Content query method and device, electronic equipment and readable storage medium

Abstract

Embodiments of the present application provide a content query method and device, electronic equipment and readable storage medium, the method comprises: receiving the user end for the initiated audit query instruction of the audit object; send the first prompt word to the analysis intelligent agent, so that the analysis intelligent agent analyzes the audit query instruction according to the first prompt word, and obtains the task execution scheme corresponding to the audit query instruction;The first prompt word comprises: first input information, first output information and analysis rule;The first input information comprises: the model information corresponding to the audit query instruction and a plurality of field models respectively;The first output information is used to represent the information of the task execution scheme;The analysis rule is used to represent the generation logic from the audit query instruction to the task execution scheme;The corresponding subtask is executed by using the target field model, and the execution result data corresponding to each subtask is obtained;According to the execution result data corresponding to a plurality of subtasks respectively, the audit result for the audit object is generated.

Description

Technical Field

[0001] This invention relates to the field of information processing technology, and in particular to a content retrieval method, apparatus, electronic device, and readable storage medium. Background Technology

[0002] With the explosive growth of the digital content industry and the increasing diversification of information dissemination methods, content moderation systems face enormous challenges in processing massive amounts of multimodal data.

[0003] Data from various fields, including videos, live streams, bullet comments, and online comments, is emerging in a complex and ever-changing manner. This places extremely high demands on the real-time performance, accuracy, and comprehensive analytical capabilities of content moderation. Traditional single-mode moderation models or static rule bases are no longer sufficient to handle such dynamic and cross-modal complex scenarios.

[0004] Therefore, there is a need for a way to achieve efficient content retrieval. Summary of the Invention

[0005] The purpose of this invention is to provide a content query method, apparatus, electronic device, and readable storage medium, capable of achieving efficient content querying. The specific technical solution is as follows: In a first aspect of this invention, a content query method is provided, the method comprising: Receive audit query commands from users for the audited object; Generate the first prompt word; A first prompt word is sent to the analysis agent, enabling the analysis agent to analyze the review query instruction based on the first prompt word and obtain a task execution plan corresponding to the review query instruction. The task execution plan includes: multiple sub-tasks and a target domain model corresponding to each sub-task. The first prompt word includes: first input information, first output information, and analysis rules. The first input information includes: model information corresponding to the review query instruction and multiple domain models respectively. The first output information is used to characterize the information of the task execution plan. The analysis rules are used to characterize the generation logic from the review query instruction to the task execution plan, determine the query intent corresponding to the review query instruction, generate multiple sub-tasks corresponding to the review query instruction based on the query intent, and determine the target domain model corresponding to each sub-task from multiple domain models based on the model information corresponding to each domain model. The target domain model is used to execute the corresponding sub-tasks, and the execution result data for each sub-task is obtained. Based on the execution result data corresponding to each of the multiple sub-tasks, an audit result is generated for the audit object.

[0006] In a second aspect of the invention, a content query device is also provided, the device comprising: The receiving module is used to receive audit query instructions initiated by the user client for the audit object; The prompt word generation module is used to generate the first prompt word; An analysis module is used to send a first prompt word to an analysis agent, so that the analysis agent analyzes the review query instruction based on the first prompt word to obtain a task execution plan corresponding to the review query instruction. The task execution plan includes: multiple sub-tasks and a target domain model corresponding to each sub-task. The first prompt word includes: first input information, first output information, and analysis rules. The first input information includes: model information corresponding to the review query instruction and multiple domain models respectively. The first output information is used to characterize the information of the task execution plan. The analysis rules are used to characterize the generation logic from the review query instruction to the task execution plan, determine the query intent corresponding to the review query instruction, generate multiple sub-tasks corresponding to the review query instruction based on the query intent, and determine the target domain model corresponding to each sub-task from multiple domain models based on the model information corresponding to multiple domain models respectively. The task execution module is used to execute corresponding sub-tasks using the target domain model and obtain execution result data for each sub-task. The result generation module is used to generate an audit result for the audit object based on the execution result data corresponding to multiple sub-tasks.

[0007] In another aspect of the present invention, a computer-readable storage medium is also provided, wherein instructions are stored therein, which, when executed on a computer, cause the computer to perform any of the data forwarding methods described above.

[0008] In another aspect of the present invention, a computer program product containing instructions is also provided, which, when run on a computer, causes the computer to execute any of the content query methods described above.

[0009] In this embodiment of the invention, in response to a user's audit query command for an audit object, the intelligent agent analyzes the audit query command to obtain a corresponding task execution plan. The task execution plan includes a target domain model corresponding to each sub-task, which can clearly identify which professional target domain model will execute each sub-task, eliminating the need for cumbersome process design and resource allocation between multiple interfaces.

[0010] Furthermore, according to the task execution plan, the embodiments of the present invention drive the target domain model to execute its corresponding sub-tasks and obtain the execution result data corresponding to each sub-task; thus enabling the multi-domain analysis tasks that originally needed to be executed serially or manually alternately to be executed in parallel, thereby shortening the overall time from planning to obtaining professional analysis results of each dimension and improving the efficiency of the task execution stage.

[0011] Furthermore, this embodiment of the invention outputs the review results to the user terminal based on the execution result data corresponding to multiple sub-tasks. Therefore, this embodiment of the invention can achieve efficient content retrieval while improving processing efficiency and the completeness of analysis dimensions. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0013] Figure 1 This is a flowchart of a content query method provided in an embodiment of the present invention; Figure 2 This is a flowchart of a content query method provided in an embodiment of the present invention; Figure 3 This is a flowchart of a content query method provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the structure of a content query device provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of the hardware structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0014] The technical solutions of the present invention will now be described with reference to the accompanying drawings in the embodiments of the present invention.

[0015] For example, the content query method provided in the embodiments of the present invention can be applied to at least the following application scenarios, which will be described below.

[0016] Real-time risk control scenario for live streaming content: Platform security auditors face an active live streaming room with frequent audience interaction and need to quickly assess its overall compliance and risk level. Traditionally, auditors need to access: a video stream monitoring system to view real-time footage; an audio-to-text and keyword alert system to monitor the streamer's statements; the backend of the bullet screen filtering system to view real-time interactive content; and the streamer's historical violation database. Auditors must manually switch between these multiple independent system interfaces, visually compare alert information from different sources, and piece together an overall risk profile based on personal experience. This process is not only slow but also prone to overlooking crucial information when handling concurrent alerts from multiple live streaming rooms, leading to misjudgments or delayed responses.

[0017] Content security analysis of trending short videos: Operations or security teams need to conduct a comprehensive impact assessment of a suddenly viral short video to determine whether it poses potential content risks or has positive value worth promoting. In the traditional process, analysts need to: review the video's content; separately examine the sentiment and keywords in the massive and rapidly growing comment section; retrieve historical reports or debunking information related to the video's theme or background; and observe its dissemination data on other social media platforms. These analytical actions are scattered across different data analysis tools, content management systems, and third-party monitoring platforms, requiring analysts to integrate multiple reports, which is time-consuming and labor-intensive.

[0018] Based on the above application scenarios, the content query method provided by the embodiments of the present invention will be described below.

[0019] Figure 1 A flowchart of a content query method provided in an embodiment of the present invention.

[0020] like Figure 1 As shown, the content query method may include steps 110-150. Optionally, this method may be executed by a central decision engine, which connects to multiple domain models. For example, the central decision engine interacts with multiple domain models deployed on physical servers or independent container instances via network interfaces through remote procedure calls or application programming interface calls. These multiple domain models may reside on the same or different physical servers or independent container instances.

