these product types, the products to be queried are determined, thus defining the query scope.
[0065] In one embodiment, the query instruction is a pre-set query instruction in a specified format; the push scheme includes the pushed product and its associated information. To improve the accuracy and speed of the large model's query results for pushed products, the query instruction can be standardized into a specified format, making it easier for the large model to accurately identify the intent category, parameters, parameter content, and other information contained therein, thereby achieving an efficient product query process. Simultaneously, to enable users to better select products based on the push scheme, the generated push scheme can include not only the queried products but also associated product information, such as browsing / purchasing links and basic product information. Furthermore, to improve the user's query experience, a reason for pushing the product can be added to the push scheme, allowing users to more intuitively understand the advantages of the pushed product. The reason could be, for example, "Suitable for use in xx scenario, capacity xx meets requirements, and can handle sudden increases in the amount of items loaded," etc.
[0066] In one embodiment, a complete session lifecycle management is performed within each user's dialogue cycle, including tracking states such as user ID, session ID, collected parameters, and current turn. The agent unit uses this state information to ensure the continuity of the dialogue and the accuracy of contextual understanding, maintaining session consistency even in complex multi-turn interactions.
[0067] In one embodiment, a service management scheme for the LLM model can ensure the stability and availability of the system under high concurrency. To address the concurrency limitations and rate limiting issues of a single LLM API (Large Language Model Interface), this invention designs a multi-API KEY management module and supports various round-robin strategies, including round_robin, random, and least_used. Round-robin distributes requests sequentially to each node, random distribution randomly distributes requests to a specific node, and least_used distributes requests to the node with the fewest pending requests. The least_used strategy dynamically senses the load of the backend LLM instances and prioritizes forwarding requests to the node with the fewest pending requests. This intelligent load balancing strategy based on real-time load is more suitable for LLM services than the traditional round_robin strategy and effectively solves the problem of uneven load caused by the randomness of the number of tokens (passwords) for different requests. In addition, this invention constructs an LLM Gateway (Large Language Model Gateway) and LLM Scheduler architecture. The Gateway is responsible for traffic entry, while the Scheduler performs intelligent scheduling based on real-time metrics of backend LLM instances (such as wait queue length, GPU (Graphics Processing Unit) cache utilization, etc.) to ensure continuous service availability. Furthermore, the system has a multi-level health check mechanism that can automatically detect anomalies and switch policies, supporting multi-level health checks and performance monitoring to ensure system availability.
[0068] The product recommendation method provided by this invention can perform intent recognition and entity extraction based on user input information through a three-layer natural language understanding structure, and generate query instructions. This achieves accurate structured understanding of natural language data from users in a vertical domain, enabling the application of agent units based on accurate query instructions. The method determines whether the push conditions are met based on the query instructions and the push conditions of the agent units. If the push conditions are not met, the query instructions are updated through expert knowledge-driven updates, improving the reliability of the query instructions. Simultaneously, adjustable push conditions manage multi-turn dialogues, adjusting the timing of product recommendations and improving accuracy. Furthermore, the multi-agent collaborative architecture allows each agent to be applicable to different professional fields and handle different user input scenarios, further improving the accuracy of product recommendations.
[0069] Figure 2 illustrates the product push process according to an embodiment of the present invention. The user inputs information through the user interaction layer and receives information returned by the system. The user input information is sent to the NLU layer for intent recognition and further entity extraction. The entity extraction results are then rewritten to obtain processed entities. A query command is generated by combining the processed entities and the intent recognition results and sent to the proxy collaboration layer. The proxy collaboration layer routes and distributes the request to the corresponding proxy unit based on the identified intent category. The proxy unit then uses the LLM model, expert knowledge base, product database, and professional suggestions to push products and returns the push results to the user interaction layer, which in turn returns them to the user. The professional suggestions can be used in conjunction with the product database to determine the final pushed products.
[0070] Figure 3 shows a schematic diagram of the main modules of the product push device provided in an embodiment of the present invention. As shown in Figure 3, the product push device 300 mainly includes a first generation module 301, a judgment module 302, an update module 303, and a second generation module 304.
