A recommendation method based on user problems

Abstract

The application discloses a recommendation method based on user problems, which integrates multi-round dialogue, user portrait and behavior data, uses a large language model to generate a semantic summary of user preferences and product facts, and then aligns the text and the recommendation feature space through a multi-expert hybrid network to output a dynamically updated candidate product list, solving the problems of insufficient semantic understanding and recall bias. The dialogue is modeled as a Markov decision process, and the strategy large model is fine-tuned based on the group strategy optimization framework. Through generating diversified decisions, multi-dimensional reward evaluation and normalized comparison, the recommendation accuracy and dialogue fluency are improved. In each round of interaction, the retrieval module is called in real time to update the candidate list, and similar cases are matched from the example library to drive the trained strategy model to generate optimal decisions and natural replies, realizing the cooperation of semantic deep understanding, strategy continuous optimization and scene adaptive decision-making, and significantly improving the overall performance and robustness of the dialogue recommendation system.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a recommendation method based on user questions. Background Technology

[0002] With the development of artificial intelligence technology, conversational recommendation systems have become an important direction for improving personalized service experiences. These systems aim to dynamically understand and respond to user needs through multi-turn interactions using natural language, ultimately providing accurate product or content recommendations. Currently, achieving this goal typically requires the comprehensive utilization of multiple information sources, including user dialogue text, static profiles, and historical behavioral data.

[0003] While existing technologies have attempted to integrate the aforementioned information, their core methods often have significant shortcomings. Mainstream solutions tend to process these heterogeneous data (dialogue text, user profiles, behavioral data) in isolation or establish shallow, static connections solely through knowledge graph entity links. This approach makes it difficult for the system to effectively model the dynamic evolution of user intent during continuous dialogue, and it also fails to deeply analyze the complex and implicit preferences and needs expressed by users through natural language. As a result, the system's recall module often performs retrieval based on incomplete or lagging user states, leading to a discrepancy between the generated candidate set and the user's true, real-time intent displayed in the dialogue, becoming the primary bottleneck restricting recommendation accuracy. Summary of the Invention

[0004] To address the aforementioned issues, this application provides a recommendation method based on user questions.

[0005] The embodiments of this application disclose the following technical solutions: The first aspect of this application provides a recommendation method based on user questions, including: In response to current user questions, obtain user preference summary text and product fact text; Based on the text embedding model, the obtained user preference summary text and product fact text are transformed into initial embedding vectors, which include user preference initial embedding vectors and product fact initial embedding vectors; based on the text embedding model, the current user question is semantically matched with a preset example library to select several dialogue examples that match the current user question; The initial embedding vector is input into a multi-expert hybrid network so that each expert network processes the input vector representation in parallel and outputs enhanced features from the perspective of the expert network. The multi-expert hybrid network contains multiple independent expert sub-networks. The expert weight distributions on the user side and the product side are calculated separately through parallel gating networks. The input vector is dynamically routed to different expert networks for parallel processing according to the expert weight distribution, so that each expert network performs a nonlinear transformation on the input and outputs intermediate features from its specific perspective. Based on the weights output by the gating network, the intermediate features processed by each expert network are weighted and summed to generate enhanced user preference embedding vectors and enhanced product fact embedding vectors, respectively. The enhanced user preference embedding vector and the enhanced product fact embedding vector are jointly input into the recall model so that the recall model can calculate the matching degree between users and products, and retrieve and generate a preliminary candidate product set from the full product pool. The enhanced product fact embedding vectors corresponding to the preliminary candidate product set and the initial relevant candidate product set are input into the reordering model to obtain the candidate product set. Based on the dialogue example, the candidate product set, and the current user question, an action prompt template is generated; The decision prompt template is input into a pre-trained strategy generation model, so that the strategy generation model outputs a recommended decision action for the current user's question and / or generates a corresponding natural language response.

[0006] In one possible implementation, the pre-trained policy model is determined in the following ways: The environmental state is represented by integrating historical dialogue records, user questions, and candidate product sets. Based on the environmental state representation, an action prompt template is generated. Based on the action prompt template, the hyperparameters are adjusted so that the large strategy model can generate multiple different action decisions and corresponding response messages in parallel. Each action decision is scored from multiple dimensions based on a reward model, and the original reward score for each action decision is calculated. For the multiple action decisions, the mean and standard deviation of the original reward scores are calculated, and the original reward scores of each action decision are normalized based on the mean and standard deviation to obtain normalized rewards. The weight parameters of the large policy model are updated with the goal of maximizing the expected value of the normalized reward.

[0007] In one possible implementation, the multiple dimensions include product click-through rate (CTR), dialogue quality, and round efficiency penalty. The product CTR is used to characterize the probability of a recommended product being clicked by a CTR prediction model. The dialogue quality is used to characterize the semantic coherence and contextual relevance of the generated responses. The round efficiency penalty is used to indicate that a negative reward is applied when the number of dialogue rounds exceeds a threshold.

[0008] In one possible implementation, obtaining the user preference summary text and product fact text includes: Acquire full user data related to users and full product data related to products. The full user data includes historical dialogue records, user profile data, and user product interaction data. The full product data includes raw product data, which includes tag fields and numeric fields. Using a large language model based on a preset sample prompt template, inference is performed on the historical dialogue data, user product interaction data, and user profile data to generate a user preference summary text that summarizes the user's long-term preferences and short-term intentions; By integrating key information from tag-type and numerical fields, and supplementing them with descriptions of core product attributes and usage scenarios, product fact text is obtained, which is used to characterize the core features of the product.

[0009] In one possible implementation, the step of semantically matching the current user question with a preset example library to filter out several dialogue examples that match the current user question includes: Encode the current user question into a first vector; Encode the user question field of each dialogue example in the example library into a corresponding second vector; Calculate the cosine similarity between the first vector and each of the second vectors; The dialogue examples are sorted from high to low according to the cosine similarity, and the top N dialogue examples are selected as the filtered dialogue examples, where N is a preset positive integer.