[0021] The method specifically includes the following steps: Step 110: Receive the audit query instruction initiated by the user terminal for the audit object; Step 120: Generate the first prompt word; Step 130: Send a first prompt word to the analysis agent so that the analysis agent can analyze the review query instruction based on the first prompt word to obtain a task execution plan corresponding to the review query instruction; the task execution plan includes: multiple sub-tasks and a target domain model corresponding to each sub-task; the first prompt word includes: first input information, first output information and analysis rules; the first input information includes: model information corresponding to the review query instruction and multiple domain models respectively; the first output information is used to characterize the information of the task execution plan; the analysis rules are used to characterize the generation logic from the review query instruction to the task execution plan, determine the query intent corresponding to the review query instruction, generate multiple sub-tasks corresponding to the review query instruction based on the query intent, and determine the target domain model corresponding to each sub-task from multiple domain models based on the model information corresponding to each domain model respectively; Step 140: Execute the corresponding sub-tasks using the target domain model to obtain the execution result data for each sub-task; Step 150: Generate an audit result for the audit object based on the execution result data corresponding to the multiple sub-tasks.

[0022] The central decision engine is software deployed on a server, which can contain several intelligent agents for understanding complex instructions and planning tasks. An agent is an intelligent proxy system with a large language model as its core controller. Essentially, it's an agent system that controls the large language model to solve user problems, relying on the large language model as its core decision-making and processing unit, and utilizing the logical reasoning capabilities of the large language model to solve user problems. In practical applications, an intelligent agent can specifically be an autonomous software or hardware entity. The intelligent agents in this embodiment of the invention can include: analytical intelligent agents and result aggregation intelligent agents, etc.

[0023] Domain models are independent artificial intelligence models that perform automated analysis and processing of content in a specific domain, such as models that specifically analyze video footage, text content, or audio content.

[0024] An audit query command is a task request submitted by the user and described in natural language. It typically involves a multi-dimensional analysis of the audit target. The audit target refers to the digital content entity that needs to be inspected, such as a live stream, a video, or an article.

[0025] A subtask is a structured task description element extracted by the analytical agent after parsing the audit query command. A task execution plan is a set of instructions generated by the analytical agent, containing a series of ordered or parallel action sequences.

[0026] The execution results data for multiple subtasks can include: results returned by different target domain models after executing the subtasks, which differ in data format and semantic level. For example, target domain model A returns structured labels, while target domain model B returns probability scores. The review result can be a comprehensive conclusion document that integrates all execution result data and is ultimately presented to the user.

[0027] In step 110, the user initiates an audit query command for a specific audit object.

[0028] In steps 120 to 130, after receiving the instruction from the audit object, the central decision engine can use its internal analytical agent to analyze the instruction and obtain the task execution plan corresponding to the instruction.

[0029] The analytical agent, based on its pre-trained world knowledge and logical reasoning capabilities, understands the true intent behind the instructions given by the auditing entity. It decomposes and transforms the natural language requests corresponding to these instructions into a series of specific, actionable sub-tasks, extracting key task information for each sub-task. Leveraging the powerful contextual understanding and instruction-following capabilities of large language models, the analytical agent transforms unstructured user needs into structured, machine-executable task descriptions (i.e., task execution plans). The task information for each sub-task can include: task intent, domain, and task type.

[0030] For example, a user's input instruction for reviewing a live stream might be "Assess whether there are any violations in this live stream." The analytical agent, through parsing, can understand that "assessment" is a complex action that needs to be broken down into multiple sub-tasks such as "analyzing the live stream feed," "monitoring current audio content," and "scanning real-time comments." The task information extracted for the sub-task of "analyzing the live stream feed" could include: task intent as "identifying inappropriate visual elements," domain as "video," and task type as "real-time stream analysis."

[0031] For example, for the instruction to "summarize the main viewpoints of all comments on a certain hot topic last week," the parsed subtask information can include: task intent as "extract viewpoints and cluster them," domain as "text," and task type as "batch data processing." The analysis of the intelligent agent's instruction to the user achieves a mapping from the user's subjective, holistic needs to objective, discrete machine tasks, solving the problem of users having to manually break down and operate multiple tasks in traditional interfaces.

[0032] In a specific implementation, a first prompt word can be sent to the analysis agent so that the analysis agent can analyze the audit query instruction based on the first prompt word and obtain the task execution plan corresponding to the audit query instruction.

[0033] In this embodiment of the invention, prompt words are text instructions input by the user into the large language model, used to clarify the task objectives and output constraints, and serving as a medium to guide the model in generating content that meets the requirements. Their function is to provide the large language model with task direction, such as text generation, question answering, and information processing, while simultaneously limiting elements such as output style, format, and length, so that the output of the large language model matches the user's expectations.

[0034] The first prompt includes: first input information, first output information, and analysis rules; the first input information includes: the audit query instruction and model information corresponding to the various domain models respectively; the first output information is used to characterize the information of the task execution plan; the analysis rules are used to characterize the generation logic from the audit query instruction to the task execution plan, determine the query intent corresponding to the audit query instruction, generate multiple sub-tasks corresponding to the audit query instruction according to the query intent, and determine the target domain model corresponding to each sub-task from the various domain models according to the model information corresponding to the various domain models respectively.

[0035] As input instructions to the analytical agent, the first prompt word contains information such as the first input information, the first output information, and analysis rules. Based on the logical reasoning and instruction-following capabilities of its pre-trained model, the analytical agent parses the audit query instruction contained in the first input information according to the analysis rules in the first prompt word to obtain the query intent. Then, based on this query intent, it generates multiple sub-tasks. Next, according to the domain model list in the first input information, it selects the corresponding target domain model for each sub-task. Finally, according to the data format specified in the first output information, it outputs a structured task execution plan.

[0036] Therefore, the process of generating the first prompt word specifically includes: obtaining the audit query instruction and model information of multiple domain models, and using them as the first input information; obtaining the first output information representing the task execution plan; obtaining the analysis rules from determining the query intent, generating sub-tasks to determining the target domain model corresponding to each sub-task; and integrating the first input information, the first output information and the analysis rules to obtain the first prompt word.

[0037] A pre-trained model refers to a basic large language model that has been pre-trained on large-scale data and possesses general language understanding, semantic parsing, and instruction processing capabilities. The training data for a pre-trained model can include unlabeled or labeled data, and the training methods for a pre-trained model can include self-supervised learning and other methods.

[0038] In one implementation, the model information may include domain information. The process of determining the target domain model for each sub-task from multiple domain models based on the model information corresponding to multiple domain models specifically includes: matching the domain information of the domain model with the domain of the sub-task, and determining the successfully matched domain model as the target domain model.

[0039] For example, for the subtask of "identifying illegal visual elements" in the "video" domain, "Video Content Moderation Model V1.0" can be selected as the target domain model based on the matching between the domain information of the domain model and the domain of the subtask. Similarly, for the subtask of "monitoring current audio content" in the "audio" domain, "Audio Illegal Word Detection Model A2.0" can be selected as the target domain model based on the matching between the domain information of the domain model and the domain of the subtask.

[0040] In another implementation, the model information may include: the number of model parameters, which is used to characterize the total number of weights and biases adjusted during the training of the corresponding domain model; the number of model parameters is used to characterize the scale and structural complexity of the model; for example, a model with 8 billion parameters contains 8 billion weights and biases obtained through training.

[0041] In this context, "weight" refers to the connection strength coefficient between two neurons in the neural network corresponding to the domain model, which is used to perform linear transformation on the input signal during forward propagation.

[0042] Bias can refer to the activation threshold of a neuron in the neural network corresponding to the domain model. It is used to add a constant offset to the weighted sum to adjust the activation reference point of the neuron.

[0043] The process of determining the target domain model for each sub-task from multiple domain models based on the model information corresponding to each domain model specifically includes: determining the target domain model for each sub-task from multiple domain models based on the amount of data of the audit object and the number of model parameters of each domain model.

[0044] By matching the amount of data of the audit object with the number of parameters of the domain model, a reasonable match between task load and model computing power can be achieved. This can avoid resource redundancy and computing power waste of large parameter models, and prevent insufficient capabilities and unsatisfactory processing results of small parameter models, thereby improving the rationality of task scheduling and resource utilization efficiency.

[0045] Specifically, if the amount of data of the audited object exceeds a first threshold, the domain model with a model parameter count exceeding a second threshold will be determined as the target domain model for each sub-task; or If the amount of data of the audit object is less than or equal to the first threshold, the domain model with the number of model parameters less than or equal to the second threshold is determined as the target domain model for each subtask.

