Agent autonomous planning multi-mode large model content creation method, device and system

Abstract

The invention provides an Agent autonomous planning multi-mode large model content creation method, device and system, and relates to the technical field of large model content creation, and the method comprises the steps: generating a venation structure skeleton of a to-be-created content according to a writing requirement and a reference file; performing related information data collection according to the plurality of to-be-retrieved questions corresponding to each piece of paragraph information, and generating paragraph content corresponding to each piece of paragraph information; generating a chart or a picture corresponding to each piece of anchor point information according to each piece of anchor point information; obtaining each sentence needing reference calculation in each paragraph content to obtain a reference sentence list A; traversing the similarity list eta, and if eta i is greater than eta ', determining that Ai is not abnormal; generating a multi-modal long text corresponding to the writing requirement input by the user; proofreading the written content; according to the method, the intelligence, the reliability and the diversity of AI writing are remarkably improved by innovatively combining dynamic knowledge acquisition, multi-modal generation, long writing planning and content credibility guarantee.

Description

Technical Field

[0001] The present invention relates to the technical field of large model content creation, and in particular to an Agent autonomously planned multi-modal large model content creation device and system. Background Art

[0002] AI writing technology mainly relies on pre-trained language models in the field of deep learning, such as the GPT (Generative Pre-trained Transformer) series. Through the training of large-scale corpora, the model has powerful language generation capabilities. Its core mechanism includes the understanding and generation of natural language, and generates text content that conforms to the context logic by inputting a small number of prompts. This type of technology is widely used in scenarios such as news writing, technical document generation, and e-commerce copywriting. The current AI writing process usually consists of the following key steps: First, the user enters the initial requirements, such as titles, topics, or keywords, and the system parses the input and encodes it into a vector that the model can process; then, the model predicts the possible output text at each step based on the input vector, and gradually generates a complete paragraph or article; finally, the system performs simple post-processing on the generated results, such as typesetting optimization or error correction. In order to improve the generation quality, some technical solutions have also introduced prompt engineering and fine-tuning, which further enhance the model's content generation capabilities by adjusting the input template or conducting additional training for specific fields.

[0003] AI writing technology has shown significant advantages in language fluency and generation efficiency, and has become an important force in promoting the development of the field of automated content generation. Although existing AI writing technology relies on large-scale pre-trained language models (such as the GPT series) and has made significant progress in the fluency and efficiency of text generation, it still has many technical defects and cannot fully meet the needs of complex writing scenarios. These defects are mainly concentrated in knowledge coverage capabilities, real-time performance, quality and length limitations of generated content, and insufficient multimodal support. The details are as follows:

[0004] 1. The model’s knowledge capacity is limited and cannot fully cover external knowledge

[0005] The knowledge reserve of pre-trained language models comes from large-scale data sets, which are static information collected at a specific point in time. Therefore, no matter how large the parameter scale of the model is, its knowledge is limited by the scope of the training data set, and it is impossible to fully cover the knowledge in all fields. For specific professional fields or the latest research progress, the model often cannot generate accurate or complete content. In addition, since the knowledge embedded in the model cannot be dynamically updated after training is completed, when faced with questions raised by users that exceed its knowledge reserves, the generated content may have obvious errors or omissions, which makes it difficult to rely solely on model-generated content to meet complex writing tasks that require high precision and extensive knowledge support. Therefore, the existing technology urgently needs to make up for the knowledge limitations of the model through real-time supplementation of external knowledge.

[0006] 2. The model is not real-time enough and the knowledge version cannot be kept up to date

[0007] After the current language model is trained, its built-in knowledge becomes a static "snapshot" that only reflects the state when the training data was collected. This "lag" in knowledge makes it difficult for the model to cope with writing tasks that require real-time updates. For example, when writing about the latest current affairs, technological trends, or real-time market reports, the results generated by the model are likely to be based on outdated information and do not match the actual situation. In addition, knowledge fields change rapidly, such as finance, healthcare, artificial intelligence, etc., where important information may be updated in hours or days, but the model cannot be synchronized. Without the ability to dynamically acquire the latest external knowledge, the existing AI writing system performs poorly in scenarios with high timeliness requirements and is difficult to meet the actual needs of users.

