User intention recognition method and system based on user behavior data

Abstract

The invention is suitable for the cross field of artificial intelligence and natural language processing, and provides a user intention recognition method and system based on user behavior data. In the embodiment, for a target user needing intention recognition, a multi-modal data set composed of behaviors, texts, audios, videos and the like of the user is firstly obtained, and then the multi-modal data set is processed through a world model used for reinforcement learning to obtain simulation data of the user. The method comprises the steps of obtaining a multi-modal data set, performing intention recognition on a user through the multi-modal data set and simulation data to obtain an initial intention recognition result and a corresponding evaluation index, and finally adjusting the initial intention recognition result through the evaluation index to determine a final target intention recognition result. According to the method and the device, for the data with relatively small data volume in the multi-modal data set, simulation expansion is carried out based on the multi-modal data set through the world model, so that the accuracy of the intention recognition model obtained by training can still be remarkably improved even under the condition that behavior samples are insufficient, and thus the accuracy of intention recognition of the user is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of artificial intelligence and natural language processing, in particular to a user intention recognition method and system based on user behavior data. BACKGROUND

[0002] The existing intention recognition method based on user behavior data usually includes three parts of behavior data capture, intention analysis and result generation. Among them, the core of the intention recognition module is the model trained by deep learning or large language model (LLM). In order to construct such a model, the processes of user behavior data collection and labeling, data preprocessing and augmentation, etc. need to be completed, and then model training is carried out. Since the intention recognition algorithm based on deep learning is usually data-driven, only when the user behavior data samples of each type are sufficient, the model can achieve ideal recognition effect. If the sample of a certain type of user behavior is insufficient, the intention recognition accuracy of the model on this type of behavior is usually low. The existing data augmentation methods, such as behavior sequence transformation, feature normalization, etc., can effectively improve the model performance when the sample is sufficient, but when the sample of a specific behavior is small (usually less than 100 records), the data generated by ordinary augmentation methods has high similarity, and the augmentation effect is limited, resulting in insufficient generalization ability of the trained model. SUMMARY

[0003] Therefore, the present application provides a user intention recognition method and system based on user behavior data to improve the accuracy of user intention recognition in the case of insufficient behavior samples.

[0004] The first aspect of the present application provides a user intention recognition method based on user behavior data, which comprises: generating a multi-modal data set according to the behavior actions of a target user, inputting the multi-modal data set into a trained world model for reinforcement learning, and obtaining simulation data of the target user through the world model, wherein the multi-modal data set includes behavior analysis results, text data and audio and video data; performing intention recognition on the target user according to the multi-modal data set and the simulation data to determine an initial intention recognition result and an evaluation index of the initial intention recognition result, wherein the evaluation index includes illusion rate, consistency score, accuracy rate and context relevance; adjusting the initial intention recognition result according to the evaluation index to obtain a target intention recognition result of the target user, and generating recommended content of the target user through the target intention recognition result.

[0005] Optionally, the simulation data of the target user obtained through the world model comprises: The target sample data is determined by an amount of data in the multi-modal data set, and the target sample data is simulated and expanded based on the multi-modal data by the world model to obtain simulated data of the target sample data.

[0006] Optionally, the intention recognition of the target user according to the multi-modal data set and the simulated data comprises: The multi-modal data set and the simulated data are analyzed and inferred by an LLM Evaluation model to obtain an initial intention recognition result of the target user, and the initial intention recognition result is analyzed by a comprehensive evaluation system composed of a plurality of LLM effect evaluation indexes to determine a corresponding evaluation index.

[0007] Optionally, the adjustment of the initial intention recognition result according to the evaluation index comprises: The comprehensive score of the evaluation index is calculated by an LLM-as-a-Judge model, and when the comprehensive score is greater than a threshold value, the initial intention recognition result is determined as the target intention recognition result; When the comprehensive score is less than the threshold value, the steps of obtaining the simulated data of the target user by the world model, and re-determining the initial intention recognition result and the corresponding evaluation index by the updated simulated data, and calculating the comprehensive score of the evaluation index by the LLM-as-a-Judge model are repeatedly executed until the comprehensive score is greater than the threshold value.