[0046] The first and second thresholds can be determined by those skilled in the art based on actual application needs. For example, when the object of review is video, the unit of data volume for the object of review can be video duration. For example, if the video duration exceeds the first threshold of 10 minutes, a domain model with more than the second threshold can be used.

[0047] The above-mentioned linkage matching mechanism using dual thresholds (first threshold and second threshold) can adaptively adapt to domain models with different parameter amounts according to the size of the data volume of the audit object; it can control the processing capacity and effect in large data volume audit scenarios, and save computing resources in small data volume scenarios, taking into account both task processing reliability and resource utilization economy.

[0048] In another implementation, embodiments of the present invention can determine the target domain model corresponding to a subtask based on the domain in the task information of the subtask and the amount of data of the audit object. Specifically, candidate domain models can be determined first based on the domain in the task information, and then the final target domain model can be determined from the candidate domain models based on the matching relationship between the amount of data of the audit object and the model parameter amounts of the candidate domain models.

[0049] This invention allows for the selective use of domain models between large-scale and small-scale language model architectures. Large-scale language model architectures refer to those with more than a second threshold of parameters, possessing strong semantic understanding and deep reasoning capabilities, and are suitable for complex tasks. Small-scale language model architectures, on the other hand, refer to those with fewer than or equal to the second threshold of parameters, exhibiting high inference speed and high throughput, and are suitable for high-concurrency, pattern-based tasks.

[0050] In summary, this invention provides a model scheduling strategy based on matching the data volume of the review object with the model's capabilities. This strategy balances processing power and response efficiency according to the data volume of the review object, thereby allocating appropriate model computing resources to sub-tasks. For review tasks with data volumes exceeding a first threshold, domain models with large language model architectures are prioritized. This is because such tasks typically involve in-depth analysis and correlation of large amounts of information, requiring models with strong semantic understanding and complex reasoning capabilities. For example, when identifying multimodal violations from long video streams or performing fine-grained opinion mining from massive amounts of text comments, large language models are suitable for processing.

[0051] For review tasks where the amount of data to be reviewed does not exceed the first threshold, a domain model with a small language model architecture will be invoked. Because these tasks process limited data, a small language model architecture is sufficient to meet their computational needs. Furthermore, its lightweight nature ensures millisecond-level response times and high concurrency processing capabilities, effectively handling scenarios with high real-time requirements. For example, tasks such as detecting violations in short video clips or filtering keywords from small amounts of text are well-suited for rapid execution by an efficient small model.

[0052] This model scheduling strategy enables the use of computationally intensive large-scale language model architectures for big data processing scenarios, while offloading a large number of small-scale tasks to lightweight small-scale language model architectures. This optimizes the overall resource utilization of the system, improves response efficiency and system throughput while ensuring processing effectiveness, and achieves a balance between service performance and operating costs.

[0053] In this embodiment of the invention, optionally, the task execution scheme may further include: the execution sequence of multiple subtasks; the analysis rule is further used to determine the execution sequence of multiple subtasks based on the dependency information between multiple subtasks and / or the priority information of each subtask.

[0054] Among them, dependency information is used to describe the execution order constraints or data flow relationships between different subtasks. For example, a subtask needs to wait for another subtask to output specific data as its input, or two subtasks cannot be executed simultaneously because they access the same protected resource.

[0055] Priority information is used to characterize the urgency and scheduling order of each subtask. For example, for a complex query that includes both "real-time live stream risk monitoring" and "historical violation record statistics", the priority of real-time monitoring will be set higher because real-time monitoring is sensitive to latency.

[0056] In determining the execution sequence, execution logic can be prioritized based on the dependencies between subtasks. Subtasks without strong dependencies can then be scheduled and ordered according to their own priorities. Subtasks that are independent of each other and have no sequential constraints can be executed in parallel, while subtasks with prerequisite dependencies are executed sequentially. High-priority subtasks can occupy computing resources first in the same batch of scheduling. This approach not only meets the logical relationships and data flow requirements between subtasks but also takes into account the timeliness requirements of business scenarios, rationally arranging the overall task scheduling order and optimizing the overall processing latency and resource scheduling rationality.

[0057] By combining subtask dependencies and priority information to determine the execution sequence, logically related subtasks can be advanced in an orderly manner. At the same time, the task scheduling order can be reasonably arranged according to priority, reducing task execution disorder and logical blockage, which helps to improve the orderliness of the overall task flow, the rationality of scheduling, and the overall processing efficiency.

[0058] It should be noted that the domain model in this embodiment of the invention may include, but is not limited to, a video moderation model for identifying sensitive objects in a video, and a text moderation model for detecting offensive words in text. The generated task execution scheme ensures that each subtask has a clearly defined responsibility model, laying the foundation for subsequent automated execution.

[0059] In one possible embodiment, the domain model includes at least two of the following: Video understanding models for understanding video content; A video moderation model used for reviewing video content; Audio recognition models used to identify audio content; An audio moderation model for reviewing audio content; A text moderation model for reviewing text content.

[0060] Domain models, as specialized components for performing specific content analysis tasks, encompass specialized analytical capabilities for different modalities of content.

[0061] The video understanding model is used to understand video content, and its output includes: video scene type, main subject in the frame, characters' actions, video content theme, and key plot information. The output of the video understanding model is typically the input of the video review model. For example, if the video understanding model identifies a video scene as an "outdoor gathering scene" and the main subject in the frame as containing an "unidentified liquid container," this output information will be directly fed into the video review model, which will then further determine whether the "unidentified liquid container" is a prohibited item and whether there are any violations in the "outdoor gathering scene."

[0062] A video review model is an AI model specifically designed to process continuous image frame sequences to identify visual elements, scenes, human actions, or specific objects contained within them. For example, it can be used to detect prohibited items or identify specific scenes. The output of a video review model includes: violation element identification results, violation content types, and violation occurrence time periods. Violation element identification results are the specific content such as people, objects, and behaviors that violate the rules, as identified by the model from the video footage. For example, a person holding "tool A" appears in the video. Violation content types are the categories into which these violation elements are classified, such as classifying "a person holding tool A" as "displaying dangerous items." Violation occurrence time periods mark the start and end times of these violation elements in the video, such as tool A appearing at 2 minutes and 10 seconds and disappearing at 2 minutes and 20 seconds.

[0063] Audio recognition models are used to identify audio content, and their outputs include: speech-to-text transcription, voice segmentation, ambient sound type, and sound source feature information. The outputs of audio recognition models are typically inputs to audio moderation models. For example, an audio recognition model might transcribe speech from a video into text containing "illegal inflammatory statements" or identify "abusive language" at a certain time. This output information is then fed into an audio moderation model, which further determines whether the transcribed text violates regulations and whether the abusive language constitutes illegal audio content.

[0064] An audio moderation model is an AI (Artificial Intelligence) model specifically designed to process sound signals to identify the speech content, specific sound effects, background noise, or analyze sound attributes. For example, it can be used to convert speech to text for keyword screening or to directly identify abnormal sound effects. Its output includes: audio violation keywords, violation sound effect types, and violation audio time positions. Specifically, audio violation keywords refer to sensitive or prohibited words identified from the audio content, such as identifying preset sensitive keywords triggered by human voices; violation sound effect types refer to the categorization of abnormal non-human voice sounds in the audio, such as identifying fighting noises or noisy shouting; and violation audio time positions indicate the start and end time intervals corresponding to the violation keywords or sound effects in the audio, such as the violation sound effect appearing between the 15th and 28th seconds of the audio.

[0065] Text moderation models are AI models specifically designed to process character sequences to understand their semantics, sentiment, intent, or to identify specific patterns and rules. For example, they are used to determine whether comments on videos contain offensive language or to detect whether text involves sensitive topics. Their output includes: sensitive content markers, violation semantic types, sentiment assessment, and text compliance review results. Sensitive content markers locate and identify the location and specific content containing sensitive information within text sentences, such as identifying text fragments containing sensitive expressions in comments. Violation semantic types classify the identified sensitive content into semantic categories, such as classifying related content as inappropriately guiding sensitive semantics. Sentiment assessment determines the overall attitude and stance presented by the text, such as determining that the text as a whole has a negative and misleading sentiment.