[0071] The first generation module 301 is used to perform intent recognition and entity extraction based on user input information, and generate a query instruction based on the intent recognition result and the entity extraction result; the judgment module 302 is used to determine whether the push conditions corresponding to the intent recognition result are met according to the query instruction; if the push conditions are not met, a guiding question is generated according to the query instruction, the push conditions, and the expert knowledge base to guide the user to the next round of input; the update module 303 is used to extract entities from the user input information in the next round, and update the query instruction according to the entity extraction result of the user input information in the next round; the second generation module 304 is used to repeatedly execute the judgment module 302 and the update module 303 until the judgment module 303 determines that the query instruction meets the push conditions, and generates a push plan to push products according to the query instruction.
[0072] The product push device provided in this embodiment of the invention can perform intent recognition and entity extraction based on user input information, and generate query instructions. This enables accurate and structured understanding of natural language data from users in a vertical domain. The device determines whether push conditions are met based on the query instructions, and if not, updates the query instructions through expert knowledge to improve their reliability. Furthermore, adjustable push conditions manage multi-turn dialogues, adjusting the timing of product pushes to improve accuracy.
[0073] In one embodiment, the product push device 300 further includes a calculation module 305 (not shown in the figure) for calculating the confidence score of the intent recognition result; a determination module 304, specifically used to determine the agent unit corresponding to the intent recognition result in response to the confidence score being greater than a preset threshold; and a judgment module 302 specifically used to: obtain the push conditions corresponding to the intent recognition result based on the agent unit corresponding to the intent recognition result; and determine whether the push conditions are met according to the query instruction.
[0074] In one embodiment, the generation module 301 can also be used to: generate a first inquiry request based on the intent recognition result in response to a confidence score not exceeding a preset threshold; and update module 303 is used to update the intent recognition result based on the first supplementary information input by the user based on the first inquiry request in response to receiving the first supplementary information input by the user based on the first inquiry request.
[0075] In one embodiment, the product push device 300 further includes a detection module 306 (not shown in the figure), which is used to: perform conflict detection on the entity recognition result based on an expert knowledge base; generate a second query request based on the conflict point in response to the detection result indicating the existence of a conflict point; and modify the entity recognition result based on the second supplementary information input by the user based on the second query request in response to receiving the second supplementary information input by the user based on the second supplementary information.
[0076] In one embodiment, the detection module 306 can also be used to: determine the conflicting entities and their priorities based on the conflict points; adjust the non-highest priority entities based on an expert knowledge base; and generate a second query request based on the highest priority entity, the non-highest priority entities, and the adjusted non-highest priority entities.
[0077] In one embodiment, the product push device 300 further includes a verification module 307 (not shown in the figure), which is used to verify the parameter range and parameter format of the entity extraction results, and to correct the entity extraction results based on the verification results.
[0078] In one embodiment, the product push device 300 further includes an extraction module 308 (not shown in the figure) for extracting user preferences from context information, which includes user input information and guidance questions in each round; and a second generation module 304, which can be used to generate a push plan for product push based on query instructions and user preferences.
[0079] In one embodiment, the second generation module 304 can also be used to: determine the product query range in the product database based on user preferences; perform similarity calculation based on the parameters in the query instruction and the parameters of products within the product query range; determine the push products that meet the preset similarity conditions based on the similarity calculation results; and generate a push scheme based on the push products to push products, wherein the push scheme includes the push products and the associated information of the promoted products.
[0080] The product push device provided in this embodiment of the invention can perform intent recognition and entity extraction based on user input information through a three-layer natural language understanding structure, and generate query instructions. This enables accurate structured understanding of natural language data from users in a vertical domain, allowing for the application of agent units based on accurate query instructions. The device determines whether push conditions are met based on the query instructions and the push conditions of the agent units. If the push conditions are not met, the query instructions are updated using expert knowledge-driven methods to improve their reliability. Simultaneously, adjustable push conditions manage multi-turn dialogues, adjusting the timing of product pushes and improving accuracy. Furthermore, the multi-agent collaborative architecture allows each agent to be applicable to different professional fields and handle different user input scenarios, further enhancing the accuracy of product pushes.
[0081] Figure 4 illustrates an exemplary system architecture 400 to which the product push method or product push device of the present invention can be applied.
[0082] As shown in Figure 4, the system architecture 400 may include terminal devices 401, 402, and 403, a network 404, and a server 405. The network 404 serves as the medium for providing communication links between the terminal devices 401, 402, and 403 and the server 405. The network 404 may include various connection types, such as wired or wireless communication links or fiber optic cables, etc.
