Digital human Agent agent live broadcast recommendation method and system based on large language model

Abstract

The invention discloses a digital human Agent agent live broadcast recommendation method and system based on a large language model, and the method comprises the steps: S1, user portrait analysis: constructing a user portrait based on the hidden information of a live broadcast room, and the basic information and behavior data of a user in the live broadcast room; s2, context analysis: analyzing dialogue content, user interaction and live broadcast attributes of a live broadcast room, and understanding and analyzing context information of a current live broadcast environment; s3, recommended commodity selection: based on commodity related information, live broadcast room hidden information, user portraits and context information, selecting corresponding commodities by using a recommendation algorithm; s4, recommended copywriting generation: according to the user portrait, the context analysis result and the selected commodity, generating a recommended copywriting for the user through a large language model; and S5, digital person display recommendation: the digital person interacts with the live broadcast room user according to the recommended copywriting, and recommends the selected commodity. The method has the advantages that the user experience and the interaction quality in the e-commerce live broadcast environment are remarkably improved, and the efficiency and the accuracy of a recommendation system are enhanced.

Description

Technical Field

[0001] The present invention relates to the field of computer science and technology, and in particular to a digital human Agent live broadcast recommendation method and system based on a large language model. Background Art

[0002] Currently, most e-commerce platforms use recommendation systems based on users' historical purchase data and browsing behavior, which predict and recommend products by analyzing users' past behavior. However, these systems often lack the ability to interact in real time and the flexibility to make personalized recommendations. Digital human technologies are mainly used to provide virtual customer service or entertainment content. Although they can simulate human images and basic interactions, they are still limited in understanding complex user needs and real-time personalized interactions. Although large language models such as the GPT series perform well in processing natural language, their application in e-commerce environments is usually limited to basic customer service query responses, lacking in-depth customization and optimization for specific e-commerce scenarios.

[0003] In summary, the current digital human live broadcast recommendation system based on large language models has the following problems:

[0004] 1. Lack of deeply personalized recommendations: Existing e-commerce recommendation systems often fail to fully utilize real-time interaction data, resulting in recommendations that are not personalized and real-time enough.

[0005] 2. Limited interactive capabilities of digital humans: Existing digital humans lack sufficient intelligence and personalized response capabilities in live broadcasts or interactive scenarios, and cannot effectively improve user engagement and purchase conversion rates.

[0006] 3. Limitations of language models in e-commerce scenarios: Existing large-scale language models lack specificity in e-commerce live broadcast applications and cannot effectively combine user portraits, product information, and live broadcast environments to make accurate recommendations. Summary of the invention

[0007] The purpose of the present invention is to provide a digital human Agent live broadcast recommendation method and system based on a large language model, so as to solve the above-mentioned problems existing in the prior art.

[0008] In order to achieve the above object, the technical solution adopted by the present invention is as follows:

[0009] A digital human agent live broadcast recommendation method based on a large language model includes the following steps:

[0010] S1. User portrait analysis: Build user portraits based on the hidden information in the live broadcast room and the basic information and behavior data of the users in the live broadcast room;

[0011] S2, context analysis: Analyze the conversation content, user interaction and live broadcast attributes of the live broadcast room to understand and analyze the context information of the current live broadcast environment;

[0012] S3, recommended product selection: Based on product-related information, hidden information in the live broadcast room, user portrait, and context information, the recommendation algorithm is used to select the corresponding product from the product database;

[0013] S4, Recommendation copy generation: Based on the user portrait, context analysis results and selected products, a large language model is used to generate appropriate recommendation copy for the user;

[0014] S5. Digital human display recommendation: The digital human interacts with the users in the live broadcast room based on the recommendation copy and recommends the selected products.

[0015] Preferably, step S1 specifically includes the following contents:

[0016] S11. Collect basic information and behavior data of users in the live broadcast room in real time;

[0017] S12. Analyze and process the collected basic information and behavior data of users using user portrait building technology to preliminarily build user portraits;

[0018] S13. Extract more hidden information from the current live broadcast environment, and integrate the hidden information with the user's basic information and behavior data to supplement and optimize the initially constructed user portrait to form a more comprehensive and accurate user portrait.