[0008] 3. The model output window length is limited, and the quality of content generation is limited

[0009] The output of language models is often limited by their architectural design, especially when generating long content. Currently, the output window length of most mainstream language models is around 2,000 words. When the output length exceeds this limit, the quality of the generated content usually drops rapidly, with confusing logic, incoherent context, and even meaningless repetitive content. In addition, the model has limited support for complex writing tasks. For example, when it is necessary to generate content with complex format, long length, and clear logic, the model cannot guarantee that the generated results are well structured and consistent, which is mostly due to the limited length of sample content during training and the design limitations of the model architecture. Users have to rely on manual editing or multiple rounds of input adjustments to make up for this shortcoming, which greatly reduces the efficiency of content generation.

[0010] 4. The model cannot directly generate full-length content

[0011] Existing language models are generally good at generating single paragraphs or short pieces of content, but when constructing complete articles or long reports, the models will expose significant shortcomings. Such shortcomings include logical jumps, lengthy content and lack of new information points, and lack of diversity in generated content. Especially when complex information needs to be integrated, models often have difficulty generating high-quality overall content. For example, tasks such as academic papers, technical documents, or business plans require content to have a clear chapter structure and logical progression, which exceeds the generation capabilities of current models. In addition, the model is mainly good at processing plain text content, and existing models are almost unable to directly generate multimodal outputs such as tables and pictures required by users. This limitation restricts the application scenarios of AI writing technology, making it difficult to cope with tasks that require a high diversity of output content forms.

[0012] 5. The credibility of the content generated by the model is low, and it is difficult to judge whether it is correct or not

[0013] Current AI writing technology has significant deficiencies in the credibility of generated content, and users cannot directly judge the accuracy and reliability of generated content. This is mainly due to the "AI hallucination" of the model, that is, the model generates content that seems to be true but is actually wrong or unfounded without sufficient support. For example, when generating content involving professional knowledge or precise data, the model may fabricate information or provide wrong answers, and these contents are easy for users to mistakenly believe that they are credible due to the fluency of language expression. In addition, the model has no built-in verification mechanism to proofread the generated results, nor can it provide information sources or references, which requires users to spend extra time and energy on manual verification when using the content generated by the model, greatly reducing writing efficiency and user experience. Especially in application scenarios that are highly dependent on accuracy (such as academic writing, legal documents, or medical reports), the credibility of model-generated content has become a key obstacle to its widespread application. Therefore, how to enhance the credibility of model-generated content, or provide reliability assurance by combining external knowledge and verification mechanisms, is an important issue that AI writing technology urgently needs to solve. Summary of the invention

[0014] In view of the above technical problems, the technical solution adopted by the present invention is:

[0015] According to a first aspect of the present application, there is provided an agent autonomously planning multimodal large model content creation method, the method comprising the following steps:

[0016] S100, generating a context structure skeleton of the content to be created according to the writing requirements and reference files input by the user; wherein the context structure skeleton includes a number of paragraph information and a number of multimodal anchor point information corresponding to the content to be created; the multimodal anchor point information is used to represent the corresponding chart or image position; each paragraph information includes a number of questions to be retrieved.

[0017] S200, traverse each paragraph information, input the reference content corresponding to each paragraph information, the paragraph content of the previous paragraph and the context structure skeleton of the full text into the preset large language model LLM for paragraph content creation, so as to obtain the paragraph content corresponding to each paragraph information; wherein, the reference content corresponding to each paragraph information is obtained by retrieving the corresponding question to be searched.

[0018] S300, traversing each anchor point information, and generating a chart or picture corresponding to each anchor point information according to each anchor point information.