[0010] In one possible implementation, the policy generation model outputs a recommended decision action for the current user question and / or generates a corresponding natural language response, including: If the matching degree between the candidate product and the user's needs is greater than the preset model judgment threshold, the recommendation action is triggered, and the product with the highest matching degree among the candidate products and a natural language response containing the product's advantages are output. If the user's request does not include key preference information, a follow-up question action is triggered, and a targeted question script is output to address the missing preference information; If the user's current question is a consultation question that includes at least one of the following: product parameters, usage scenarios, or functional features, then a response action is triggered, and a natural language response directly corresponding to the question is output.

[0011] In one possible implementation, updating the weight parameters of the policy model with the objective of maximizing the expected value of the normalized reward includes: Define an objective function with the expected value of the normalized reward as the core optimization term. At the same time, introduce a KL divergence regularization term to constrain the deviation between the new strategy and the old strategy before the update. The KL divergence regularization term adjusts the constraint strength through a preset constraint coefficient to avoid excessive policy update amplitude leading to model performance oscillation. Based on the multiple action decisions obtained from sampling and the corresponding normalized rewards, the gradient of the objective function with respect to the weight parameters of the policy model is calculated. The calculated gradient is truncated using the L2 norm to keep the L2 norm of the gradient within a preset threshold. Based on the trimmed gradient, the weight parameters of the large model of the policy are iteratively updated according to a preset learning rate, wherein the learning rate ranges from one ten-thousandth to one thousandth. When the absolute value of the difference between two adjacent rounds of the objective function value does not exceed one ten-thousandth, or when the number of iterations reaches the preset maximum number of iterations, the weight parameter update stops. The weight parameters at this time are the weight parameters of the trained policy model.

[0012] In one possible implementation, the integration of the environmental state representation includes: Historical dialogue records are transformed into text sequence vectors, current user questions are transformed into query vectors, and candidate product sets are transformed into product feature matrices. Through vector concatenation and matrix flattening, an environment state vector with uniform dimensions is generated.

[0013] In one possible implementation, the re-ranking model is used to concatenate the enhanced product fact embedding vector with the product's inherent attribute features and user interaction behavior features, capture the correlation weights between features through a multi-layer attention mechanism, and then input the result into a fully connected layer to output a ranking score.

[0014] In one possible implementation, each dialogue example in the preset example library is stored as a record, and each record contains at least a user question field, an action decision field, and a response text field; the dialogue examples in the example library cover a variety of typical dialogue recommendation scenarios, including product recommendation, attribute filtering, and requirement clarification.

[0015] A second aspect of this application provides an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, it implements the user-question-based recommendation method as described in the first aspect above.

[0016] A third aspect of this application provides a computer program product that, when run on a computer, executes the user-question-based recommendation method as described in the first aspect above.

[0017] A fourth aspect of this application provides a computer-readable storage medium storing instructions that, when executed on a terminal device, cause the terminal device to perform the user-question-based recommendation method as described in the first aspect above.

[0018] Compared with the prior art, this application has the following beneficial effects: By leveraging user preference summary text generated using a large language model, deep fusion and reasoning of historical dialogues, user profiles, and behavioral data are achieved. This enables modeling the dynamic evolution of user intent and parsing implicit needs in natural language, overcoming the shortcomings of information isolation and shallow association. Secondly, by enhancing and aligning text embedding vectors through a multi-expert hybrid network, deep dialogue semantic knowledge is effectively integrated into the vector space of the recommendation model, significantly improving the system's ability to understand complex semantics. This allows the recall model to match based on accurate, context-rich user representations, ensuring that the candidate product set closely matches the user's true intent. Finally, by combining dynamically retrieved dialogue examples with real-time generated candidate product lists to construct decision prompts, the strategy model can make decisions based on historical successes in specific contexts. This allows it to respond to explicit questions while also properly handling ambiguous or implicit user needs, ultimately outputting precise recommendation actions and / or natural language responses, achieving accurate capture and satisfaction of the user's true intent. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 The overall process architecture diagram provided for the embodiments of this application; Figure 2 This is a first flowchart illustrating an embodiment of the present application; Figure 3 This is a second flowchart provided as an embodiment of the present application; Figure 4 This is a third flowchart provided for an embodiment of this application; Figure 5 This is a fourth flowchart provided for an embodiment of this application. Detailed Implementation

[0021] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present application.

[0022] A conversational recommender system (CRS) is an intelligent system that combines natural language interaction with recommendation algorithms. It aims to dynamically understand user preferences through multi-turn dialogues and provide personalized recommendations. Unlike traditional recommender systems that rely on static user profiles, conversational recommender systems can capture the evolution of user intent in real time, refining recommendation results progressively through interactive formats such as question-and-answer, clarification, and feedback, significantly improving recommendation accuracy and user experience.

[0023] To facilitate understanding of the technical solutions provided in the embodiments of this application, the background technology involved in the embodiments of this application will be described below.

[0024] As mentioned earlier, existing technical solutions have significant limitations in multi-turn dialogue information fusion. They typically treat dialogue text, user static profiles, and historical behavior data as isolated information fragments, relying solely on superficial static connections such as knowledge graph entity links to establish superficial relationships, lacking the ability to model the dynamic evolution of user intent during the dialogue. Therefore, the system struggles to deeply analyze the implicit needs and preference shifts expressed through natural language in the dialogue context, resulting in a discrepancy between the final product list and the user's true intent.

[0025] Furthermore, in traditional architectures, the recall and policy decision modules are typically trained in stages and independently. This design causes errors generated in the recall stage to be directly transmitted to the decision stage, limiting the policy model to optimization only on biased or restricted candidate sets, and preventing the use of decision feedback to correct the recall process. Due to the lack of a joint representation learning mechanism between the two modules, the policy model cannot directly utilize the deep semantics of the product learned in the recall stage, requiring re-feature extraction and alignment, resulting in attenuation and loss of semantic information during transmission.

