End-cloud cooperative media all-in-one machine interaction control method and device, equipment and medium

Abstract

This application relates to an interactive control method, apparatus, device, and medium for an edge-cloud collaborative media all-in-one machine. The method includes: acquiring user request text data, and performing intent parsing on the user request text data based on natural language processing to obtain the intent of the requesting object; retrieving corresponding relevant information from a vectorized knowledge base based on the intent of the requesting object to obtain enhanced context information; inputting the enhanced context information and the intent of the requesting object into a specialized large-scale model for reasoning to obtain preliminary content results; performing compliance checks on the preliminary content results based on content security filtering rules to obtain compliant content results; and synthesizing the compliant content results into a virtual avatar broadcast video based on virtual avatar driving to obtain a digital human broadcast video stream. This method can quickly respond to sudden content demands and effectively improve the stability and real-time performance of the interaction.

Description

Technical Field

[0001] This invention belongs to the field of interactive control, and in particular relates to interactive control methods, devices, equipment and media for end-to-end cloud collaborative media all-in-one machines. Background Technology

[0002] With the deep integration and application of artificial intelligence technology in the media field, intelligent content generation and interaction technologies based on large models have emerged. This technology possesses natural language understanding, multimodal information processing, and continuous learning capabilities, enabling precise analysis of complex semantics and anthropomorphic output, bringing revolutionary possibilities to media production and dissemination models. Traditional media content production and external services mainly rely on a combination of manual news gathering and editing, fixed scripts, and pre-produced materials. In the content creation stage, reporters and editors need to manually search for information, write manuscripts, and conduct multiple rounds of proofreading; in the external presentation stage, they rely on presenters memorizing key points or playing pre-recorded video materials, ensuring the standardization and consistency of information transmission through standardized processes. However, traditional methods have significant limitations: content production efficiency is limited by manpower input, making it difficult to cope with the sudden, large-scale, and immediate generation needs of content; manual searching and writing are prone to oversights, and the accuracy of professional terminology highly depends on personal experience; fixed broadcasting and presentation lack interactivity and flexibility, making it difficult to meet the public's in-depth needs for real-time Q&A and personalized interpretation; furthermore, in terms of content compliance, relying entirely on manual review also carries the risks of response delays and inconsistent standards. Summary of the Invention

[0003] Therefore, it is necessary to provide a method, device, equipment, and medium for interactive control of a cloud-based media all-in-one machine that can quickly generate content and provide flexible interaction, addressing the aforementioned technical issues.

[0004] Firstly, this application provides an interactive control method for a cloud-edge collaborative media all-in-one machine, including:

[0005] The system acquires user request text data and performs intent parsing on the user request text data based on natural language processing to obtain the intent of the request object; the intent of the request object includes question and answer writing requests and topic planning requests.

[0006] Based on the intent of the request object, relevant information is retrieved from the vectorized knowledge base to obtain enhanced contextual information;

[0007] The enhanced contextual information and the intent of the requesting object are input into a specialized large model for reasoning to obtain preliminary content results;

[0008] Based on content security filtering rules, the preliminary content results are checked for compliance to obtain compliant content results;

[0009] Driven by virtual avatars, compliant content is synthesized into a video broadcast by a virtual avatar, resulting in a digital human broadcast video stream.

[0010] Furthermore, the vectorized knowledge base is constructed using the following methods:

[0011] Acquire multi-format material data, and based on optical character recognition, extract the text contained in the image data in the multi-format material data to obtain the first converted text;

[0012] Based on language recognition, the audio and video data in multi-format material data are converted into text to obtain the second converted text.

[0013] By integrating the original text from the first transformed text, the second transformed text, and the multi-format material data, a plain text database is obtained.

[0014] Based on a plain text database, semantic analysis is performed on the text to obtain text segments, and each text segment is converted into a vector to obtain a set of vector pairs;

[0015] Based on the set of vector pairs, similarity comparison and optimization are performed to obtain a vectorized knowledge base.

[0016] Furthermore, the specialized large model is constructed through the following methods:

[0017] Based on a plain text database, domain-specific texts are filtered to obtain a refined domain-specific corpus.

[0018] Based on a domain-refined corpus, input commands and prompt words are designed to obtain a domain-adapted command template library; the domain-adapted command template library is used to train the model to follow the specifications of the corresponding domain.

[0019] The examples from the domain adaptation instruction template library are input into the base model, and the base model is subjected to context learning to obtain the transition model;

[0020] Based on a domain-refined corpus and efficient parameter fine-tuning, a portion of the parameters of the transition model are lightly adjusted to obtain the fine-tuned model.

[0021] Based on the test case set and security rules, the output of the fine-tuned model is calibrated to obtain a specialized large model; the test case set includes routine task test cases and potential risk test cases.

[0022] Furthermore, based on the intent of the requesting object, relevant information is retrieved from the vectorized knowledge base to obtain enhanced contextual information, including:

[0023] Based on the intent of the requesting object, document fragments containing keywords are searched from the vectorized knowledge base to obtain keyword search results;

[0024] The request intent is converted into a vector to obtain the request vector. Then, the document fragment with the highest cosine similarity to the request vector is retrieved from the vectorized knowledge base using the following formula to obtain the vector semantic retrieval result:

[0025]

[0026] in, For cosine similarity, For query vector, Document vector;

[0027] Deduplication is performed on the keyword search results and the vector semantic search results to obtain a set of candidate document fragments;

[0028] Based on the reordering model, the candidate document fragment set is sorted twice to obtain a reordered document fragment list.

[0029] Based on the reordered list of document fragments, the document fragments are combined into contextual text blocks to obtain enhanced contextual information.

[0030] Furthermore, if the request object's intent is a topic selection and planning request, then based on the request object's intent, before searching for document fragments containing keywords from the vectorized knowledge base and obtaining keyword search results, the process also includes:

[0031] Based on the intent of the requesting object, a topic planning request signal is generated and sent to the cloud; the topic planning request signal is used to instruct the cloud to carry out topic planning and to provide a topic planning report.

[0032] Receive topic planning reports, and based on these reports, transform hot topics into specific questions to obtain a set of search intents;

[0033] Add the set of search intents to the request object intent to obtain the updated request object intent.

[0034] Furthermore, after receiving the topic selection and planning request signal in the cloud, it also includes:

[0035] The original internet dataset was obtained and preprocessed to obtain high-quality text corpus; the original internet dataset includes text data, image data and video data.

