Recommended method, device, electronic equipment and medium

Abstract

Embodiments of the present application disclose a recommendation method and device, electronic equipment and medium, relating to the technical field of artificial intelligence. A specific embodiment of the method includes: performing intent recognition on user input query information to determine an intent label of the query information; performing task analysis on the query information based on the intent label to generate a task planning data set corresponding to the query information; calling at least one sub-agent and / or at least one function module according to the task planning data set to obtain at least one execution result fed back by the at least one sub-agent and / or the at least one function module after executing the task planning data set; and generating recommendation information matched with the query information based on the at least one execution result and feeding back the recommendation information to the user. The method realizes intent understanding and dynamic task planning for complex queries, and significantly improves the accuracy of recommendations, interactive flexibility and personalized experience through multi-agent collaborative execution.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, specifically to a recommended method, apparatus, electronic device, and medium. Background Technology

[0002] With the development of artificial intelligence technology, users have placed higher demands on recommendation systems for accurate understanding and personalized interaction in complex task scenarios.

[0003] Currently, most recommendation systems in related technologies use static tag matching, which recommends content based on a single-dimensional feature of the user's historical behavior. This results in insufficient recommendation accuracy in complex task scenarios, affecting the user's interaction experience. Summary of the Invention

[0004] This application provides a recommended method, apparatus, electronic device, and medium.

[0005] In a first aspect, embodiments of this application provide a recommended method applied to a target intelligent agent, comprising:

[0006] Perform intent recognition on the query information entered by the user to determine the intent label of the query information; Based on intent tags, the query information is parsed to generate a task planning dataset corresponding to the query information; Based on the task planning dataset, at least one sub-agent and / or at least one functional module are invoked to obtain at least one execution result fed back after at least one sub-agent and / or at least one functional module executes the task planning dataset; Based on at least one execution result, generate recommendation information that matches the query information and then feed the recommendation information back to the user.

[0007] In some embodiments, intent recognition is performed on the query information input by the user, and the intent tags of the query information are determined as follows: Based on value assessment prompts and value assessment samples, the query information is scored to determine whether the query information is valid. If the query information is valid, the intent of the valid query information is identified to obtain the intent label corresponding to the valid query information.

[0008] In some embodiments, based on intent tags, the query information is parsed to generate a task planning dataset corresponding to the query information, including: Based on intent labels and task planning models, combined with task prompts and task constraints, the query information is parsed to generate a task planning dataset corresponding to the query information.

[0009] In some embodiments, before performing task parsing on the query information based on intent tags and task planning models, combined with task prompts and task constraints, to generate the task planning dataset corresponding to the query information, the method includes: Obtain query samples and input them into the model to be trained to generate the training task planning dataset corresponding to the query samples; Based on format verification rules, process verification rules, and result verification rules, the correctness of the training task planning dataset's format, the correctness of its subtask annotations, and its consistency with the reference plan are verified respectively, and the verification results corresponding to the query samples are obtained. Based on the validation results, the parameters of the model to be trained are adjusted until the model meets the convergence condition, and the model that meets the convergence condition is used as the task planning model.

[0010] In some embodiments, at least one sub-agent includes at least one of a query analysis agent and a user behavior agent; at least one functional module includes at least one of a time acquisition module, a location acquisition module, a weather acquisition module, a hot event acquisition module, and a content search module. Based on the task planning dataset, at least one sub-agent and / or at least one functional module is invoked to obtain at least one execution result returned by the at least one sub-agent and / or at least one functional module after executing the task planning dataset, including: When the query analysis agent is invoked, explicit and / or implicit tags extracted from the query information are obtained; and / or, When the user behavior intelligent agent is invoked, a dynamic preference profile of the user is obtained. The dynamic preference profile includes at least one preference label and a score corresponding to each preference label in the search at least one preference label; and / or, When the time acquisition module is invoked, the query time information corresponding to the query information is obtained; and / or, When the location acquisition module is invoked, the user's geographic location information is obtained; and / or, When the weather acquisition module is invoked, retrieve the corresponding weather information for the query; and / or, When the hot event acquisition module is invoked, hot event information related to the query information is retrieved; and / or, When the content search module is invoked, a list of content items matching the query information is obtained.

[0011] In some embodiments, generating recommended content that matches the query information based on at least one execution result, and then feeding the recommended content back to the user includes: Based on at least one execution result, at least one target content item matching the query information is selected from the target content library, so as to generate a recommended content list based on at least one target content item; Based on the recommended content list and target persona prompts, generate recommendation information containing emotional expression to provide the recommended content list and / or recommendation information back to the user.

[0012] Secondly, embodiments of this application provide a recommendation device applied to a target intelligent agent, the device comprising: The recognition unit is used to identify the intent of the query information input by the user and determine the intent label of the query information; The planning unit is used to perform task parsing on query information based on intent labels in order to generate a task planning dataset corresponding to the query information. The acquisition unit is used to call at least one sub-agent and / or at least one functional module based on the task planning dataset, so as to obtain at least one execution result fed back by at least one sub-agent and / or at least one functional module after executing the task planning dataset; The recommendation unit is used to generate recommendation information that matches the query information based on at least one execution result, and to feed the recommendation information back to the user.

[0013] Thirdly, embodiments of this application provide an electronic device, including: a processor and a memory for storing a computer program capable of running on the processor, wherein, when the processor runs the computer program, it performs the method described in any embodiment of the first aspect.

[0014] Fourthly, embodiments of this application provide a non-transitory computer-readable storage medium storing computer instructions, wherein the computer instructions are used to cause a computer to perform the methods described in any embodiment of the first aspect.

[0015] Fifthly, embodiments of this application provide a computer program product including a computer program that, when executed by a processor, performs the method described in any embodiment of the first aspect.

