Dynamic generation method, device and equipment for supply features

Abstract

The embodiment of the invention discloses a dynamic generation method, device and equipment for supply features, relates to the technical field of artificial intelligence, and is used for solving the problem that an existing supply feature is disjointed from a user intention. The method comprises the steps of receiving demand description of a demand terminal based on an agent interaction interface; analyzing the demand description to generate a supply understanding task; executing the supply understanding task to obtain a supply reasoning result of target supply; the step of executing the supply understanding task comprises the sub-steps of obtaining multi-dimensional feature data of each target supply in a target supply set; based on the to-be-generated feature semantic dimension, constructing a reasoning prompt; reasoning the multi-dimensional feature data according to reasoning prompts through a large language model to generate a supply reasoning result; and structuring a supply reasoning result into digital supply features, and carrying out persistent storage.

Description

Technical Field

[0001] This specification relates to the field of artificial intelligence technology, and in particular to a method, apparatus, and device for dynamically generating supply features. Background Technology

[0002] In various recommendation systems, such as intelligent recommendation, supply and demand matching, and personalized distribution, the supply characteristics of goods, content, services, and benefits are the key factors that determine accurate matching and personalized recommendations.

[0003] In traditional solutions, the generation of supply features mainly relies on predefined static feature systems such as categories, brands, and price ranges, or on characterizing supply based on historical user behavior. However, static tags are difficult to cover the fine-grained semantic attributes of long-tail supply. Furthermore, for cold-start supply with only sparse information such as titles, the feature representation is severely insufficient, leading to ineffective understanding and recommendation, resulting in supply-side blind spots that directly affect the subsequent distribution accuracy and user experience of the recommendation system.

[0004] To alleviate these problems, current methods include manual annotation, rule-based extraction, or batch processing based on fixed models. However, relying on annotation teams to semantically annotate the supply is costly, time-consuming, and lacks scalability, making it difficult to meet the rapid coverage needs of massive new supply. Rule-based extraction or batch processing based on fixed models, on the other hand, cannot dynamically adjust the feature generation dimensions according to real-time user needs, resulting in generated features that are disconnected from the user's current intent.

[0005] Therefore, there is a need for a way to dynamically generate semantically aligned supply features based on user needs. Summary of the Invention

[0006] This specification provides one or more embodiments of a method, apparatus, and medium for dynamically generating supply features, which addresses the following technical problems: To solve the above-mentioned technical problems, one or more embodiments of this specification are implemented as follows: This specification provides a method for dynamically generating supply characteristics through one or more embodiments, comprising: The requirements description of the terminal receiving the requirements based on the intelligent agent interaction interface; The demand description is parsed to generate a supply understanding task; the supply understanding task is used to dynamically determine the target supply set related to the semantics of the demand description, as well as the feature semantic dimensions to be generated. The supply understanding task is executed to obtain the supply reasoning result of the target supply; the execution of the supply understanding task includes: obtaining multi-dimensional feature data of each target supply in the target supply set; constructing reasoning prompts based on the feature semantic dimensions to be generated; and using a large language model, reasoning on the multi-dimensional feature data according to the reasoning prompts to generate the supply reasoning result. The supply reasoning results are structured into digital supply characteristics and stored persistently.

[0007] This specification provides one or more embodiments of a dynamic generation apparatus for supply features, comprising: The receiving module is used to receive the requirement description from the requesting terminal based on the intelligent agent interaction interface; The task generation module is used to parse the demand description to generate a supply understanding task; the supply understanding task is used to dynamically determine the target supply set related to the semantics of the demand description, as well as the feature semantic dimensions to be generated. The reasoning module is used to execute the supply understanding task and obtain the supply reasoning result of the target supply. The execution of the supply understanding task includes: acquiring multi-dimensional feature data of each target supply in the target supply set; constructing reasoning prompts based on the feature semantic dimensions to be generated; and using a large language model to reason about the multi-dimensional feature data according to the reasoning prompts to generate the supply reasoning result. The storage module is used to structure the supply reasoning results into digital supply characteristics and store them persistently.

[0008] This specification provides one or more embodiments of a dynamic generation device for supply features, comprising: At least one processor; and, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions executable by the at least one processor, which, when executed by the at least one processor, enable the at least one processor to: The requirements description of the terminal receiving the requirements based on the intelligent agent interaction interface; The demand description is parsed to generate a supply understanding task; the supply understanding task is used to dynamically determine the target supply set related to the semantics of the demand description, as well as the feature semantic dimensions to be generated. The supply understanding task is executed to obtain the supply reasoning result of the target supply; the execution of the supply understanding task includes: obtaining multi-dimensional feature data of each target supply in the target supply set; constructing reasoning prompts based on the feature semantic dimensions to be generated; and using a large language model, reasoning on the multi-dimensional feature data according to the reasoning prompts to generate the supply reasoning result. The supply reasoning results are structured into digital supply characteristics and stored persistently.

[0009] The above-described at least one technical solution adopted in the embodiments of this specification can achieve the following beneficial effects: The system directly receives natural language demand descriptions from end-users via an intelligent agent interaction interface and automatically parses them into structured supply understanding tasks. This reduces the technical requirements on end-users and ensures real-time generation of digital supply features based on demand descriptions. The parsing of demand descriptions dynamically generates supply understanding tasks adapted to the current demand side's description. These tasks dynamically determine the target supply set semantically related to the demand description, as well as the semantic dimensions of the features to be generated, ensuring that the subsequently generated feature dimensions correspond to the user's intent. Reasoning prompts constructed from the semantic dimensions of the features to be generated by a large language model are used to reason about multi-dimensional feature data, solving the problem of cold-start supply, where insufficient feature representation leads to ineffective understanding and recommendation, resulting in supply-side blind spots. Attached Figure Description

[0010] To more clearly illustrate the technical solutions in the embodiments or prior art of this specification, the drawings used in the description of the embodiments or prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this specification. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In the drawings: Figure 1 A flowchart illustrating a method for dynamically generating supply features as provided in an embodiment of this specification; Figure 2 This specification provides a schematic diagram of an existing process for digital supply of manual marking, as an embodiment of the present invention. Figure 3 A schematic diagram of an existing process for a digital supply labeling standard operating procedure provided for embodiments of this specification; Figure 4 This is a schematic diagram of the overall link for the dynamic generation of supply characteristics in an application scenario provided by an embodiment of this specification. Figure 5 A schematic diagram of a domain model for the dynamic generation of supply characteristics in an application scenario provided in this specification embodiment; Figure 6 This is a schematic diagram of the design of an intelligent agent in an application scenario provided by an embodiment of this specification; Figure 7 A schematic diagram of the structure of a dynamic generation device for supply features provided in an embodiment of this specification; Figure 8This is a schematic diagram of the structure of a dynamic generation device for supply features provided in an embodiment of this specification. Detailed Implementation

[0011] This specification provides a method, apparatus, and device for dynamically generating supply features through embodiments.

[0012] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this specification, and not all embodiments. Based on the embodiments of this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this specification.

[0013] In the scenarios mentioned in the background technology, various recommendation systems, such as intelligent recommendation, supply and demand matching, and personalized distribution, all rely on a deep semantic understanding of the digital supply of goods, content, services, and benefits. Among these, the quality of the generated supply feature tags directly determines whether the recommendation system can accurately match user intent to achieve personalized recommendations, and is a key link in achieving efficient matching.

[0014] Traditional solutions primarily rely on static feature systems based on predefined human definitions such as categories, brands, and price ranges for compliance, or on statistical representation of supply based on historical user behavior data such as click-through rates and conversion rates. However, in practical applications, static tags have limited coverage and struggle to capture the fine-grained semantic attributes of long-tail supply. Furthermore, for cold-start supply with only sparse information like titles, their feature representation is severely inadequate, leading to ineffective understanding and recommendation, creating supply-side blind spots, and directly impacting the subsequent distribution accuracy and user experience of the recommendation system. Statistical representation based on historical user behavior data, in the context of current recommendation systems emphasizing personalization, real-time performance, and long-tail coverage, relies on historical average feedback from all users, failing to reflect the specific needs of current users. This can easily result in high-click-rate products being continuously pushed while niche but highly relevant supply is overlooked. Moreover, historical user behavior data is highly lagging, making it difficult to respond to semantic changes or contextual needs, and unable to dynamically adjust feature dimensions.