[0008] Optionally, the method further comprises: The multi-modal data set, the simulated data, and the target intention recognition result of the target user are stored in a memory database, and the LLM Evaluation model and the LLM-as-a-Judge model are cyclically iteratively optimized by the data in the memory database, wherein the memory database is intelligently searched by a Memory R1 framework.

[0009] The second aspect of the present application provides a user intention recognition system based on user behavior data, which comprises a data acquisition simulation module, an intention recognition module, and a user content generation module. The data acquisition simulation module is used to generate a multi-modal data set according to the behavior actions of a target user, input the multi-modal data set into a trained world model for reinforcement learning, and obtain simulated data of the target user by the world model, wherein the multi-modal data set comprises behavior analysis results, text data, and audio and video data. The intention recognition module is configured to perform intention recognition on the target user according to the multi-modal data set and the simulation data to determine an initial intention recognition result of the target user and an evaluation index of the initial intention recognition result, wherein the evaluation index comprises an illusion rate, a consistency score, an accuracy rate, and a context relevance. The user content generation module is configured to adjust the initial intention recognition result according to the evaluation index to obtain a target intention recognition result of the target user, and generate recommended content of the target user through the target intention recognition result.

[0010] Optionally, the simulation data of the target user obtained through the world model in the data collection simulation module comprises: Target sample data is determined according to the amount of data in the multi-modal data set, and simulation data of the target sample data is obtained by simulating expansion of the target sample data based on the multi-modal data through the world model.

[0011] Optionally, the intention recognition module comprises: The multi-modal data set and the simulation data are analyzed and inferred through an LLM Evaluation model to obtain an initial intention recognition result of the target user, and the initial intention recognition result is analyzed through a comprehensive evaluation system composed of a plurality of LLM effect evaluation indexes to determine a corresponding evaluation index.

[0012] Optionally, the user content generation module comprises: A comprehensive score of the evaluation index is calculated through an LLM-as-a-Judge model, and when the comprehensive score is greater than a threshold value, the initial intention recognition result is determined as the target intention recognition result. When the comprehensive score is less than the threshold value, the steps of obtaining simulation data of the target user through the world model, and determining an initial intention recognition result and a corresponding evaluation index through the updated simulation data are repeated until the comprehensive score is greater than the threshold value.

[0013] Optionally, the system further comprises: The model optimization module is configured to store the multi-modal data set of the target user, the simulation data, and the target intent recognition result into a memory database, and perform cyclic iteration optimization on the LLM Evaluation model and the LLM-as-a-Judge model through the data in the memory database, wherein the memory database is intelligently searched through a Memory R1 framework.

[0014] In the embodiments provided in the present application, for a target user who needs to perform intent recognition, a multi-modal data set composed of behaviors, texts, audio and video of the user is first obtained, and then the multi-modal data set is processed through a world model for reinforcement learning to obtain simulation data of the user. The user is then subjected to intent recognition through the multi-modal data set and the simulation data to obtain an initial intent recognition result and a corresponding evaluation index. Finally, the initial intent recognition result is adjusted through the evaluation index to determine a final target intent recognition result. For data with a small amount of data in the multi-modal data set, the world model is used to simulate expansion based on the multi-modal data set, so that even in the case of insufficient behavior samples, the accuracy of the intent recognition model trained can still be significantly improved, thereby improving the accuracy of intent recognition of the user. BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1 Method flowchart provided for embodiments of the present application; Figure 2 System module diagram provided for embodiments of the present application; Figure 3 Computer device internal structure schematic diagram provided for embodiments of the present application. DETAILED DESCRIPTION

[0016] The exemplary embodiments will be described in detail herein with reference to the attached drawings. The following description is only exemplary and is not intended to limit the present application. In the following description, same numbers refer to same elements in all the drawings. The following exemplary embodiments described in the following exemplary embodiments are not intended to represent all embodiments consistent with the present application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of the present application as detailed in the appended claims.