[0016] This application provides a recommendation method that identifies the intent of user-input query information to determine the intent tag of the query information; based on the intent tag, performs task parsing on the query information to generate a task planning dataset corresponding to the query information; according to the task planning dataset, calls at least one sub-agent and / or at least one functional module to obtain at least one execution result fed back after the at least one sub-agent and / or at least one functional module executes the task planning dataset; based on at least one execution result, generates recommendation information matching the query information and feeds the recommendation information back to the user. This method achieves intent understanding and dynamic task planning for complex queries, and significantly improves the accuracy, interaction flexibility, and personalized experience of recommendations through multi-agent collaborative execution.

[0017] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are merely embodiments of this application. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort, and this application can be applied to other similar scenarios based on the provided drawings.

[0019] Figure 1 A flowchart illustrating a recommended method provided in an embodiment of this application; Figure 2 A flowchart of another recommended method provided for embodiments of this application; Figure 3 A schematic diagram illustrating a specific recommended method provided in an embodiment of this application; Figure 4 A schematic diagram illustrating intent recognition and task planning for a target intelligent agent, as provided in an embodiment of this application; Figure 5 A schematic diagram of the structure of a recommended device 500 provided in an embodiment of this application; Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0020] The present application will now be described in further detail with reference to the accompanying drawings and embodiments. It is to be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. The described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0021] It should be noted that the terms "system," "device," "unit," and / or "module" used in this application are methods of distinguishing different components, elements, parts, sections, or assemblies at different levels. However, if other terms can achieve the same purpose, they can be replaced by other expressions.

[0022] Hereinafter, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first," "second," or "third" may explicitly or implicitly include one or more of that feature.

[0023] Most recommendation systems in related technologies use static tag matching to recommend content based on single-dimensional features in users' historical behavior, resulting in insufficient accuracy and personalized interactive experience in complex task scenarios.

[0024] Specifically, taking music recommendation systems as an example, most music recommendation systems in related technologies are based on static tag matching for content delivery. These systems generate a recommendation list by matching preset tags (such as song type, artist, and genre) with the user's historical behavior tags. However, these methods generally only consider single-dimensional features (such as the types of songs the user has listened to), ignoring changes in the user's music preferences over different time periods, environmental conditions, and emotional states.

[0025] In terms of interaction, music recommendation systems typically use fixed dialogue templates or simple keyword matching to generate responses, lacking the ability to recognize multiple intentions and coordinate the use of multiple tools. This results in insufficient personalization and accuracy of recommendation results and interactive experience in complex task scenarios (such as "recommend songs suitable for listening to on a rainy night and help me create a playlist").

[0026] It is evident that recommender systems in related technologies have the following drawbacks: 1) Single analysis dimension: It is impossible to simultaneously integrate multi-dimensional information such as user behavior time sequence characteristics (such as listening habits at different time periods), environmental dynamic factors (such as weather and location), and emotional state, resulting in a lack of context adaptability in recommendation results.

[0027] 2) Insufficient interaction capabilities: It lacks the ability to recognize multiple intents, cannot break down complex tasks, and cannot call multiple external tools to complete tasks collaboratively. The interaction process is rigid and lacks personalization.

[0028] 3) Delayed user profile updates: User profiles that are usually static or updated infrequently cannot reflect recent changes in user interests in a timely manner, resulting in recommendations that are out of touch with the user’s current needs.

[0029] To address the aforementioned issues, this application employs a target intelligent agent to perform intent recognition and structured task parsing of user queries, generating an executable planning dataset. Based on this dataset, multiple sub-intelligent agents or functional modules are dynamically invoked to collaboratively obtain multi-dimensional execution results, which are then fused to generate personalized recommendation information. This achieves intelligent recommendation from static tag matching to multi-intent understanding and multi-tool collaboration, significantly improving the accuracy of recommendations, the flexibility of interaction, and the scalability of the system in complex task scenarios.

[0030] The following section provides a detailed description of a recommended method provided in this application, with reference to the accompanying drawings.

[0031] Figure 1 A flowchart illustrating a recommended method provided in an embodiment of this application is shown. Figure 1 As shown, this recommendation method is applied to a target agent, which performs tasks such as identifying the user's query intent, planning tasks, calling sub-agents or functional modules, and generating and providing recommendation information. The recommendation method includes steps 101-104.

[0032] Step 101: Perform intent recognition on the query information entered by the user to determine the intent label of the query information.

[0033] In embodiments of this application, the query information entered by the user (such as "recommend songs suitable for listening to on rainy nights") is typically natural language text, which may implicitly contain multiple objectives. This application can map this natural language text to one or more predefined or dynamically generated intent categories (such as recommendation, search, emotional companionship, etc.) through intent recognition, providing direction for subsequent task decomposition.

[0034] The query information can be the original request or question entered by the user in the form of voice, text, etc.

[0035] Intent recognition is a natural language understanding technology used to determine the purpose or category of a user's query.

[0036] Intent tags are category markers that identify user intent, and can have a hierarchical structure (such as first-level intent "recommendation" and second-level intent "scenario recommendation").

[0037] Step 102: Based on the intent label, perform task parsing on the query information to generate the task planning dataset corresponding to the query information.

[0038] In the embodiments of this application, after obtaining the user's intent label, the application also needs to transform the abstract intent into a specific, executable sequence of operations. Task parsing is similar to formulating an action plan: for example, for "recommending songs for rainy nights," it may be necessary to sequentially execute subtasks such as "obtaining the current weather," "analyzing the user's recent listening preferences," and "searching for matching songs." The task planning dataset records this series of subtasks and their invocation order in a structured form.

[0039] Task parsing is the process of breaking down user intent into multiple sequential and dependent subtasks.