[0015] To alleviate the aforementioned problems, two approaches have been proposed: a digital supply labeling method involving semantic tagging by a human annotation team, and a digital supply labeling standard operating procedure (SOP) method that automatically extracts features using a fixed rule engine or standard operating procedure workflow. However, manual annotation relies on technicians manually achieving the desired results, which is costly, time-consuming, and lacks scalability, making it difficult to meet the rapid coverage needs of massive new supply. In contrast, the digital supply labeling SOP method... Its rule-based or fixed-model-based batch processing cannot dynamically adjust the feature generation dimensions according to real-time user needs, resulting in generated features that are out of touch with the current intent. Furthermore, in scenarios with complex semantics of demand, traditional methods cannot parse multi-level intents, nor can they make reasonable judgments about whether new supply meets the composite conditions, causing new supply to be ignored for extended periods. In other words, after receiving user demand, the system only uses the raw metadata of the supply to call a large language model for general semantic labeling, generating a set of fixed-dimensional labels. However, this approach does not integrate the context of the user demand into the reasoning process, resulting in generated features with strong generalization but weak task relevance.

[0016] Moreover, in the currently widely deployed recommendation system architecture, the feature generation module is mostly an offline batch processing flow, decoupled from online requirements, and cannot support real-time dynamic feature inference. If the existing system is modified to support requirement-driven feature generation, the feature production chain needs to be reconstructed and complex scheduling and caching mechanisms need to be introduced, resulting in high engineering costs and long implementation cycles.

[0017] To address the aforementioned issues, this application proposes a dynamic method for generating supply features. This method receives user demand descriptions through an intelligent agent interaction interface, dynamically parses and generates supply understanding tasks, constructs reasoning prompts based on the semantic dimensions of the features to be generated, drives a large language model to perform context-aware semantic reasoning on the multi-dimensional feature data of the target supply, and then structures the reasoning results into digital supply features and stores them persistently. Without relying on historical user behavior data, it can generate supply features that match the semantic capabilities of the target supply, including cold-start supply, and align with the user's current intent, based on the demand description and the target supply at the moment the user initiates a demand.

[0018] Based on this overall approach, the following detailed description of the proposed solution will be provided.

[0019] Figure 1 This document provides a flowchart illustrating the dynamic generation of supply features according to one or more embodiments. This method can be applied to various fields requiring semantic understanding and feature construction of digital supply, such as intelligent recommendation, content distribution, local life services, and e-commerce platforms. The process can be executed using computing devices relevant to the field (e.g., servers in recommendation systems). Certain input parameters or intermediate results in the process can be manually adjusted to help improve the accuracy and business adaptability of feature generation. Furthermore, it should be noted that supply features are attributes or information that can be used to describe and characterize a supply. Supply tags are a common form of representation for supply features. Some embodiments in the following description use supply tags for intuitive explanation, but are not limited to this form.

[0020] Figure 1The process may include the following steps: S102: Requirements description of the terminal receiving the requirements based on the intelligent agent interaction interface.

[0021] like Figure 2 and Figure 3 The diagram illustrates the existing process of digital supply labeling for one type of supply characteristic, and the existing process of the standard operating procedure (SOP) for digital supply labeling. In the current traditional method, demand description is based on the demander directly inputting supply understanding requirements into the supply understanding platform, which then processes this information to generate a supply understanding task. Therefore, the current supply characteristic determination process requires the demander to fill out a structured form or configure Standard Operating Procedure (SOP) rules before inputting them into the supply understanding platform. This means the demander must translate complex business intentions into instructions that the platform can understand. At this point, ordinary users or terminals cannot directly trigger the supply understanding task, resulting in real-time feature generation not being driven by genuine user intent. Furthermore, the process of inputting into the supply understanding platform is a one-time, one-way action. When the supply understanding platform cannot understand the content input by the demander or the information is insufficient, it cannot proactively initiate multi-round dialogues to confirm or guide the content. The demander needs to guess the problem, revise, and resubmit, leading to low efficiency. In addition, manually entered requirements are usually static batch data with a wide range of data, which cannot capture real-time user intent based on this content, and can easily lead to a disconnect between the subsequently generated supply characteristics and actual demand.

[0022] Therefore, to address the problems caused to the supply understanding platform by the existing demand side directly inputting supply understanding demand, such as... Figure 4 As shown in the embodiments of this application, the user's requirement description is automatically received through an intelligent agent interaction interface. The intelligent agent can extract the natural language text received from the user and convert it into a structured requirement description.

[0023] Furthermore, to improve the accuracy of requirement descriptions and avoid the problem of miscommunication due to single descriptions, this process can be specifically based on the following method: Receiving dialogue information uploaded by the requirement terminal through the intelligent agent interaction interface. Dialogue information refers to the natural language text input by the user of the requirement terminal in one or more interaction rounds; it is the original semantic content in which the user expresses their business needs, answers questions, or clarifies. In response to this dialogue information, multi-round dialogue state tracking is performed on the requirement terminal. That is, dialogue information uploaded from the requirement terminal is continuously received, each newly received dialogue information is parsed, and the understanding of the user's overall intent is updated based on the historical dialogue context. This state tracking continues until the end of the dialogue is detected or a reasoning instruction is triggered, at which point the current dialogue state is obtained. This process ensures that the recommendation system only initiates subsequent reasoning processes when the requirements are fully clear, avoiding invalid calculations. After obtaining the complete current dialogue state, the requirement content of the requirement terminal is extracted from the current dialogue state and structured into a requirement description. Structured processing refers to extracting the requirements content from the current dialogue state, removing the ambiguity of natural language and redundant information in the interaction process, and converting the extracted requirements content into a formatted requirements description that conforms to machine processing standards. This requirements description includes at least: the screening conditions for the target supply and the semantic dimensions of the features to be analyzed.

[0024] In this process, the demand side only needs to upload its demand description, expressing its true intent, to the intelligent agent through a natural language interaction interface. This allows the intelligent agent to transform the unstructured demand into a supply understanding task, facilitating the dynamic generation of task-oriented supply understanding tasks driven by real user intent. This reduces the high technical requirements on the demand side compared to traditional manual or SOP-based labeling methods. Furthermore, compared to the one-way submission of traditional methods, this application enables multi-turn dialogue based on the intelligent agent to refine the demand description, reducing rework and inefficient retries caused by miscommunication of demand content. Receiving the demand description from the demand terminal allows for real-time determination of feature semantic dimensions based on current user needs, transforming the offline batch processing of traditional supply understanding in the aforementioned context into online on-demand inference. This eliminates the need to wait for manual inclusion in labeling tasks or the accumulation of behavioral data, thus significantly improving the real-time performance of supply features.

[0025] Furthermore, in the process of understanding requirements based on natural language dialogue information, relying solely on user input may result in semantic ambiguity, incomplete expression, or inconsistency with the true intent. Therefore, to dynamically correct requirement semantics by integrating user interface behavior and dialogue content, thereby improving the accuracy and robustness of requirement understanding, in one embodiment of this specification, the aforementioned multi-turn dialogue state tracking of the requirement terminal to obtain the current dialogue state when the dialogue ends or a reasoning instruction is triggered is specifically performed in the following manner: While the user expresses their needs through dialogue, the interface behavior characteristics of the corresponding interactive interface of the monitoring agent are observed during the dialogue.

[0026] Interface behavior features refer to the explicit operational behaviors generated by the intelligent agent on the human-computer interaction interface when users interact with it. Therefore, in one approach, the explicit operational behaviors generated on the interaction interface are directly acquired as interface behavior features. These features include: recording the user's gaze or cursor trajectory as it stays and moves across different areas of the interface, reflecting a sequence of user focus points reflecting changes in attention; recording the sequence of user operations on interactive controls such as clicking, selecting, dragging, and inputting; and the temporal characteristics of the operation behaviors representing their temporal attributes. In another approach, interface behavior features can also be obtained by: real-time tracking of the user's visual or operational focus on the interaction interface to determine the currently focused supply object. This focused supply object can be a digitally supplied unit that can be recommended or understood, such as product cards, content items, or service entry points on the interaction interface. For each identified focused supply object, the attribute features corresponding to the focused supply object during the dialogue are identified. The corresponding structured attribute features can be obtained from a supply database or cache, including but not limited to: category, price, brand, specifications, historical ratings, and semantic description text. By aligning the sequence of user control operations during the conversation with the attribute characteristics of each focused object in a temporal sequence, we can obtain the interface behavior characteristics.