[0026] On the other hand, conversational recommendation systems have a large action space, while user feedback is often sparse and delayed. In this scenario, traditional deep Q-network algorithms are prone to amplifying noise signals due to overestimation bias, resulting in low learning efficiency and slow convergence. Meanwhile, the experience replay mechanism relied upon by DQN also faces challenges in conversational recommendation environments: because dialogue states and user policies are updated rapidly, the historical experience data in the cache may not be consistent with the data distribution of the current policy, leading to unstable or even divergent model training.

[0027] To address the aforementioned issues, this application's embodiments revolve around three key modules: dialogue-enhanced product retrieval and re-ranking, model fine-tuning based on group strategy optimization, and a dialogue recommendation task integrating dynamic examples and candidate products. In the product retrieval and re-ranking stage, the system integrates multi-turn dialogue records, user profiles, and interaction behaviors to generate structured user preference summaries and product feature descriptions. Subsequently, a multi-expert hybrid network achieves deep fusion between text semantic representation and the recommendation model's vector space, significantly enhancing the system's ability to understand complex dialogue contexts, improving recall accuracy, and overcoming the semantic limitations of traditional recommendation models. In the model fine-tuning stage, the dialogue recommendation process is constructed as a Markov decision process and employs a group strategy optimization method. This framework supports multi-action decision generation and a group reward evaluation mechanism, enabling dynamic and efficient optimization of the strategy model. In the dialogue recommendation execution stage, the system retrieves examples relevant to the current context from the example library through semantic matching and constructs dynamic prompts by combining them with a real-time updated candidate product list. This mechanism enables large language models to have a reference basis in cold start scenarios and maintain stable generalization ability in complex scenarios, thereby greatly improving the system's adaptability and robustness in actual business.

[0028] Overall, this application constructs a closed-loop training framework for dialogue recommendation that integrates knowledge enhancement and reinforcement learning. By deeply combining the knowledge enhancement capabilities of multi-expert hybrid networks with group policy learning methods, it achieves synergistic optimization among dialogue state perception, product recall accuracy, and policy decision-making efficiency.

[0029] It should be noted that the user-question-based recommendation method, system, product, device, and medium provided in this application can be applied to the field of computer technology or conversational recommendation systems. The above are merely examples and do not limit the application areas of the user-question-based recommendation method, system, product, device, and medium provided in this application. Furthermore, the embodiments of this application may not limit the executing entity of the user-question-based recommendation. For example, the user-question-based recommendation method of this application can be applied to data processing devices such as terminal devices or servers. The terminal device can be an electronic device such as a computer or a personal digital assistant (PDA). The server can be a standalone server, a cloud server, or a cluster server composed of multiple servers.

[0030] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0031] The following is an example of how this application breaks down the conversational recommendation task into three sub-tasks. See [link to example]. Figure 1 , Figure 1 The overall process architecture diagram provided for the embodiments of this application includes: (1) Product retrieval and re-ranking task: Based on historical dialogues and other product recommendation-related data, the most relevant products are retrieved from the candidate product pool, and the top 5 candidate products are obtained using a re-ranking model; (2) Model fine-tuning based on group reinforcement learning: Based on the GRPO training framework, the conversational recommendation process is reshaped into a large model that takes corresponding action decisions based on changes in the dialogue environment, outputs a reply dialogue or recommended products, and obtains corresponding reward scores based on dialogue quality, product click-through rate and round penalty, and updates the weight parameters of the strategy large model, so that the large model can better make corresponding action decisions based on changes in the dialogue environment, uses the candidate products obtained from the product retrieval and re-ranking task as background information, and infers the degree of matching between the candidate products and user preferences; (3) Conversational recommendation task based on dynamic examples and candidate products: Based on the user's current question, relevant examples are obtained from the manually constructed example library through semantic matching, and prompt templates are constructed based on dynamically updated examples and candidate products, so that the large model decides the next action, improving the robustness and recommendation accuracy of the large language model dialogue recommendation.

[0032] Task 1 is a product retrieval and re-ranking task, mainly divided into three stages: dialogue knowledge generation, dialogue knowledge enhancement, and retrieval and re-ranking. The aim is to build a retrieval and ranking engine capable of deeply understanding users' dynamic intent and accurately matching candidate products. This stage does not simply rely on traditional keyword matching or static user profiles, but rather systematically integrates multi-turn dialogues, user behavior, and product knowledge through the three core stages of knowledge generation, knowledge enhancement, and product recall and re-ranking. This forms an enhanced semantic representation that is highly synergistic with downstream recommendation tasks, providing a high-quality, highly relevant set of product candidates for subsequent dialogue decisions.

[0033] The first stage is the knowledge generation task, which aims to transform fragmented, multimodal raw data into structured textual knowledge that can be understood and reasoned about by a large language model. See also Figure 2 , Figure 2 The first flowchart provided for the embodiments of this application is specifically divided into two parallel tasks: The user preference summary text generation system integrates three main categories of data: user profiles describing long-term stable user characteristics, user-product interaction behaviors reflecting historical interests, and multi-turn historical dialogues demonstrating the dynamic intent of the current conversation. By designing a dedicated zero-shot prompt template, the system guides a large language model to analyze user preferences, comprehensively reasoning and analyzing the above information to ultimately generate a concise user preference summary report. This report not only restates the facts but also extracts users' style tendencies, core needs, and potential latent demands, achieving a fusion modeling of long-term user preferences and short-term intentions.

[0034] Product fact text generation involves standardizing structured fields in the product database (such as numerical prices and sizes, and enumerated brands, colors, and categories) based on product-side information, and then integrating them into a coherent and complete product fact description text according to business logic. This is equivalent to creating a standardized instruction manual for each product, converting machine-readable data into a semantic description that is understandable to both humans and machines.

[0035] The second stage involves enhancing dialogue knowledge; see [link / reference]. Figure 3 , Figure 3 This is a second flowchart illustrating an embodiment of this application. The knowledge enhancement task aims to bridge the semantic gap between the text semantic space and the feature space of the recommendation model. Directly inputting text embedding vectors into the recommendation model often yields poor results because the representation targets and distributions of the two differ. To address this, this application introduces a multi-expert hybrid model architecture to achieve cross-modal deep alignment and information purification.