[0019] Preferably, step S2 specifically includes the following contents:

[0020] S21, real-time capture and analysis of the conversation content in the live broadcast room, identification and extraction of user text input, detection and recording of activity information in the live broadcast room, and implementation of basic analysis of the live broadcast room context information;

[0021] S22. Based on basic analysis, use NLP models to dig deeper into the contextual information of the live broadcast room conversation content and monitor the frequency and emotions of user interactions;

[0022] S23, tracking the key nodes of the conversation content and the user's interaction, and determining the main topics and interests of the user;

[0023] S24. Combine the identified activity information, dynamically update and associate other potential contextual information, access external data sources, capture major news, festivals and social media trends in real time, and associate them with the interactive information in the live broadcast room.

[0024] Preferably, step S3 is specifically to use a recommendation algorithm to comprehensively score the products in the product database according to product popularity, recent exposure frequency, user preference matching, inventory, supply chain status, user portrait, contextual information and hidden information, and select the product with the highest comprehensive score as the selected product.

[0025] Preferably, the recommendation algorithm is specifically:

[0026] Comprehensive score = user preference score * user preference weight + context association score * context association weight +

[0027] Hidden information score * hidden information weight

[0028] User preference score = Σ(user interest tag weight * product tag matching degree)

[0029] Contextual relevance score = Σ(contextual keyword weight * product keyword matching degree)

[0030] Hidden information score = Σ(hidden information keyword weight * product attribute matching degree)

[0031] Among them, product tag matching degree = weight 1 * similarity between product tag and user interest tag; product keyword matching degree = weight 2 * similarity between product name keyword and user consultation context; product attribute matching degree = weight 3 * similarity between product attribute and user consultation context.

[0032] Preferably, step S4 specifically includes the following contents:

[0033] S41, after cleaning and sorting the user portrait, context analysis results, selected products and hidden information, multi-level data fusion and deep association annotation are performed to convert them into a structured input data stream that can be understood by the large language model;

[0034] S42. The large language model uses a multi-level attention mechanism to consider the needs of specific users, the current interaction context, and implicit factors, and generates personalized, real-time, and attractive recommendation copy for the corresponding users.

[0035] Preferably, the method further comprises,

[0036] S6. Recommendation algorithm feedback loop: Set weight adjustment factors based on user responses to recommended products and purchase results, use the weight adjustment factors to adjust the weights of relevant influencing factors, use the new weight values ​​to recalculate the comprehensive scores of each product in the product database, and select the product with the highest comprehensive score as the selected product.

[0037] Preferably, step S6 specifically includes the following steps:

[0038] S61, record the user's behavior data in the live broadcast room in real time, and analyze the effect of the recommended copy;

[0039] S62. Calculate the user interaction score of each recommended copy and the purchase conversion rate of the products associated with each recommended copy based on the user's behavior data in the live broadcast room and the effect of the recommended copy;

[0040] S63, setting an optimization threshold standard according to the user interaction score and the purchase conversion rate of the product associated with each recommendation copy, and when the threshold adjustment standard is reached, the weight of the influencing factor needs to be adjusted;

[0041] S64. Set a weight adjustment factor, and use the weight adjustment factor to adjust the weight of the corresponding influencing factor; new weight = old weight + weight adjustment factor * (actual result - expected result) / expected result;

[0042] S65. Continuously adjust the weight values ​​of influencing factors based on real-time monitoring and analysis of live broadcast room related information to ensure that the recommendation algorithm can always maintain the best state.

[0043] The present invention also aims to provide a digital human agent live broadcast recommendation system based on a large language model, the system can implement the above-mentioned method, the system includes:

[0044] User portrait analysis module: builds user portraits based on hidden information in the live broadcast room and basic information and behavior data of users in the live broadcast room;

[0045] Context analysis module: Analyzes the conversation content, user interaction and live broadcast attributes of the live broadcast room, and understands and analyzes the context information of the current live broadcast environment;

[0046] Recommended product selection module: Based on product-related information, hidden information in the live broadcast room, user portraits, and context information, the recommendation algorithm is used to select corresponding products from the product database;

[0047] Recommendation copy generation module: Based on the user portrait, context analysis results and selected products, the module generates appropriate recommendation copy for the corresponding user through a large language model;

[0048] Digital human display recommendation module: The digital human interacts with the corresponding users in the live broadcast room based on the recommendation copy and recommends the selected products.