[0027] Based on methods such as semantic similarity, the semantic relationship between interface behavior features and dialogue information is obtained. This relationship is then used to correct the dialogue information, resulting in a revised version. In other words, by establishing the semantic relationship between interface behavior features and dialogue information, corresponding interface behavior features are obtained. These features then clarify ambiguous or denotative content in the dialogue, generating semantically clear and specific dialogue information. For example, when a user says in a dialogue, "Lower the price of this," the system will combine the specific supply items semantically associated with the interface behavior features to correct the ambiguous pronoun "this" into a clear supply identifier.

[0028] Based on the corrected dialogue information and interface behavior features, the dialogue state corresponding to the dialogue information is dynamically updated to obtain the current dialogue state. At this time, the current dialogue state is dynamically updated based on each round of interaction, which not only integrates the user's semantic intent, but also combines the interface behavior features of the interactive interface, realizing dynamic correction of the semantics of the requirements and improving the accuracy and robustness of subsequent requirement understanding.

[0029] S104: Parse the demand description to generate a supply understanding task; the supply understanding task is used to dynamically determine the target supply set related to the semantics of the demand description, as well as the feature semantic dimensions to be generated.

[0030] based on Figure 2 and Figure 3 Taking supply tag acquisition as an example, the current method of acquiring supply characteristics shows that after the supply understanding platform inputs supply understanding requirements, it generates new supply understanding tasks. Therefore, the existing method generates fixed supply understanding tasks based on these instruction-based supply understanding requirements. The content and scope of this task are rigidly defined when the requirements are entered, and cannot be dynamically adjusted according to the deeper semantics of the requirement description. In this case, generating a supply understanding task merely fills in the preset parameters contained in the supply understanding requirements, failing to respond to the personalized and real-time needs of the demand side.

[0031] To solve this existing problem, such as Figure 4 As shown, this application executes all SOP steps based on a supply understanding task generated by the agent after receiving the demand description. After the agent's interaction interface successfully receives and structures the user's demand description, it does not simply map it to a pre-set task template. Instead, the agent performs semantic parsing and task planning on the demand description, dynamically generating a supply understanding task adapted to the current demander's demand description. This supply understanding task can dynamically determine the target supply set related to the semantics of the demand description, as well as the feature semantic dimensions to be generated.

[0032] Specifically, in one or more embodiments of this specification, the task of parsing demand descriptions to generate supply understanding includes: The requirement description is subjected to intent identification and key semantic dimension extraction to obtain multi-level intents and key semantic dimensions. Intent identification can be based on classification or sequence labeling models. Multi-level intents include: the user's idea graph corresponding to the task objective and nested sub-intents of specific sub-objectives or constraints required by the idea graph. Key semantic dimensions are key entities, attributes, and constraints related to the multi-level intents, used to constitute the specific components of the task.

[0033] Based on the acquired multi-level intents and the key semantic dimensions, a feasibility diagnosis is performed on the requirement description to dynamically determine the target supply set and the feature semantic dimensions to be generated, thus generating a structured supply understanding task. This feasibility diagnosis includes: diagnosing the completeness of the information dimensions, i.e., diagnosing whether the key semantic dimension supports the realization of the multi-level intents; knowledge boundary diagnosis, comparing the key semantic dimensions with the system's knowledge base and business rule base to ensure the requirement is within a reasonable business and cognitive scope; and system capability diagnosis, assessing whether the current system resources and processes can meet the multi-level intents. Based on the diagnostic results, the target supply set and the feature semantic dimensions to be generated are dynamically determined, generating a structured supply understanding task. This dynamic determination process is as follows: based on the completeness diagnosis of the information dimensions and the knowledge boundary diagnosis, the target supply set is dynamically determined. Combining the knowledge boundary diagnosis and the system capability diagnosis, a list of feature dimensions to be used for reasoning is determined as the feature semantic dimensions to be generated.

[0034] In another embodiment, the process of parsing demand descriptions to generate supply understanding tasks can also be implemented as follows: Intent recognition and key semantic dimension extraction are performed on the demand descriptions to obtain multi-level intents and key semantic dimensions. The key semantic dimensions extracted from the demand descriptions are converted into structured query conditions, and a pre-built multi-dimensional index is used to retrieve the supply database, obtaining a set of candidate supplies that meet the constraints as the target supply set. Here, the multi-dimensional index refers to an inverted index structure pre-built based on the structured attributes of the supply, supporting efficient joint retrieval of multiple constraints. Simultaneously, the system maps the corresponding list of feature dimensions to be generated from a predefined feature semantic dimension registry based on the attribute fields involved in the query conditions. The target supply set and the feature dimension list are encapsulated into a structured supply understanding task for subsequent execution. This method also enables on-demand generation of demand descriptions.

[0035] S106: Execute the supply understanding task to obtain the supply reasoning result of the target supply; the execution of the supply understanding task includes: obtaining multi-dimensional feature data of each target supply in the target supply set; constructing reasoning prompts based on the feature semantic dimensions to be generated; and using a large language model to reason about the multi-dimensional feature data according to the reasoning prompts to generate the supply reasoning result.

[0036] like Figure 2 As shown, taking supply tags in supply characteristics as an example, in the current digital supply manual tagging method, after the supply understanding platform defines the characteristics of supply candidate tags, the task and candidate characteristics are handed over to the manual tagging team. The team then tags the supply by referring to the supply information and understanding, and obtains the supply reasoning results for the supply understanding platform. Figure 3Taking supply tags as an example in supply characteristics, the current standard operating procedure (SOP) for digital supply tagging involves the supply understanding platform calling the supply reasoning SOP steps to infer new features, then sending them to the supply understanding SOP to execute all steps, and finally returning the obtained supply reasoning results to the supply understanding platform. It is evident that manual tagging relies on human review of each supply piece of information, which cannot handle massive amounts of new supply. Furthermore, the strong subjectivity of human input leads to poor consistency in supply reasoning results and an inability to dynamically adjust the results based on real-time user intent. In contrast, the SOP-based tagging method relies on preset rules, has poor scalability, typically only processes structured fields and cannot handle unstructured information, and is unable to handle cold-start supply.

[0037] To solve this problem, such as Figure 4 As shown, in one or more embodiments of this application, the recommendation system controls the supply reasoning engine to execute a supply understanding task and obtain the supply reasoning result of the target supply. Specifically, the supply understanding task is executed in the following way: based on the dynamically determined set of target supplies in the supply understanding task, multi-dimensional related data for each target supply is obtained from a feature library or relevant data source. The multi-dimensional related data includes at least the supply's basic description, category static information, and user behavior data. Contextual information and instruction requirements related to this dimension are extracted or generated, and this information is organized into reasoning prompts that conform to the input specifications of a large language model. The reasoning prompts and multi-dimensional feature data are submitted to the large language model, which analyzes and reasons about the input information to generate a supply reasoning result corresponding to the semantic dimension of the feature to be generated. Based on this embodiment, even if the target supply lacks historical user behavior data and only has its basic description and category static information, high-level semantic features can be generated through reasoning by the large language model, avoiding the problem of cold-start supply with severely insufficient feature expression, leading to ineffective understanding and recommendation. The inference hints incorporate the semantic dimensions of features specified by the current requirements, enabling large language models to infer based on task-related attributes and ensure consistency between output features and user intent. Constructing multi-dimensional feature data for inference also enhances the comprehensiveness of the reasoning. This process allows for the addition of new feature dimensions simply by adjusting the inference hints, avoiding the problem of manual annotation teams or complex rule configurations lacking dynamic generation based on real-time user needs.