[0066] This invention integrates and collaboratively invokes these domain models, which are deeply optimized for different information modalities, enabling decoupling and specialized analysis of the diverse information constituting digital content. Each domain model is trained on large-scale data within its specific domain, possessing accuracy, speed, and efficiency far exceeding general-purpose models within that domain.

[0067] For example, when processing a short video containing visuals, background music, and user comments, the video moderation model analyzes the video content frame by frame, the audio moderation model separates and analyzes the background music and any audio, and the text moderation model processes the video title and the user comments below. These domain models work in parallel within their respective areas of expertise, extracting structured risk signals or content features from pixels, sound waves, and characters, respectively. As another example, in an online audio live streaming scenario, the audio moderation model is the primary analysis tool, processing the audio stream in real time; simultaneously, the accompanying real-time subtitles can be analyzed concurrently by the text moderation model; while the static cover image or the host's avatar in the live stream can be handled by the video moderation model for security checks.

[0068] This invention, by integrating and utilizing domain models of video, audio, and text, lays the technical foundation for refined and professional analysis of multimodal digital content. This allows the central decision engine to flexibly allocate the most suitable professional tools to process corresponding data based on task requirements, improving the accuracy and authority of the analysis results for each modality. By decomposing the complex multimodal content review problem into multiple mature, parallel-processable single-modal analysis sub-problems, it not only significantly improves processing efficiency but also ensures the quality of each analysis dimension through the application of professional models. These heterogeneous yet professional analysis results converge, providing high-quality, multi-dimensional input data for the backend to generate comprehensive and in-depth integrated review judgments.

[0069] Heterogeneity refers to the difference in input or output between data processing modalities of models in different domains. For example, video understanding models and video review models both take continuous image frames as input, but the former outputs understanding results such as video scenes, while the latter outputs review results such as violation elements. The input is the same but the output is different, which is heterogeneous. Similarly, audio review models take sound signals as input and output audio violation information, which are different from video models in both input and output, and are also heterogeneous.

[0070] It should be noted that the domain model can be a model that has undergone fine-tuning training; the fine-tuning training process of the domain model includes: obtaining domain training samples; the domain training samples include: domain input samples and domain outputs; inputting the domain input samples into the pre-trained model to obtain the predicted output; determining the loss value based on the matching information between the predicted output and the domain output; and adjusting the model parameters of the pre-trained model through backpropagation according to the loss value.

[0071] The principle of fine-tuning training is as follows: Based on the general knowledge representation and reasoning capabilities of the pre-trained model, labeled training samples (i.e., domain training samples) from the target domain are introduced. During the fine-tuning process, the domain input samples are fed into the pre-trained model to obtain the predicted output. The loss is calculated based on the difference between the predicted output and the true label (i.e., domain output). The model parameters are adjusted through the backpropagation algorithm. In this way, while maintaining the original general understanding capabilities of the model, the fine-tuned pre-trained model can adapt to the task characteristics and business needs of a specific domain.

[0072] For the same pre-trained model, fine-tuning can be performed using training samples from different domains to derive domain models suitable for different business scenarios. For example, fine-tuning the pre-trained model using video understanding task samples yields a video domain model with deep video content analysis capabilities; fine-tuning the pre-trained model using audio review task samples yields an audio domain model with audio violation recognition capabilities; and fine-tuning the pre-trained model using text classification and review task samples yields a text domain model suitable for text compliance judgment. Thus, multiple heterogeneous domain models with similar structures but different functions and applicable domains can be derived from the same pre-trained model.

[0073] Of course, different domain models can also be obtained based on different pre-trained models. For example, video understanding models can be obtained by fine-tuning a multimodal model pre-trained on image-text pair data, while text moderation models can be obtained by fine-tuning a language model pre-trained on pure text corpora. Video understanding models and text moderation models can originate from pre-trained models with different architectures and training objectives.

[0074] In one possible embodiment, before step 140, which utilizes the target domain model to execute the corresponding sub-task, the following steps may also be included: Verify the task execution plan; the verification includes at least one of the following: verifying the logical consistency in the task execution plan, verifying the availability of the target domain model specified in the task execution plan, and verifying the completeness of the task execution plan; If the verification fails, the task execution plan will be regenerated.

[0075] Verification refers to the automated multi-dimensional review of the generated task execution plan by the verification agent within the central decision engine to assess its rationality and feasibility.

[0076] Among them, logical consistency verification refers to checking whether there are self-contradictions or errors that violate preset logical rules within the plan, such as whether there is circular dependency, that is, task A depends on the result of task B, while task B in turn depends on the result of task A, thus leading to logical deadlock.

[0077] Model availability verification refers to verifying whether each domain model specified in the plan is currently in a normal service state, such as whether the server it resides on is online and whether its service interface is responsive.

[0078] Completeness verification refers to confirming whether the generated task execution plan fully covers all sub-tasks parsed by the analysis agent, and whether the execution elements of each sub-task have been clearly specified, such as whether each sub-task has been successfully associated with a specific model, and whether the necessary input parameters are complete.

[0079] Validating the task execution plan allows for a preliminary rehearsal before execution, enabling the early detection and interception of flawed plans. This proactively exposes and resolves risks that could lead to subsequent execution failures or resource waste. For example, during logical consistency validation, the dependency graph in the task plan is traversed, and graph theory algorithms are used to detect cycles, preventing infinite waiting during execution. During model availability validation, a lightweight heartbeat or status query request is sent to each model service listed in the plan, and their health is assessed based on the responses.

[0080] For example, if the task execution plan specifies the invocation of "Video Sensitive Scene Recognition Model V3," but during verification it is found that the model service is temporarily offline due to an upgrade, this verification will fail. During plan integrity verification, the task list in the plan is compared with the parsed original subtask list, and the key fields of each plan entry are checked for emptiness. For example, if five subtasks are parsed, but the generated task execution plan only contains the scheduling of four tasks, the integrity verification will report an error.

[0081] If any of the above checks fails, a feedback loop will be triggered, which will send the specific reasons for the failure to the central decision engine, driving it to re-plan and allocate resources according to the new constraints, thereby generating a new and revised task execution plan.

[0082] This invention embodiment utilizes a first verification agent within a central decision engine to perform logical consistency checks. A second prompt word can be input to the first verification agent. This second prompt word includes second input information, second output information, and a first verification rule, which specifies whether there are self-contradictions or errors violating preset logical rules within the task execution plan. Based on the second prompt word, the first verification agent constructs a directed topology graph and performs loop traversal detection on the sub-task dependencies of the task execution plan in the second input information, and outputs a logical compliance judgment result according to the format requirements of the second output information.

[0083] This invention can utilize a second verification agent within a central decision engine to perform model availability verification. A third prompt can be input to the second verification agent. This third prompt includes third input information, third output information, and a second verification rule. The second verification rule specifies whether each domain model to be invoked in the task execution plan is currently in a serviceable state. Based on the third prompt, the second verification agent performs service probe detection and health assessment on the target domain model list and real-time status data in the third input information, and outputs the model schedulability determination result according to the format of the third output information.

[0084] This invention embodiment utilizes a third verification agent within a central decision engine to perform integrity verification, wherein a fourth prompt word can be input to the third verification agent. The fourth prompt word includes fourth input information, fourth output information, and a third verification rule. The third verification rule specifies whether the generated task execution plan fully covers all parsed subtasks and that the execution elements of each subtask are clearly defined. Based on the fourth prompt word, the third verification agent performs set matching and field completeness verification on the complete set of preceding parsed subtasks in the fourth input information and the current plan's subtask arrangement entries, and outputs the plan integrity judgment result according to the format of the fourth output information.

[0085] In this embodiment of the invention, when a task execution plan fails the verification in any dimension, the corresponding verification agent can output a verification result that includes the anomaly type, anomaly location, and specific reason, and feed the verification result back to the central decision engine in the form of constraints.