[0036] Identify high-frequency words in high-quality text corpora, categorize the high-frequency words, and obtain a list of hot topics;

[0037] Based on the list of trending topics, a sentiment analysis is conducted to obtain a trend analysis report;

[0038] By integrating the list of trending topics and trend analysis reports, a topic planning report is obtained.

[0039] Secondly, this application also provides an interactive control device for a cloud-edge collaborative media all-in-one machine, comprising:

[0040] The intent module is used to acquire user request text data and, based on natural language processing, perform intent parsing on the user request text data to obtain the intent of the request object; the intent of the request object includes question and answer writing requests and topic planning requests.

[0041] The enhancement module is used to retrieve relevant information from the vectorized knowledge base based on the intent of the request object to obtain enhanced context information;

[0042] The reasoning module is used to input enhanced contextual information and the intent of the requesting object into a specialized large model for reasoning, and to obtain preliminary content results;

[0043] The verification module is used to perform compliance verification on the preliminary content results based on content security filtering rules, and obtain compliant content results.

[0044] The broadcasting module is used to synthesize compliant content results into a virtual avatar broadcasting video based on the virtual avatar, resulting in a digital human broadcasting video stream.

[0045] Thirdly, this application also provides a computer device including a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement any step of the method provided in the first aspect of this application.

[0046] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements any step of the method provided in the first aspect of this application.

[0047] The aforementioned edge-cloud collaborative media all-in-one machine interactive control method, device, equipment, and medium acquire user request text data and, based on natural language processing, perform intent parsing on the user request text data to obtain the request object's intent. The request object's intent includes question-and-answer writing requests and topic planning requests. Based on the request object's intent, relevant information is retrieved from a vectorized knowledge base to obtain enhanced context information. The enhanced context information and the request object's intent are input into a specialized large-scale model for reasoning to obtain preliminary content results. Based on content security filtering rules, the preliminary content results are subjected to compliance verification to obtain compliant content results. Based on virtual avatar driving, the compliant content results are synthesized into a virtual avatar broadcast video to obtain a digital human broadcast video stream. It can build a local dedicated knowledge base to ensure the accuracy and authority of generated content, and introduce a large-scale model for content generation, meeting the needs for real-time generation of recurring and large-scale content. It flexibly displays content to meet the public's needs for real-time Q&A and personalized interpretation, effectively realizing a flexible and rapid media all-in-one machine interactive control method. Attached Figure Description

[0048] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0049] Figure 1 This is a schematic diagram of the flow of an edge-cloud collaborative media all-in-one machine interactive control method provided in an embodiment of the present invention;

[0050] Figure 2 A schematic diagram of the structure of an end-to-cloud collaborative media all-in-one machine interactive control device provided in an embodiment of the present invention; Detailed Implementation

[0051] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0052] In one embodiment, such as Figure 1 As shown, a method for interactive control of a media all-in-one machine with end-to-end cloud collaboration is provided. This embodiment illustrates the application of this method to a local terminal. It is understood that this method can also be applied to a server, and further to a system including both a terminal and a cloud, and is implemented through interaction between the terminal and the cloud. In this embodiment, the method includes the following steps:

[0053] Step 101: Obtain user request text data and perform intent parsing on the user request text data based on natural language processing to obtain the intent of the request object; the intent of the request object includes question and answer writing requests and topic planning requests.

[0054] In this context, user request text data refers to the textual form of the original question or instruction input by the user into the system via voice or text. Natural Language Processing (NLP) is a technology that enables computers to understand, interpret, and manipulate human language; in this embodiment, it is used as a tool for analyzing text. The request object intent refers to the fundamental purpose or category of the user request after analysis, and is explicitly divided into two types: Question-and-Answer Writing Requests, where the user hopes to obtain specific information or generate a standardized text, with the intent to seek facts or complete document creation; and Topic Planning Requests, where the user hopes to obtain suggestions on content direction or analysis of hot topics, with the intent to seek inspiration or strategic planning. The terminal utilizes NLP technology to perform in-depth analysis of the user request text data. It understands the semantics, context, and potential purpose of the text through algorithmic models, identifies key entities, action verbs, and sentence structures in the request, and thus determines which predefined intent it best fits. The analysis results are structured to generate a clear request object intent identifier, along with extracted key information elements.

[0055] Step 102: Based on the intent of the requesting object, retrieve the corresponding relevant information from the vectorized knowledge base to obtain enhanced context information.

[0056] Specifically, a vectorized knowledge base is a special database storing massive amounts of knowledge fragments. Its unique feature is that each knowledge fragment is converted into a set of high-dimensional mathematical vectors, ensuring that semantically similar texts are also located close together in the vector space, facilitating semantic similarity retrieval by computers. Enhanced contextual information refers to a set of one or more knowledge fragments retrieved from the vectorized knowledge base that are highly relevant to the user's request intent. This provides the factual basis and background material needed for specialized large-scale models to generate answers. The terminal also converts the request intent into a query vector and searches the vectorized knowledge base. By calculating the similarity between the query vector and all knowledge fragment vectors in the database, the most similar knowledge fragments are identified. To improve accuracy, keyword retrieval can be used as an aid, and the initial search results are deduplicated and reordered based on relevance. The most relevant fragments are then combined into a coherent piece of enhanced contextual information.

[0057] Step 103: Input the enhanced contextual information and the requester's intent into the specialized large model for reasoning to obtain preliminary content results.

[0058] Specifically, a specialized large-scale model is an AI model pre-trained on a large amount of data and further optimized for a specific domain. It is better than a general-purpose large-scale model at understanding and generating text that conforms to the domain's norms, terminology, and style. Preliminary content results refer to the raw text content generated by the specialized large-scale model based on the received information; this is a complete answer or draft to the user's request, but it has not yet undergone final security verification. The terminal submits the requester's intent and enhanced contextual information as input to the specialized large-scale model. The model comprehensively understands the user's intent and the provided background information, utilizing its powerful language generation and logical reasoning capabilities to organize language, construct logic, and generate a text that directly responds to the request.

[0059] Step 104: Based on the content security filtering rules, perform compliance checks on the preliminary content results to obtain compliant content results.