[0038] like Figure 5As shown, when performing a supply understanding task in a certain application scenario, the process includes the following steps: First, obtain supply instances related to the current task from the supply information entity (i.e., the template entity), such as supply ID, supply name, and supply source, and use their associated multi-dimensional feature data as input. Second, based on the semantic dimension of the features to be generated specified by the supply understanding task, extract task configuration information from the current task context, such as task status, scene code, operator, and task link. Combine this with supply interaction features (i.e., feature semantic dimension metadata), such as feature Chinese name, feature type, and configuration information, to dynamically construct a reasoning prompt to drive the large language model's reasoning. The large language model receives this reasoning prompt and the multi-dimensional feature data of the target supply, performs semantic understanding and reasoning, and outputs a structured supply reasoning result. It should be noted that a supply understanding task can contain multiple target supplies; a target supply can participate in multiple supply understanding tasks; a supply understanding task can process multiple target supplies; a target supply can have multiple supply features; and a supply feature can be used by multiple target supplies.

[0039] Specifically, in one or more embodiments of this specification, obtaining multidimensional feature data of each target supply in the target supply set specifically includes: For each target supply in the target supply set, such as Figure 6When performing supply understanding based on intelligent agents, it can be executed within the intelligent agent architecture based on sub-agents with corresponding functions, and the supply understanding toolset and corresponding database intelligent agents can be directly invoked. Therefore, in this embodiment, it is possible to obtain corresponding global supply metadata from multiple data sources such as the product master database, content management system, third-party access platform, and web crawler collection service. After standardizing the global supply metadata, the global supply metadata is aligned, and a unified supply identifier is assigned to each global supply metadata based on the supply entity corresponding to each global supply metadata. This identifier is used to ensure that data from different sources can be accurately aggregated onto the same supply object. That is, the global supply metadata is standardized, including unified field naming, unit normalization, text cleaning, category mapping, and missing value filling, to eliminate format and semantic differences between different data sources and obtain standardized global supply metadata. The process involves acquiring key attributes corresponding to standardized global supply metadata, performing similarity calculations on these attributes, or determining whether standardized global supply metadata from different data sources points to the same supply entity based on simultaneous exposure within the same user session or supply chain relationships. For multiple standardized global supply metadata entries confirmed to belong to the same supply entity, a globally unique unified supply identifier is assigned. User behavior data from demand terminals is parsed based on contextual information to determine the supply entity corresponding to the user behavior data, thereby binding the user behavior data to the unified supply identifier corresponding to that supply entity. Based on this unified supply identifier, global supply metadata and user behavior data are aggregated, using global supply metadata and user behavior data belonging to the same supply identifier as multi-dimensional feature data of the target supply.

[0040] By introducing a unified supply identifier, the alignment problem of multi-source heterogeneous supply data was effectively solved, and the integration of static attributes and dynamic user behavior data was realized. This significantly improved the integrity and consistency of supply representation and laid a reliable data foundation for high-precision supply understanding tasks.

[0041] Specifically, to address the problem that existing fixed inference instructions cannot dynamically adapt to specific tasks, in one or more embodiments of this specification, inference prompts are constructed based on the semantic dimension of the features to be generated, so that the inference prompts can accurately guide the large language model. The specific process of constructing inference prompts includes: Based on the semantic dimension of the feature to be generated, a basic prompt template is generated. This process involves parsing the semantic dimension to be generated, identifying its dimension type (e.g., attribute, rating, Boolean, descriptive), and recognizing the unique business semantics of that dimension. Templates matching the dimension type and semantics are retrieved from a predefined prompt template library as basic prompt word templates. Each template includes placeholders and task description logic. If the semantic dimension is a new item or not covered by the predefined prompt template library, a basic prompt template can be dynamically generated based on a general template. For example, a rule engine can automatically fill the name and definition of the semantic dimension into a standard sentence to obtain a basic prompt word template. Essentially, the basic prompt template describes the analysis task for the semantic dimension to the large language model, explicitly defining the analysis role, objectives, and preliminary analysis guidance. To overcome the problem that the output content is difficult for subsequent programs to extract directly, the output format requirements for the inference results can be embedded into the basic prompt word template as instructions to obtain an inference prompt framework. The multidimensional feature data of the target supply obtained above is combined with the inference hint framework, that is, the multidimensional feature data is filled into the corresponding placeholder positions of the inference hint framework to obtain inference hints.

[0042] In this process, based on the dynamic requirement of the semantic dimension of the features to be generated, inference prompts containing specific tasks, output format requirements, and multi-dimensional feature data are automatically constructed in real time. This helps to transform the requirement description into instructions that can be stably executed by the large language model, reducing the manual dependence on prompt engineering and the cost of trial and error. Moreover, compared with the static and fixed inference rules in traditional methods, prompts can be dynamically generated according to the specific requirements of each task, making the inference process correspond to the user's intent and improving the accuracy of subsequent feature generation. By embedding the output format requirements as mandatory instructions, not only are the ambiguities and post-processing costs caused by the large language model outputting free text avoided, but the inference results can also be directly used by downstream systems.

[0043] Furthermore, for new supply, determining supply characteristics manually requires time attributes, which leads to long time cycles. Moreover, when user behavior data is lacking, this manual method is highly subjective and difficult to ensure the accuracy of supply characteristics. Determining supply characteristics based on standard operating procedures relies on historical behavior data or rich structured fields for calculation. For new supply with only basic information and lacking any user interaction data, this new supply cannot be effectively recalled and distributed, creating a supply-side blind spot. In one or more embodiments of this specification, corresponding to this part of the target supply, the method further includes: If the target supply is a new target supply, then during the supply understanding task, the full-domain supply metadata of the new target supply is obtained. This full-domain supply metadata includes: machine-readable structured attribute data and semantically rich unstructured descriptive text. Based on the feature semantic dimensions to be generated, a basic prompt template is generated, and the output format requirements of the supply reasoning result are embedded into the basic prompt word template in the form of instructions to obtain a reasoning prompt framework. This process corresponds to the process in the previous embodiment. After obtaining the reasoning prompt framework, the multi-level intents in the demand description are parsed, and the main intent and its nested sub-intents are identified. Intent keywords for each sub-intent are extracted; this extraction process can be completed based on part-of-speech tagging, dependency parsing, or lightweight named entity recognition. The full-domain supply metadata is semantically sliced; for example, structured attribute data can be converted into key-value pair sequences as supply fragments, and unstructured descriptive text can be segmented based on sentences or semantic units to obtain supply fragments. Based on the cosine similarity between the supply fragments and the intent keywords, the intent relevance value of each supply fragment is determined to filter intent-related candidate supply fragments. The selected supply fragments are combined with the multi-level intents in a logical order to obtain a fused fragment. This fused fragment is then filled into the corresponding placeholder position in the reasoning hint framework to generate a reasoning hint.

[0044] This process eliminates the need to wait for the accumulation of user behavior data. Once a new target supply is identified, high-quality supply features can be generated based on its metadata, enabling subsequent recommendation and distribution and eliminating supply-side blind spots. Furthermore, by combining selected supply fragments with multi-level intents in a logical order to obtain fused fragments, the supply tags generated based on inference prompts are not generic tags but are associated with the user intent of the current demand terminal. Only a description of the demand in natural language is needed to generate corresponding inference prompts for new target supplies, improving operational efficiency during the cold start phase.

[0045] In a certain application scenario, an example of an inference prompt generated by the virtual assistant (robot, or BOT) of the intelligent agent is: "Suppose you are an operations expert in the field of internet content. Please think deeply about the given content and then label it with the following six aspects for the recommendation system to train the model and make recommendations."

[0046] The first aspect is the theme category. You need to select one appropriate tag number from the following 19 tags: 1 - Sports Events and Exercises, 2 - Health and Medical Care, 3 - Law and Policy, 4 - Life Skills and Home, 5 - Technology and Digital, 6 - News and Events, 7 - Economy and Consumption, 8 - Safety Education and Emergency Response, 9 - Education and Development, 10 - Tourism and Culture, 11 - Security and Employment, 12 - Environment and Meteorology, 13 - Food Safety and Nutrition, 14 - Affairs and Diplomacy, 15 - Agriculture and Rural Areas, 16 - Transportation and Travel, 17 - Culture, Entertainment and Arts, 18 - Finance and Wealth Management, 19 - Animals and Nature.

[0047] The second aspect is timeliness. You need to select one appropriate tag number from the following nine tags: 1 - Breaking News / Event Updates, 2 - Legal Interpretation, 3 - Practical Life Guide, 4 - Health and Medical Science Popularization, 5 - Safety Warnings and Reminders, 6 - Hot Topic Tracking, 7 - Technology and Innovation Applications, 8 - Seasonal / Holiday Tips, and 9 - Long-Term Knowledge Popularization.