[0040] A task planning dataset is a collection of data that describes the execution sequence of subtasks. It can include the name, type (calling an agent or a functional module), parameters, and execution order of each subtask.

[0041] Step 103: Based on the task planning dataset, invoke at least one sub-agent and / or at least one functional module to obtain at least one execution result fed back after at least one sub-agent and / or at least one functional module executes the task planning dataset.

[0042] In the embodiments of this application, the corresponding capability units (sub-agents or functional modules) are invoked sequentially or in parallel according to the order in the task planning dataset. Each unit returns an execution result after completing its assigned sub-task (e.g., the weather module returns "moderate rain", and the user behavior agent returns "preference: rock, score 8.5"). All execution results are aggregated for use in the next step.

[0043] Among them, sub-agents are modules with independent reasoning or decision-making capabilities, such as query analysis agents (extracting tags) and user behavior agents (generating dynamic profiles).

[0044] A functional module is a stateless or simple logical unit that performs a specific function, such as a tool for obtaining time, location, weather, or searching for content.

[0045] The execution result is the data output by the sub-agent or functional module after completing the task, which can be in the form of tags, profiles, values, lists, etc.

[0046] Step 104: Based on at least one execution result, generate recommendation information that matches the query information and feed the recommendation information back to the user.

[0047] In some embodiments, the multi-dimensional execution results collected in the previous step can be merged and filtered to find the query information that best matches the user's input and the target content items in the current context, and then presented in a user-friendly format. To enhance the interactive experience, language models are often used to generate natural language responses with emotional expression, making the recommendation results more human-like.

[0048] Recommendations are the final results returned to the user and can include a list of content (such as songs and videos), natural language responses, and operation suggestions.

[0049] The feedback in this application can convey recommendation information to users through interface display, voice broadcast, or other means, and there are no restrictions on this.

[0050] In summary, the recommendation method proposed in this application can achieve deep intent understanding and automated task decomposition of complex user queries. By obtaining multi-dimensional contextual information through multi-agent collaborative execution, it can ultimately generate personalized and emotional recommendation results. This effectively solves the problems of single dimension, rigid interaction, and lagging profiling in related technologies, and significantly improves the recommendation accuracy and user experience in complex scenarios.

[0051] Figure 2 This is a flowchart illustrating another recommended method provided in an embodiment of this disclosure. Figure 2 based on Figure 1 The illustrated embodiment further defines steps 101 and 102. Figure 2 In the illustrated embodiment, step 102 includes steps 201 and 202, and step 102 includes step 203. For example... Figure 2 As shown, the method may include: Step 201: Based on the value assessment prompts and value assessment samples, score the query information entered by the user to determine whether the query information is valid based on the scoring results.

[0052] In the embodiments of this application, not all user-input query information is worthy of subsequent complex processing by the system (such as meaningless characters, malicious requests, and idle chat unrelated to business). This application can use a value assessment mechanism to quantify and score query information, and only queries with scores exceeding a preset scoring threshold are considered valid. This mechanism is typically implemented using a Language Model (LLM) combined with prompt words and sample samples: prompt words define scoring dimensions (semantic completeness, intent clarity, security, etc.), samples provide scoring references, and the model outputs a score for the current query based on these criteria.

[0053] Value assessment prompts are a set of structured instructions that include scoring criteria, dimensional weights, and output format requirements. They are used to guide the language model to score the query information from multiple dimensions.

[0054] The value assessment sample consists of a small number of pre-labeled query examples (few-shots) to help the model understand the scoring scale.

[0055] Valid query information refers to query information that, after being scored, achieves a comprehensive score that reaches a preset scoring threshold. Such queries are considered to have processing value and can proceed to the subsequent intent recognition process. The preset scoring threshold can be set according to actual needs and is not limited in this embodiment.

[0056] For example, the valuation tips are as follows: "You are a professional query value assessment expert. Please score the user-input query from multiple dimensions according to the following scoring criteria."

[0057] Scoring dimensions and weights: 1) Semantic completeness (weight 0.3): Evaluates whether the query contains complete semantic information; 2) Clarity of Intent (weight 0.25): Assessing the clarity of the user's intent; 3) Music Relevance (Weight 0.2): Evaluates the relevance of the query to the music domain; 4) Security (weight 0.15): Evaluate whether the query contains inappropriate content; 5) Interaction value (weight 0.1): Evaluate the interaction value of the query.

[0058] The following is a sample of the valuation: Query: "Recommend songs suitable for listening on rainy nights"; Rating: 9.25 points; Reasons: Semantic integrity, clear intent, strong musical relevance, safe and interactive.

[0059] Then enter the current query information to rate the current query information, for example: Please rate the following query and provide a detailed rating and reason; Query: [query to be rated]

[0060] Step 202: If the query information is valid, perform intent recognition on the valid query information to obtain the intent label corresponding to the valid query information.

[0061] In some embodiments, after filtering out low-value queries, this application performs intent recognition on the remaining valid query information. This process maps natural language queries to predefined or dynamically generated intent categories, supporting single-intent or multi-intent output. For example, generative intent recognition based on a language model fine-tuned by LoRA can output first-level intents (such as "recommendation") and second-level intents (such as "scenario recommendation").

[0062] For example, after determining that the query information is valid, this application can combine a pre-trained intent recognition model to identify the intent involved in each query, and obtain {intent1, intent2, …, intentM}. The intent recognition model primarily employs a large model fine-tuned based on LoRA (Low-Rank Adaptation) for generative intent recognition. Through an architecture of a pre-trained large model + LoRA adapter + intent generation head, it achieves flexible recognition of user intents. This application's intent recognition model supports multi-intent recognition, generating new intent labels that are not limited to predefined categories, while possessing good interpretability and context awareness. The intent label system can include primary intents (recommendation, search, emotional support, trend tracking, personalization) and secondary intents (scene recommendation, emotional recommendation, style recommendation, time-based recommendation). For the query "recommend songs suitable for listening on a rainy night," the intent recognition model can identify multi-dimensional intent labels such as {recommendation: 0.95, scene recommendation: 0.88, emotional recommendation: 0.72}, providing important basis for subsequent task planning and agent scheduling.