[0036] First, the user preference summary text and product description obtained in the dialogue knowledge generation stage are used to obtain the corresponding user preference embedding and product fact embedding through a text embedding model. Then, a multi-expert hybrid model (MoE) architecture is adopted to route the user preference embedding and product fact embedding to different expert networks for processing through a gating mechanism. Finally, the expert network input is summarized and passed through the user representation layer and product representation layer to obtain the enhanced user preference embedding and product fact embedding.

[0037] The third stage involves retrieval and re-ranking tasks; see [link / reference] Figure 4 , Figure 4 The third flowchart provided in this application embodiment incorporates the two types of enhancement embeddings optimized in the second stage as additional key reference information into the recommendation recall model and the re-ranking model. The model performs joint calculations based on these enhancement information to ultimately select a list of the top 5 candidate products. This Top 5 candidate product list is not output all at once. The subsequent two tasks, "model fine-tuning based on reinforcement learning" and "conversational recommendation based on dynamic examples and candidate products," will repeatedly call this list and will also trigger this task to update the candidate product list again based on new situations such as the progress of the dialogue and changes in user needs, to ensure that the recommendation basis is always accurate.

[0038] Task 2 of this application aims to train a large-scale language model into an intelligent agent capable of making optimal decisions based on dynamic dialogue environments. This stage formalizes the conversational recommendation process as a Markov decision process and, through an innovative swarm policy optimization framework, enables the model not only to learn to generate natural language responses but also to learn to select the best strategy in a complex decision space (follow-up questions, answers, recommendations).

[0039] The process is mainly divided into four parts: environment update, action decision, group reward feedback, and weight update. Each round of dialogue begins with an environment update, updating the previous round's dialogue content to the historical dialogue, and invoking a product retrieval and re-ranking task to update the current round's Top 5 candidate product list. Then, a dialogue recommendation action prompt template is constructed. The prompt strategy language model selects one action from three options: follow-up question, answer, and recommendation, and generates the corresponding dialogue. For the same set of dialogue recommendation action prompts, different hyperparameters are adjusted to generate N different action decision results. The reward model then scores the action decision results based on factors such as product click-through rate, dialogue quality, and penalties for excessively long dialogue rounds, calculating the group's average score and standard deviation. Finally, the scores of the group's N action decision results are normalized before gradient calculation and weight update of the strategy model. Gradient clipping and KL regularization are used to prevent excessively large updates to the strategy model. A reinforcement learning-based model fine-tuning task is used to improve the product recommendation accuracy and dialogue context fluency of the strategy model in conversational recommendation tasks.

[0040] Task 3 of this application is a conversational recommendation task based on dynamic examples and candidate products. In each round of dialogue, the user poses a question to the dialogue recommendation system. The system calls the product retrieval and re-ranking task to update the Top 5 candidate product list for the current round, and then calls the dynamic example module to semantically match the five examples closest to the user's question from a manually constructed example pool. The examples and candidate product list, together with the user's question, construct a dialogue recommendation action example prompt template. The prompt is used by a strategy model trained by the model fine-tuning task to make action decisions and generate responses, and finally outputs the response text and recommended products to the user. In each round of dialogue, the product retrieval and recommendation task is called to update the candidate product list. As the dialogue deepens and the user's needs become richer, the candidate product list will be more precisely narrowed down to the range of user preferences.

[0041] The following provides further explanation of each task.

[0042] Task 1 involves product retrieval and re-ranking based on dialogue knowledge enhancement. This task enhances recommendation recall and re-ranking accuracy by collecting knowledge relevant to multi-turn dialogue recommendations, thereby improving performance in subsequent dialogue recommendation tasks. The task is divided into three stages: knowledge generation, knowledge enhancement, and product recall and re-ranking.

[0043] The knowledge generation stage is mainly divided into user preference summary text generation and product fact text generation. For user preference summary text generation, historical dialogues, user-product interaction data, and user profile data are collected. A zero-shot prompting inference method is used. The prompt template is a structured instruction framework designed to guide the large language model in user preference analysis and to perform in-depth inference and summarization of multi-source heterogeneous user data. Its design logic lies in systematically organizing three types of key information to simulate the thought process of human experts making comprehensive judgments.

[0044] The template first injects a static background of the user, providing profile tags such as age, gender, and region through {{User Profile}}, establishing a basic context and long-term perspective for understanding user preferences. Next, it presents dynamic behavioral evidence, listing the user's historical browsing, clicks, purchases, and other interaction records in {{User-Item Interaction}}. This data reveals stable, long-term user interest patterns and consumption tendencies, serving as core factual evidence for preference analysis. Finally, it incorporates real-time dialogue context, providing multiple rounds of the current conversation through {{History Dialogue}}. This section captures the user's latest expressions of intent, clarifications of needs, and feedback, reflecting their short-term and potentially evolving specific needs.

[0045] After integrating the above information, the template concludes with a clear instruction: "Analyze user preferences. Please provide a clear explanation based on their browsing history and other relevant factors." This instruction not only requires the model to generalize facts but also emphasizes that it must perform causal relationships and reasoning, that is, combine long-term behavioral patterns with short-term dialogue context to deduce users' explicit and implicit preferences and generate a structured, interpretive text summary.

[0046] Among them, user profile data and user-product interaction data provide personalized background information to help the large language model understand long-term user preferences; historical dialogue data, by collecting users' current questions and needs descriptions, helps the large language model understand short-term user demand shifts. For product factual text, this study processes product data and combines it with business knowledge, integrating numerical and enumerable product data into the product description to obtain the product factual text.