[0048] The third aspect is emotional inclination, which requires you to choose the most appropriate label number from the following 7 labels: 1 - Positive (positive / supportive), 2 - Negative (criticism / exposure), 3 - Neutral (information / notification), 4 - Warning / reminder, 5 - Surprise / excitement, 6 - Anger / controversy, 7 - Tension / crisis.

[0049] The fourth aspect is content quality, which requires selecting one appropriate tag number from the following nine tags: 1-Authoritative policies and regulations, 2-In-depth analysis and commentary, 3-User-generated content (UGC), 4-Practical life guides, 5-Social hot topics and livelihood news, 6-Health and science popularization, 7-Safety warnings and risk prevention, 8-Consumer trends and market dynamics, and 9-Regional services and local information.

[0050] The fifth aspect is interactive potential, which requires selecting one or more appropriate tag numbers from the following four tags: 1-Policy and people's livelihood hot topics (high comment potential), 2-Practical skills dissemination (high sharing rate), 3-Controversial topics that ignite controversy (easily cause controversy), 4-Emotional resonance and diffusion (high dissemination potential).

[0051] The sixth aspect is the target audience, which requires selecting one or more appropriate tag numbers from the following 19 tags: 1 - General household users, 2 - Policy researchers / civil servants, 3 - Health and wellness groups, 4 - Technology and car enthusiasts, 5 - Parents / education professionals, 6 - Tourists and local residents, 7 - Legal and security enthusiasts, 8 - Sports / entertainment enthusiasts, 9 - Finance and investors, 10 - Students, 11 - Science enthusiasts / popular science readers, 12 - Middle-aged and elderly users, 13 - Working professionals / job seekers, 14 - Consumer rights protectors, 15 - Agricultural / rural residents, 16 - Those concerned with local affairs, 17 - Environmental protection and public safety advocates, 18 - Special interest groups (pets / animals), 19 - Culture and tradition enthusiasts.

[0052] The given topic is: What to do about psychological distress and emotional problems? Hotline is here. Accurately tag the content according to the seven aspects mentioned above, ensuring only the tag numbers are output, and guaranteeing accurate JSON format. Check and correct any errors in the output format. Furthermore, in traditional supply feature generation schemes, whether manual or based on fixed standard operating procedures, the reasoning process is usually singular and fixed, making it difficult to adaptively adjust to the complexity of the supply understanding task, resulting in a trade-off between reasoning efficiency and quality. To address this, in one or more embodiments of this specification, a large language model is used to reason about the multidimensional feature data based on reasoning prompts to generate supply reasoning results, specifically including: The inference prompts are input into the large language model to obtain multi-dimensional inference output. Specifically, the large language model performs preliminary semantic understanding and logical reasoning based on the task description, format requirements, and multi-dimensional feature data in the prompts, and outputs structured multi-dimensional inference output. After obtaining this multi-dimensional inference output, in order to determine the nature of the task and dynamically decide whether to initiate and what kind of further inference analysis to achieve deep satisfaction of complex needs and efficient response to simple needs, the embodiments in this specification determine the task type based on at least one task element in the supply understanding task: the level of abstraction of the semantic dimension to be generated, the output mode specified by the supply understanding task, the size and status of the target supply set, or the execution mode of historically similar tasks.

[0053] Then, based on the task type corresponding to the supply understanding task, it is determined whether the supply inference engine should execute at least one enhanced inference subtask. Enhanced inference subtasks include: tag dimension expansion, batch tag generation, or target audience selection. If an enhanced inference subtask is executed, the supply inference result is obtained based on the multi-dimensional inference output and the corresponding output of the enhanced inference subtask. If no enhanced inference subtask is executed, the multi-dimensional inference output is used as the supply inference result.

[0054] The abstraction level of the feature semantic dimension is used to measure whether the semantic dimension can be directly extracted from structured data or requires external knowledge and semantic reasoning. Specifically, the system maintains a predefined dictionary of specific attributes. If the semantic dimension to be generated does not belong to this dictionary and its semantic similarity with any attribute in the dictionary is lower than a preset threshold, it is judged as a high abstraction dimension; otherwise, it is a low abstraction dimension.

[0055] The output pattern specified by the supply understanding task is determined by the intent parsing results in the task generation phase, representing the expected form of reasoning results, including: returning Boolean or multi-valued labels, returning a list of target user IDs or audience package identifiers, returning a label table of the full supply, returning a natural language report with evidence and reasons, etc.

[0056] The size of the target supply set is used to represent the quantity of the target supply. The status includes whether it is a new product, whether there is user behavior data, inventory status, etc. This information is obtained from the global supply metadata database.

[0057] The execution mode for historically similar tasks is that if any historically similar task successfully executes an enhanced subtask, the current task inherits that enhancement strategy.

[0058] Based on the above task elements, in a feasible embodiment, the task type can be determined in the following ways: if the output mode is to return a list of target user IDs or a group identifier, then the task type is determined to be a target group selection type; if the output mode is to return a full supply of tag tables, or the target supply quantity is greater than the preset supply quantity, then the task type is a batch tag generation type; if the feature semantic dimension is a high abstract dimension and the supply is in a cold start state, then the task type is a tag dimension expansion type, otherwise it is a basic tag generation type.

[0059] Based on the determined task type, execute at least one enhanced reasoning subtask corresponding to each task type. For example, if the task type is target audience selection, then trigger the target audience selection subtask. This subtask, based on the semantic tags in the multi-dimensional reasoning output, performs the following steps: Based on the semantic tags in the multi-dimensional reasoning output, calculate historical supplies similar to the target supply in function, composition, or scenario through SQL query, large language model reasoning, and knowledge base retrieval; Based on the associated content of the target supply, identify similar content platforms with similar style or audience; Aggregate interactive users of similar supplies and active users of similar platforms to form an initial candidate target user pool; Perform a comprehensive score on the candidate users, such as scoring based on the matching degree between user profile and supply semantic tags, the interaction intensity between users and similar platforms, and the suitability of user basic attributes with supply; Sort by comprehensive score in descending order to generate target user group packages.

[0060] If the task type is batch label generation, a batch label generation subtask is triggered, which executes the main inference in parallel with the full supply or performs batch inference based on the Span SOP-built batch inference task. If the task type is label dimension expansion, a label dimension expansion subtask is triggered, which expands the dimensions of the labels by using the large model output or calls external knowledge bases to supplement evidence and generate more granular labels. If the task type is basic label generation, no enhanced inference subtasks are executed.

[0061] S108: The supply reasoning results are structured into digital supply characteristics and stored persistently.

[0062] To address the problems of chaotic output forms and lack of unified management in the current system, such as... Figure 4 As shown, in this embodiment of the application, after performing the supply understanding task and obtaining the supply reasoning result of the target supply, the supply reasoning result is structured into digital supply features and persistently stored.

[0063] In traditional recommendation systems, supply features are often stored in unstructured or semi-structured forms, lacking a unified representation. This causes the supply profile to lag behind the actual supply status. When facing dynamic supply understanding scenarios triggered by natural language demand, this easily leads to the waste of reusable supply features and prevents cross-scenario feature sharing. Specifically, in one or more embodiments of this specification, the supply reasoning results are structured into digital supply features and persistently stored, specifically including: The supply reasoning results are parsed and standardized to generate feature records corresponding to the target supply. For example, the supply reasoning results are parsed to obtain key-value pairs corresponding to each semantic dimension of the feature to be generated; the extracted values ​​are standardized, including but not limited to unit unification and format structuring. The standardized feature records corresponding to the target supply are then used. Based on the unified supply identifier of the target supply, existing feature records of the target supply are obtained. These feature records are compared semantically to achieve deduplication and fusion, resulting in a feature vector representation of the target supply. This feature vector representation is organized in key-value pairs, where the key is the semantic dimension of the feature, and the value is the standardized feature value and related metadata. This obtained feature vector representation is then persistently stored as a digital supply feature in a supply feature database. The digital supply features obtained through this process retain semantic interpretability and support efficient querying and cross-scenario reuse.