[0063] Step 203: Based on intent labels and task planning models, combined with task prompts and task constraints, perform task parsing on the query information to generate the task planning dataset corresponding to the query information.

[0064] In some embodiments, this application can transform a user's intent tags into a sequence of executable subtasks. The task planning model (the task planning module in the target agent fine-tuned through reinforcement learning) is based on its extracted intent tags, query information, and additional task prompts (e.g., requiring an output format of...). <think> ...< / think> <plan> ...< / plan> The task planning dataset is generated by combining task constraints (e.g., 3-8 steps, only registered agents or tools can be called) with task constraints.

[0065] Task planning models can be large language models that have been pre-trained and fine-tuned for task planning (such as GRPO reinforcement learning).

[0066] Task cues are instruction texts that guide the model on how to break down tasks, express thought processes, and plan content.

[0067] Task constraints are rules that limit the generation of plans, such as the maximum number of steps and the types of capabilities that can be invoked.

[0068] The task planning dataset is structured data containing thought content and planning content. The planning content consists of ordered subtasks, each of which includes a task name and the type of capability to be invoked (agent or tool).

[0069] Before generating the task planning dataset, a task planning model needs to be trained first. The specific process is as follows: Obtain query samples (collected real user questions), input the query samples into the model to be trained (the initial language model that has not been fine-tuned or has not completed the planning training), and generate the corresponding training task planning dataset.

[0070] Based on format verification rules, process verification rules, and result verification rules, the correctness of the training task planning dataset's format, the correctness of its subtask annotations, and its consistency with the reference plan are verified, respectively, and the verification results corresponding to each query sample are obtained.

[0071] Format validation rules: Check whether the output conforms to a predefined format (e.g., by matching with regular expressions).

[0072] Process validation rules: Check the correctness of the name and type matching in each subtask.

[0073] Result verification rules: Compare the model output with the manually annotated reference plan (usually scored by the evaluation model).

[0074] Based on the validation results, reinforcement learning (such as GRPO) is used to adjust the parameters of the model to be trained until the model meets the convergence conditions (loss stabilizes and reward reaches the threshold). The model that meets the convergence conditions is then used as the task planning model.

[0075] For example, this application uses a task planning model to perform overall planning based on the user-input query, obtaining several task planning data in the following form: <think> This is the content of my thoughts.< / think> <plan> (Problem Analysis, Intelligent Agent)##(Weather Acquisition, Tool)##(IP Acquisition, Tool)...< / plan> .

[0076] in, <think>The content within the tags represents the model's thinking and analysis process, improving the accuracy of task planning through a thought chain model; <plan>The content within the label represents the execution steps based on the query plan, which are used for subsequent fine-tuning and training of the target agent model.

[0077] Specific task prompts are shown below: "You are a music task planning expert. Please generate an executable task plan based on the input. Output only in the following format:" <think> Your analysis< / think> <plan> (Problem Analysis, Intelligent Agent)##(Weather Acquisition, Tool)##(User Profile Analysis, Intelligent Agent)##(Sentiment Analysis, Intelligent Agent)##(Music Recommendation, Intelligent Agent)< / plan> .

[0078] enter: query: "{query}"; intents: {multi-intent JSON}; context: {time / weather / location}; capabilities: {List of registered agents / tools}.

[0079] Task constraints: Use only the capabilities listed; Number of steps: 3~8; Each step is strictly formatted as (task name, type), with the type limited to either "agent" or "tool". Do not output any content other than the format shown above.

[0080] For each query, G=3~5 candidate plans are generated, using different temperatures, top-p values, and random seeds (e.g., temp∈{0.3,0.7}, top_p∈{0.8,0.95}, seed∈{11,13,17}), while preserving the decoding configuration. This application pre-generates a task planning dataset using a task planning model, and then manually verifies and modifies the data to ensure it conforms to the correct task planning dataset.

[0081] For example, the target agent in this application is implemented through a fine-tuned language model (i.e., a task planning model). Fine-tuning training is performed around the goals of "reliable planning, executable steps, and usable results." The core consists of two parts: a verifiable reward function and reinforcement learning training. The approach is as follows: first, using automatically verifiable format, process, and result signals as supervision, a computable reward is defined; then, a GRPO-based policy optimization is employed to continuously improve the task planning output of the target agent in the actual execution process.

[0082] For the setting of verifiable rewards, three parts of verifiable rewards are set: format reward, process reward, and result reward.

[0083] The format reward calculation method is as follows: , Here, `response` represents the response generated by the task planning module of the target agent, and its format is the same as the data format of any training task planning data in the training task planning dataset. `check_format` uses a regular expression to match whether the response meets the expected format. The regular expression for format matching is as follows: , The process reward is calculated as follows: , The name represents the name of the subtask generated in each step of the response, such as "weather acquisition", while the type represents the type of each step, such as "agent" or "tool".

[0084] The reward is calculated as follows: , This represents the reward calculation between the outcome and the target result of each training task planning data (plan) in the training task planning dataset, measuring the model output. <plan>Consistency with the labeled reference planning target. The approach is to normalize both and then have them scored by the frozen evaluation model E, obtaining a score in the interval [0,1], which is directly used as the result reward.

[0085] The final model's reward is a weighted sum of the three rewards mentioned above, as shown below: , Where w1, w2, and w3 represent the weights of the three parts of the reward, respectively. This value is a pre-designed empirical value, typically set as w1=0.2, w2=0.3, and w3=0.5.