[0049] Preferably, the system further comprises:

[0050] Feedback loop module: Set weight adjustment factors based on users' responses to recommended products and purchase results, use the weight adjustment factors to adjust the weights of relevant influencing factors, use the new weight values ​​to recalculate the comprehensive scores of each product in the product database, and select the product with the highest comprehensive score as the selected product.

[0051] The beneficial effects of the present invention are: 1. Improving the accuracy of personalized recommendations: By analyzing user portraits and live broadcast room contexts in real time, the present invention can provide more personalized and accurate product recommendations. This not only increases the user's purchase probability, but also improves the user's satisfaction with the live broadcast content. 2. Enhance user participation and interactive experience: The digital human in the present invention can interact with users in a more natural and intelligent way to create a more attractive shopping experience. This increases user participation and helps to improve user loyalty to brands and products. 3. Optimize the timeliness of product recommendations: The present invention can dynamically adjust the recommended content based on the real-time data and user feedback of the live broadcast room, so that the recommendation is not only accurate, but also extremely timely, especially in promotional activities and special holidays. More effective. 4. Improve conversion rate and sales performance: The present invention significantly improves the conversion rate of the live broadcast room through accurate and personalized recommendations, combined with optimized user interaction, and brings higher sales performance to merchants. 5. Continuous optimization and learning ability: Thanks to the feedback loop mechanism, the present invention can continuously learn and optimize the recommendation strategy, thereby continuously improving performance over time. 6. Reduce labor costs and resource investment: Compared with the traditional live sales model, the present invention reduces the dependence on anchors and sales staff through automation and intelligence, thereby reducing operating costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 4 is a flow chart of a live broadcast recommendation method in an embodiment of the present invention. DETAILED DESCRIPTION

[0053] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings. It should be understood that the specific implementation methods described herein are only used to explain the present invention and are not used to limit the present invention.

[0054] like Figure 1 As shown, in this embodiment, in order to solve the limitations of traditional e-commerce recommendation systems in real-time interaction and personalized recommendations, a digital human Agent intelligent body live broadcast recommendation method based on a large language model is provided. This method integrates a large language model with real-time user interaction data, such as user portraits, product information, live broadcast room dialogues and attribute information, to provide highly personalized and real-time changing product recommendations. The present invention uses advanced language processing capabilities to enhance the interaction efficiency and quality of digital humans in e-commerce live broadcasts, enabling them to respond to user needs and behaviors more intelligently and flexibly, thereby improving user engagement and purchase conversion rates. The present invention focuses on optimizing and customizing large language models in order to more effectively combine the specific needs and environment of e-commerce live broadcasts to achieve accurate product recommendations and sales promotion. The method of the present invention specifically includes the following six parts:

[0055] 1. User portrait analysis

[0056] Build user portraits based on the hidden information in the live broadcast room and the basic information and behavior data of the users in the live broadcast room. Specifically, it includes the following contents:

[0057] 1.1. Collect basic information (age, gender, purchase history) and behavioral data (interaction information) of users in the live broadcast room in real time.

[0058] 1.2. Utilize user portrait construction technology to analyze and process the collected basic information and behavior data of users, and preliminarily construct user portraits.

[0059] 1.3. Extract more hidden information from the current live broadcast environment and integrate the hidden information with the user's basic information and behavior data to supplement and optimize the initially constructed user portrait to form a more comprehensive and accurate user portrait.

[0060] In this embodiment, the user portrait of the live broadcast room is dynamically constructed and updated, and the recommendation strategy can be adjusted in real time according to the user portrait to ensure the personalization and timeliness of the recommendation.

[0061] 2. Contextual Analysis

[0062] Analyze the conversation content, user interaction and live broadcast attributes (such as holiday promotions) in the live broadcast room, and understand and analyze the context information of the current live broadcast environment. Specifically, it includes the following:

[0063] 2.1. Capture and parse the conversation content of the live broadcast room in real time, identify and extract user text input (including questions, comments, feedback), detect and record activity information in the live broadcast room (such as ongoing promotions, special products, and the host's recommended content), identify preset activity tags and trigger conditions, and realize basic analysis of the live broadcast room context information.