[0060] Content security filtering rules are a pre-defined set of standards and rules regarding content security and factual accuracy, defining what content can be output and what content must be prohibited or modified. Compliant content results refer to the final text content that has been confirmed as safe and compliant after the initial content results have undergone security filtering. The terminal loads all activated content security filtering rules, which may be categorized into different levels, including core security rules and quality specification rules. The initial content result text is simultaneously sent to multiple detection engines for analysis. The keyword and pattern matching engine quickly scans the text for explicitly prohibited sensitive words, erroneous expressions, or specific combinations of patterns; it is fast and used to intercept the most obvious risks. The semantic understanding and context analysis engine uses more complex natural language processing models to understand the overall meaning and context of the text, identifying negative or dangerous viewpoints expressed through metaphors, irony, etc., even if they do not contain sensitive words. The fact-checking engine compares the factual statements claimed in the text with authoritative data from a vectorized knowledge base. The system performs rapid comparisons of information sources, identifying potentially factual errors or outdated information. The outputs of each detection engine, based on preset risk thresholds and decision logic, determine the entire content. Optionally, the determination results include: Pass (no risk, directly marked as compliant); Reject (violates core security rules, unacceptable risk, logs are recorded, and a preset security warning may be returned to the user instead of the original content); Correct (automatically replaces words or adjusts wording for automatically fixable quality issues or minor risks, marking the processed version as compliant). Regardless of the result, detailed logs of this inspection are fully recorded in the audit system, ensuring traceability throughout the entire process.

[0061] Step 105: Based on the virtual avatar, the compliant content results are synthesized into a virtual avatar broadcast video to obtain a digital human broadcast video stream.

[0062] Virtual avatar driving is a technological process that uses generated text or voice to control the appearance and behavior of a virtual digital human, enabling it to perform corresponding lip movements, facial expressions, and body movements. Digital human broadcasting video stream refers to a synthesized video file or real-time data stream containing a virtual avatar synchronously broadcasting compliant content. The terminal synthesizes speech from compliant content text using a deep neural network to generate a highly natural and emotionally expressive audio clip. It analyzes the text to determine appropriate speaking speed, intonation, and pauses, generating corresponding phoneme sequences. Using these phoneme sequences and time information, a lip-sync animation model drives the virtual avatar's face. This model calculates the most suitable lip shape for each moment of pronunciation. An emotional expression model analyzes the text's emotional tone, generating corresponding micro-expressions for the virtual avatar. To make the broadcast more natural, actions from a pre-defined body movement library are matched to the virtual avatar based on the text's rhythm and semantics. Optionally, slight gestures are added at the beginning of sentences or at points of emphasis, and natural body weight shifts and blinking are added during broadcast intervals to avoid a stiff appearance. All driving signals are sent to a graphics rendering engine, which renders the virtual avatar model in real-time according to these parameters, synthesizes it with possible background images, and precisely synchronizes the final image sequence with the audio stream, outputting a continuous digital human broadcast video stream that can be played directly on the local all-in-one terminal screen.

[0063] This embodiment provides a cloud-edge collaborative media all-in-one machine interactive control method, device, equipment, and medium. It acquires user request text data and, based on natural language processing, performs intent parsing on the user request text data to obtain the request object's intent. The request object's intent includes question-and-answer writing requests and topic planning requests. Based on the request object's intent, it retrieves corresponding relevant information from a vectorized knowledge base to obtain enhanced context information. The enhanced context information and the request object's intent are input into a specialized large-scale model for reasoning to obtain preliminary content results. Based on content security filtering rules, the preliminary content results are subjected to compliance verification to obtain compliant content results. Based on virtual avatar driving, the compliant content results are synthesized into a virtual avatar broadcast video to obtain a digital human broadcast video stream. Through the above methods, a local dedicated knowledge base can be constructed to ensure the accuracy and authority of the generated content. Introducing a large-scale model for content generation can meet the needs of real-time generation of recurring and large-scale content, flexibly displaying content to meet the public's needs for real-time Q&A and personalized interpretation, effectively realizing a flexible and rapid media all-in-one machine interactive control method.

[0064] In one embodiment, the vectorized knowledge base is constructed using the following methods:

[0065] Step 201: Obtain multi-format material data, and extract the text contained in the image data in the multi-format material data based on optical character recognition to obtain the first converted text.

[0066] Multi-format material data refers to the collected collection of raw materials, which takes various forms and typically includes image files such as scanned documents, charts, and promotional posters; audio data such as interview recordings; video data such as news videos and promotional films; and raw text data such as electronic documents and web articles. Optical character recognition (OCR) is a technology specifically designed to analyze images containing text, identify and extract characters from the images, and convert them into editable and searchable text formats. The first-converted text refers to the text content successfully extracted from the image data within the multi-format material data using OCR technology. The terminal filters out all image format files from the multi-format material data and calls the OCR engine to process each image. This engine performs image preprocessing, including noise reduction, contrast enhancement, layout analysis, character positioning, and character recognition, converting the visual symbols in the images into corresponding digital text characters and outputting structured text data, i.e., the first-converted text.

[0067] Step 202: Based on language recognition, convert the audio and video data in the multi-format material data into text to obtain the second converted text.

[0068] Specifically, speech recognition, also known as voice recognition technology, is a technology that automatically converts human speech signals into corresponding text. The second-converted text refers to the text content generated by converting audio data and audio track information contained in multi-format material data using speech recognition technology. The terminal separates audio and video files from the multi-format material data. For video files, audio-video separation is performed first, extracting the audio track. The pure audio data and the extracted audio data are then fed into a speech recognition engine. This engine performs noise reduction and segmentation on the audio, and identifies the speech content using acoustic and language models, ultimately generating the corresponding text record, i.e., the second-converted text.

[0069] Step 203: Integrate the original text from the first converted text, the second converted text, and the multi-format material data to obtain a plain text database.

[0070] Specifically, raw text refers to data in multi-format material that is already in text format and requires no conversion. The plain text database is a structured collection that integrates the first-converted text, the second-converted text, and the raw text. It serves as the central storage pool for all text data, preparing it for subsequent in-depth processing. The terminal gathers the first-converted text, the second-converted text, and the directly usable raw text from the original data, performing preliminary data cleaning tasks, including standardizing encoding formats, removing obvious garbled characters or irrelevant characters, and adding metadata identifiers to each data entry, ultimately constructing a complete and ordered plain text database.