[0017] The terms used in the present application are merely for the purpose of describing specific embodiments and are not intended to limit the present application. The singular forms "a", "an" and "the" used in the present application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term "and / or" used herein means and includes any or all possible combinations of one or more associated listed items.

[0018] It should be understood that although the terms first, second, third, etc., may be used in this application to describe various information, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, without departing from the scope of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to determination."

[0019] This application provides a user intent recognition method and system based on user behavior data to improve the accuracy of user intent recognition.

[0020] First, some technical terms used in this invention will be explained: Consistency Score: The consistency score is a metric that quantifies the stability of a model's output. It is evaluated by calculating the cosine similarity between the results of multiple runs, ensuring the consistency of the model's response under similar inputs.

[0021] Hallucination Rate: The hallucination rate is a metric for evaluating the content generated by a large language model. It quantifies the proportion of the model's output that does not conform to reality. It is often calculated by matching the model's output with real data and is used to measure the model's reliability.

[0022] Bayesian inference is a probabilistic inference method based on Bayes' theorem. It updates posterior beliefs through prior distribution and likelihood function and is suitable for uncertainty modeling and parameter estimation.

[0023] The Memory R1 framework employs two dedicated agents for proactive memory management: the MemoryManager learns structured operations, including add, update, delete, and null operations, to maintain and evolve the external memory library; the Answer Agent pre-selects relevant entries, filters and enhances the generated results through memory distillation strategies, and then performs inference to generate the answer. This framework primarily enhances the persistence and generalization capabilities of LLM agents in multi-session dialogues and long-context tasks, such as filtering noise and integrating relevant facts in multi-hop inference, temporal problems, or open-domain QA, thereby improving performance.

[0024] The technical solutions of this application will be described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0025] like Figure 1As shown, a flowchart of a user intention recognition method based on user behavior data provided by the present application can include the following steps: Step S101, generating a multi-modal data set according to the behavior actions of the target user, inputting the multi-modal data set into the trained world model for reinforcement learning, and obtaining simulation data of the target user through the world model.

[0026] In this embodiment, a reinforcement learning-based world model needs to be built first to simulate and predict user behavior. The data sources of the world model mainly come from the collected historical data of various users, including user behavior analysis results (such as click sequences and intention labels), text content (such as conversation logs and descriptive narratives), generated videos (such as demonstration videos simulating user interaction), etc. These data are integrated into a multi-modal data set, covering text, image and video, etc. to capture the complexity of user intention. After training the world model with these historical data, a trained world model for reinforcement learning is obtained.

[0027] When simulating data, the images, texts, sensor signals, etc. in the multi-modal data of the target user need to be aligned and encoded in the time dimension to convert them into state representation sequences that the world model can understand. Then, these state sequences are input into the pre-trained reinforcement learning world model, and multi-step reasoning and dynamic prediction are performed in the latent space through forward propagation. Thus, simulation data can be obtained, which can be one or more of the analysis results, text data, and audio and video data in the multi-modal data set.

[0028] In another embodiment, the simulation data of the target user obtained through the world model includes: The target sample data is determined by the amount of data in the multi-modal data set, and the target sample data is simulated and expanded based on the multi-modal data through the world model to obtain simulation data of the target sample data.

[0029] In this embodiment, the target sample data that needs to be simulated and expanded is determined by the amount of data in the multi-modal data set, and then it is simulated and expanded. For example, when a certain sample data, such as an abnormal navigation path of a specific product page, is insufficient (less than a preset data amount threshold), the world model will perform variational inference on other data in the multi-modal data set to generate simulation data for the abnormal navigation path, ensuring the uniformity and diversity of data distribution.

[0030] Step S102, according to the multi-modal data set and the simulation data, performing intention recognition on the target user to determine the initial intention recognition result and the evaluation index of the initial intention recognition result.