[0064] 2.2. Based on the basic analysis, the NLP model is used to dig deeper into the contextual information of the live broadcast room conversation content and monitor the frequency and emotions of user interactions. The NLP model can more accurately capture the emotional changes, intention recognition and potential needs in the conversation.

[0065] 2.3. Track the key nodes of the conversation content and the user's interaction to determine the main topics and interests of the user. For example, if a user mentions "winter skin care" in multiple comments, the system identifies this main demand and adjusts the recommendation strategy.

[0066] 2.4. Combine the identified activity information to dynamically update and associate other potential contextual information such as real-time events, weather conditions, and market trends. Connect to external data sources to capture major news, festivals, and social media trends in real time and associate them with interactive information in the live broadcast room.

[0067] In this embodiment, through context analysis, recommendations can be made more accurate and timely, thereby improving user engagement and purchase intention.

[0068] 3. Recommended product selection

[0069] Based on product-related information, hidden information in the live broadcast room, user portraits, and contextual information, a recommendation algorithm is used to select corresponding products from the product database.

[0070] Product related information includes:

[0071] (1) Product popularity: Calculate the popularity index of each product based on historical sales data and user reviews.

[0072] (2) Profit margin: Calculate the profit margin of each product based on the product’s sales price and cost data.

[0073] (3) Recent exposure frequency: Analyze the number of times a product has been recently exposed on the platform to avoid recommending products that appear too frequently and that users may already be familiar with or have no interest in.

[0074] (4) User preference matching: Real-time analysis of user profiles, including age, gender, purchase history, and browsing behavior, to calculate the matching degree between each product and the current user’s preferences.

[0075] (5) Inventory and supply chain status: Check the current inventory level and replenishment cycle of the goods to ensure that the recommended goods can be delivered in time.

[0076] Specifically: Based on product popularity, recent exposure frequency, user preference matching, inventory, supply chain status, user portrait, contextual information and hidden information, a recommendation algorithm is used to comprehensively score the products in the product database, and the product with the highest comprehensive score is selected.

[0077] The recommendation algorithm is specifically:

[0078] (1) User preference score: User preference score = Σ(user interest tag weight * product tag matching degree). Example: User interest tags: like skin care products (weight value can be set to 3), frequent purchase (weight value can be set to 2). Product tag matching degree = weight 1 * similarity between product tag and user interest tag. Product tags: moisturizer, skin care products.

[0079] (2) Context relevance score: Context relevance score = Σ(context keyword weight * product keyword matching degree). Example: Context keywords: dryness (weight value can be set to 2), moisturizing (weight value can be set to 1). Product keywords: moisturizing cream. Product keyword matching degree = weight 2 * product name keyword and user inquiry context similarity.

[0080] (3) Hidden information score: Hidden information score = Σ(hidden information keyword weight * product attribute matching degree). Example: Weather information: cold weather (weight value can be set to 3). Product attributes: Skin care products suitable for cold weather. Product attribute matching degree = weight 3 * similarity between product attributes and user inquiry context.

[0081] (4) Comprehensive score = user preference score * user preference weight + context relevance score * context relevance weight + hidden information score * hidden information weight.

[0082] The above-mentioned relevant weight values ​​involved in the calculation can be preliminarily set according to actual conditions to better meet actual needs.

[0083] In this embodiment, the product-related information is mainly used for product recommendations in the live broadcast room when the live broadcast is in a carousel or without user interaction. User portraits, context information, and hidden information are mainly used for product recommendations when there is user interaction in the live broadcast room.

[0084] In this embodiment, the recommendation algorithm can dynamically select the products that are most suitable for the current user, while also taking into account the sales and inventory management goals of the merchant, to ensure that the recommended products not only meet user needs but also have high commercial value.

[0085] 4. Generate Recommended Copy

[0086] Based on the user portrait, context analysis results and selected products, a large language model is used to generate appropriate recommendation copy for the corresponding user. Specifically, it includes the following:

[0087] 4.1. After cleaning and organizing user portraits, context analysis results, selected products and hidden information (current time, historical anniversaries, recent major news events, etc.), multi-level data fusion and deep association annotation are performed to convert them into structured input data streams that can be understood by the large language model.