[0064] Furthermore, in one or more embodiments of this specification, the method further includes: like Figure 4As shown, by providing supply understanding services, digital supply characteristics are provided to recommendation scenarios, so as to generate supply profiles based on digital supply characteristics and distribute personalized recommendations.

[0065] Providing the supply reasoning results to the recommendation scenario includes: In response to a request from a recommendation scenario, return the corresponding supply inference result, or synchronize the supply inference result to the feature storage system on which the recommendation scenario depends.

[0066] Furthermore, based on the characteristics of digital supply, a supply profile is generated for personalized recommendation distribution, specifically including: Based on the characteristics of digital supply, construct or update the supply profile of the target supply; The supply profile and user profile are matched in multiple dimensions to determine the target user group corresponding to the supply profile, and the target users of the target user group are ranked based on the matching degree to obtain the recommended target users. The target supply corresponding to the supply profile is used as the recommended content to generate a personalized recommendation list; The personalized recommendation list is pushed to the user terminal corresponding to the target user, thus realizing the distribution of personalized recommendations.

[0067] The process of building or updating the supply profile includes: obtaining the currently generated digital supply characteristics, determining whether the target supply already has a historical supply profile, and if a historical supply profile exists, then incrementally or fully updating the supply profile based on the digital supply characteristics, that is, replacing the historical digital supply characteristics of the historical supply profile with the digital supply characteristics. If it is the first time to build, then the supply profile of the target supply is initialized with the current digital supply characteristics.

[0068] The process of selecting target users by performing multi-dimensional matching between supply profiles and user profiles can be achieved. This can be done by matching supply profiles and user profiles according to various semantic dimensions to be generated, obtaining target user groups, and then ranking target users in the target user group according to the matching degree to obtain recommended target users.

[0069] By using the target supply corresponding to the supply profile as recommended content, and sorting it based on relevance, diversity, or business strategies, a personalized recommendation list can be generated. This personalized recommendation list is then pushed to the user terminals corresponding to the target users via message queues or APIs. This process achieves supply and demand matching based on semantic understanding, which not only improves the relevance of recommendations and user experience, but also enables cold-start supply that originally lacked behavioral data to be effectively recalled and distributed based on its supply profile, significantly enhancing the personalized recommendation of the recommendation system.

[0070] Based on the same idea, one or more embodiments of this specification also provide apparatus and devices corresponding to the above methods, such as... Figure 7 , Figure 8 As shown.

[0071] Figure 7 This is a schematic diagram of the structure of a dynamic generation device for supply features provided in one or more embodiments of this specification. The device includes: The receiving module 702 is used to receive the requirement description of the requesting terminal based on the intelligent agent interaction interface; The task generation module 704 is used to parse the demand description to generate a supply understanding task; the supply understanding task is used to dynamically determine the target supply set related to the semantics of the demand description, as well as the feature semantic dimensions to be generated. The reasoning module 706 is used to execute the supply understanding task and obtain the supply reasoning result of the target supply; the execution of the supply understanding task includes: acquiring multi-dimensional feature data of each target supply in the target supply set; constructing reasoning prompts based on the feature semantic dimensions to be generated; and using a large language model to reason about the multi-dimensional feature data according to the reasoning prompts to generate the supply reasoning result. The storage module 708 is used to structure the supply reasoning results into digital supply characteristics and store them persistently.

[0072] Optionally, the inference module 706 specifically includes: For each target supply in the target supply set, obtain its corresponding global supply metadata, align the global supply metadata, and assign a unified supply identifier; The user behavior data of the demand terminal is bound to the unified supply identifier based on context information; Based on the unified supply identifier, the global supply metadata and the user behavior data are aggregated to generate multidimensional feature data of the target supply.

[0073] Optionally, the inference module 706 specifically includes: generating a basic prompt template based on the feature semantic dimension to be generated; the basic prompt template is used to describe the analysis task for the feature semantic dimension to the large language model; The output format requirements of the supply reasoning results are embedded into the basic prompt word template in the form of instructions to obtain the reasoning prompt framework; The multidimensional feature data of the target supply is combined with the reasoning hint framework to obtain reasoning hints.

[0074] Optionally, the reasoning module 706 specifically includes: using a large language model to reason about the multidimensional feature data based on the reasoning prompts, and generating the supply reasoning result, specifically including: The inference prompts are input into the large language model to obtain multi-dimensional inference outputs. Based on the task type corresponding to the supply understanding task, determine whether the supply inference engine should execute at least one enhanced inference subtask; If so, the supply reasoning result is obtained based on the multi-dimensional reasoning output and the output corresponding to the enhanced reasoning subtask; If not, the multi-dimensional reasoning output will be used as the supply reasoning result.

[0075] Optionally, the task type is determined based on at least one task element in the supply understanding task: the level of abstraction of the feature semantic dimension to be generated, the output mode specified by the supply understanding task, the size and status of the target supply set, or the execution mode of historically similar tasks. The enhanced reasoning subtasks include: expanding the label dimensions, generating batch labels, or selecting the target audience.

[0076] Optionally, the receiving module 702 specifically includes: receiving dialogue information uploaded by the receiving terminal based on the intelligent agent interaction interface; In response to the dialogue information, the terminal in demand is tracked for multiple rounds of dialogue status to obtain the current dialogue status when the dialogue ends or a reasoning instruction is triggered. Based on the current dialogue state, extract the structured requirement description of the requesting terminal.

[0077] Optionally, the receiving module 702 specifically includes: monitoring the interface behavior characteristics of the interactive interface corresponding to the intelligent agent's interactive interface during the dialogue; the interface behavior characteristics include: user focus sequence, control operation sequence, and operation timing characteristics; Obtain the semantic association between the interface behavior features and the dialogue information, and modify the dialogue information according to the semantic association to obtain the modified dialogue information; Based on the corrected dialogue information and the interface behavior characteristics, the dialogue state corresponding to the dialogue information is dynamically updated to obtain the current dialogue state.

[0078] Optionally, the task generation module 704 specifically includes: performing intent recognition and key semantic dimension extraction on the requirement description to obtain multi-level intents and key semantic dimensions; wherein, the multi-level intents include: user intent graphs and nested sub-intents; Based on the multi-level intent and the key semantic dimensions, a feasibility diagnosis is performed on the demand description to dynamically determine the target supply set and the feature semantic dimensions to be generated, thus generating a structured supply understanding task.

[0079] Optionally, the storage module specifically includes: parsing and standardizing the supply reasoning results to generate feature records corresponding to the target supply; The feature records are deduplicated and fused with the existing feature records of the target supply to obtain the feature vector representation of the target supply; The feature vector is represented as the digital supply feature and persistently stored in the supply feature library.

[0080] Optionally, the device further includes: a distribution module 710; The distribution module is used to provide the digital supply characteristics to the recommendation scenario through the supply understanding service, so as to generate a supply profile based on the digital supply characteristics and distribute personalized recommendations. The step of providing the supply inference results to the recommendation scenario includes: In response to the call request of the recommended scenario, the supply inference result is returned, or the supply inference result is synchronized to the feature storage system on which the recommended scenario depends.

[0081] Optionally, the distribution module 710 specifically includes: Based on the characteristics of digital supply, construct or update the supply profile of the target supply; Personalized recommendations are distributed through multi-dimensional matching of supply profiles and user profiles.

[0082] Optionally, the device further includes: a new processing module 712; The newly added processing module is used to obtain the global supply metadata of the new target supply when performing the supply understanding task if the target supply is a new target supply; wherein, the global supply metadata includes: structured attribute data and unstructured descriptive text; Based on the semantic dimension of the feature to be generated, a basic prompt template is generated, and the output format requirements of the inference result are embedded into the basic prompt word template in the form of instructions to obtain the inference prompt framework. The global supply metadata of the newly added target supply and the multi-level intent of the demand description are integrated and populated into the reasoning prompt framework to generate reasoning prompts.

[0083] Figure 8 A schematic diagram of a dynamic generation device for supply features, provided for one or more embodiments of this specification, the device comprising: At least one processor; and, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions executable by the at least one processor, which, when executed by the at least one processor, enable the at least one processor to: The requirements description of the terminal receiving the requirements based on the intelligent agent interaction interface; The demand description is parsed to generate a supply understanding task; the supply understanding task is used to dynamically determine the target supply set related to the semantics of the demand description, as well as the feature semantic dimensions to be generated. The supply understanding task is executed to obtain the supply reasoning result of the target supply; the execution of the supply understanding task includes: obtaining multi-dimensional feature data of each target supply in the target supply set; constructing reasoning prompts based on the feature semantic dimensions to be generated; and using a large language model, reasoning on the multi-dimensional feature data according to the reasoning prompts to generate the supply reasoning result. The supply reasoning results are structured into digital supply characteristics and stored persistently.