[0019] S400, obtaining each quoted sentence in each paragraph content to obtain a quoted sentence list A=(A 1 , A 2 , …, A i , …, A n ), i=1, 2,...,n; among them, A i is the i-th quoted sentence obtained, and n is the number of quoted sentences obtained.

[0020] S500, obtaining the similarity between each quoted sentence in A and the corresponding reference content, so as to obtain a similarity list corresponding to A η = (η 1 , η 2 ,…,η i ,…,η n ), where η i A i The similarity between the content and the corresponding reference content.

[0021] S600, traverse η, if η i >η', then determine A i No exception; otherwise, confirm A i Exception; wherein η' is a preset sentence similarity threshold.

[0022] S700, if each quoted sentence in A is normal, then the content of each paragraph is verified to generate a multimodal long text corresponding to the writing requirement input by the user.

[0023] According to another aspect of the present application, there is also provided an agent autonomously planning multimodal large model content creation device, the device comprising:

[0024] The context structure skeleton generation agent is used to generate the context structure skeleton of the content to be created according to the writing requirements and reference files input by the user; wherein the context structure skeleton includes a number of paragraph information and a number of multimodal anchor point information corresponding to the content to be created; the multimodal anchor point information is used to represent the corresponding chart or image position; each paragraph information includes a number of questions to be retrieved.

[0025] The paragraph traversal writing agent traverses each paragraph information, and inputs the reference content corresponding to each paragraph information, the paragraph content of the previous paragraph, and the context structure skeleton of the full text into the preset large language model LLM for paragraph content creation, so as to obtain the paragraph content corresponding to each paragraph information; among which, the reference content corresponding to each paragraph information is obtained by retrieving the corresponding question to be searched.

[0026] The multimodal content processing agent is used to traverse each anchor point information and generate a chart or picture corresponding to each anchor point information according to each anchor point information.

[0027] The quoted sentence acquirer is used to acquire each quoted sentence in each paragraph content to obtain a quoted sentence list A=(A 1 , A 2 , …, A i , …, A n ), i=1, 2,...,n; where, A i is the i-th quoted sentence obtained, and n is the number of quoted sentences obtained.

[0028] The citation calculation agent is used to obtain the similarity between each quoted sentence in A and the corresponding reference content, so as to obtain the similarity list corresponding to A η = (η 1 , η 2 ,…,η i ,…,η n ), where η i A i The similarity between the content and the corresponding reference content.

[0029] Exception determiner, traverse η, if η i >η', then determine A i No exception; otherwise, confirm A i Exception; wherein η' is a preset sentence similarity threshold.

[0030] The text verification agent is used to verify the content of each paragraph if there is no abnormality in each quoted sentence in A, so as to generate a multimodal long text corresponding to the writing requirements input by the user.

[0031] According to another aspect of the present application, there is also provided an Agent autonomously planned multimodal large model content creation system, the system comprising: a processor and a storage medium; the storage medium stores at least one instruction or at least one program, the at least one instruction or the at least one program is loaded and executed by the processor to implement the Agent autonomously planned multimodal large model content creation method as described in the first aspect.

[0032] The present invention has at least the following beneficial effects:

[0033] The Agent autonomously plans the multimodal large model content creation method of the present invention, generates the context structure skeleton of the content to be created according to the writing requirements and reference files input by the user; wherein the context structure skeleton includes a plurality of paragraph information and a plurality of multimodal anchor information corresponding to the content to be created; the multimodal anchor information is used to characterize the corresponding chart or picture position; each paragraph information includes a plurality of questions to be retrieved; the reference content corresponding to each paragraph information, the paragraph content of the previous paragraph and the context structure skeleton of the full text are input into the preset large language model LLM for paragraph content creation to obtain the paragraph content corresponding to each paragraph information; traverse each anchor information, and generate the chart or picture corresponding to each anchor information according to each anchor information; obtain each quoted sentence in each paragraph content to obtain the quoted sentence list A; obtain the similarity between each quoted sentence in A and the corresponding reference content to obtain the similarity list η corresponding to A; traverse η, if η i >η', then determine A i No exception; otherwise, confirm A i Abnormal; if each quoted sentence in A is normal, then the content of each paragraph is verified to generate a multimodal long text corresponding to the writing requirements input by the user; in the present invention, by innovatively combining dynamic knowledge acquisition, multimodal generation, long-form writing planning and content credibility assurance, the intelligence, reliability and diversity of AI writing are significantly improved. Compared with the prior art, the present invention not only solves the key problems of traditional AI writing in knowledge coverage, real-time, multimodal support and generation quality, but also provides users with an efficient, reliable, and illustrated intelligent content creation tool. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0035] Figure 1A flowchart of the Agent autonomous planning multi-modal large model content creation method provided in an embodiment of the present invention;

[0036] Figure 2 A schematic diagram of the implementation process of the Agent autonomous planning multimodal large model content creation method provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0037] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

[0038] It should be noted that, based on the present disclosure, those skilled in the art should understand that an aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number of aspects described herein can be used to implement a device and / or practice a method. In addition, other structures and / or functionalities other than one or more of the aspects described herein can be used to implement this device and / or practice this method.

[0039] The following will refer to Figure 1 The flowchart of the Agent autonomous planning multimodal large model content creation method shown is an introduction to an Agent autonomous planning multimodal large model content creation method.

[0040] The agent autonomously plans the multimodal large model content creation method, which may include the following steps:

[0041] S100, generating a context structure skeleton of the content to be created according to the writing requirements and reference files input by the user; wherein the context structure skeleton includes a number of paragraph information and a number of multimodal anchor point information corresponding to the content to be created; the multimodal anchor point information is used to represent the corresponding chart or image position; each paragraph information includes a number of questions to be retrieved.

[0042] In this embodiment, if a user wants to create an article, he needs to input the writing requirements and the reference files provided by the user; in order to solve the problem that the traditional model is difficult to plan the content of a long article, this embodiment proposes to use the generalization ability of the model and use the preset context structure skeleton to generate the context structure skeleton of the article by the intelligent agent. The context structure skeleton not only includes the outline of the full text, but also defines the specific paragraph structure, search plan and multimodal anchor planning. When generating the context structure skeleton, the model first determines the logical framework and main content distribution of the full text, and plans 5-10 questions to be retrieved for each paragraph. These questions are optimized and designed by the model and have the characteristics of "atomicization", that is, the content is clear and cannot be divided. The skeleton also further marks the location (anchor point) where the chart or picture needs to be inserted, and generates description information for each anchor point, such as the theme of the picture content or the comparison content of the chart data. Through this scheme, the structure of the article is clear and the logic is rigorous, which is convenient for subsequent generation and can also make preliminary planning for inserting multimodal content.

[0043] S200, traverse each paragraph information, input the reference content corresponding to each paragraph information, the paragraph content of the previous paragraph and the context structure skeleton of the full text into the preset large language model LLM for paragraph content creation, so as to obtain the paragraph content corresponding to each paragraph information; wherein, the reference content corresponding to each paragraph information is obtained by retrieving the corresponding question to be searched.

[0044] Furthermore, the reference content corresponding to each paragraph of information is obtained by the following steps:

[0045] S210, obtaining a number of to-be-searched questions corresponding to each paragraph of information, so as to obtain a to-be-searched question list set B = (B 1 , B 2 , …, B j , …, B m ), j = 1, 2, ..., m; where B j is the list of questions to be retrieved corresponding to the jth paragraph information, and m is the number of paragraph information; B j =(B j，1 , B j，2 , …, B j，r , …, B j，f(j) ), r = 1, 2, ..., f(j); where B j，r is the rth question to be retrieved corresponding to the jth paragraph information, and f(j) is the number of questions to be retrieved corresponding to the jth paragraph information.