[0086] In this embodiment of the invention, after receiving the verification result, the central decision engine converts the verification result into a constraint description that can be integrated into the first prompt word, and reconstructs the first prompt word based on the constraint description. The reconstruction includes: adding the constraint description to the first input information, so that it, together with the original audit query instruction and domain model information, constitutes an updated input; and adding mandatory instructions that satisfy the constraint description to the analysis rules. Subsequently, the central decision engine sends the reconstructed first prompt word to the analysis agent, driving it to re-analyze the audit query instruction based on the new input and rules, thereby generating a new task execution plan.

[0087] In this embodiment of the invention, the newly generated task execution plan will be submitted to each verification agent for verification, thus forming a closed-loop iterative process of verification-feedback-reconstruction-regeneration. This closed-loop iterative process will continue until the generated task execution plan passes verification in all dimensions or reaches the system's preset iteration limit, thereby proactively identifying and correcting plan defects in the task planning stage, reducing the risk of failure and resource waste in the subsequent execution process.

[0088] In summary, the embodiments of the present invention effectively prevent runtime failures, resource waste, or task execution interruptions caused by defects in the task execution scheme itself through an automated verification mechanism, reduce the fault tolerance requirements of downstream execution modules, and advance the timing of error handling from expensive runtime to low-cost planning.

[0089] In step 140, the central decision engine can use a predefined application programming interface to invoke various target domain models in the task execution plan. Each invoked target domain model independently processes its assigned subtask, executes the subtask using its built-in specific algorithm, and returns the execution result data to the central decision engine.

[0090] Optionally, when the execution sequence of subtasks is sequential, the central decision engine calls each target domain model sequentially through standardized service interfaces according to the order determined by the analysis rules. After the domain model corresponding to the previous subtask completes processing and returns the result, the model call of the next subtask is initiated, and all subtasks are executed sequentially in sequence, adapting to business scenarios where there are prerequisites, resource mutual exclusion, etc., which cannot run in parallel.

[0091] Optionally, when the execution sequence of subtasks is parallel, the central decision engine can utilize distributed parallel processing capabilities to drive multiple target domain models to work simultaneously through standard service interfaces, replacing the waiting time caused by manual serial operation of multiple tasks in the traditional approach.

[0092] For example, analysis requests can be sent simultaneously to the video moderation model, audio moderation model, and bullet comment text model. The video moderation model can return data containing a "no obvious violations found" label and a set of screenshot confidence scores; while the bullet comment text model can return an alert list containing "3 suspected personal attack comments detected" and its original text. These results differ in data format, dimensions, and meaning, constituting the execution result data that needs to be integrated subsequently.

[0093] In step 150, after the central decision engine collects multiple execution result data returned by all domain models, it performs comprehensive processing and generates the final audit result output to the user.

[0094] Optionally, the central decision engine can set up a result aggregation and presentation layer, which is responsible for integrating the specialized and fragmented execution result data given by multiple target domain models in the backend into a unified and coherent narrative (i.e., review result) for human decision-makers.

[0095] The aforementioned review results may include a review report. For example, the "no violations" conclusion from the video review model, the "1 sensitive word detected" alert from the audio model, and the "3 offensive comments" list from the text review model can be integrated to generate a review report in natural language. The content of the review report may be: "Based on comprehensive analysis, the overall risk of the current live stream is low, but there are a few violations: Audio segment A mentions the sensitive word 'XXX' at the 15th second; 3 comments involving personal attacks were found in the bullet comments (the user identifiers corresponding to the comments are...).

[0096] In one possible embodiment, step 150 may specifically include the following steps: Step A1: Input multiple execution result data into the result aggregation intelligent agent, and use the result aggregation intelligent agent to fuse the multiple execution result data to obtain the fused information; Step A2: Convert the merged information into the review results in natural language.

[0097] In simple fusion scenarios, the fusion information can include a list of execution result data output by each target domain model. This list is a summary of the execution result data of each target domain model. For example, the video model's "Detected suspected illegal props, confidence level 85%", the audio model's "Mentioned suspected illegal words", and the text model's "User A's comment contained suspected illegal expressions" are all included in the list.

[0098] In the event of a conflict between any two sets of execution result data, the fused information includes: the conflicting execution result data, the conflict type identifier, and the review conclusion based on the fusion analysis rules. Examples of conflicting execution result data are as follows: the video review model's execution result data is "no sensitive content," while the bullet screen review model's execution result data is "three unfriendly comments exist"; the conflict type identifier is "the execution result data for video content and bullet screen content are inconsistent"; the final review conclusion is, for example, "there is an overall risk of violation" or "the main text is compliant, but the comment section contains unfriendly remarks," representing a comprehensive judgment on whether a violation has occurred.

[0099] Therefore, in simple fusion scenarios, the fusion information includes a list of execution result data (e.g., summarizing "suspected inappropriate props (confidence 85%)" detected by the video model, "mentions of suspected inappropriate words" identified by the audio model, and "user A's comment contains suspected inappropriate expressions" extracted by the text model). In conflict scenarios, the fusion information includes the execution result data that caused the conflict (e.g., the video review result is "no sensitive content" while the bullet screen review result is "three unfriendly comments exist"), conflict type identifiers (e.g., "the execution results of video content and bullet screen content are inconsistent"), and the review conclusion after judgment according to the fusion analysis rules (e.g., "there is an overall risk of violation" or "the main text is compliant but the comment section contains unfriendly remarks").

[0100] This invention allows multiple execution result data from different domain models to be input into a dedicated result aggregation agent. The result aggregation agent is a large language model that has undergone targeted training or prompting engineering adjustments. Its core function is to understand, associate, and integrate input data with varying formats and semantics, rather than directly generating new content. Information fusion refers to the process by which the model comprehensively processes multi-source input data, including but not limited to identifying the correlations between different data, resolving potential contradictions in conclusions, extracting commonly referred entities and events, and logically merging and reorganizing information. The fused information is a structured or semi-structured data representation with inherent consistency generated after the information fusion process. It integrates all key findings and eliminates ambiguities and conflicts between the original data.

[0101] Leveraging the powerful contextual understanding, reasoning, and generation capabilities of large language models, an intelligent information integration and summarization layer is constructed to digest, process, and extract a unified "comprehensive report." For example, after reviewing a live stream segment, the video review model can return a structured alert stating "Suspected illegal props detected, confidence level 85%," the audio review model can return a text segment containing the word "betting," and the bullet screen text model can return an entry such as "User A commented 'Can we win this game?'"

[0102] As a result, after receiving execution result data from multiple target domain models, the aggregation agent understands that the execution result data all point to the theme of "gambling risk," and connects the scattered evidence to form a fusion judgment: "The live broadcast showed suspected illegal props, and the anchor's voice and bullet screen discussions involved illegal-related words, leading to a comprehensive judgment of gambling risk." This is the fused information. For example, for a news article and its comment section, a text moderation model might return "no sensitive content" for the article itself, while a comment moderation model might return "three comments containing personal attacks." The aggregation model would then combine these two pieces of information to generate a unified conclusion: "The article content is compliant, but the comment section contains unfriendly remarks, requiring attention to the community atmosphere."

[0103] Information fusion can manifest as conflict resolution of multi-source information, where, when conclusions from different models contradict each other, a final judgment is generated by selecting or weighting the results according to pre-defined rules. Alternatively, it can manifest as structured summarization, extracting a summary from massive amounts of raw identification results, ordered by risk type, timeline, or importance. The resulting fused information can be a structured data summary, such as a JSON object containing risk type, evidence list, and overall confidence level; or it can be an intermediate text description that has had internal contradictions resolved and is ready for report generation.

[0104] This invention can leverage the text generation capabilities of a large language model to organize fused and structured information into coherent and complete natural language paragraphs or documents according to a narrative logic and report format easily understood by humans. For example, the aforementioned fused information can be automatically expanded to generate a complete paragraph report containing "risk characterization," "specific evidence citation," "time point of occurrence," and "handling recommendations." As another example, a concise "content overview and community status" description can be generated from the unified conclusions regarding news comment sections.