[0083] Users can use terminal devices 401, 402, and 403 to interact with server 405 via network 404 to receive or send messages, etc. Various communication client applications can be installed on terminal devices 401, 402, and 403, such as shopping applications, web browser applications, search applications, instant messaging tools, email clients, social media platform software, etc. (for example only).
[0084] Terminal devices 401, 402, and 403 can be various electronic devices with displays that support web browsing, including but not limited to smartphones, tablets, laptops, and desktop computers.
[0085] Server 405 can be a server that provides various services, such as a backend management server that supports shopping websites browsed by users using terminal devices 401, 402, and 403 (for example only). The backend management server can analyze and process data such as received product push requests, and feed back the processing results (such as pushing products - for example only) to the terminal device.
[0086] It should be noted that the product push method provided in this embodiment of the invention is generally executed by server 405, and correspondingly, the product push device is generally set in server 405.
[0087] It should be understood that the number of terminal devices, networks, and servers shown in Figure 4 is merely illustrative. Depending on implementation needs, any number of terminal devices, networks, and servers can be included.
[0088] Referring now to Figure 5, a schematic diagram of a computer system 500 suitable for implementing a terminal device or server according to embodiments of the present invention is shown. The terminal device or server shown in Figure 5 is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments of the present invention.
[0089] As shown in Figure 5, the computer system 500 includes a central processing unit (CPU) 501, which can perform various appropriate actions and processes based on programs stored in read-only memory (ROM) 502 or programs loaded from storage section 508 into random access memory (RAM) 503. The RAM 503 also stores various programs and data required for the operation of the system 500. The CPU 501, ROM 502, and RAM 503 are interconnected via a bus 504. An input / output (I / O) interface 505 is also connected to the bus 504.
[0090] The following components are connected to I / O interface 505: an input section 506 including a keyboard, mouse, etc.; an output section 507 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and speakers, etc.; a storage section 508 including a hard disk, etc.; and a communication section 509 including a network interface card such as a LAN card, modem, etc. The communication section 509 performs communication processing via a network such as the Internet. A drive 510 is also connected to I / O interface 505 as needed. A removable medium 511, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on drive 510 as needed so that computer programs read from it can be installed into storage section 508 as needed.
[0091] In particular, according to the embodiments disclosed in this invention, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this invention include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication section 509, and / or installed from removable medium 511. When the computer program is executed by central processing unit (CPU) 501, it performs the functions defined above in the system of this invention.
[0092] It should be noted that the computer-readable medium shown in this invention can be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. A computer-readable storage medium can be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this invention, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In this invention, a computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. Computer-readable signal media can also be any computer-readable medium other than computer-readable storage media, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to: wireless, wire, optical fiber, RF, etc., or any suitable combination thereof.
[0093] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.
[0094] The units or modules described in the embodiments of the present invention can be implemented in software or hardware. The described units or modules can also be housed in a processor; for example, a processor can be described as including a generation module, a judgment module, an update module, and a determination module. The names of these units or modules do not necessarily limit the specific unit or module itself. For example, a generation module can also be described as "a module for performing intent recognition and entity extraction based on user input information, and generating query instructions based on the intent recognition results and entity extraction results."
[0095] In another aspect, the present invention also provides a computer-readable medium, which may be included in the device described in the above embodiments; or it may exist independently and not assembled into the device. The computer-readable medium carries one or more programs, which, when executed by the device, cause the device to include: step S101, performing intent recognition and entity extraction based on user input information, and generating a query instruction based on the intent recognition result and the entity extraction result; step S102, determining whether the push conditions corresponding to the intent recognition result are met according to the query instruction, and if the push conditions are not met, generating a guiding question based on the query instruction, the push conditions, and an expert knowledge base to guide the user to the next round of input; step S103, performing entity extraction on the user input information of the next round, and updating the query instruction based on the entity extraction result of the user input information of the next round; step S104, repeating steps 2 and 3 until the query instruction meets the push conditions, and generating a push scheme based on the query instruction to push products.