[0071] Step 204: Based on the plain text database, perform semantic analysis and segmentation on the text to obtain text fragments, and convert each text fragment into a vector to obtain a set of vector pairs.

[0072] Semantic segmentation is a text processing technique that intelligently divides text based on semantic units to ensure that each text segment is semantically independent and complete. A text segment is the output of semantic segmentation, referring to a small piece of text with complete semantic meaning extracted from a plain text database. In the field of artificial intelligence, a vector represents an object using a set of high-dimensional numbers; semantically similar texts have vectors that are closer together in high-dimensional space. A vector pair set is a set where each element is a pair structure containing a text segment and its corresponding vector representation. The terminal loads text from a plain text database and uses a pre-trained natural language processing model to perform deep semantic analysis on the text, understanding its grammatical structure, contextual logic, and core semantic units. The model identifies natural pauses, thematic turning points, and semantically complete sub-parts in the text. Based on the results of the semantic analysis, a semantic slicing operation is performed. Optionally, a long report is intelligently cut into multiple semantically complete text segments according to its inherent reporting structure, ensuring that each segment is semantically self-consistent and avoiding splitting a complete semantic into two segments or forcibly merging unrelated semantics. For each segmented text, a specialized text embedding model is invoked. This model takes the text segment as input and outputs a fixed-length high-dimensional vector through complex neural network calculations. The transformation process essentially maps the deep semantic features of the text to a specific coordinate point in a mathematical space. Each text segment and its generated vector are stored in a one-to-one correspondence, forming a set of vector pairs. At this point, the original materials of the knowledge base are ready, but the optimal retrieval efficiency and quality have not yet been achieved.

[0073] Step 205: Based on the set of vector pairs, perform similarity comparison optimization to obtain a vectorized knowledge base.

[0074] Similarity matching optimization is an optimization process that improves the query efficiency and accuracy of the knowledge base by calculating the similarity between different vectors in a vector pair set, identifying and merging semantically highly repetitive segments, or adjusting the vector space distribution. A vectorized knowledge base is a high-performance database system specifically designed for storing and quickly retrieving vector pair sets. After optimization, it includes deduplicated, denoised, well-organized text segments and their vector representations. The terminal performs large-scale similarity calculations on vectors in the vector pair set, using algorithms such as cosine similarity to calculate the similarity score between each pair of vectors. The closer the score is to 1, the more semantically similar the text segments represented by the two vectors are. A similarity threshold is set. When the similarity between two vectors exceeds this threshold, the corresponding text segments are determined to be highly semantically repetitive. For duplicate items, a deduplication operation is performed, retaining only the segment with the most complete and highest quality information, while deleting or archiving other duplicate segments. For content with high similarity but not reaching the duplication threshold, strategies such as merging or creating related indexes are adopted to avoid returning a large number of identical results during retrieval, ensuring the diversity and coverage of the returned information. After deduplication and merging are completed, vector database technology is used to create an efficient index for all remaining high-quality vectors. The index structure allows the database to quickly find the few results most similar to the query vector from millions of vectors. The deduplicated and optimized text segments and their vectors, along with the created efficient indexes, are stored in the database to form the final vectorized knowledge base.

[0075] This embodiment constructs a vectorized knowledge base to provide accurate and relevant knowledge fragments for enhanced retrieval generation, directly ensuring the accuracy and authority of question-answering and content generation services.

[0076] In one embodiment, the specialized large model is constructed through the following methods:

[0077] Step 301: Based on the plain text database, perform domain text filtering to obtain a refined domain corpus.

[0078] The plain text corpus refers to a unified collection of text data obtained after preliminary processing. Its content may cover multiple fields, and the quality varies. Domain text filtering is a data filtering and selection process. Its core is to identify and extract the most suitable subset of texts for the target domain from the plain text corpus based on preset domain relevance and quality indicators. The domain-refined corpus is a high-quality, highly clean collection of texts, whose content is closely related to the target domain and meets established quality standards. The process involves defining the scope and core keywords of the target domain, employing multiple strategies for parallel filtering, and quickly filtering out obviously relevant documents based on a defined keyword list and regular expression rules. A text classification model is used to classify the text in the plain text corpus by domain, automatically determining whether it belongs to the target domain and discovering more implicitly relevant texts. The filtered text undergoes quality evaluation, removing texts that are too short, logically incoherent, contain a large amount of garbled text, or contain irrelevant advertisements. The high-quality domain texts filtered through various strategies are then aggregated and deduplicated to form a clean, refined domain corpus.

[0079] Step 302: Based on the domain-refined corpus, design input instructions and prompt words to obtain a domain-adapted instruction template library; the domain-adapted instruction template library is used to train the model to follow the specifications of the corresponding domain.

[0080] Specifically, input instructions and prompts refer to carefully designed text prompts to guide large models in completing specific tasks. These include task descriptions, contextual information, and output format requirements, serving as the interface specifications for user-model interaction. The domain-adapted instruction template library is a structured collection containing numerous paired examples of instructions and ideal outputs designed for various typical tasks within the target domain. Based on a refined domain corpus and business requirements, the terminal identifies all common task types within the domain and automatically generates a large number of high-quality instruction templates for each type. Each template includes the instruction (a clear task description), the input (simulated or real task input content), and the ideal output (a standard answer example conforming to domain norms and styles). The template library systematically organizes all the designed instruction templates to form the domain-adapted instruction template library, essentially serving as model training material for that domain.

[0081] Step 303: Input the examples from the domain adaptation instruction template library into the base model, perform context learning on the base model, and obtain the transition model.

[0082] Specifically, a base model refers to a large language model that has been pre-trained on massive amounts of general-purpose data, possesses strong basic language capabilities, and has broad knowledge, but has not yet been optimized for a specific domain. Contextual learning is a training technique that guides the model to generate responses by providing a small number of task examples in the input prompts without changing the model's internal parameters, allowing it to imitate the style and format of the examples. An interim model, after extensive contextual learning training, is a transitional model whose behavior, while unchanged, is better able to understand and execute domain-specific tasks, and whose performance is closer to the target. The terminal extracts several high-quality instruction, input, and ideal output examples from the domain-adaptive instruction template library. These examples cover one or more typical tasks within the target domain. The examples are then concatenated according to a specific format to form a complete prompt text. This concatenated long prompt text is input into the base model. Based on its powerful attention mechanism, the model analyzes the entire context, identifies the implicit task patterns, output formats, language styles, and technical terms in the examples, and generates a response by imitating the style and logic of the previous examples based on its understanding of the last current task instruction. The model is repeatedly practiced using a large amount of data from the template library, allowing it to continuously encounter various correct ways to complete tasks, thereby strengthening its behavior patterns within the domain.