[0105] By introducing a result-aggregating agent to handle information fusion and report generation, the core challenge of output integration in multi-model collaborative work is effectively solved. It automates the cognitively demanding tasks of information comparison, correlation analysis, and report writing that previously required manual intervention. Through deep reasoning using large language models, it can uncover the intrinsic connections between disparate pieces of evidence and even reveal complex risks that individual models cannot detect, thereby enhancing the depth and insight of comprehensive judgment.

[0106] For example, a fifth prompt word can be input into the result aggregation agent. The fifth prompt word includes fifth input information, fifth output information, and fusion analysis rules. The fifth input information includes execution result data output by multiple domain models. The fifth output information is used to specify the unified output format after information fusion and the standard for converting the results into natural language review results. The fusion analysis rules are used to specify the integration processing logic of multi-source execution results, including identifying semantic relationships between different results, resolving conclusion conflicts, extracting related entity and event information, and completing the merging and reorganization of information according to preset rules.

[0107] In one possible embodiment, step A1 above may specifically include the following steps: For the same audit object or audit attribute in the audit object, identify whether there are conflicts in the execution result data corresponding to different sub-tasks; If a conflict exists, the basis for conflict resolution is obtained; the basis for conflict resolution includes at least one of the following: the confidence score of the execution result data, the task weight of the subtask, and the predefined business rules related to the audit object; Based on the resolution criteria, final judgment data is selected or generated from conflicting execution result data; The final judgment data replaces the conflicting execution result data to obtain the merged information.

[0108] A conflict refers to a situation where different domain models output execution result data that directly contradicts or is significantly inconsistent in their judgment conclusions or key semantics regarding the same entity, attribute, or event within the audit object. Conflict resolution criteria refer to a set of objective standards or data relied upon when making arbitration decisions, aiming to provide an authoritative, reasonable, or prioritizing basis for resolving conflicts. Final judgment data refers to authoritative data generated after the conflict resolution process, representing the sole conclusion for that audit point.

[0109] For example, in a live streaming scenario, a video moderation model outputs "No violent behavior detected (90% confidence)" based on image analysis, while an audio moderation model identifies painful screams and impact sounds from concurrent audio, thus outputting "Suspected violent audio event detected (75% confidence)". In this case, the execution results of the two models conflict regarding the attribute of "whether violence exists".

[0110] The resulting aggregation agent possesses the ability to perform cross-modal association and semantic alignment, understanding that data from different models, presented in various forms, actually points to the same focus of review. For example, the model needs to associate the "scene from second 120 to second 125" reported by the video model with the "audio segment from second 121 to second 124" reported by the audio model, identifying their high overlap on the timeline and pointing to the same event, thus determining whether the conclusions conflict. Conflicts can manifest in various forms, such as direct opposition (one classifies it as "violation" while the other classifies it as "not violation"); significant differences in degree (one classifies it as "high risk" while the other classifies it as "low risk"); or contradictions in the description of key facts.

[0111] Once a conflict is identified, the basis for conflict resolution is obtained. This basis comes from multiple dimensions: First, the confidence score of the target domain model itself for the output. This is a quantitative value of the target domain model's evaluation of its judgment result, such as a 90% confidence score for a video model versus a 75% confidence score for an audio model. Second, the task weights of the sub-tasks. This reflects the priority given to different types of evidence during comprehensive judgment. For example, in a review scenario, the weight of the sub-task for "visual content compliance" is usually preset to be higher than that for the sub-task for "background audio compliance." Third, predefined business rules related to the review object or scenario. This is a set of explicit rules, such as "If any model detects specific sensitive content involving minors, then regardless of the results of other models, it will be considered the highest risk." These bases together constitute a quantitative or logical framework for multi-factor decision-making.

[0112] This invention employs a pre-defined conflict resolution strategy for adjudication. For example, one strategy could be to "compare the confidence scores of both conflicting parties and adopt the conclusion of the one with the higher score." Based on this strategy, the "non-violent" conclusion of the video model would be adopted in the example above. Another strategy could be to "prioritize the adherence to pre-defined business rules when they exist." For example, if a business rule stipulates that "in cases of audio-visual information conflict, audio content is the primary basis for risk assessment," then the conclusion of the audio model would be adopted. Yet another strategy is to comprehensively consider confidence level and task weight, calculate a comprehensive score, and then make a decision. By implementing such strategies, an unambiguous final judgment can be automatically generated from the conflict data.

[0113] The confidence score is a quantified value between 0 and 1, representing the reliability of the execution result data output by the model. The confidence score is obtained as follows: during the generation of the execution result data, the model takes the feature representation obtained during its internal forward propagation and a predefined baseline feature distribution as inputs. A preset mapping function is used to calculate the degree of matching between the two, and this degree of matching is converted into a value between 0 and 1, which is then used as the confidence score corresponding to the execution result data and output synchronously after being bound to the execution result data.

[0114] Task weight is a preset quantitative value used to determine the relative importance of each subtask in the final judgment. This value is predefined based on business rules and domain knowledge during the task execution generation phase and associated with the corresponding subtask. It is then invoked in the review scenario to participate in subsequent conflict resolution calculations.

[0115] Based on the aforementioned conflict resolution strategy, the process of selecting or generating final judgment data from conflicting execution result data (i.e., the processing logic of the conflict resolution strategy) specifically includes: If the first execution result data among the conflicting execution result data conforms to the predefined business rules, then the final judgment data is generated based on the first execution result data; or If none of the conflicting execution results conform to the predefined business rules, the final judgment data will be generated based on the one with the higher confidence score among the conflicting execution results; or If conflicting execution result data do not conform to the predefined business rules, a comprehensive score is calculated based on the confidence score and the task weight of the subtask, and the execution result data with the highest comprehensive score is selected to generate the final judgment data.

[0116] In this embodiment of the invention, the above-mentioned conflict resolution strategy and strategy analysis process can be carried in the fusion analysis rule of the fifth prompt word. After the result aggregation agent reads the fifth prompt word, it resolves the conflicting execution result data according to the processing logic defined in the fusion analysis rule, thereby generating the final judgment data.

[0117] Replacing conflicting original execution result data with final judgment data means that in subsequent information fusion and report generation stages, only the unified conclusion after arbitration will be presented regarding the disputed point, thereby improving the consistency and authority of the internal logic of the final audit result output.

[0118] Therefore, this invention, by automating the handling of unavoidable conflicting conclusions in multi-model collaboration and incorporating the resolution criteria into the decision-making process, makes the entire process more controllable and predictable, and compliant with business logic and security standards. This provides crucial technical support for achieving high-quality, highly reliable automated content review decisions, thereby improving the reliability of intelligent review output results.

[0119] In one example, the review result of this embodiment of the invention includes video metadata, final judgment data, and the reasoning process for the final judgment data.

[0120] Video metadata, such as video identifier, title, and uploader, is used to identify the review target. The final judgment data is the review conclusion generated by fusing and resolving conflicts from the execution results data of multiple domain models. The reasoning process of the final judgment data records the logical steps and basis from the execution result data output by each model, conflict identification, to conflict resolution based on confidence scores, task weights of subtasks, and predefined business rules, to finally generating the final judgment data.

[0121] In summary, this embodiment of the invention responds to a user's audit query command for an audited object. The central decision engine uses a large language model to parse the query command, obtaining task information for each sub-task. This task information includes at least one of the following: task intent information, audit domain, and task type. A task execution plan is generated based on this information. The execution plan includes the domain model corresponding to each sub-task, clearly specifying which professional model will execute each sub-task, eliminating the need for cumbersome manual process design and resource allocation across multiple sub-tasks. Based on the execution plan, the domain models are driven to execute their respective sub-tasks, yielding multiple execution result data for each sub-task. This allows multi-domain analysis tasks that previously required sequential or manual alternation to be executed in parallel, thereby shortening the overall time from planning to obtaining professional analysis results across various dimensions and improving efficiency during task execution. Based on the multiple execution result data, audit results are output to the user. Thus, efficient content querying is achieved while ensuring processing efficiency and the completeness of analysis dimensions.