[0088] 4.2. The large language model uses a multi-level attention mechanism to consider the needs of specific users, the current interaction context, and implicit factors to generate personalized, real-time, and attractive recommendation copy for the corresponding users.

[0089] In this embodiment, based on detailed user portraits and context information, clear prompts are provided to the large language model to guide it to generate recommendation texts with rich context and user preference information. The large language model uses the GPT-4 model. The generation process ensures that the generated recommendation texts are not only highly personalized and real-time, but also can attract users' attention and enhance interactivity, thereby increasing purchase intention and conversion rate.

[0090] 5. Digital Human Display Recommendation

[0091] The digital human interacts with the corresponding users in the live broadcast room based on the recommendation copy and recommends the selected products.

[0092] In this embodiment, the digital human is a 3D rendered virtual character with basic facial expressions and speech synthesis capabilities.

[0093] 6. Recommendation Algorithm Feedback Loop

[0094] According to the user's response to the recommended product and the purchase result, the weight adjustment factor is set, the weight of the relevant influencing factors is adjusted using the weight adjustment factor, and the comprehensive score of each product in the product database is recalculated using the new weight value, and the product with the highest comprehensive score is selected as the selected product. Specifically, it includes the following contents:

[0095] 6.1. Record the user’s behavior data in the live broadcast room in real time (including clicks, browsing, dwell time, comment interactions and purchase behavior), and analyze the effectiveness of the recommended copy (such as click-through rate, browsing depth, page dwell time and actual purchase conversion rate).

[0096] 6.2. Based on the user’s behavioral data in the live broadcast room and the effectiveness of the recommended copy, calculate the user interaction score (such as clicks, views and comments, etc.) for each recommended copy to evaluate the user’s interest level, and calculate the purchase conversion rate (number of orders / number of clicks) of the products associated with each recommended copy.

[0097] The relevant data obtained by calculation needs to be cleaned, such as removing noise and outliers, to ensure the accuracy and reliability of the data.

[0098] 6.3. Set the weight adjustment factor based on the user interaction score and the purchase conversion rate of the products associated with each recommended copy.

[0099] Set optimization threshold standards, such as when the click-through rate is lower than a certain value or the purchase conversion rate is lower than a certain percentage, the recommendation algorithm needs to be optimized and adjusted. By default, each influencing factor (such as product popularity, profit margin, exposure frequency, user preference matching and inventory status, and related weights used in the comprehensive score) has an initial set weight, and an adjustment factor is set to increase or decrease the weight of each factor.

[0100] 6.4. According to the performance in the feedback data, use the weight adjustment factor to adjust the weight of the corresponding influencing factor. The new weight = old weight + adjustment factor * (actual result - expected result) / expected result.

[0101] 6.5. Continuously adjust the weight values ​​of influencing factors based on real-time monitoring and analysis of live broadcast room related information to ensure that the recommendation algorithm can always maintain the best state.

[0102] In this embodiment, the recommendation algorithm loop feedback mechanism is used to collect users' responses to recommendations and purchase results, and continuously optimize the recommendation algorithm. The loop feedback mechanism can ensure that the algorithm continues to learn and improve, continuously improve the accuracy and conversion rate of recommendations, and ultimately improve the commercial performance and user satisfaction of the live broadcast room.

[0103] In this embodiment, a digital human agent live broadcast recommendation system based on a large language model is provided. The system can implement the above-mentioned method. The system includes:

[0104] (1) User portrait analysis module: Build user portraits based on hidden information in the live broadcast room and basic information and behavior data of users in the live broadcast room;

[0105] (2) Context analysis module: Analyzes the conversation content, user interaction, and live broadcast attributes of the live broadcast room to understand and analyze the context information of the current live broadcast environment;

[0106] (3) Recommended product selection module: Based on product-related information, hidden information in the live broadcast room, user portrait, and context information, the recommendation algorithm is used to select the corresponding product;

[0107] (4) Recommendation copy generation module: Based on the user portrait, context analysis results, and selected products, a large language model is used to generate appropriate recommendation copy for the corresponding user;

[0108] (5) Digital human display recommendation module: The digital human interacts with the corresponding users in the live broadcast room based on the recommendation copy and recommends the selected products.