[0083] Step 304: Based on the domain-refined corpus and efficient parameter fine-tuning, a portion of the parameters of the transition model are lightly adjusted to obtain the fine-tuned model.

[0084] Among them, parameter-efficient fine-tuning is a type of fine-tuning technique. Its core idea is to adjust the model's behavior by introducing a small number of additional trainable parameters or updating only a portion of the key parameters in the model, which can significantly reduce computational costs and resource consumption. The fine-tuned model refers to the model obtained after parameter-efficient fine-tuning of the transitional model. Its internal parameters have been updated, giving it stronger and more stable domain-specific capabilities. The terminal uses the content from the domain-refined corpus and the domain-adapted instruction template library as training data for fine-tuning. A parameter-efficient fine-tuning method is selected to inject some trainable low-rank adapter parameters into the core modules of the model. The model is trained using the training data. During training, most of the original parameters of the model are frozen, and only the newly added few adapter parameters are updated. Through learning, these adapter parameters are adjusted to a state that makes its output more consistent with the requirements of the domain corpus and instruction templates. After training, the adapter parameters are merged back into the original model to obtain the fine-tuned model.

[0085] Step 305: Based on the test case set and security rules, calibrate the output of the fine-tuned model to obtain a specialized large model; the test case set includes regular task test cases and potential risk test cases.

[0086] The test case set is a dataset specifically designed to evaluate model performance. It contains two types of test cases: regular task test cases, which test the model's performance on normal domain tasks, such as question-answering accuracy and writing quality; and potential risk test cases, which are carefully designed test questions that may induce the model to produce harmful, biased, unsafe, or inaccurate responses, used to test the model's security defenses. The security rules are a set of explicit content security specifications that define what content is unacceptable. The specialized large model is a domain-optimized, rigorously security-calibrated, and practically usable secure and reliable large model. The terminal will conduct large-scale, multi-round automated testing and manual evaluation on the test case set after fine-tuning the model. It will record the model's performance on all test cases, especially its failures on potentially risky test cases. The test results will be analyzed in depth to accurately identify which types of tasks or specific topics the model is prone to producing non-compliant or risky outputs. Based on the discovered deviations and risk points, targeted calibration measures will be taken. For risky test cases where the model fails, more safe positive examples and dangerous negative examples will be constructed for additional, targeted fine-tuning of the model. During each inference, an invisible prompt emphasizing security and compliance requirements will be added before or after the user's instruction to constantly remind the model to follow specifications. A rule-based or lightweight model-based security filter will be deployed at the model's output to scan the generated content in real time. Once high-risk content is detected, it will be intercepted, replaced, or marked. This process will be iterated multiple times, with retesting after each calibration, until the model maintains high performance on regular tasks while stably and reliably passing all potential risk tests and providing safe and appropriate responses to various sensitive or dangerous questions.

[0087] This embodiment ensures that the model can resist malicious guidance in complex real-world application environments, avoids generating content that may cause public opinion or compliance risks, and improves the security of interaction.

[0088] In one embodiment, based on the intent of the requesting object, relevant information is retrieved from a vectorized knowledge base to obtain enhanced contextual information, including:

[0089] Step 401: Based on the intent of the requesting object, search for document fragments containing keywords from the vectorized knowledge base to obtain keyword search results.

[0090] The request intent refers to the core purpose and category of the user request obtained after natural language processing parsing, which includes keywords or core concepts that can represent that intent. Although named vectorized, the vectorized knowledge base still stores original text fragments at its core; it is a knowledge base storing massive amounts of text fragments and their corresponding vector representations. Keyword retrieval results refer to the set of one or more document fragments containing keywords found in the original text of the vectorized knowledge base by matching the keywords in the request intent. The terminal extracts core keywords from the request intent and uses these keywords to perform matching queries in the text index of the vectorized knowledge base. The matching is literal matching, which can be supplemented by basic natural language processing techniques such as stemming or synonym expansion to improve recall. All document fragments containing the keywords or their synonyms will be retrieved, forming preliminary keyword retrieval results.

[0091] Step 402: Convert the request object intent into a vector to obtain the request vector. Then, using the following formula, retrieve the document fragment with the highest cosine similarity to the request vector from the vectorized knowledge base to obtain the vector semantic retrieval result:

[0092]

[0093] in, For cosine similarity, For query vector, This is a document vector.

[0094] Specifically, a request vector refers to a high-dimensional numerical vector that transforms the complete textual description of the request's intent into an embedding model. Mathematically, this vector represents the deep semantic features of the user's request. Cosine similarity is a mathematical method for measuring the directional difference between two vectors, with a value ranging from -1 to 1. The closer the value is to 1, the more consistent the directions of the two vectors, meaning the texts they represent are semantically similar. The formula is the dot product of the two vectors divided by the product of their magnitudes. Vector semantic retrieval results refer to the set of the top K document fragments with the highest cosine similarity scores, selected by calculating the cosine similarity between the request vector and the vectors of all document fragments in the vectorized knowledge base. The terminal uses the same embedding model as when building the knowledge base to transform the complete textual description of the request's intent into a request vector, calculates the cosine similarity between the request vector and each document vector in the knowledge base, sorts them according to the similarity score, and selects the document fragments with the highest scores as the vector semantic retrieval results.

[0095] Step 403: Deduplicate the keyword search results and vector semantic search results to obtain a set of candidate document fragments.

[0096] The candidate document fragment set refers to the unified and deduplicated set of document fragments obtained by merging the keyword search results and vector semantic search results, and then removing completely duplicate or highly duplicated fragments. The terminal merges the keyword search results and vector semantic search results sets, and then performs deduplication on the merged set. Deduplication can be based on the unique identifier of the document fragment, or by calculating text fingerprints to determine whether the content is completely identical or highly similar, ensuring that each document fragment in the set is unique.

[0097] Step 404: Based on the reordering model, perform a secondary sort on the candidate document fragment set to obtain a reordered document fragment list.