[0122] Reference Figure 2 The diagram illustrates a content query method according to an embodiment of the present invention. This method is executed by a central decision engine and specifically includes the following steps: Step 201: Receive the user's audit query instruction for the audit object; Step 202: Using the analytical agent, perform intent parsing and task decomposition on the audit query command to obtain a task execution plan; Step 203: Use the verification agent to verify the task execution plan; if the verification fails, proceed to step 204; if the verification passes, proceed to step 205. Step 204: Generate constraints and send them back to step 202 to re-trigger the task parsing and planning process in step 202; Step 205: Dynamically distribute the verified task execution plan to the corresponding target domain models such as video review, live broadcast review, and bullet screen review. Step 206: Utilize each target domain model to perform content review corresponding to the sub-tasks and output their respective execution result data; wherein, the architecture of the target domain model can be LLM (Large Language Model) / SLM (Small Language Model), etc.

[0123] Step 207: Generate the audit result for the audit object based on the execution result data corresponding to the multiple sub-tasks.

[0124] Reference Figure 3 This diagram illustrates a flowchart of a content query method according to an embodiment of the present invention. The processing architecture of this content query method specifically includes: an intelligent decision-making layer, a domain model routing layer, and a result aggregation and output layer. First, the large language model decision engine of the intelligent decision-making layer receives the natural language request submitted by the user, identifies the query intent, and breaks down the complex natural language request into multiple sub-tasks. Next, the domain model routing layer selects a large language model, a small language model, or a data query tool corresponding to the domain for each sub-task. Finally, the result aggregation and output layer collects the execution results of each sub-task, integrates them, and resolves conflicts to generate an audit result presented to the user. The audit result may include: an audit report in natural language form and structured audit result data, etc.

[0125] For example, if a user submits a natural language request to "analyze the comprehensive risk of live stream room A over the past hour," the large language model decision engine of the intelligent decision layer first identifies the intent and then breaks it down into sub-tasks such as video image analysis, bullet screen sentiment analysis, and historical record query. Then, the domain model routing layer selects the corresponding large or small language model for video image analysis and bullet screen sentiment analysis, respectively, and selects a data query tool for historical record query. After each sub-task is executed, the result aggregation and output layer collects the video review results, bullet screen review results, and historical record data. After integration and conflict resolution, it generates a natural language report and structured data to be presented to the user.

[0126] Based on the above content query method, this embodiment of the invention also provides a content query device, applicable to any electronic device, such as... Figure 4As shown, the content query device may include: The receiving module 401 is used to receive the audit query instruction initiated by the user terminal for the audit object; Prompt word generation module 402 is used to generate the first prompt word; Analysis module 403 is used to send a first prompt word to the analysis agent, so that the analysis agent analyzes the review query instruction according to the first prompt word to obtain a task execution plan corresponding to the review query instruction; the task execution plan includes: multiple sub-tasks and a target domain model corresponding to each sub-task; the first prompt word includes: first input information, first output information and analysis rules; the first input information includes: model information corresponding to the review query instruction and multiple domain models respectively; the first output information is used to characterize the information of the task execution plan; the analysis rules are used to characterize the generation logic from the review query instruction to the task execution plan, determine the query intent corresponding to the review query instruction, generate multiple sub-tasks corresponding to the review query instruction according to the query intent, and determine the target domain model corresponding to each sub-task from multiple domain models according to the model information corresponding to multiple domain models respectively; Task execution module 404 is used to execute corresponding sub-tasks using the target domain model and obtain execution result data for each sub-task. The result generation module 405 is used to generate an audit result for the audit object based on the execution result data corresponding to multiple sub-tasks.

[0127] In one possible embodiment, the task execution scheme further includes: the execution sequence of multiple subtasks; the analysis rule is also used to determine the execution sequence of multiple subtasks based on the dependency information between the multiple subtasks and / or the priority information of each subtask.

[0128] Optionally, the model information includes: the number of model parameters; the number of model parameters is used to characterize the total number of weights and biases adjusted during the training process of the corresponding domain model; The step of determining the target domain model corresponding to each sub-task from multiple domain models based on the model information corresponding to each domain model includes: determining the target domain model corresponding to each sub-task from multiple domain models based on the amount of data of the audit object and the number of model parameters of each domain model.

[0129] Optionally, determining the target domain model corresponding to each sub-task from multiple domain models based on the data volume of the audit object and the number of model parameters for each domain model includes: If the amount of data of the audited object exceeds the first threshold, the domain model with more than the second threshold of model parameters will be determined as the target domain model for each sub-task; or If the amount of data of the audit object is less than or equal to the first threshold, the domain model with the number of model parameters less than or equal to the second threshold is determined as the target domain model for each subtask.

[0130] Optionally, before the task execution module 404 executes the corresponding subtask using the target domain model, the apparatus further includes: A verification module is used to verify the task execution plan; the verification includes at least one of the following: verifying the logical consistency in the task execution plan, verifying the availability of the target domain model in the task execution plan, and verifying the integrity of the task execution plan; The regeneration module is used to regenerate the task execution plan if the verification fails.

[0131] Optionally, the domain model is a model that has been fine-tuned and trained; the fine-tuning and training process of the domain model includes: acquiring domain training samples; the domain training samples include: domain input samples and domain outputs; inputting the domain input samples into the pre-trained model to obtain a predicted output; determining a loss value based on the matching information between the predicted output and the domain output; and adjusting the model parameters of the pre-trained model through backpropagation according to the loss value.

[0132] Optionally, the domain model includes at least two of the following: Video understanding models for understanding video content; A video moderation model used for reviewing video content; Audio recognition models used to identify audio content; An audio moderation model for reviewing audio content; A text moderation model for reviewing text content.

[0133] Optionally, the result generation module 405 includes: The result aggregation module is used to input the execution result data corresponding to multiple subtasks into the result aggregation intelligent agent, which then performs information fusion on the execution result data corresponding to the multiple subtasks to obtain fused information. The conversion module is used to convert the fused information into audit results in natural language form.

[0134] Optionally, the information fusion of the execution result data corresponding to the plurality of subtasks includes: For the same audit object or audit attribute in the audit object, identify whether there are conflicts in the execution result data corresponding to different sub-tasks; If a conflict exists, the basis for conflict resolution is obtained; the basis for conflict resolution includes at least one of the following: the confidence score of the execution result data, the task weight of the subtask, and the predefined business rules related to the audit object; Based on the aforementioned resolution criteria, final judgment data is selected or generated from conflicting execution result data; The fusion information is obtained by replacing the conflicting execution result data with the final determination data.

[0135] Optionally, the step of selecting or generating final judgment data from conflicting execution result data based on the resolution criteria includes: If the first execution result data among the conflicting execution result data conforms to the predefined business rules, then the final judgment data is generated based on the first execution result data; or If none of the conflicting execution results conform to the predefined business rules, the final judgment data will be generated based on the one with the higher confidence score among the conflicting execution results; or If conflicting execution result data do not conform to the predefined business rules, a comprehensive score is calculated based on the confidence score and the task weight of the subtask, and the execution result data with the highest comprehensive score is selected to generate the final judgment data.

[0136] In this embodiment of the invention, in response to a user's audit query command for an audit object, the central decision engine parses the audit query command using a large language model to obtain task information for each subtask. The task information includes at least one of the following: task intent information, audit domain, and task type. A task execution plan is generated based on the task information. The task execution plan includes the domain model corresponding to each subtask, clearly specifying which professional model will execute each subtask, eliminating the need for cumbersome manual process design and resource allocation across multiple subtasks. Based on the task execution plan, the domain models are driven to execute their respective subtasks, obtaining multiple execution result data for each subtask. This allows multi-domain analysis tasks that previously required sequential or manual alternation to be executed in parallel, thereby shortening the overall time from planning to obtaining professional analysis results across various dimensions and improving the efficiency of the task execution phase. Based on the multiple execution result data, the audit results are output to the user. Thus, while ensuring processing efficiency and the completeness of analysis dimensions, efficient content querying is achieved. This invention also provides an electronic device, such as... Figure 5As shown, it includes a processor 701, a communication interface 702, a memory 703, and a communication bus 704, wherein the processor 701, the communication interface 702, and the memory 703 communicate with each other through the communication bus 704. Memory 703 is used to store computer programs; The processor 701, when executing a program stored in the memory 703, performs the following steps: receiving a review query instruction initiated by a user terminal for a review object; generating a first prompt word; sending the first prompt word to an analysis agent, so that the analysis agent analyzes the review query instruction based on the first prompt word to obtain a task execution plan corresponding to the review query instruction; the task execution plan includes: multiple sub-tasks and a target domain model corresponding to each sub-task; the first prompt word includes: first input information, first output information, and analysis rules; the first input information includes: model information corresponding to the review query instruction and multiple domain models respectively; the first output information is used to characterize the information of the task execution plan; the analysis rules are used to characterize the generation logic from the review query instruction to the task execution plan, determine the query intent corresponding to the review query instruction, generate multiple sub-tasks corresponding to the review query instruction based on the query intent, and determine the target domain model corresponding to each sub-task from multiple domain models based on the model information corresponding to each domain model; execute the corresponding sub-task using the target domain model to obtain execution result data corresponding to each sub-task; and generate a review result for the review object based on the execution result data corresponding to multiple sub-tasks respectively.