[0047] In the knowledge enhancement stage, this study first employs the Sentence Embedding Model (Sentence-BERT) to convert the user preference summary text and product fact text obtained in the previous stage into vector representations. Specifically, the resulting labels for the user preference summary text *u* and the product fact text *p* after embedding are as follows:

[0048] Because the vector spaces of text embeddings and recommendation ranking models differ, directly using product preference embeddings as input to the recommendation ranking model yields poor results. This study employs a multi-expert hybrid model (MoE) architecture, designing E = 3 independent expert networks to integrate and process the embedded information. Each expert network is an independent fully connected network layer. The expert networks organize common and independent information from user preference texts and product fact texts, and finally, through a gating mechanism, summarize the expert network outputs. The enhanced user preference embeddings and product fact embeddings are then obtained through user representation layers and product representation layers. (Enhanced user preference embeddings are then described.) and product fact embedding The calculation formula is as follows:

[0049] in, Refers to the network of experts. and A gating network that refers to user preferences and product facts. and These are the weights of the expert network output by the gated network, used for the final embedded weighted integration.

[0050] As an additional input domain, user preferences are added to the existing recommendation recall and ranking models. During training, the expert network and the recommendation model work together in the product recall and re-ranking stages. This study embeds the enhanced user preferences. Product fact embedding optimization is used to ensure that the embedding vectors enhanced by the expert network are fused with the vector distribution of the recommendation model. Furthermore, the handling of the additional input domain allows us to freely switch between different recommendation models to adapt to different recommendation tasks and compare the performance of different models during practice.

[0051] Task 2 is a model fine-tuning task based on group policy optimization. This task aims to model the conversational recommendation process as a Markov Decision Process (MDP) using a group policy optimization framework. This allows the large policy model to select the optimal action (follow-up question, answer, or recommendation) based on the dynamic dialogue environment, thereby improving product recommendation accuracy and conversational coherence. The overall process consists of four stages: environment update, action decision, group reward feedback, and weight update.

[0052] During the environment update phase, the main elements of dynamic environment updates in this framework are three: the user question in the current round, historical dialogue records, and the candidate product list. In the t-th round of dialogue, the product recall and reordering tasks are invoked to update the candidate product list, outputting the Top 5 candidate products for round t (denoted as...). Dialogue with History and user issues Piecing together to construct the environmental state :

[0053] During the action decision-making phase, a policy network is built based on a pre-trained large language model. This paper employs a zero-shot suggestion method to construct a dynamic suggestion template for action generation. The suggestion template is a structured decision instruction framework designed to guide a policy network based on a large language model to perform contextualized reasoning and action selection in the dynamic environment of dialogue recommendation. It systematically integrates the environmental information, user state, and explicit action specifications necessary for decision-making to simulate the judgment process of human customer service or recommendation experts in similar scenarios.

[0054] The template first defines the resources and context for decision-making: by providing the Top 5 most relevant candidate products retrieved in the current round through {{Candidates}}, it clarifies the scope of recommendations that the system can make; by providing the complete history of the current dialogue through {{History Dialogue}}, it ensures that the decision-making can follow the context and understand the evolution of the needs.

[0055] The decision-making process then focused on the direct trigger point: the user's current problem was clearly identified through the {{Query}}, which was the most immediate input driving the decision.

[0056] After fully explaining the context, the template constrains the output with explicit instructions: requiring the model to choose one of three predefined actions—follow-up question (actively clarifying when information is insufficient), answer (directly responding to the user's explicit question), or recommend (providing explanatory recommendations when matching products exist)—and generate the corresponding response text. This design both gives the model the flexibility to make decisions based on the context and constrains its behavioral scope through action definitions, ensuring that the output conforms to business logic and dialogue objectives. Overall, the template transforms the open-ended language generation task into a controllable and evaluable sequential decision problem through a progressive structure of background—current situation—task—constraints.

[0057] For the same prompt template, this study uses a hyperparameter adjustment method (such as temperature, top-k, etc.) to generate N = 4 sets of differentiated action decisions. And the corresponding reply text.

[0058] In the group reward feedback reward phase, this study constructs a reward model. Action decisions are evaluated based on product click-through rate, conversation flow level, and recommendation round penalty. Among these factors, product click-through rate... The probability of users clicking on recommended products is assessed using an auxiliary CTR prediction model; conversation fluency. The system calculates semantic coherence scores for responses to context based on a large language model; round-based penalty. If the number of dialogue rounds exceeds a threshold T, a negative reward is applied to promote efficient decision-making. The reward model calculation formula is as follows:

[0059] in, The manually set reward weights are dynamically adjusted based on the model's final fine-tuning performance to prevent the model from overly favoring the generation of fluent responses while neglecting the recommendation process. Then, this study calculates the average reward for action decisions within the same group. and standard deviation The reward for each action is normalized:

[0060] During the weight update phase, this study sets an optimization objective. To maximize the expected value of the normalized action reward:

[0061] This study also employs KL divergence regularization and gradient pruning to prevent the updated policy model from deviating too much from the initial model and to prevent catastrophic regression. Therefore, the formula for calculating the policy model weight update is as follows:

[0062] in, As a regularization constraint, the update magnitude is dynamically adjusted during training. This represents the KL divergence between the policy model and the initial model.

[0063] By combining GRPO for model fine-tuning training, this study avoids local optima in single-action decision-making strategies through hyperparameter adjustment of multi-action decision-making strategies. The normalized group reward mechanism feedback reduces the impact of differences in reward scales on strategy updates. Furthermore, this task is deeply coupled with product recall and re-ranking tasks, achieving closed-loop optimization.

[0064] For user question vectors Calculate its relationship with each example problem vector. Calculate the cosine similarity and select the Top N = 5 most relevant examples:

[0065] This study employs a few-shot suggestion method for constructing suggestion templates. In each round of dialogue, the product recall and re-ranking task is invoked to update the candidate product list, ensuring that the recommendation results evolve in real-time with user needs. Then, based on semantic matching, the most relevant examples are matched from the example library, allowing the model to still refer to similar cases for decision-making in the cold start problem. Finally, historical dialogues and the user's current question are integrated to generate the final dialogue product recommendation suggestions. The suggestion template is shown in task 2.