[0084] In the 1990s, improvements to a technology could be clearly distinguished as either hardware improvements (e.g., improvements to the circuit structure of diodes, transistors, switches, etc.) or software improvements (improvements to the methodology). However, with technological advancements, many methodological improvements today can be considered direct improvements to the hardware circuit structure. Designers almost always obtain the corresponding hardware circuit structure by programming the improved methodology into the hardware circuit. Therefore, it cannot be said that a methodological improvement cannot be implemented using hardware physical modules. For example, a Programmable Logic Device (PLD) (such as a Field Programmable Gate Array (FPGA)) is such an integrated circuit whose logic function is determined by the user programming the device. Designers can program and "integrate" a digital system onto a PLD themselves, without needing chip manufacturers to design and manufacture dedicated integrated circuit chips. Furthermore, nowadays, instead of manually manufacturing integrated circuit chips, this programming is mostly implemented using "logic compiler" software. Similar to the software compiler used in program development, the original code before compilation must also be written in a specific programming language, called a Hardware Description Language (HDL). There are many HDLs, such as ABEL (Advanced Boolean Expression Language), AHDL (Altera Hardware Description Language), Confluence, CUPL (Cornell University Programming Language), HDCal, JHDL (Java Hardware Description Language), Lava, Lola, MyHDL, PALASM, and RHDL (Ruby Hardware Description Language). Currently, VHDL (Very-High-Speed ​​Integrated Circuit Hardware Description Language) and Verilog are commonly used. Those skilled in the art should also understand that by simply performing some logic programming on the method flow using one of these hardware description languages ​​and programming it into an integrated circuit, the hardware circuit implementing the logical method flow can be easily obtained.

[0085] The controller can be implemented in any suitable manner. For example, it can take the form of a microprocessor or processor and a computer-readable medium storing computer-readable program code (e.g., software or firmware) executable by the (micro)processor, logic gates, switches, application-specific integrated circuits (ASICs), programmable logic controllers, and embedded microcontrollers. Examples of controllers include, but are not limited to, the following microcontrollers: ARC 625D, Atmel AT91SAM, Microchip PIC18F26K20, and Silicon Labs C8051F320. A memory controller can also be implemented as part of the control logic of the memory. Those skilled in the art will also recognize that, in addition to implementing the controller in purely computer-readable program code form, the same functionality can be achieved by logically programming the method steps to make the controller take the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers. Therefore, such a controller can be considered a hardware component, and the means included therein for implementing various functions can also be considered as structures within the hardware component. Alternatively, the means for implementing various functions can be considered as both software modules implementing the method and structures within the hardware component.

[0086] The systems, devices, modules, or units described in the above embodiments can be implemented by computer chips or entities, or by products with certain functions. A typical implementation device is a computer. Specifically, a computer can be, for example, a personal computer, laptop computer, cellular phone, camera phone, smartphone, personal digital assistant, media player, navigation device, email device, game console, tablet computer, wearable device, or any combination of these devices.

[0087] For ease of description, the above devices are described in terms of function, divided into various units. Of course, in implementing this specification, the functions of each unit can be implemented in one or more software and / or hardware.

[0088] Those skilled in the art will understand that the embodiments of this specification can be provided as methods, systems, or computer program products. Therefore, the embodiments of this specification can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the embodiments of this specification can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0089] This specification is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this specification. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0090] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0091] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0092] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0093] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0094] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information by any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0095] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0096] This specification can be described in the general context of computer-executable instructions that are executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform a specific task or implement a specific abstract data type. This specification can also be practiced in distributed computing environments, where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0097] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the embodiments of apparatus, devices, and non-volatile computer storage media are basically similar to the method embodiments, and therefore described more simply; relevant parts can be referred to the descriptions of the method embodiments.

[0098] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0099] The above description is merely one or more embodiments of this specification and is not intended to limit this specification. Various modifications and variations can be made to the one or more embodiments of this specification by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of one or more embodiments of this specification should be included within the scope of the claims of this specification.

Claims

1. A method for dynamically generating supply characteristics, the method comprising: The requirements description of the terminal receiving the requirements based on the intelligent agent interaction interface; Parse the demand description to generate a supply understanding task; The supply understanding task is used to dynamically determine the target supply set related to the semantic description of the demand, as well as the feature semantic dimension to be generated; The supply understanding task is executed to obtain the supply reasoning result of the target supply; the execution of the supply understanding task includes: obtaining multi-dimensional feature data of each target supply in the target supply set; constructing reasoning prompts based on the feature semantic dimensions to be generated; and using a large language model, reasoning on the multi-dimensional feature data according to the reasoning prompts to generate the supply reasoning result. The supply reasoning results are structured into digital supply characteristics and stored persistently.

2. The method as described in claim 1, specifically including obtaining multi-dimensional feature data of each target supply in the target supply set, includes: For each target supply in the target supply set, obtain its corresponding global supply metadata, align the global supply metadata, and assign a unified supply identifier; The user behavior data of the demand terminal is bound to the unified supply identifier based on context information; Based on the unified supply identifier, the global supply metadata and the user behavior data are aggregated to generate multidimensional feature data of the target supply.

3. The method as described in claim 1, wherein, based on the semantic dimension of the feature to be generated, a reasoning prompt is constructed, specifically including: Based on the semantic dimensions of the features to be generated, a basic prompt template is generated; The basic prompt template is used to describe the analysis task for the feature semantic dimension to the large language model; The output format requirements of the supply reasoning results are embedded into the basic prompt word template in the form of instructions to obtain the reasoning prompt framework; The multidimensional feature data of the target supply is combined with the reasoning hint framework to obtain reasoning hints.

4. The method as described in claim 3, wherein the inference results are generated by reasoning about the multidimensional feature data based on the inference prompts using a large language model, specifically including: The inference prompts are input into the large language model to obtain multi-dimensional inference outputs. Based on the task type corresponding to the supply understanding task, determine whether the supply inference engine should execute at least one enhanced inference subtask; If so, the supply reasoning result is obtained based on the multi-dimensional reasoning output and the output corresponding to the enhanced reasoning subtask; If not, the multi-dimensional reasoning output will be used as the supply reasoning result.

5. The method of claim 4, wherein the task type is determined based on at least one task element in the supply understanding task: the level of abstraction of the feature semantic dimension to be generated, the output mode specified by the supply understanding task, the size and status of the target supply set, or the execution mode of historically similar tasks; The enhanced reasoning subtasks include: Expanding tag dimensions, generating tags in batches, or selecting target audiences.

6. The method as described in claim 1, wherein receiving the demand description from the demand terminal based on the intelligent agent interaction interface specifically includes: Based on the intelligent agent interaction interface, the dialogue information uploaded by the receiving terminal is received; In response to the dialogue information, the terminal in demand is tracked for multiple rounds of dialogue status to obtain the current dialogue status when the dialogue ends or a reasoning instruction is triggered. Based on the current dialogue state, extract the structured requirement description of the requesting terminal.

7. The method as described in claim 6, wherein multi-turn dialogue state tracking is performed on the requesting terminal to obtain the current dialogue state when the dialogue ends or a reasoning instruction is triggered, specifically includes: Monitor the interface behavior characteristics of the interactive interface corresponding to the intelligent agent's interaction interface during the dialogue. The interface behavior features include: user focus sequence, control operation sequence, and operation timing features; Obtain the semantic association between the interface behavior features and the dialogue information, and modify the dialogue information according to the semantic association to obtain the modified dialogue information; Based on the corrected dialogue information and the interface behavior characteristics, the dialogue state corresponding to the dialogue information is dynamically updated to obtain the current dialogue state.