[0086] The training of the task planning model in this application will be based on reinforcement learning using a reward calculation method combined with GRPO, training the target agent task planning model so that the model has output capabilities. <think> Your analysis< / think> <plan> (Problem Analysis, Intelligent Agent)##(Weather Acquisition, Tool)##(User Profile Analysis, Intelligent Agent)##(Sentiment Analysis, Intelligent Agent)##(Music Recommendation, Intelligent Agent)< / plan> Such content, in <plan>This part is the task planning data in the task planning dataset.

[0087] The loss function during model training is shown below: , Where Q represents the set of query samples, o represents the task planning data response generated by the task planning model, and G represents sampling G responses for each query. This represents the task planning model used for sampling. This represents the training model before optimization. This represents the currently updated task planning model. Through continuous iterative optimization until the model loss converges, the final fine-tuned task planning model is obtained and used for task planning in the target agent.

[0088] Step 204: Based on the task planning dataset, invoke at least one sub-agent and / or at least one functional module to obtain at least one execution result fed back after at least one sub-agent and / or at least one functional module executes the task planning dataset.

[0089] In some embodiments, this application can parse the subtask sequence in the task planning dataset and call the corresponding capability units sequentially or in parallel. After each unit is executed, an execution result is returned, and all results are collected for further use.

[0090] Sub-agents are modules with independent reasoning, analysis, or decision-making capabilities. For example: a query analysis agent extracts explicit tags (words directly mentioned by the user) and implicit tags (inferred emotions, scenarios, etc.) from query information. A user behavior agent generates a dynamic preference profile based on the user's historical interaction data. This profile includes multiple preference tags and their scores (the scores reflect how well the tags fit the current user).

[0091] Functional modules are tool modules that perform deterministic functions, such as: time acquisition module (returns the current time or queries related time information), location acquisition module (gets the user's geographical location (such as city, latitude and longitude)), weather acquisition module (gets the weather conditions of the query location or the current environment), hot event acquisition module (gets real-time hot events related to the query information (such as "a singer releases a new album")), and content search module (retrieves matching content items (such as songs, videos, articles, etc.) based on the query information or intermediate results and returns a list of content items).

[0092] The execution result is the specific data returned by each module, which can be tags, profiles, numbers, strings, lists, etc.

[0093] For example, after obtaining the user's multi-dimensional intent labels and the task planning dataset obtained based on the user's query, this application can call the corresponding intelligent agent or functional module (also known as a tool module) based on the planning dataset, and work together according to the task planning dataset to complete each step of the sub-task.

[0094] Taking music recommendation as an example, this application can build three sub-agents / tools for the music assistant: a query analysis agent, a user behavior agent, and a tool library module.

[0095] The main function of the query analysis agent is to extract explicit and implicit tags from the user's input query. Explicit tags identify the tags explicitly mentioned in the user's input. For example, for the query "recommend songs suitable for listening to on a rainy night", the explicit tags could be "rainy night" or "night". Implicit tags are tags related to the query in terms of emotions, such as "relaxation" or "loneliness", which can reflect the user's current emotions and mood.

[0096] Specifically, for explicit label extraction, this application can adopt a few-shot prompt word strategy, using a language model to directly extract and parse the display labels based on the input query; Then, for implicit label extraction, this application can adopt a multi-model ensemble strategy for extraction, because implicit labels often do not have a standard answer, and a single language model may have cognitive limitations. Therefore, this application can use other types of open source language models in parallel, and combine the extracted multi-dimensional intent labels as supplementary information to extract privacy labels, and then merge and deduplicate the labels to obtain the final user implicit labels. Implicit and explicit tags are primarily used to provide corresponding tag searches in subsequent tool calls (such as song search), facilitating the search for related song content. They also serve as input for user profiling, used to determine user style preferences, etc.

[0097] Among them, the main function of the user behavior intelligent agent is to analyze the user's historical music listening data and the previously generated tag content, and generate a profile. The user profile mainly includes the following dimensions: style preference (e.g., folk music), time preference (e.g., listening to music late at night), and personality attributes (gender / age). Specifically, since users' listening preferences shift over time, this application constructs a dynamic user profile update module based on a sliding window, which mainly includes the following steps: First, set up a sliding window with a fixed time size, such as [2025.4.1-2025.5.1), [2025.5.1-2025.6.1), etc. Then, obtain the user's listening data in each window, including song name, song tags, artist information, playlist tags, playback time, number of plays, etc. The multi-dimensional user profile generation based on a large model first pre-sets a candidate set of style preference tags and a candidate set of time period preference tags. Then, the song data in each window obtained earlier and the pre-set sets of tags of different dimensions are combined into user profile prompts using prompt word templates. The language model (also known as the large model) then generates user profile tags for that time period and scores the degree of relevance of different tags (1-10 points). Finally, the set of tags for user style preference and time period preference and the corresponding scores {(tag1, score1), (tag2, score2), …} are obtained. The dynamic user profile update based on moving average is performed as follows: The tag matching score of the user profile within each time period is updated in the following way:

[0098] in This represents the matching score of the i-th tag within the t-th window. If the user does not have the tag within the t-th window, the score is 0. This represents the update parameter, used to control the degree to which a user's recent music listening data affects the overall user profile. Clip(x, a,b) means cropping the value of x to between a and b.

[0099] The end-user behavior intelligent agent can output the latest user profile of the user at the current time.

[0100] Among them, the main function of the tool module is to provide tools for obtaining real-time time, geolocation (Shanghai), weather (moderate rain), hot events (a singer releasing a new album), and song search. These tools are used to solve sub-tasks that large models cannot handle.