[0098] The re-ranking model is a machine learning model specifically designed to measure the relevance between queries and documents. It is more complex and accurate than cosine similarity calculation, but also more computationally expensive. It can comprehensively consider more subtle semantic relationships, therefore it is not used for full-database scanning, but only for accurately scoring a small number of recalled candidate documents. The re-ranked document fragment list refers to an ordered list of document fragments, arranged from highest to lowest relevance, obtained after the candidate document fragment set has been recalculated and sorted by the re-ranking model. The terminal uses the user's request intent as a constant query endpoint. It sequentially extracts each document fragment from the candidate document fragment set as a document endpoint, forming a (query, document) pair for each candidate fragment. Each pair is simultaneously input into the re-ranking model, which allows for deep interaction between the query and document texts within the model. The model leverages its powerful attention mechanism to analyze the semantic relationships between each word in the query and each word in the document. The re-ranking model outputs a precise relevance score for each pair, which is not simply a probabilistic judgment about whether the document adequately answers the query, but rather a score derived from training on massive amounts of data. The re-ranking scores of all candidate fragments are collected, and based on this new, more precise score, all candidate fragments are sorted in descending order. The fragment with the highest score, indicating that the re-ranking model determines it is most relevant to the user's intent, is therefore placed at the top of the list. After sorting, a final re-ranked document fragment list is generated. The order in this list is considered to reflect the true relevance more accurately than the initial retrieval results.

[0099] Step 405: Based on the reordered list of document fragments, combine the document fragments into contextual text blocks to obtain enhanced contextual information.

[0100] Specifically, a contextual text block refers to a coherent text formed by concatenating the top-ranked fragments from the reordered document fragment list in a logical order. Enhanced contextual information is the contextual text block, which contains background knowledge most relevant to the user's request, obtained through layers of filtering and optimization. Starting from the top of the reordered document fragment list, the terminal sequentially selects a specified number of document fragments, simply concatenates them end-to-end, and adds line breaks and other separators between fragments to form a complete text block. This text block, the enhanced contextual information, will be used as key input and sent along with the user request to a specialized large-scale model.

[0101] This embodiment addresses the semantic gap and improves context quality by enhancing context, reducing interference from irrelevant information. It provides high-quality, highly relevant external knowledge for specialized large models, enabling them to reason and create answers based on solid factual information, thus ensuring the accuracy and authority of the output content.

[0102] In one embodiment, if the request object's intent is a topic selection and planning request, before searching for document fragments containing keywords from the vectorized knowledge base based on the request object's intent and obtaining keyword search results, the process further includes:

[0103] Step 501: Based on the intent of the requesting object, generate a topic planning request signal and send the topic planning request signal to the cloud; the topic planning request signal is used to instruct the cloud to carry out topic planning and to provide a topic planning report.

[0104] The request object intent refers to the core purpose of the user request obtained after natural language processing parsing; in this embodiment, it has been identified as a topic planning request. The topic planning request signal is a structured data packet or instruction signal that encapsulates the key information required for topic planning, including the core theme, keywords, time frame requirements for analysis, and contextual information extracted from the request object intent. The cloud refers to an analysis platform located on a remote server with powerful computing capabilities and access to massive amounts of internet data, responsible for performing large-scale data capture and analysis tasks that cannot be completed by a local all-in-one machine. The terminal parses and encapsulates the request object intent, packages it together with the request's metadata, generates a standardized topic planning request signal, and sends the request signal to the designated cloud service interface through a secure network communication protocol, ensuring accurate delivery of the request and initiating the cloud analysis task.

[0105] Step 502: Receive the topic planning report, and based on the topic planning report, transform the hot topics into specific questions to obtain a set of search intents.

[0106] Specifically, the topic selection and planning report is a structured analysis result that includes a list of trending topics identified from the internet within a specified time and theme, the popularity index of each topic, sentiment analysis, and relevant representative articles or abstracts. Trending topics refer to specific events or themes identified in the topic selection and planning report that have received widespread public attention and discussion within a certain period. The search intent set refers to the set formed by converting each trending topic in the topic selection and planning report into a standard question format that the system can understand and process. The terminal receives the topic selection and planning report returned from the cloud, parses the report, extracts the core parts, iterates through each trending topic, and uses a rule engine model to transform the topic into one or more clear, answer-oriented questions. All generated questions are then aggregated to form the search intent set. The essence of this set is a precise set of search instructions that directly correspond to relevant knowledge fragments that may exist in the local knowledge base.

[0107] Step 503: Add the set of search intents to the request object intent to obtain the updated request object intent.

[0108] Specifically, the updated request object intent is an enhanced and comprehensive request intent formed by incorporating the retrieval intent set into the original request object intent. The terminal logically merges the retrieval intent set with the request object intent, and the updated intent serves as a new starting point for step 401.

[0109] This embodiment ensures that the topic selection and planning schemes provided to users reflect their initial direction and closely integrate with the latest social concerns, so that the generated content is both in-depth, timely, and relevant to current events, thereby improving the targeted nature of the interaction.

[0110] In one embodiment, after the cloud receives the topic selection and planning request signal, it further includes:

[0111] Step 601: Obtain the original Internet dataset and preprocess it to obtain high-quality text corpus; the original Internet dataset includes text data, image data and video data.

[0112] The raw internet dataset refers to the collection of raw data obtained in real-time or periodically from public network sources such as news websites, social media platforms, forums, and video websites through various technologies such as web crawlers and data interfaces. It takes various forms, including webpage text, images, short videos, blog posts, and comments. Preprocessing is a complete set of data cleaning and standardization processes aimed at filtering out noise from the raw internet dataset and transforming it into a clean, consistent text format. High-quality text corpus is a collection of pure text data that has undergone rigorous cleaning, deduplication, and format standardization; it serves as the foundational material for subsequent accurate natural language analysis and mining. Based on the theme and time range specified in the received topic planning request signal, the cloud mobilizes the corresponding data collection modules to selectively crawl data from multiple internet sources, forming an initial raw internet dataset. For text data, the core content is directly extracted, and irrelevant information such as advertisements, navigation bars, and script code is removed. For image data, optical character recognition technology is used to extract text from the images. For video data, the audio track is first separated, and then the audio is converted into text subtitles using speech recognition technology. Completely duplicated or highly similar entries are identified and deleted, and meaningless characters, garbled text, emoticons, and dialogue watermarks such as forwards and replies are filtered out. From unstructured text, the core metadata such as the main text, publication time, author, and source is identified and extracted. All text is uniformly encoded and subjected to necessary paragraph division and sentence segmentation. Based on rules such as text length, language fluency, and information completeness, low-quality segments are filtered out, forming a high-quality text corpus that can be analyzed.