[0137] The communication bus mentioned above can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus can be divided into address bus, data bus, control bus, etc. For ease of illustration, only one thick line is used to represent it in the diagram, but this does not mean that there is only one bus or one type of bus.

[0138] The communication interface is used for communication between the aforementioned terminal and other devices.

[0139] The memory may include random access memory (RAM) or non-volatile memory, such as at least one disk storage device. Optionally, the memory may also be at least one storage device located remotely from the aforementioned processor.

[0140] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0141] In another embodiment of the present invention, a computer-readable storage medium is also provided, which stores instructions that, when executed on a computer, cause the computer to perform any of the content query methods described in the above embodiments.

[0142] In another embodiment of the present invention, a computer program product containing instructions is also provided, which, when run on a computer, causes the computer to execute any of the content query methods described in the above embodiments.

[0143] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid state disk (SSD)).

[0144] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0145] The various embodiments in this specification are described in a related 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 system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0146] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention are included within the scope of protection of the present invention.

Claims

1. A content query method, characterized by, The method includes: Receive audit query commands from users for the audited object; Generate the first prompt word; A first prompt word is sent to the analysis agent, enabling the analysis agent to analyze the review query instruction based on the first prompt word and obtain a task execution plan corresponding to the review query instruction. The task execution plan includes: multiple sub-tasks and a target domain model corresponding to each sub-task. The first prompt word includes: first input information, first output information, and analysis rules. The first input information includes: model information corresponding to the review query instruction and multiple domain models respectively. The first output information is used to characterize the information of the task execution plan. The analysis rules are used to characterize the generation logic from the review query instruction to the task execution plan, determine the query intent corresponding to the review query instruction, generate multiple sub-tasks corresponding to the review query instruction based on the query intent, and determine the target domain model corresponding to each sub-task from multiple domain models based on the model information corresponding to each domain model. The target domain model is used to execute the corresponding sub-tasks, and the execution result data for each sub-task is obtained. Based on the execution result data corresponding to each of the multiple sub-tasks, an audit result is generated for the audit object.

2. The method according to claim 1, characterized in that, The task execution scheme further includes: the execution sequence of multiple subtasks; the analysis rules are also used to determine the execution sequence of multiple subtasks based on the dependency information between multiple subtasks and / or the priority information of each subtask.

3. The method according to claim 1, characterized in that, The model information includes: the number of model parameters; the number of model parameters is used to characterize the total number of weights and biases adjusted during the training process of the corresponding domain model; The step of determining the target domain model corresponding to each sub-task from multiple domain models based on the model information corresponding to each domain model includes: determining the target domain model corresponding to each sub-task from multiple domain models based on the amount of data of the audit object and the number of model parameters of each domain model.

4. The method according to claim 3, characterized in that, The step of determining the target domain model corresponding to each sub-task from multiple domain models based on the data volume of the audit object and the model parameter volume of each domain model includes: If the amount of data of the audited object exceeds the first threshold, the domain model with more than the second threshold of model parameters will be determined as the target domain model for each sub-task; or If the amount of data of the audit object is less than or equal to the first threshold, the domain model with the number of model parameters less than or equal to the second threshold is determined as the target domain model for each subtask.

5. The method according to claim 1, characterized in that, Before executing the corresponding subtask using the target domain model, the method further includes: The task execution plan is validated; the validation includes at least one of the following: validating the logical consistency of the task execution plan, validating the usability of the target domain model in the task execution plan, and validating the completeness of the task execution plan. If the verification fails, the task execution plan will be regenerated.

6. The method according to any one of claims 1 to 5, characterized in that, The domain model is a model that has been fine-tuned and trained. The fine-tuning training process of the domain model includes: acquiring domain training samples; the domain training samples include: domain input samples and domain outputs; inputting the domain input samples into the pre-trained model to obtain a predicted output; determining a loss value based on the matching information between the predicted output and the domain output; and adjusting the model parameters of the pre-trained model through backpropagation according to the loss value.

7. The method according to any one of claims 1 to 5, characterized in that, The domain model includes at least two of the following: Video understanding models for understanding video content; A video moderation model used for reviewing video content; Audio recognition models used to identify audio content; An audio moderation model for reviewing audio content; A text moderation model for reviewing text content.

8. The method according to any one of claims 1 to 5, characterized in that, The step of generating an audit result for the audit object based on the execution result data corresponding to multiple sub-tasks includes: The execution result data corresponding to multiple subtasks are input into the result aggregation intelligent agent, which then performs information fusion on the execution result data corresponding to the multiple subtasks to obtain fused information. The fused information is converted into the audit results in natural language.

9. The method according to any one of claims 1 to 5, characterized in that, The information fusion of the execution result data corresponding to the multiple sub-tasks includes: For the same audit object or audit attribute in the audit object, identify whether there are conflicts in the execution result data corresponding to different sub-tasks; If a conflict exists, the basis for conflict resolution is obtained; the basis for conflict resolution includes at least one of the following: the confidence score of the execution result data, the task weight of the subtask, and the predefined business rules related to the audit object; Based on the aforementioned resolution criteria, final judgment data is selected or generated from conflicting execution result data; The fusion information is obtained by replacing the conflicting execution result data with the final determination data.

10. The method according to claim 9, characterized in that, The step of selecting or generating final judgment data from conflicting execution result data based on the resolution criteria includes: If the first execution result data among the conflicting execution result data conforms to the predefined business rules, then the final judgment data is generated based on the first execution result data; or If none of the conflicting execution results conform to the predefined business rules, the final judgment data will be generated based on the one with the higher confidence score among the conflicting execution results; or If conflicting execution result data do not conform to the predefined business rules, a comprehensive score is calculated based on the confidence score and the task weight of the subtask, and the execution result data with the highest comprehensive score is selected to generate the final judgment data.

11. A content query device, characterized in that, The device includes: The receiving module is used to receive audit query instructions initiated by the user client for the audit object; The prompt word generation module is used to generate the first prompt word; An analysis module is used to send a first prompt word to an analysis agent, so that the analysis agent analyzes the review query instruction based on the first prompt word to obtain a task execution plan corresponding to the review query instruction. The task execution plan includes: multiple sub-tasks and a target domain model corresponding to each sub-task. The first prompt word includes: first input information, first output information, and analysis rules. The first input information includes: model information corresponding to the review query instruction and multiple domain models respectively. The first output information is used to characterize the information of the task execution plan. The analysis rules are used to characterize the generation logic from the review query instruction to the task execution plan, determine the query intent corresponding to the review query instruction, generate multiple sub-tasks corresponding to the review query instruction based on the query intent, and determine the target domain model corresponding to each sub-task from multiple domain models based on the model information corresponding to multiple domain models respectively. The task execution module is used to execute corresponding sub-tasks using the target domain model and obtain execution result data for each sub-task. The result generation module is used to generate an audit result for the audit object based on the execution result data corresponding to multiple sub-tasks.

12. An electronic device, characterized in that, It includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, used to store computer programs; A processor, when executing a program stored in memory, implements the steps of the method described in any one of claims 1-10.

13. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1-10.

Images