[0101] It should be noted that the at least one sub-agent and at least one functional module listed above are merely illustrative examples. This application is not limited to these specific types, and other sub-agents or functional modules can be flexibly configured according to requirements in practical applications.

[0102] Step 205: Based on at least one execution result, generate recommended content that matches the query information and send the recommended content back to the user.

[0103] In some embodiments, this application can fuse and filter the collected multi-dimensional execution results to select the target content items most suitable for the user's current query and context, and organize them into recommendation information. To improve the naturalness and personalization of the interaction, this application can also combine preset persona prompts to allow the language model to generate natural language responses with emotional expression, which are returned to the user along with the recommended content. That is, this application can filter at least one target content item matching the query information from the target content library based on at least one execution result, generate a recommended content list based on at least one target content item, and generate recommendation information containing emotional expression based on the recommended content list and target persona prompts, so as to feed back the recommended content list and / or recommendation information to the user.

[0104] Recommended content is the final result returned to the user, which can be a list of content (such as a song list), natural language responses, or a combination of both.

[0105] The target content library is a database that stores content items to be recommended (such as a music library or a video library).

[0106] The target content item is a single recommended object selected from the content library.

[0107] The recommended content list is a collection of multiple target content items.

[0108] Target persona prompts are instruction texts that describe the personality, tone, and emotional tendencies of an intelligent agent, used to guide the language model to generate responses that conform to a specific persona.

[0109] Emotional expression is natural language statements that contain emotional overtones such as empathy, encouragement, and companionship (e.g., "A rainy night is perfect for listening to this song quietly...").

[0110] Recommendations are the messages ultimately displayed to the user and may include a list of recommended content and / or emotional natural language responses.

[0111] For example, taking music recommendation as an example, this application can be based on the generated task planning dataset, i.e. <plan> (Problem Analysis, Intelligent Agent)##(Weather Acquisition, Tool)##(IP Acquisition, Tool)...< / plan> The content of the plan is parsed according to a specific format to obtain the specific intelligent agent capabilities and names required for each step. In each subtask, the corresponding sub-agent or tool module is invoked, ultimately producing output that matches the user's query. For example, for the query "recommend songs suitable for listening on rainy nights," a corresponding music recommendation list is obtained. To achieve more accurate music recommendations and user-friendly interaction between the user and the intelligent agent, this application can perform dynamic music recommendation decisions and emotional output.

[0112] Specifically, this application first combines the user query and the returned song list, and then uses a prompt word strategy to make the language model output the song list that best meets the user's needs; Then, based on the previously filtered song list, emotional output will be generated to better align with the pre-defined AI agent persona. Here, a prompt word strategy is used to allow the large model to respond to user queries based on the filtered song list, generating natural language responses that incorporate emotional expressions (such as "Rainy nights are perfect for companionship, I recommend this song 'XX'—a healing folk song from the new album, similar to the style you often listen to").

[0113] In summary, this application achieves accurate parsing and dynamic scheduling of complex queries through a technical approach that integrates value filtering, intent recognition, task planning, multi-capability collaborative execution, and emotional recommendation. It integrates multi-dimensional contextual information to generate personalized and emotional recommendation information, effectively overcoming the shortcomings of related technologies such as single dimension, rigid interaction, and lagging profiling, and significantly improving the accuracy of the recommendation system and user experience.

[0114] Furthermore, for ease of understanding, such as Figure 3 The diagram shown is a schematic representation of a specific recommended method provided in this application.

[0115] Reference Figure 3 Taking music recommendation as an example, after a user inputs query information (e.g., "recommend songs suitable for listening to on a rainy night"), the main agent (i.e., the target agent) first performs intent recognition and task planning, generating a planning dataset containing multiple sub-tasks. Subsequently, the main agent sequentially calls each sub-agent and functional module according to the plan: the query analysis agent extracts explicit tags (e.g., "rainy day," "night") and implicit tags (e.g., "lonely," "relax") from the query; the weather acquisition tool obtains weather information for the current or specified location; the location acquisition tool determines the user's location; the hot topic acquisition tool obtains real-time hot topics related to music (e.g., a singer releasing a new album); and the user behavior agent dynamically generates a user profile (including style preferences, time preferences, etc.) based on a sliding window. Based on the above multi-dimensional information, the song search tool retrieves a matching list of songs. Finally, the dynamic song decision module filters out the recommendation results that best match the user's current context and preferences, and the emotional output module generates recommendation information with emotional expression (e.g., "This folk song is perfect for listening to quietly on a rainy night; it's similar to the style you often listen to"), providing feedback to the user.

[0116] Specifically, the intent recognition and task planning of the target intelligent agent can be as follows: Figure 4 As shown, refer to Figure 4 In the dynamic intent parsing and task planning section, this application receives user-input query information (e.g., "recommend songs suitable for listening to on a rainy night") through a main intelligent agent. Through a coupled intent recognition and task planning process, it identifies multiple user intents (such as recommendations, emotional support, trend tracking, etc.) on one hand, and simultaneously plans the execution steps required to complete the user's task and the corresponding tool or sub-agent invocation order on the other. That is, intent recognition and task planning are intertwined and iteratively optimized to generate a task planning dataset that better meets user needs.

[0117] In order to implement the above embodiments, this application also provides a recommended device. Figure 5 This is a schematic diagram of the structure of a recommended device 500 provided in an embodiment of this application. For example... Figure 5 As shown, the device includes: The identification unit 510 is used to identify the intent of the query information input by the user and determine the intent label of the query information; Planning unit 520 is used to perform task parsing on query information based on intent labels in order to generate a task planning dataset corresponding to the query information. The acquisition unit 530 is used to call at least one sub-agent and / or at least one functional module according to the task planning dataset, so as to obtain at least one execution result fed back by at least one sub-agent and / or at least one functional module after executing the task planning dataset; Recommendation unit 540 is used to generate recommendation information that matches the query information based on at least one execution result, and to feed the recommendation information back to the user.