[0096] According to the technical solution of this invention, intent recognition and entity extraction can be performed based on user input information to generate query instructions. This achieves accurate and structured understanding of natural language data from users in a vertical domain. The system determines whether push conditions are met based on the query instructions. If the conditions are not met, the query instructions are updated using expert knowledge to improve their reliability. Furthermore, adjustable push conditions manage multi-turn dialogues, adjusting the timing of product pushes to improve accuracy.
[0097] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can occur depending on design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.
Claims
1. A product recommendation method, characterized in that, include: Step S101: Perform intent recognition and entity extraction based on user input information, and generate query instructions based on intent recognition results and entity extraction results; Step S102: Determine whether the push conditions corresponding to the intent recognition result are met based on the query instruction. If the push conditions are not met, generate a guiding question based on the query instruction, the push conditions, and the expert knowledge base to guide the user to the next round of input. Step S103: Extract entities from the user input information in the next round and update the query instruction based on the entity extraction results. Step S104: Repeat steps S102 and S103 until the query instruction meets the push conditions, and generate a push plan based on the query instruction to push products.
2. The method according to claim 1, characterized in that, Before determining whether the push conditions corresponding to the intent recognition result are met according to the query instruction, the method further includes: calculating the confidence score of the intent recognition result; determining the agent unit corresponding to the intent recognition result in response to the confidence score being greater than a preset threshold; the step of determining whether the push conditions corresponding to the intent recognition result are met according to the query instruction includes: obtaining the push conditions corresponding to the intent recognition result based on the agent unit corresponding to the intent recognition result; and determining whether the push conditions are met according to the query instruction.
3. The method according to claim 2, characterized in that, The method further includes: generating a first inquiry request based on the intent recognition result in response to the confidence score not being greater than the preset threshold; and updating the intent recognition result based on the first supplementary information input by the user based on the first inquiry request in response to receiving the first supplementary information input by the user based on the first supplementary information.
4. The method according to claim 1, characterized in that, Before generating the query instruction and updating the query instruction, the method further includes: performing conflict detection on the entity recognition result based on the expert knowledge base; in response to the detection result indicating the existence of a conflict point, generating a second query request based on the conflict point; and in response to receiving second supplementary information input by the user based on the second query request, modifying the entity recognition result based on the second supplementary information.
5. The method according to claim 4, characterized in that, The step of generating a second query request based on the conflict point includes: determining the conflicting entities and their priorities based on the conflict point; adjusting the non-highest priority entities based on the expert knowledge base; and generating the second query request based on the highest priority entity, the non-highest priority entities, and the adjusted non-highest priority entities.
6. The method according to claim 1, characterized in that, Before generating a query instruction based on the intent recognition result and the entity extraction result, the method further includes: validating the parameter range and parameter format of the entity extraction result, and correcting the entity extraction result based on the validation result.
7. The method according to any one of claims 1-6, characterized in that, Before generating a push notification scheme based on the query instruction to push products, the method further includes: extracting user preferences from context information, the context information including user input information and guidance questions in each round; generating a push notification scheme based on the query instruction to push products includes: generating a push notification scheme based on the query instruction and the user preferences to push products.
8. The method according to claim 7, characterized in that, The step of generating a push notification scheme based on the query instruction and the user preferences to push products includes: determining the product query range in the product database based on the user preferences; calculating the similarity between the parameters in the query instruction and the parameters of the products in the product query range; determining the push products that meet the preset similarity conditions based on the similarity calculation results; and generating a push notification scheme based on the push products to push products, wherein the push notification scheme includes the push products and the association information of the push products.
9. A product pushing device, characterized in that, include: The first generation module is used to perform intent recognition and entity extraction based on user input information, and to generate query instructions based on the intent recognition results and entity extraction results; The judgment module is used to determine whether the push conditions corresponding to the intent recognition result are met based on the query instruction. If the push conditions are not met, the module generates guiding questions based on the query instruction, the push conditions, and the expert knowledge base to guide the user to make the next round of input. The update module is used to extract entities from the user input information in the next round and update the query instruction based on the entity extraction results of the user input information in the next round. The second generation module is used to repeatedly execute the judgment module and the update module until the judgment module determines that the query instruction meets the push conditions, and generates a push plan based on the query instruction to push the product.
10. An electronic device, characterized in that, include: One or more processors; A storage device for storing one or more programs that, when executed by one or more processors, cause the one or more processors to perform the method as described in any one of claims 1-8.
11. A computer-readable medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1-8.
12. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1-8.