[0066] In the action decision-making and response generation stages, this study uses a large policy model trained on task2. The inference prompt template outputs the action type and a natural language response. Its action constraint logic is as follows: if the match between a candidate product and the user's needs is greater than the model's judgment threshold, a recommendation action is triggered to suggest one of the candidate products to the user; if the user's needs are ambiguous, a follow-up question action is triggered to inquire about the product attributes the user prefers; if the user raises a question, a response action is triggered to answer the user's question. After outputting the response and recommended product generated by the strategy model to the user, the content of this round of dialogue is updated to the dialogue history, and the next round of dialogue begins.

[0067] Dynamic examples provide cross-scenario generalization capabilities, reducing decision-making errors of the strategy model in edge cases, and the real-time updated candidate product list enables the recommendation scope to shrink to the user's true preferences as the conversation progresses.

[0068] In one possible implementation, updating the weight parameters of the large policy model begins with designing the objective function. This function comprises two key parts: first, it is driven by maximizing the expected normalized reward, directly guiding the model to learn decisions that yield higher evaluation scores; second, it introduces a KL divergence regularization term. This regularization term measures and constrains the difference in output distribution between the updated policy model and the old policy model. By adjusting the strength of this term with a preset constraint coefficient, the system can achieve a balance between actively exploring better policies and maintaining relative policy stability, thereby effectively preventing model forgetting or drastic fluctuations in decision quality caused by excessively large single updates.

[0069] After clarifying the optimization objective, the process enters the gradient calculation and processing stage. Based on multiple sets of action decisions generated from the current model sampling and their corresponding normalized rewards, the system calculates the gradient of the objective function relative to the model weights. This gradient indicates the direction and intensity of parameter updates. Next, the calculated original gradient is truncated using the L2 norm, i.e., gradient pruning, forcibly limiting its magnitude to a preset safety threshold. This operation is crucial for training stability, preventing excessively large gradient values ​​from "pushing" the model parameters into undesirable regions in cases of individual abnormal samples or inaccurate reward estimation.

[0070] Finally, parameter iteration and convergence determination are performed. The system uses the pruned gradients, combined with a small preset learning rate, to iteratively update the weight parameters of the large policy model. The learning rate is set on the order of one ten-thousandth to one thousandth, ensuring a robust update pace. The training process continues until preset convergence conditions are met: either the objective function value changes negligibly over multiple iterations, with the difference between adjacent iterations not exceeding one ten-thousandth, indicating that the optimization has stabilized; or the preset maximum number of iterations is reached. At this point, updates stop, and the resulting model weights are the final parameters of the trained large policy model.

[0071] In practical applications, Task 3 represents the final user-facing service form and real-time evolution interface of the entire system. It is not a static dialogue script, but a dynamically perceptive, real-time decision-making, and self-iterable intelligent agent. Its core design philosophy is to simultaneously integrate the latest perception results (candidate products) from within the system with prior external knowledge (successful examples) in each round of interaction, driving the optimized strategy model to make the best response and forming a closed loop.

[0072] The process begins with the user posing the current question, which triggers the system's real-time response mechanism. The system initiates two key retrieval tasks in parallel: First, it invokes the retrieval and re-ranking engine from Task 1, taking the complete dialogue history incorporating this new question as input, and re-executes deep semantic retrieval to ensure that the resulting Top 5 candidate product list is strictly aligned with the user's latest comprehensive intent. Second, it initiates dynamic example matching, calculating the similarity between the current question and historical questions in a manually constructed example library using a semantic embedding model, quickly filtering out the most relevant successful dialogue cases. These cases provide a reference strategy paradigm for the current decision.

[0073] Subsequently, the system enters the information fusion and decision-making stage. It intelligently integrates information from four dimensions—the real-time acquisition of the candidate product list, matching examples, complete dialogue history, and the current user question—to construct a highly information-dense few-sample prompt template.

[0074] Next, the strategy model performs inference and output. The prompt is input into the strategy language model, which has been fine-tuned through reinforcement learning in Task 2. Based on internalized decision-making logic, the model comprehensively analyzes the current situation and outputs the final selected action (follow-up question, answer, or recommendation) and a corresponding natural and fluent natural language response, completing this round of response to the user.

[0075] Specifically, such as Figure 5 As shown, Figure 5 The fourth flowchart provided for the embodiments of this application includes the following implementation process: S101. In response to the current user question, obtain the user preference summary text and product fact text. S102. Based on the text embedding model, transform the obtained user preference summary text and product fact text into an initial embedding vector.

[0076] S101 and S102 constitute a parallel dual-channel system for information perception and preparation. S101 responds to the current user question, invoking and integrating the knowledge generation capabilities of Task 1's product retrieval and reordering. Based on a complete history including the latest dialogue, it ensures the timeliness of basic semantic understanding. Simultaneously, S102 performs semantic matching of the current user question with a pre-set example library, thus acquiring external prior knowledge, i.e., successful examples, in parallel. This dual-path information acquisition mechanism ensures that the system possesses both a deep understanding of the current context and the ability to quickly draw upon historical experience when making decisions.

[0077] S103. Input the initial embedded vector into a multi-expert hybrid network so that each expert network processes the input vector representation in parallel and outputs the enhanced features from the perspective of the expert network.

[0078] S104. Based on the weights output by the gating network, the intermediate features processed by each expert network are weighted and summed to generate enhanced user preference embedding vectors and enhanced product fact embedding vectors, respectively.

[0079] S105. Input the enhanced user preference embedding vector and the enhanced product fact embedding vector into the recall model so that the recall model can calculate the matching degree between users and products, and retrieve and generate a preliminary candidate product set from the full product pool.

[0080] S106. Input the enhanced product fact embedding vectors corresponding to the preliminary candidate product set and the initial related candidate product set into the reordering model to obtain the candidate product set.

[0081] S103 and S104 enhance and align the initial embedding vectors using a multi-expert hybrid network, representing the specific implementation of the knowledge enhancement stage in Task 1, aiming to improve the quality of semantic representations. S105 and S106 perform recall and reordering respectively, ultimately outputting a set of candidate products strictly aligned with the current dialogue state in S106. This series of steps encapsulates and invokes the capabilities of Task 1, and its output forms the factual basis and optional action space for subsequent decisions.

[0082] S107. Based on the dialogue example, the candidate product set, and the current user question, generate an action prompt template.