[0031] In this embodiment, it is necessary to first preprocess the multi-modal data set and the simulation data into structured input, such as sequence events in JSON format, and then input the LLM Evaluation model for inference to obtain the initial intent recognition result. The model will simulate the potential distribution of user intent, use Bayesian inference combined with prior knowledge such as historical user patterns to infer the current intent. For example, if the target user repeatedly views product specifications on an e-commerce website but does not add them to the shopping cart, the model may evaluate as "comparison intent" rather than "purchase intent".

[0032] After determining the initial intent recognition result of the target user, the initial intent recognition result is evaluated through a comprehensive evaluation system composed of LLM effect evaluation indexes. The evaluation process is performed by a large language model, such as a variant based on GPT-4, and analyzes user interactions from multiple channels, such as text-based intent, such as dialog robot conversations, interactive media, such as user click feedback on web interactive elements, and video agents, such as pausing or fast forwarding in video playback. To obtain the evaluation index of the initial intent recognition result, the evaluation index includes hallucination rate, consistency score, accuracy rate, and context relevance. This embodiment introduces a multi-dimensional index system to avoid the bias of a single index.

[0033] Step S103, adjusting the initial intent recognition result according to the evaluation index to obtain the target intent recognition result of the target user, and generating recommended content for the target user through the target intent recognition result.

[0034] In this embodiment, an evaluation value can be preset, and then the initial intent recognition result needs to be quantitatively evaluated according to the preset evaluation index. If the evaluation result does not reach the evaluation value, it means that the multi-modal data set or simulation data obtained in S101 needs to be adjusted. At this time, it can jump to the step of generating a multi-modal data set in S101, or it can jump to the step of obtaining simulation data of the target user through the world model. After re-determining the multi-modal data set and simulation data through the above steps, the initial intent recognition result and the evaluation index are re-determined again. Until the evaluation value of the re-determined initial intent recognition result reaches the above evaluation value, it can be determined that the adjustment of the initial intent recognition result is successful, and the initial intent recognition result at this time is determined as the target intent recognition result.

[0035] After determining the target intent recognition result, a personalized user content can be generated based on the target intent recognition result using a multi-modal based LLM Multi-Agent. The Multi-Agent system includes multiple specialized AI Agents: User Analysis Agent to analyze user behavior, Creative Writing Agent to generate text, and Video Generation Agent to create personalized videos. After inputting the intent result, the Agents collaborate with each other: for example, User Analysis extracts features -> Creative Writing drafting narratives -> Video Generation synthesizes visualizations. The generation process supports multi-modal fusion, such as text-to-video conversion, using Google Veo3 and other video models to generate high-fidelity videos. The innovation lies in adaptive personalization: AI Agents adjust parameters through RL to generate optimized content for different users, such as high-conversion personalized content recommendations. The output includes chatbot responses, interactive media, and Voice Agent interactions, ensuring improved user experience. Overall, this step realizes an intent-to-content closed loop, with significant conversion rate improvement.

[0036] In another embodiment, the adjusting the initial intent recognition result according to the evaluation index comprises: calculating a comprehensive score of the evaluation index by an LLM-as-a-Judge model, and when the comprehensive score is greater than a threshold, determining the initial intent recognition result as the target intent recognition result; when the comprehensive score is less than the threshold, repeating the steps of obtaining the simulation data of the target user by the world model, re-determining the initial intent recognition result and the corresponding evaluation index by the updated simulation data, and calculating the comprehensive score of the evaluation index by the LLM-as-a-Judge model until the comprehensive score is greater than the threshold.

[0037] This embodiment further introduces an LLM-as-a-Judge model based on LLM Evaluation, as an automatic evaluation "referee" mechanism, for secondary verification and arbitration of the evaluation result of the previous step. The model is essentially a specialized LLM variant trained on the Meta-Evaluation task, i.e., evaluating the output quality of other models.