[0118] In some embodiments, the identification unit 510 is configured to: Based on value assessment prompts and value assessment samples, the query information is scored to determine whether the query information is valid. If the query information is valid, the intent of the valid query information is identified to obtain the intent label corresponding to the valid query information.

[0119] In some embodiments, the planning unit 520 is used for: Based on intent labels and task planning models, combined with task prompts and task constraints, the query information is parsed to generate a task planning dataset corresponding to the query information.

[0120] In some embodiments, the planning unit 520 is configured to: obtain a query sample and input the query sample into the model to be trained before performing task parsing on the query information based on intent labels and task planning models, combined with task prompts and task constraints, to generate a task planning dataset corresponding to the query information, in order to generate a training task planning dataset corresponding to the query sample. Based on format verification rules, process verification rules, and result verification rules, the correctness of the training task planning dataset's format, the correctness of its subtask annotations, and its consistency with the reference plan are verified respectively, and the verification results corresponding to the query samples are obtained. Based on the validation results, the parameters of the model to be trained are adjusted until the model meets the convergence condition, and the model that meets the convergence condition is used as the task planning model.

[0121] In some embodiments, at least one sub-agent includes at least one of a query analysis agent and a user behavior agent; at least one functional module includes at least one of a time acquisition module, a location acquisition module, a weather acquisition module, a hot topic event acquisition module, and a content search module; the acquisition unit 530 is used for: When the query analysis agent is invoked, explicit and / or implicit tags extracted from the query information are obtained; and / or, When the user behavior intelligent agent is invoked, a dynamic preference profile of the user is obtained. The dynamic preference profile includes at least one preference label and a score corresponding to each preference label in the search at least one preference label; and / or, When the time acquisition module is invoked, the query time information corresponding to the query information is obtained; and / or, When the location acquisition module is invoked, the user's geographic location information is obtained; and / or, When the weather acquisition module is invoked, retrieve the corresponding weather information for the query; and / or, When the hot event acquisition module is invoked, hot event information related to the query information is retrieved; and / or, When the content search module is invoked, a list of content items matching the query information is obtained.

[0122] In some embodiments, the recommended unit 540 is configured to: Based on at least one execution result, at least one target content item matching the query information is selected from the target content library, so as to generate a recommended content list based on at least one target content item; Based on the recommended content list and target persona prompts, generate recommendation information containing emotional expression to provide the recommended content list and / or recommendation information back to the user.

[0123] The methods and apparatus provided in the embodiments of this application have been described above. To implement the functions of the methods provided in the embodiments of this application, the electronic device may include a hardware structure and software modules, and may implement the above functions in the form of a hardware structure, software modules, or a hardware structure plus software modules. One of the above functions may be executed in the form of a hardware structure, software modules, or a hardware structure plus software modules.

[0124] Figure 6 This is a block diagram illustrating an electronic device 600 for implementing the recommended method described above, according to an exemplary embodiment. For example, the electronic device 600 may be a mobile phone, computer, messaging device, game console, tablet device, medical device, fitness device, personal digital assistant, etc.

[0125] Reference Figure 6 The electronic device 600 may include a communication interface 601, capable of interacting with other devices; a processor 602, connected to the communication interface 601 to enable interaction with other devices, used to execute the methods provided by one or more of the above-described technical solutions when running a computer program; and a memory 603, on which the computer program is stored. Specifically, the specific processing procedure of the processor 602 can refer to the recommended methods described in the above embodiments of this disclosure.

[0126] Of course, in practical applications, the various components in electronic device 600 are coupled together through bus system 604. It can be understood that bus system 604 is used to realize the connection and communication between these components. In addition to a data bus, bus system 604 also includes a power bus, a control bus, and a status signal bus. However, for the sake of clarity, in... Figure 6 The general designated all buses as Bus System 604.

[0127] The memory 603 in this embodiment is used to store various types of data to support the operation of the electronic device 600. Examples of such data include any computer program used to operate on the electronic device 600.

[0128] The methods disclosed in the embodiments of this application can be applied to processor 602, or implemented by processor 602. Processor 602 may be an integrated circuit chip with signal processing capabilities. In the implementation process, each step of the above method can be completed by the integrated logic circuit of the hardware in processor 602 or by instructions in the form of software. The processor 602 may be a general-purpose processor, a digital signal processor (DSP), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Processor 602 can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this application. A general-purpose processor may be a microprocessor or any conventional processor, etc. The steps of the methods disclosed in the embodiments of this application can be directly manifested as being executed by a hardware decoding processor, or being executed by a combination of hardware and software modules in the decoding processor. The software modules may be located in a storage medium, which is located in memory 603. Processor 602 reads the information in memory 603 and combines its hardware to complete the steps of the aforementioned method.

[0129] In an exemplary embodiment, the electronic device 600 may be implemented by one or more application-specific integrated circuits (ASICs), DSPs, programmable logic devices (PLDs), complex programmable logic devices (CPLDs), field-programmable gate arrays (FPGAs), general-purpose processors, controllers, microcontrollers (MCUs), microprocessors, or other electronic components to perform the aforementioned method.

[0130] Embodiments of this disclosure also propose a non-transitory computer-readable storage medium storing computer instructions, wherein the computer instructions are used to cause a computer to perform the recommended methods described in the above embodiments of this disclosure.

[0131] Embodiments of this disclosure also provide a computer program product, including a computer program that is executed by a processor using the recommended methods described in the above embodiments of this disclosure.