[0083] Based on the dialogue example obtained in S102, the candidate product set obtained in S106, and the original current user question, a contextualized prompt template is dynamically constructed. This template intelligently integrates information from three dimensions: "what can be pushed," "how similar situations were handled in the past," and "what the user is specifically asking now," providing the strategy model with a complete and clear decision context.

[0084] S108. Input the decision prompt template into the pre-trained strategy generation model so that the strategy generation model outputs recommended decision actions and / or generates corresponding natural language responses for the current user problem.

[0085] S108 inputs the prompt template constructed in S107 into the pre-trained policy generation model, which is the model fine-tuned from Task 2. This model integrates all information and, based on its internalized optimization strategy, outputs the final decision action and natural language response, completing this round of interaction with the user.

[0086] The formation of closed loops and iterations is implicit in the cyclical execution of the entire process. The response output of S108 and any further user feedback that may arise will be recorded and updated in the dialogue history. When the next round of user questions is triggered, this updated history will become the new input for step S101, thereby driving a new round of decision-making loops with deeper understanding and more accurate candidates. Therefore, S101-S108 not only defines the complete path of an interaction, but also, through its cyclical execution mechanism, enables the system to achieve self-evolutionary capabilities, becoming more accurate and understanding with each use in real-world services.

[0087] In one possible implementation, the process is cyclically executed within an online recommendation dialogue. For each new question posed by the user in the current round, all steps in claim 1, from "obtaining the user preference summary text and product fact text" to "outputting a natural language response," are re-executed based on the updated historical dialogue records. The candidate product set is dynamically updated as the number of dialogue rounds increases and the user's needs become richer.

[0088] In one possible implementation, a dialogue history update step is also included, which is performed after the policy generation model outputs a recommended decision action and / or a natural language response, specifically including: The structured information of this round of interaction is associated and stored in the dialogue history database. The structured information includes the current user question, the type of recommendation decision action, the natural language response content, the candidate product set (if the recommendation action is triggered), and the user's real-time feedback information (click / not click, whether to continue the consultation). The updated data in the dialogue history database is used to generate user preference summary text in the next round of interaction, as well as to construct the environmental state representation during the training of the large strategy model, to ensure that the next round of decision is aligned with the user's latest needs.

[0089] One possible implementation also includes an iterative update step for summarizing user preferences in the text, specifically including: After each round of interaction, the structured information of this round, stored in the dialogue history update steps, is added to the "historical dialogue records" and "user product interaction data" in the user's full data. This triggers the large language model to re-execute the inference process, integrate the updated user full data, iteratively optimize the original user preference summary text, add or correct the user's short-term intent and latest preference tendencies, and generate an updated user preference summary text. The updated user preference summary text replaces the original text and serves as the basis for generating the initial embedding vector of user preferences in the next round of interaction.

[0090] In one possible implementation, a dynamic update step for the example library is also included, specifically: setting an update cycle for the example library (ranging from 7 to 30 days), or triggering an update when the cumulative number of newly added valid dialogue data reaches a preset threshold (ranging from 100 to 500); filtering valid dialogue examples from the newly added real interaction data, with the filtering criteria being: user feedback satisfaction ≥ 80%, dialogue rounds ≤ preset efficiency threshold, and action decision matching degree ≥ 0.7; structuring the filtered valid dialogue examples and adding them to the example library, while removing invalid examples from the example library that meet the criteria of "no matching call records in the past 6 months" or "user feedback satisfaction < 30%"; the dynamic update step is used to maintain the timeliness and effectiveness of the example library and improve the accuracy of subsequent semantic matching.

[0091] In one possible implementation, the retrieval subsystem, composed of the multi-expert hybrid network, the recall model, and the re-ranking model, and the decision-making subsystem, composed of the policy generation model, are collaboratively updated online or offline. This includes: generating training data for optimizing the retrieval subsystem based on the decision results generated by the policy generation model in the dialogue and user feedback; and using the training data to periodically or triggerfully update the multi-expert hybrid network, the recall model, and / or the re-ranking model in the retrieval subsystem to optimize product recall accuracy.

[0092] This application also provides corresponding devices and computer storage media for implementing the user-question-based recommendation scheme provided in this application.

[0093] The device includes a memory and a processor. The memory stores instructions or code, and the processor executes the instructions or code to cause the device to perform the user-question-based recommendation method described in any embodiment of this application.

[0094] The computer storage medium stores code, and when the code is run, the device running the code implements the user-question-based recommendation method described in any embodiment of this application.

[0095] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems or apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple, and relevant parts can be referred to the method section.

[0096] It should be understood that in this application, "at least one" refers to one or more items, and "more" refers to two or more items. "And / or" is used to describe the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: only A exists, only B exists, and both A and B exist simultaneously, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one" or similar expressions refer to any combination of these items, including any combination of singular or plural items. For example, "at least one" of a, b, or c can represent: a, b, c, a and b, a and c, b and c, or a and b and c, where a, b, and c can be single or multiple.

[0097] It should be understood that the terms center, longitudinal, transverse, up, down, front, back, left, right, vertical, horizontal, top, bottom, inside, outside, etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the present invention.

[0098] It should be noted that, unless otherwise explicitly specified and limited, the terms installation, connection, and linking should be interpreted broadly. For example, they can refer to fixed connections, detachable connections, or integral connections; they can refer to mechanical connections or electrical connections; they can refer to direct connections or indirect connections through an intermediate medium; and they can refer to the internal communication between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0099] It should also be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the statement "comprising a..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0100] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented directly by hardware, a software module executed by a processor, or a combination of both. The software module can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art.