[0038] The input includes the initial intent recognition result generated by the LLM Evaluation and the index value of the evaluation index. The LLM-as-a-Judge adopts zero-shot or few-shot prompting technology to simulate the judgment process of human experts. For example, through a prompt template such as "Based on the following indicators, judge whether the intent recognition is reliable: hallucination rate <5%, accuracy rate >90%", the model calculates a comprehensive score. When the comprehensive score is less than the preset evaluation value, jump to the step of generating a multi-modal data set or determining simulation data in S101, and then re-determine the initial intent recognition result and the evaluation index. Until the comprehensive score of the re-determined initial intent recognition result determined by the LLM-as-a-Judge model reaches the above evaluation value, it can be determined that the adjustment of the initial intent recognition result is successful, and the initial intent recognition result at this time is determined as the target intent recognition result.

[0039] At this point, the process shown in Figure 1 is completed.

[0040] In the embodiments of the present application, for a target user who needs to perform intent recognition, a multi-modal data set composed of the user's behavior, text, audio and video, etc. is first obtained, and then the multi-modal data set is processed by a world model for reinforcement learning to obtain simulation data of the user. The user is then intent-recognized through the multi-modal data set and the simulation data to obtain an initial intent recognition result and a corresponding evaluation index. Finally, the initial intent recognition result is adjusted through the evaluation index to determine the final target intent recognition result. The present application simulates expansion based on the multi-modal data set through the world model for the data with less data amount in the multi-modal data set. Even in the case of insufficient behavior samples, the intent recognition model trained can still significantly improve the accuracy, thereby improving the accuracy of intent recognition of the user.

[0041] In another embodiment, the above method further comprises: storing the multi-modal data set, simulation data, and target intent recognition result of the target user into a memory database, and performing cyclic iterative optimization of the LLM Evaluation model and the LLM-as-a-Judge model through the data in the memory database, wherein the memory database is intelligently searched through a Memory R1 framework.

[0042] In this embodiment, after dual evaluation by LLM Evaluation and LLM-as-a-Judge, the collected multimodal datasets, simulated data, and target intent recognition results can be processed and injected into the memory database for reinforcement learning optimization. The collected data is first preprocessed, such as intent vector embedding and behavior sequence normalization, and then used as input for reinforcement learning. The database can be modeled as a knowledge graph, where nodes represent intent entities, such as "purchase intent," and edges represent behavior associations, such as "click-through rate -> conversion." Reinforcement learning algorithms, such as Deep Q-Network and DQN, run on this graph, and the agent learns the optimal update strategy: the reward function is defined as the improvement in intent recognition accuracy, and actions include data augmentation, redundancy removal, or merging similar items. For example, for rare behaviors, the agent generates synthetic paths and reuses historical experience through simulated replay.

[0043] Furthermore, to improve the efficiency of context processing, this embodiment introduces the Memory R1 framework for intelligent retrieval of relevant data in the database. This framework is an advanced memory retrieval module that, based on a vector database and semantic search algorithms, achieves fast and accurate context extraction. The retrieval process begins with the user's current behavior, generates a query vector, then matches similar items in the database, and uses Memory R1's filtering mechanism to remove noisy data, making the output more refined. For example, if the user's query involves historical intent, Memory R1 will retrieve relevant sequences and apply an attention mechanism to highlight key context.

[0044] like Figure 2 As shown, this application also provides a user intent recognition system based on user behavior data, the system including a data acquisition simulation module, an intent recognition module, and a user content generation module; The data acquisition simulation module is used to generate a multimodal dataset based on the target user's behavior and input the multimodal dataset into a trained world model for reinforcement learning. The simulated data of the target user is obtained through the world model. The multimodal dataset includes behavior analysis results, text data, and audio and video data. The intent recognition module is used to perform intent recognition on the target user based on the multimodal dataset and the simulation data, so as to determine the initial intent recognition result and the evaluation index of the initial intent recognition result, wherein the evaluation index includes illusion rate, consistency score, accuracy and context relevance; The user content generation module is used to adjust the initial intent recognition result according to the evaluation index to obtain the target intent recognition result of the target user, and then generate recommended content for the target user based on the target intent recognition result.