[0109] (6) Feedback loop module: Set a weight adjustment factor based on the user interaction score and the purchase conversion rate of the products associated with each recommendation copy, use the weight adjustment factor to adjust the weight of the relevant influencing factors, use the new weight value to recalculate the comprehensive score of each product in the product database, and select the product with the highest comprehensive score as the selected product.

[0110] By adopting the above technical solution disclosed in the present invention, the following beneficial effects are obtained:

[0111] The present invention provides a method and system for live broadcast recommendation based on a large language model, which improves the accuracy of personalized recommendation: by analyzing user portraits and live broadcast room context in real time, the present invention can provide more personalized and accurate product recommendations. This not only increases the user's purchase probability, but also improves the user's satisfaction with the live broadcast content. Enhance user participation and interactive experience: the digital human in the present invention can interact with users in a more natural and intelligent way to create a more attractive shopping experience. This increases user participation and helps to improve user loyalty to brands and products. Optimize the timeliness of product recommendations: the present invention can dynamically adjust the recommended content according to the real-time data and user feedback of the live broadcast room, so that the recommendation is not only accurate, but also extremely timely, especially more effective in promotional activities and special holidays. Improve conversion rate and sales performance: the present invention significantly improves the conversion rate of the live broadcast room through accurate and personalized recommendations, combined with optimized user interaction, and brings higher sales performance to merchants. Continuous optimization and learning ability: thanks to the feedback loop mechanism, the present invention can continuously learn and optimize the recommendation strategy, thereby continuously improving performance over time. Reduce labor costs and resource investment: Compared with the traditional live sales model, the present invention reduces the dependence on anchors and sales staff through automation and intelligence, thereby reducing operating costs.

[0112] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be considered as the scope of protection of the present invention.

Claims

1. A digital human agent live broadcast recommendation method based on a large language model, characterized by: The following steps are included: S1. User portrait analysis: Build user portraits based on the hidden information in the live broadcast room and the basic information and behavior data of the users in the live broadcast room; S2, context analysis: Analyze the conversation content, user interaction and live broadcast attributes of the live broadcast room to understand and analyze the context information of the current live broadcast environment; S3, recommended product selection: Based on product-related information, hidden information in the live broadcast room, user portrait, and context information, the recommendation algorithm is used to select the corresponding product from the product database; S4, Recommendation copy generation: Based on the user portrait, context analysis results and selected products, a large language model is used to generate appropriate recommendation copy for the user; S5. Digital human display recommendation: The digital human interacts with the users in the live broadcast room based on the recommendation copy and recommends the selected products.

2. The method for recommending live broadcasts by a digital human agent based on a large language model according to claim 1 is characterized by: Step S1 specifically includes the following contents: S11. Collect basic information and behavior data of users in the live broadcast room in real time; S12. Analyze and process the collected basic information and behavior data of users using user portrait building technology to preliminarily build user portraits; S13. Extract more hidden information from the current live broadcast environment, and integrate the hidden information with the user's basic information and behavior data to supplement and optimize the initially constructed user portrait to form a more comprehensive and accurate user portrait.

3. The method for recommending live broadcasts by a digital human agent based on a large language model according to claim 1 is characterized by: Step S2 specifically includes the following contents: S21, real-time capture and analysis of the conversation content in the live broadcast room, identification and extraction of user text input, detection and recording of activity information in the live broadcast room, and implementation of basic analysis of the live broadcast room context information; S22. Based on basic analysis, use NLP models to dig deeper into the contextual information of the live broadcast room conversation content and monitor the frequency and emotions of user interactions; S23, tracking the key nodes of the conversation content and the user's interaction, and determining the main topics and interests of the user; S24. Combine the identified activity information, dynamically update and associate other potential contextual information, access external data sources, capture major news, festivals and social media trends in real time, and associate them with the interactive information in the live broadcast room.