[0113] Step 602: Identify high-frequency words in high-quality text corpora and categorize them to obtain a list of hot topics.

[0114] Specifically, high-frequency words refer to words or phrases that appear significantly more frequently than in everyday life in high-quality text corpora, and are used to refer to specific events, entities, or emotions. The hot topic list is a list of topics with clear themes, formed after merging and semantic understanding. Each topic in the list contains a core theme statement and its related keyword group. The terminal performs word segmentation on the high-quality text corpus and uses statistical or deep learning models to identify high-frequency words and phrases with abnormally high frequency and attention within a specified time window. By analyzing co-occurrence relationships and semantic associations, clustering algorithms are used to automatically merge related words and text fragments into the same semantic cluster. Each cluster represents a potential hot topic. For each semantic cluster, the algorithm automatically extracts a theme name that summarizes its core content and lists its core keywords, representative statements, and discussion volume, forming a structured hot topic list.

[0115] Step 603: Based on the list of trending topics, conduct sentiment analysis to obtain a trend analysis report.

[0116] Specifically, sentiment analysis is a natural language processing technique used to determine whether the subjective emotion expressed in a text is positive, negative, or neutral, and sometimes further subdivided into more specific emotions such as joy, anger, disappointment, and expectation. A trend analysis report is a statistical summary and description of the overall sentiment across all related discussions under a trending topic. The report quantifies the distribution ratio of various sentiments and may describe their trends over time. In the cloud, for each topic in the trending topic list, all relevant texts under that topic are extracted and input into the sentiment analysis model. Based on deep learning technology, the model can accurately understand the true emotions behind complex expressions by combining context, summarize the judgment results of all text fragments under that topic, perform statistical calculations, analyze the change curve of sentiment intensity over time, and generate a descriptive analysis of the public opinion field for that topic based on the statistical data.

[0117] Step 604: Integrate the list of hot topics and the trend analysis report to obtain the topic planning report.

[0118] Specifically, the topic planning report is a structured document that integrates all the key findings from the list of trending topics and the trend analysis report, aiming to provide direct and actionable insights for topic planning on local terminal all-in-one devices. The cloud links and merges the list of trending topics with the trend analysis report. For example, each trending topic is treated as an independent chapter, containing three sub-sections: popularity analysis, sentiment analysis, and opinion summary. Following a predefined, machine-readable format, the structured information is encapsulated into a complete topic planning report. The report includes metadata such as an execution summary, analysis timeframe, and data source descriptions. The cloud then sends this generated report back to the original requesting local terminal all-in-one device via a data interface, completing a full edge-cloud collaborative analysis task.

[0119] This embodiment enables local terminals to quickly and accurately form specific search intents through end-to-cloud collaboration, ultimately helping to generate high-quality content that is timely and sensitive to trending topics.

[0120] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0121] Based on the same inventive concept, this application also provides an end-to-end cloud-coordinated media all-in-one machine interactive control device for implementing the aforementioned end-to-end cloud-coordinated media all-in-one machine interactive control method. The solution provided by this device is similar to the implementation scheme described in the above method. Therefore, the specific limitations of one or more end-to-end cloud-coordinated media all-in-one machine interactive control device embodiments provided below can be found in the limitations of the end-to-end cloud-coordinated media all-in-one machine interactive control method described above, and will not be repeated here.

[0122] In one exemplary embodiment, such as Figure 2 As shown, a cloud-edge collaborative media all-in-one machine interactive control device 700 is provided, including:

[0123] The intent module 701 is used to acquire user request text data and perform intent parsing on the user request text data based on natural language processing to obtain the intent of the request object; the intent of the request object includes question and answer writing requests and topic planning requests.

[0124] Enhancement module 702 is used to retrieve relevant information from the vectorized knowledge base based on the intent of the request object to obtain enhanced context information;

[0125] The reasoning module 703 is used to input enhanced contextual information and the intent of the requesting object into a specialized large model for reasoning to obtain preliminary content results;

[0126] The verification module 704 is used to perform compliance verification on the preliminary content results based on content security filtering rules to obtain compliant content results.

[0127] The broadcasting module 705 is used to synthesize compliant content results into a virtual avatar broadcasting video based on the virtual avatar, thus obtaining a digital human broadcasting video stream.

[0128] Furthermore, the device also includes a building module for:

[0129] Acquire multi-format material data, and based on optical character recognition, extract the text contained in the image data in the multi-format material data to obtain the first converted text;

[0130] Based on language recognition, the audio and video data in multi-format material data are converted into text to obtain the second converted text.

[0131] By integrating the original text from the first transformed text, the second transformed text, and the multi-format material data, a plain text database is obtained.

[0132] Based on a plain text database, semantic analysis is performed on the text to obtain text segments, and each text segment is converted into a vector to obtain a set of vector pairs;

[0133] Based on the set of vector pairs, similarity comparison and optimization are performed to obtain a vectorized knowledge base.

[0134] Furthermore, building modules are also used for:

[0135] Based on a plain text database, domain-specific texts are filtered to obtain a refined domain-specific corpus.

[0136] Based on a domain-refined corpus, input commands and prompt words are designed to obtain a domain-adapted command template library; the domain-adapted command template library is used to train the model to follow the specifications of the corresponding domain.

[0137] The examples from the domain adaptation instruction template library are input into the base model, and the base model is subjected to context learning to obtain the transition model;

[0138] Based on a domain-refined corpus and efficient parameter fine-tuning, a portion of the parameters of the transition model are lightly adjusted to obtain the fine-tuned model.

[0139] Based on the test case set and security rules, the output of the fine-tuned model is calibrated to obtain a specialized large model; the test case set includes routine task test cases and potential risk test cases.

[0140] Furthermore, enhancement module 702 is also used for:

[0141] Based on the intent of the requesting object, document fragments containing keywords are searched from the vectorized knowledge base to obtain keyword search results;

[0142] The request intent is converted into a vector to obtain the request vector. Then, the document fragment with the highest cosine similarity to the request vector is retrieved from the vectorized knowledge base using the following formula to obtain the vector semantic retrieval result:

[0143]

[0144] in, For cosine similarity, For query vector, Document vector;

[0145] Deduplication is performed on the keyword search results and the vector semantic search results to obtain a set of candidate document fragments;

[0146] Based on the reordering model, the candidate document fragment set is sorted twice to obtain a reordered document fragment list.