[0045] In another embodiment, the obtaining simulation data of the target user through the world model in the data collection simulation module comprises: determining target sample data through the amount of data in the multi-modal data set, and simulating expansion of the target sample data based on the multi-modal data through the world model to obtain simulation data of the target sample data.

[0046] In another embodiment, the intent recognition of the target user according to the multi-modal data set and the simulation data in the intent recognition module comprises: analyzing and reasoning the multi-modal data set and the simulation data through the LLM Evaluation model to obtain an initial intent recognition result of the target user, and analyzing the initial intent recognition result through a comprehensive evaluation system composed of a plurality of LLM effect evaluation indexes to determine a corresponding evaluation index.

[0047] In another embodiment, the adjusting the initial intent recognition result according to the evaluation index in the user content generation module comprises: calculating a comprehensive score of the evaluation index through the LLM-as-a-Judge model, and when the comprehensive score is greater than a threshold value, determining the initial intent recognition result as the target intent recognition result; when the comprehensive score is less than the threshold value, repeating the steps of obtaining simulation data of the target user through the world model, and re-determining the initial intent recognition result and the corresponding evaluation index through the updated simulation data, and calculating the comprehensive score of the evaluation index through the LLM-as-a-Judge model until the comprehensive score is greater than the threshold value.

[0048] In another embodiment, the system further comprises: a model optimization module for storing the multi-modal data set, simulation data and target intent recognition result of the target user into a memory database, and performing cyclic iterative optimization of the LLM Evaluation model and the LLM-as-a-Judge model through the data in the memory database, wherein the memory database is intelligently searched through the Memory R1 framework.

[0049] Further, the above system can also refer to a Content Engine integrated with multiple tools and frameworks.

[0050] As the core integration layer of the system, Content Engine realizes the content production capability of LLMMulti-Agent by calling various tools and frameworks (such as MCP for multi-context processing, RAG for enhanced generation, Memory for historical reference) and interacting with external API interfaces. The engine adopts a modular design, first aggregates data retrieved from Memory R1, and then calls tools through APIs: for example, uses large language models such as ChatGPT to generate text descriptions, video models such as Kling to synthesize videos, and inference models such as Deepseek to perform deep analysis. The integration process is based on frameworks such as LangGraph to realize the orchestration of tool chains, such as serially calling RAG to enhance the intent description, and then generating multi-modal output in parallel.

[0051] The above embodiment of the present application provides a user intention recognition method based on user behavior data, and a user intention recognition system based on user behavior data is provided based on the method. Through the above method, the behavior data of the user can be simulated and expanded, so that even in the case of insufficient behavior samples, the accuracy of the intention recognition model trained can be significantly improved, thereby improving the accuracy of the user's intention recognition.

[0052] The embodiment also discloses a computer device, as shown in the figure Figure 3 The computer device includes a processor and a memory, and the memory stores at least one instruction which is loaded and executed by the processor to implement the user intention recognition method based on user behavior data as described above.

[0053] The above is only a preferred embodiment of the present application, and is not intended to limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included within the scope of protection of the present application.

Claims

1. A method for recognizing user intention based on user behavior data, characterized by, The method comprises: generating a multi-modal data set according to the behavior action of the target user, inputting the multi-modal data set into a trained world model for reinforcement learning, obtaining simulation data of the target user through the world model, wherein the multi-modal data set comprises behavior analysis results, text data and audio and video data; performing intent recognition on the target user according to the multi-modal data set and the simulation data to determine an initial intent recognition result and an evaluation index of the initial intent recognition result, wherein the evaluation index comprises an illusion rate, a consistency score, an accuracy rate and a context relevance; adjusting the initial intent recognition result according to the evaluation index to obtain a target intent recognition result of the target user, and generating recommended content of the target user through the target intent recognition result.