8. The method as described in claim 1, wherein parsing the demand description to generate a supply understanding task, specifically includes: The requirement description is subjected to intent recognition and key semantic dimension extraction to obtain multi-level intents and key semantic dimensions; wherein, the multi-level intents include: user intent graphs and nested sub-intents; Based on the multi-level intent and the key semantic dimensions, a feasibility diagnosis is performed on the demand description to dynamically determine the target supply set and the feature semantic dimensions to be generated, thus generating a structured supply understanding task.

9. The method as described in claim 1, wherein the supply reasoning result is structured into digital supply characteristics and persistently stored, specifically includes: The supply reasoning results are analyzed and standardized to generate feature records corresponding to the target supply; The feature records are deduplicated and fused with the existing feature records of the target supply to obtain the feature vector representation of the target supply; The feature vector is represented as the digital supply feature and persistently stored in the supply feature library.

10. The method of claim 1, further comprising: Through supply understanding services, the digital supply characteristics are provided to recommendation scenarios to generate supply profiles based on the digital supply characteristics for personalized recommendation distribution. The step of providing the supply inference results to the recommendation scenario includes: In response to the call request of the recommended scenario, the supply inference result is returned, or the supply inference result is synchronized to the feature storage system on which the recommended scenario depends.

11. The method as described in claim 10, wherein generating a supply profile based on the digital supply characteristics for personalized recommendation distribution specifically includes: Based on the characteristics of digital supply, construct or update the supply profile of the target supply; Personalized recommendations are distributed through multi-dimensional matching of supply profiles and user profiles.

12. The method of claim 8, further comprising: If the target supply is a new target supply, then when performing the supply understanding task, the global supply metadata of the new target supply is obtained; wherein, the global supply metadata includes: structured attribute data and unstructured descriptive text; Based on the semantic dimension of the feature to be generated, a basic prompt template is generated, and the output format requirements of the inference result are embedded into the basic prompt word template in the form of instructions to obtain the inference prompt framework. The global supply metadata of the newly added target supply and the multi-level intent of the demand description are integrated and populated into the reasoning prompt framework to generate reasoning prompts.

13. A dynamic generation device for supply characteristics, the device comprising: The receiving module is used to receive the requirement description from the requesting terminal based on the intelligent agent interaction interface; The task generation module is used to parse the demand description to generate supply understanding tasks; The supply understanding task is used to dynamically determine the target supply set related to the semantic description of the demand, as well as the feature semantic dimension to be generated; The reasoning module is used to execute the supply understanding task and obtain the supply reasoning result of the target supply. The execution of the supply understanding task includes: acquiring multi-dimensional feature data of each target supply in the target supply set; constructing reasoning prompts based on the feature semantic dimensions to be generated; and using a large language model to reason about the multi-dimensional feature data according to the reasoning prompts to generate the supply reasoning result. The storage module is used to structure the supply reasoning results into digital supply characteristics and store them persistently.

14. The apparatus of claim 13, wherein the inference module specifically comprises: For each target supply in the target supply set, obtain its corresponding global supply metadata, align the global supply metadata, and assign a unified supply identifier; The user behavior data of the demand terminal is bound to the unified supply identifier based on context information; Based on the unified supply identifier, the global supply metadata and the user behavior data are aggregated to generate multidimensional feature data of the target supply.

15. The apparatus of claim 13, wherein the inference module specifically comprises: Based on the semantic dimensions of the features to be generated, a basic prompt template is generated; The basic prompt template is used to describe the analysis task for the feature semantic dimension to the large language model; The output format requirements of the supply reasoning results are embedded into the basic prompt word template in the form of instructions to obtain the reasoning prompt framework; The multidimensional feature data of the target supply is combined with the reasoning hint framework to obtain reasoning hints.

16. The apparatus of claim 15, wherein the inference module specifically comprises: Using a large language model, inference is performed on the multidimensional feature data based on the inference prompts to generate the supply inference result, specifically including: The inference prompts are input into the large language model to obtain multi-dimensional inference outputs. Based on the task type corresponding to the supply understanding task, determine whether the supply inference engine should execute at least one enhanced inference subtask; If so, the supply reasoning result is obtained based on the multi-dimensional reasoning output and the output corresponding to the enhanced reasoning subtask; If not, the multi-dimensional reasoning output will be used as the supply reasoning result.

17. The apparatus of claim 16, wherein the task type is determined based on at least one task element in the supply understanding task: the level of abstraction of the feature semantic dimension to be generated, the output mode specified by the supply understanding task, the size and status of the target supply set, or the execution mode of historically similar tasks. The enhanced reasoning subtasks include: Expanding tag dimensions, generating tags in batches, or selecting target audiences.

18. The apparatus of claim 13, wherein the receiving module specifically comprises: Based on the intelligent agent interaction interface, the dialogue information uploaded by the receiving terminal is received; In response to the dialogue information, the terminal in demand is tracked for multiple rounds of dialogue status to obtain the current dialogue status when the dialogue ends or a reasoning instruction is triggered. Based on the current dialogue state, extract the structured requirement description of the requesting terminal.

19. The apparatus of claim 18, wherein the receiving module specifically comprises: Monitor the interface behavior characteristics of the interactive interface corresponding to the intelligent agent's interaction interface during the dialogue. The interface behavior features include: user focus sequence, control operation sequence, and operation timing features; Obtain the semantic association between the interface behavior features and the dialogue information, and modify the dialogue information according to the semantic association to obtain the modified dialogue information; Based on the corrected dialogue information and the interface behavior characteristics, the dialogue state corresponding to the dialogue information is dynamically updated to obtain the current dialogue state.

20. The apparatus of claim 13, wherein the task generation module specifically comprises: The requirement description is subjected to intent recognition and key semantic dimension extraction to obtain multi-level intents and key semantic dimensions; wherein, the multi-level intents include: user intent graphs and nested sub-intents; Based on the multi-level intent and the key semantic dimensions, a feasibility diagnosis is performed on the demand description to dynamically determine the target supply set and the feature semantic dimensions to be generated, thus generating a structured supply understanding task.

21. The apparatus of claim 13, wherein the storage module specifically comprises: The supply reasoning results are analyzed and standardized to generate feature records corresponding to the target supply; The feature records are deduplicated and fused with the existing feature records of the target supply to obtain the feature vector representation of the target supply; The feature vector is represented as the digital supply feature and persistently stored in the supply feature library.

22. The apparatus of claim 13, further comprising: Distribution module; The distribution module is used to provide the digital supply characteristics to the recommendation scenario through the supply understanding service, so as to generate a supply profile based on the digital supply characteristics and distribute personalized recommendations. The step of providing the supply inference results to the recommendation scenario includes: In response to the call request of the recommended scenario, the supply inference result is returned, or the supply inference result is synchronized to the feature storage system on which the recommended scenario depends.

23. The apparatus of claim 22, wherein the distribution module specifically comprises: Based on the characteristics of digital supply, construct or update the supply profile of the target supply; Personalized recommendations are distributed through multi-dimensional matching of supply profiles and user profiles.

24. The apparatus of claim 20, further comprising: Add a new processing module; The newly added processing module is used to obtain the global supply metadata of the new target supply when performing the supply understanding task if the target supply is a new target supply; wherein, the global supply metadata includes: structured attribute data and unstructured descriptive text; Based on the semantic dimension of the feature to be generated, a basic prompt template is generated, and the output format requirements of the inference result are embedded into the basic prompt word template in the form of instructions to obtain the inference prompt framework. The global supply metadata of the newly added target supply and the multi-level intent of the demand description are integrated and populated into the reasoning prompt framework to generate reasoning prompts.

25. A device for dynamically generating supply characteristics, the device comprising: At least one processor; as well as, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions executable by the at least one processor, which, when executed by the at least one processor, enable the at least one processor to: The requirements description of the terminal receiving the requirements based on the intelligent agent interaction interface; The demand description is parsed to generate a supply understanding task; the supply understanding task is used to dynamically determine the target supply set related to the semantics of the demand description, as well as the feature semantic dimensions to be generated. The supply understanding task is executed to obtain the supply reasoning result of the target supply; the execution of the supply understanding task includes: obtaining multi-dimensional feature data of each target supply in the target supply set; constructing reasoning prompts based on the feature semantic dimensions to be generated; and using a large language model, reasoning on the multi-dimensional feature data according to the reasoning prompts to generate the supply reasoning result. The supply reasoning results are structured into digital supply characteristics and stored persistently.

Images