[0147] Based on the reordered list of document fragments, the document fragments are combined into contextual text blocks to obtain enhanced contextual information.

[0148] Furthermore, enhancement module 702 is also used for:

[0149] Based on the intent of the requesting object, a topic planning request signal is generated and sent to the cloud; the topic planning request signal is used to instruct the cloud to carry out topic planning and to provide a topic planning report.

[0150] Receive topic planning reports, and based on these reports, transform hot topics into specific questions to obtain a set of search intents;

[0151] Add the set of search intents to the request object intent to obtain the updated request object intent.

[0152] Furthermore, the device also includes a public opinion monitoring module, located in the cloud, for:

[0153] The original internet dataset was obtained and preprocessed to obtain high-quality text corpus; the original internet dataset includes text data, image data and video data.

[0154] Identify high-frequency words in high-quality text corpora, categorize the high-frequency words, and obtain a list of hot topics;

[0155] Based on the list of trending topics, a sentiment analysis is conducted to obtain a trend analysis report;

[0156] By integrating the list of trending topics and trend analysis reports, a topic planning report is obtained.

[0157] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps of the aforementioned end-to-cloud collaborative media all-in-one machine interactive control method.

[0158] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0159] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0160] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A method for interactive control of an all-in-one media machine with end-to-end cloud collaboration, characterized in that, The method includes: The system acquires user request text data and performs intent parsing on the user request text data based on natural language processing to obtain the intent of the request object; the intent of the request object includes question-and-answer writing requests and topic planning requests. Based on the intent of the requesting object, relevant information is retrieved from the vectorized knowledge base to obtain enhanced context information; The enhanced context information and the intent of the requesting object are input into a specialized large model for reasoning to obtain preliminary content results; Based on content security filtering rules, the preliminary content results are subjected to compliance checks to obtain compliant content results; Driven by a virtual avatar, the compliant content results are synthesized into a virtual avatar broadcast video, resulting in a digital human broadcast video stream.

2. The method according to claim 1, characterized in that, The vectorized knowledge base is constructed using the following methods, including: Acquire multi-format material data, and based on optical character recognition, extract the text contained in the image data in the multi-format material data to obtain the first converted text; Based on language recognition, the audio and video data in the multi-format material data are converted into the text to obtain the second converted text; By integrating the first converted text, the second converted text, and the original text from the multi-format material data, a plain text database is obtained. Based on the plain text database, the text is semantically analyzed and sliced ​​to obtain text segments, and each text segment is converted into a vector to obtain a set of vector pairs. Based on the set of vector pairs, similarity comparison and optimization are performed to obtain the vectorized knowledge base.

3. The method according to claim 2, characterized in that, The specialized large model is constructed using the following methods, including: Based on the aforementioned plain text database, domain-specific text filtering is performed to obtain a refined domain-specific corpus; Based on the refined corpus of the aforementioned domain, input instructions and prompt words are designed to obtain a domain-adapted instruction template library; the domain-adapted instruction template library is used to train the model to follow the specifications of the corresponding domain. The examples in the domain adaptation instruction template library are input into the base model, and the base model is subjected to context learning to obtain the transition model; Based on the refined corpus of the aforementioned domain and efficient parameter fine-tuning, a portion of the parameters of the transition model are lightly adjusted to obtain the fine-tuned model; Based on the test case set and security rules, the output of the fine-tuned model is calibrated to obtain the specialized large model; the test case set includes regular task test cases and potential risk test cases.

4. The method according to claim 1, characterized in that, The step of retrieving relevant information from a vectorized knowledge base based on the intent of the request object to obtain enhanced context information includes: Based on the intent of the requesting object, document fragments containing keywords are searched from the vectorized knowledge base to obtain keyword search results; The intent of the request object is converted into a vector to obtain a request vector. Then, the document fragment with the highest cosine similarity to the request vector is retrieved from the vectorized knowledge base using the following formula to obtain the vector semantic retrieval result: in, For cosine similarity, For query vector, Document vector; The keyword search results and the vector semantic search results are deduplicated to obtain a set of candidate document fragments; Based on the reordering model, the candidate document fragment set is sorted a second time to obtain a reordered document fragment list; Based on the reordered list of document fragments, the document fragments are combined into contextual text blocks to obtain the enhanced contextual information.

5. The method according to claim 4, characterized in that, If the intent of the request object is a topic selection and planning request, then before searching for document fragments containing keywords from the vectorized knowledge base based on the intent of the request object to obtain keyword search results, the process further includes: Based on the intent of the requesting object, a topic selection planning request signal is generated and sent to the cloud; the topic selection planning request signal is used to instruct the cloud to perform topic selection planning and to provide a topic selection planning report. Receive the topic selection and planning report, and based on the topic selection and planning report, transform hot topics into specific questions to obtain a set of search intents; The search intent set is added to the request object intent to obtain the updated request object intent.

6. The method according to claim 5, characterized in that, After receiving the topic selection and planning request signal in the cloud, it also includes: Obtain the raw internet dataset and preprocess it to obtain high-quality text corpus; the raw internet dataset includes text data, image data, and video data. High-frequency words in the high-quality text corpus are identified and categorized to obtain a list of hot topics; Based on the list of trending topics, a sentiment analysis is performed to obtain a trend analysis report; The topic selection and planning report is obtained by integrating the list of hot topics and the trend analysis report.

7. A cloud-edge collaborative media all-in-one machine interactive control device, characterized in that, The device includes: The intent module is used to acquire user request text data and, based on natural language processing, perform intent parsing on the user request text data to obtain the intent of the request object; the intent of the request object includes question-and-answer writing requests and topic planning requests. The enhancement module is used to retrieve relevant information from the vectorized knowledge base based on the intent of the request object to obtain enhanced context information; The reasoning module is used to input the enhanced context information and the intent of the requesting object into a specialized large model for reasoning to obtain preliminary content results; The verification module is used to perform compliance verification on the preliminary content results based on content security filtering rules to obtain compliant content results; The broadcasting module is used to synthesize the compliant content results into a virtual avatar broadcasting video based on the virtual avatar, thereby obtaining a digital human broadcasting video stream.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images