2. The method of claim 1, wherein, The simulation data of the target user through the world model comprises: determining target sample data through the amount of data in the multi-modal data set, and simulating and expanding the target sample data based on the multi-modal data through the world model to obtain simulation data of the target sample data.

3. The method of claim 1, wherein, The intent recognition on the target user according to the multi-modal data set and the simulation data comprises: analyzing and reasoning the multi-modal data set and the simulation data through an LLM Evaluation model to obtain an initial intent recognition result of the target user, and analyzing the initial intent recognition result through a comprehensive evaluation system composed of a plurality of LLM effect evaluation indexes to determine the corresponding evaluation index.

4. The method of claim 3, wherein, The adjustment of the initial intent recognition result according to the evaluation index comprises: calculating a comprehensive score of the evaluation index through an LLM-as-a-Judge model, and when the comprehensive score is greater than a threshold value, determining the initial intent recognition result as the target intent recognition result; when the comprehensive score is less than the threshold value, repeating the steps of obtaining simulation data of the target user through the world model, and determining an initial intent recognition result and a corresponding evaluation index through updated simulation data, and calculating a comprehensive score of the evaluation index through the LLM-as-a-Judge model until the comprehensive score is greater than the threshold value.

5. The method of claim 4, wherein, The method further comprises: storing the multi-modal data set, simulation data and target intent recognition result of the target user into a memory database, and performing cyclic iterative optimization of the LLM Evaluation model and the LLM-as-a-Judge model through the data in the memory database, wherein the memory database is intelligently searched through a Memory R1 framework. 6.A user intent recognition system based on user behavior data, characterized by, The system comprises a data acquisition and simulation module, an intent recognition module and a user content generation module; The data collection simulation module is configured to generate a multi-modal data set according to a behavior action of a target user, input the multi-modal data set into a trained world model for reinforcement learning, and obtain simulation data of the target user through the world model, wherein the multi-modal data set includes behavior analysis results, text data, and audio and video data. The intent recognition module is configured to perform intent recognition on the target user according to the multi-modal data set and the simulation data, determine an initial intent recognition result of the target user and an evaluation index of the initial intent recognition result, wherein the evaluation index includes an illusion rate, a consistency score, an accuracy rate, and a context relevance. The user content generation module is configured to adjust the initial intent recognition result according to the evaluation index, obtain a target intent recognition result of the target user through the target intent recognition result, and generate recommended content for the target user through the target intent recognition result.

7. The system of claim 6, wherein, The simulation data of the target user obtained through the world model in the data collection simulation module includes: Target sample data is determined according to the amount of data in the multi-modal data set, and the target sample data is simulated and expanded based on the multi-modal data through the world model to obtain simulation data of the target sample data.

8. The system of claim 6, wherein, The intent recognition module includes: The multi-modal data set and the simulation data are analyzed and inferred through an LLM Evaluation model to obtain an initial intent recognition result of the target user, and the initial intent recognition result is analyzed through a comprehensive evaluation system composed of a plurality of LLM effect evaluation indexes to determine a corresponding evaluation index.

9. The method of claim 8, wherein, The user content generation module includes: The comprehensive score of the evaluation index is calculated through an LLM-as-a-Judge model, and when the comprehensive score is greater than a threshold value, the initial intent recognition result is determined as the target intent recognition result; When the comprehensive score is less than the threshold value, the steps of obtaining simulation data of the target user through the world model and recalculating the initial intent recognition result and the corresponding evaluation index through the updated simulation data are repeated until the comprehensive score is greater than the threshold value.

10. The system of claim 9, wherein, The system further includes: A model optimization module is configured to store the multi-modal data set, simulation data, and target intent recognition result of the target user into a memory database, and perform cyclic iterative optimization on the LLM Evaluation model and the LLM-as-a-Judge model through the data in the memory database, wherein the memory database is intelligently searched through a Memory R1 framework.

Images