4. The method for recommending live broadcasts by a digital human agent based on a large language model according to claim 1 is characterized by: Step S3 specifically includes using a recommendation algorithm to comprehensively score the products in the product database based on product popularity, recent exposure frequency, user preference matching, inventory, supply chain status, user portrait, contextual information and hidden information, and selecting the product with the highest comprehensive score as the selected product.

5. The method for recommending live broadcasts by a digital human agent based on a large language model according to claim 4 is characterized in that: The recommendation algorithm is specifically: Comprehensive score = user preference score * user preference weight + contextual relevance score * contextual relevance weight + hidden information score * hidden information weight User preference score = Σ(user interest tag weight * product tag matching degree) Contextual relevance score = Σ(contextual keyword weight * product keyword matching degree) Hidden information score = Σ(hidden information keyword weight * product attribute matching degree) Among them, product tag matching degree = weight 1 * similarity between product tag and user interest tag; product keyword matching degree = weight 2 * similarity between product name keyword and user consultation context; product attribute matching degree = weight 3 * similarity between product attribute and user consultation context.

6. The method for recommending live broadcasts by a digital human agent based on a large language model according to claim 1 is characterized by: Step S4 specifically includes the following contents: S41, after cleaning and sorting the user portrait, context analysis results, selected products and hidden information, multi-level data fusion and deep association annotation are performed to convert them into a structured input data stream that can be understood by the large language model; S42. The large language model uses a multi-level attention mechanism to consider the needs of specific users, the current interaction context, and implicit factors, and generates personalized, real-time, and attractive recommendation copy for the corresponding users.

7. The method for recommending live broadcasts by a digital human agent based on a large language model according to claim 1 is characterized by: The method also includes, S6. Recommendation algorithm feedback loop: Set weight adjustment factors based on user responses to recommended products and purchase results, use the weight adjustment factors to adjust the weights of relevant influencing factors, use the new weight values ​​to recalculate the comprehensive scores of each product in the product database, and select the product with the highest comprehensive score as the selected product.

8. The method for recommending live broadcasts by a digital human agent based on a large language model according to claim 7 is characterized by: Step S6 specifically includes the following steps: S61, record the user's behavior data in the live broadcast room in real time, and analyze the effect of the recommended copy; S62. Calculate the user interaction score of each recommended copy and the purchase conversion rate of the products associated with each recommended copy based on the user's behavior data in the live broadcast room and the effect of the recommended copy; S63, setting an optimization threshold standard according to the user interaction score and the purchase conversion rate of the product associated with each recommendation copy, and when the threshold adjustment standard is reached, the weight of the influencing factor needs to be adjusted; S64. Set a weight adjustment factor, and use the weight adjustment factor to adjust the weight of the corresponding influencing factor; new weight = old weight + weight adjustment factor * (actual result - expected result) / expected result; S65. Continuously adjust the weight values ​​of influencing factors based on real-time monitoring and analysis of live broadcast room related information to ensure that the recommendation algorithm can always maintain the best state.

9. A digital human agent live broadcast recommendation system based on a large language model, characterized by: The system can implement the method described in any one of claims 1 to 8 above, and the system includes: User portrait analysis module: builds user portraits based on hidden information in the live broadcast room and basic information and behavior data of users in the live broadcast room; Context analysis module: Analyzes the conversation content, user interaction and live broadcast attributes of the live broadcast room, and understands and analyzes the context information of the current live broadcast environment; Recommended product selection module: Based on product-related information, hidden information in the live broadcast room, user portraits, and context information, the recommendation algorithm is used to select corresponding products from the product database; Recommendation copy generation module: Based on the user portrait, context analysis results and selected products, the module generates appropriate recommendation copy for the corresponding user through a large language model; Digital human display recommendation module: The digital human interacts with the corresponding users in the live broadcast room based on the recommendation copy and recommends the selected products.

10. The digital human agent live broadcast recommendation system based on a large language model according to claim 9 is characterized by: The system also includes, Feedback loop module: Set weight adjustment factors based on users' responses to recommended products and purchase results, use the weight adjustment factors to adjust the weights of relevant influencing factors, use the new weight values ​​to recalculate the comprehensive scores of each product in the product database, and select the product with the highest comprehensive score as the selected product.

Images