[0101] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A recommendation method based on user questions, characterized in that, include: In response to current user questions, obtain user preference summary text and product fact text; Based on the text embedding model, the obtained user preference summary text and product fact text are transformed into initial embedding vectors, which include user preference initial embedding vectors and product fact initial embedding vectors. Based on the text embedding model, the current user question is semantically matched with a preset example library to select several dialogue examples that match the current user question; The initial embedding vector is input into a multi-expert hybrid network so that each expert network processes the input vector representation in parallel and outputs enhanced features from the perspective of the expert network. The multi-expert hybrid network comprises multiple independent expert sub-networks. Expert weight distributions on the user side and product side are calculated separately through a parallel gating network. Based on the expert weight distribution, the input vector is dynamically routed to different expert networks for parallel processing, so that each expert network performs a nonlinear transformation on the input and outputs intermediate features from its specific perspective. Based on the weights output by the gating network, the intermediate features processed by each expert network are weighted and summed to generate enhanced user preference embedding vectors and enhanced product fact embedding vectors, respectively. The enhanced user preference embedding vector and the enhanced product fact embedding vector are jointly input into the recall model so that the recall model can calculate the matching degree between users and products, and retrieve and generate a preliminary candidate product set from the full product pool. The enhanced product fact embedding vectors corresponding to the preliminary candidate product set and the initial relevant candidate product set are input into the reordering model to obtain the candidate product set. Based on the dialogue example, the candidate product set, and the current user question, an action prompt template is generated; The decision suggestion template is input into a pre-trained strategy generation model, so that the strategy generation model outputs a recommended decision action for the current user's question and / or generates a corresponding natural language response.

2. The method according to claim 1, characterized in that, The methods for determining the pre-trained policy model include: The environmental state is represented by integrating historical dialogue records, user questions, and candidate product sets. Based on the environmental state representation, an action prompt template is generated. Based on the action prompt template, the hyperparameters are adjusted so that the large strategy model can generate multiple different action decisions and corresponding response messages in parallel. Each action decision is scored from multiple dimensions based on a reward model, and the original reward score for each action decision is calculated. For the multiple action decisions, the mean and standard deviation of the original reward scores are calculated, and the original reward scores of each action decision are normalized based on the mean and standard deviation to obtain normalized rewards. The weight parameters of the large policy model are updated with the goal of maximizing the expected value of the normalized reward.

3. The method according to claim 2, characterized in that, The multiple dimensions include product click-through rate (CTR), dialogue quality, and round efficiency penalty. The product CTR is used to characterize the probability of a recommended product being clicked by the auxiliary CTR prediction model. The dialogue quality is used to characterize the semantic coherence and contextual relevance of the generated response. The round efficiency penalty is used to indicate that a negative reward is applied when the number of dialogue rounds exceeds a threshold.

4. The method according to claim 1, characterized in that, The acquisition of user preference summary text and product fact text includes: Acquire full user data related to users and full product data related to products. The full user data includes historical dialogue records, user profile data, and user product interaction data. The full product data includes raw product data, which includes tag fields and numeric fields. Using a large language model based on a preset sample prompt template, inference is performed on the historical dialogue data, user product interaction data, and user profile data to generate a user preference summary text that summarizes the user's long-term preferences and short-term intentions; By integrating key information from tag-type and numerical fields, and supplementing them with descriptions of core product attributes and usage scenarios, product fact text is obtained, which is used to characterize the core features of the product.

5. The method according to claim 1, characterized in that, The step of semantically matching the current user question with a preset example library to select several dialogue examples that match the current user question includes: Encode the current user question into a first vector; Encode the user question field of each dialogue example in the example library into a corresponding second vector; Calculate the cosine similarity between the first vector and each of the second vectors; The dialogue examples are sorted from high to low according to the cosine similarity, and the top N dialogue examples are selected as the filtered dialogue examples, where N is a preset positive integer.

6. The method according to claim 1, characterized in that, The strategy generation model outputs recommended decision actions and / or generates corresponding natural language responses for the current user's question, including: If the matching degree between the candidate product and the user's needs is greater than the preset model judgment threshold, the recommendation action is triggered, and the product with the highest matching degree among the candidate products and a natural language response containing the product's advantages are output. If the user's request does not include key preference information, a follow-up question action is triggered, and a targeted question script is output to address the missing preference information; If the user's current question is a consultation question that includes at least one of the following: product parameters, usage scenarios, or functional features, then a response action is triggered, and a natural language response directly corresponding to the question is output.

7. The method according to claim 2, characterized in that, The step of updating the weight parameters of the large policy model with the objective of maximizing the expected value of the normalized reward includes: Define an objective function with the expected value of the normalized reward as the core optimization term. At the same time, introduce a KL divergence regularization term to constrain the deviation between the new strategy and the old strategy before the update. The KL divergence regularization term adjusts the constraint strength through a preset constraint coefficient to avoid excessive policy update amplitude leading to model performance oscillation. Based on the multiple action decisions obtained from sampling and the corresponding normalized rewards, the gradient of the objective function with respect to the weight parameters of the policy model is calculated. The calculated gradient is truncated using the L2 norm to keep the L2 norm of the gradient within a preset threshold. Based on the trimmed gradient, the weight parameters of the large model of the policy are iteratively updated according to a preset learning rate, wherein the learning rate ranges from one ten-thousandth to one thousandth. When the absolute value of the difference between two adjacent rounds of the objective function value does not exceed one ten-thousandth, or when the number of iterations reaches the preset maximum number of iterations, the weight parameter update stops. The weight parameters at this time are the weight parameters of the trained policy model.

8. The method according to claim 2, characterized in that, The integration methods for the environmental state representation include: Historical dialogue records are transformed into text sequence vectors, current user questions are transformed into query vectors, and candidate product sets are transformed into product feature matrices. Through vector concatenation and matrix flattening, an environment state vector with uniform dimensions is generated.

9. The method according to claim 1, characterized in that, The re-ranking model is used to concatenate the enhanced product fact embedding vector with the product's inherent attribute features and user interaction behavior features. It captures the correlation weights between features through a multi-layer attention mechanism and then inputs them into a fully connected layer to output a ranking score.

10. The method according to claim 5, characterized in that, Each dialogue example in the preset example library is stored as a record, and each record contains at least a user question field, an action decision field, and a response text field. The dialogue examples in the example library cover a variety of typical dialogue recommendation scenarios, including product recommendation, attribute filtering, and requirement clarification.

Images