[0132] Embodiments of this disclosure also propose a chip including one or more interface circuits and one or more processors; the interface circuits are used to receive signals from the memory of an electronic device and send signals to the processors, the signals including computer instructions stored in the memory, which, when executed by the processor, cause the electronic device to perform the recommended methods described in the above embodiments of this disclosure.

[0133] It should be noted that the terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this disclosure are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this disclosure described herein can be implemented in orders other than those illustrated or described herein. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this disclosure. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this disclosure as detailed in the appended claims.

[0134] In the description of this specification, the terms "one embodiment," "some embodiments," "illustrative embodiment," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with an embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0135] Any process or method description in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing a particular logical function or process, and the scope of the preferred embodiments of the invention includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as will be understood by those skilled in the art to which embodiments of the invention pertain.

[0136] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a system including a processing module, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of computer-readable media include: an electrical connection having one or more wires (control method), a portable computer disk drive (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic device, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which programs can be printed, because programs can be obtained electronically, for example, by optically scanning the paper or other media, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0137] It should be understood that various parts of the embodiments of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0138] Those skilled in the art will understand that all or part of the steps of the methods described in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

[0139] Furthermore, the functional units in the various embodiments of the present invention can be integrated into a processing module, or each unit can exist physically separately, or two or more units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium. The storage medium mentioned above can be a read-only memory, a disk, or an optical disk, etc.

[0140] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.< / plan> < / plan> < / plan> < / think>

Claims

1. A recommendation method, characterized in that, The method is applied to a target intelligent agent, including: The intent of the query information input by the user is identified to determine the intent tag of the query information; Based on the intent tag, the query information is parsed to generate a task planning dataset corresponding to the query information; Based on the task planning dataset, at least one sub-agent and / or at least one functional module are invoked to obtain at least one execution result fed back by the at least one sub-agent and / or the at least one functional module after executing the task planning dataset; Based on the at least one execution result, recommended information matching the query information is generated, and the recommended information is fed back to the user.

2. The method according to claim 1, characterized in that, The process of identifying the intent of the user-input query information and determining the intent tag of the query information includes: Based on value assessment prompts and value assessment samples, the query information is scored to determine whether the query information is valid. If the query information is valid, intent recognition is performed on the valid query information to obtain the intent tag corresponding to the valid query information.

3. The method according to claim 1 or 2, characterized in that, The step of parsing the query information based on the intent tag to generate the task planning dataset corresponding to the query information includes: Based on the intent label and task planning model, combined with task prompts and task constraints, the query information is parsed to generate the task planning dataset corresponding to the query information.

4. The method according to claim 3, characterized in that, Before performing task parsing on the query information based on the intent label and task planning model, combined with task prompts and task constraints, to generate the task planning dataset corresponding to the query information, the method includes: Obtain a query sample and input the query sample into the model to be trained to generate the training task planning dataset corresponding to the query sample; Based on format verification rules, process verification rules, and result verification rules, the format correctness, subtask annotation correctness, and consistency with the reference plan of the training task planning dataset are verified respectively, and the verification results corresponding to the query sample are obtained. Based on the verification results, the parameters of the model to be trained are adjusted until the model to be trained meets the convergence condition, and the model to be trained that meets the convergence condition is used as the task planning model.

5. The method according to claim 1, characterized in that, The at least one sub-agent includes at least one of a query analysis agent and a user behavior agent; the at least one functional module includes at least one of a time acquisition module, a location acquisition module, a weather acquisition module, a hot event acquisition module, and a content search module. Based on the task planning dataset, at least one sub-agent and / or at least one functional module is invoked to obtain at least one execution result fed back by the at least one sub-agent and / or the at least one functional module after executing the task planning dataset, including: When the query analysis agent is invoked, explicit tags and / or implicit tags extracted from the query information are obtained; and / or, When the user behavior agent is invoked, a dynamic preference profile of the user is obtained, the dynamic preference profile including at least one preference label and a score corresponding to each preference label in the search at least one preference label; and / or, When the time acquisition module is invoked, the query time information corresponding to the query information is obtained; and / or, When the location acquisition module is invoked, the user's geographic location information is obtained; and / or, When the weather acquisition module is invoked, the corresponding weather information is obtained; and / or, When the hotspot event acquisition module is invoked, hotspot event information related to the query information is acquired; and / or, When the content search module is invoked, a list of content items matching the query information is obtained.

6. The method according to claim 1, characterized in that, The step of generating recommended content matching the query information based on the at least one execution result and feeding the recommended content back to the user includes: Based on the at least one execution result, at least one target content item matching the query information is selected from the target content library, so as to generate a recommended content list based on the at least one target content item; Based on the recommended content list and target persona prompts, recommended information containing emotional expressions is generated to feed back the recommended content list and / or the recommended information to the user.

7. A recommended device, characterized in that, The device is applied to a target intelligent agent, and the device includes: The identification unit is used to identify the intent of the query information input by the user and determine the intent tag of the query information; The planning unit is used to perform task parsing on the query information based on the intent label, so as to generate a task planning dataset corresponding to the query information; The acquisition unit is configured to, based on the task planning dataset, invoke at least one sub-agent and / or at least one functional module to obtain at least one execution result fed back by the at least one sub-agent and / or the at least one functional module after executing the task planning dataset; The recommendation unit is used to generate recommendation information that matches the query information based on the at least one execution result, and to feed the recommendation information back to the user.

8. An electronic device, characterized in that, include: The processor and the memory used to store computer programs that can run on the processor. When the processor is used to run the computer program, it performs the method according to any one of claims 1 to 6.

9. A non-transitory computer-readable storage medium storing computer instructions, characterized in that, The computer instructions are used to cause the computer to perform the method according to any one of claims 1 to 6.

10. A computer program product, characterized in that, Includes a computer program that, when executed by a processor, performs the method according to any one of claims 1 to 6.

Images