[0046] In this embodiment, when the context structure skeleton is generated, the paragraph information corresponding to each paragraph is included. Each paragraph information includes several questions to be retrieved corresponding to the paragraph. Each question to be retrieved corresponding to each paragraph information can be obtained to obtain B; it can be understood that the questions to be retrieved are the questions involved in the corresponding paragraph. By searching the questions to be retrieved, the relevant paragraph content is obtained, and further creation is carried out.

[0047] S220, traverse B j , if B j，r For the first type of problem, the default local large model is used to calculate B j，r Search is performed; among which, the results corresponding to the first type of questions are not time-sensitive.

[0048] S230, if B j，r For the preset second type of question, the network is used to j，r Search to obtain the reference content list set DB corresponding to B = (DB 1 , DB 2 , …, DB j , …, DB m ), where DB j For B j Corresponding reference content list; DB j =(DB j，1 , DB j，2 , …, DB j，r , …, DB j，f(j) );DB j，r For B j，r The references retrieved;

[0049] In this embodiment, the questions to be retrieved are divided into fixed questions that are not time-sensitive and questions that are time-sensitive; questions that are not time-sensitive can be understood as fixed knowledge, for example: what is pi; such questions are directly obtained through the preset local large model; and the answers to the questions to be retrieved that are time-sensitive will change with time, for example: news that happened on the day; such questions need to be retrieved on the Internet through the network; thereby, it can be ensured that the retrieved paragraph content takes into account both accuracy and timeliness.

[0050] In this embodiment, a retriever is defined and implemented. In order to make up for the limitations of the knowledge of the pre-trained language model and the lack of real-time performance, this system designs a "retrieval" for multi-source knowledge acquisition. The retriever dynamically obtains the latest information related to user needs by integrating search engines, private domain knowledge bases and external databases. The retriever not only supports text retrieval, but also has the ability to search for pictures with text, and can find pictures related to the content from external networks or local resources. In addition, in conjunction with the retriever, a "reader" module is also designed to be responsible for content parsing and extracting the search results. The reader is implemented by a large model, which organizes the original information extracted from web page content, file content and private domain knowledge base into highly concentrated core content fragments, namely reference content. These fragments become important reference materials in the subsequent writing process. Through this solution, the model can dynamically obtain the latest knowledge, while introducing text and multimodal data (such as pictures) into content creation, improving the breadth and depth of generated content.

[0051] In this embodiment, writing is traversed by paragraphs. In order to solve the problem of limited length of the model output window, a paragraph-based step-by-step writing method is adopted in this embodiment. Specifically, the generated context structure skeleton is traversed, and specific content is generated one by one by paragraph. Each paragraph calls the model generation separately, avoiding the problem of reduced content quality due to output window limitations. Through this method of paragraph-by-paragraph generation and splicing, the system can easily create long texts with a length of tens of thousands of words. At the same time, in order to ensure the logical consistency and contextual coherence of the content, when the model window meets the requirements, the already written content and the "context structure skeleton" are passed to the model at the same time as part of the instruction. The generated content has better contextual coherence.

[0052] S300, traversing each anchor point information, and generating a chart or picture corresponding to each anchor point information according to each anchor point information.

[0053] Further, step S300 includes the following steps:

[0054] S310, obtaining each anchor point information to obtain an anchor point information list C=(C 1 , C 2 , …, C p , …, C q ), p = 1, 2, ..., q; where C p is the pth anchor point information obtained, and q is the number of anchor point information obtained.

[0055] S320, traverse C, if C p For the preset data display anchor information, the preset multimodal model is used to convert C p The corresponding data are plotted into graphs.

[0056] S330, if C p For the preset image display anchor information, the preset multimodal model is used. p Generate an image or search existing resources for images that match the description.

[0057] In this embodiment, a multimodal content processing agent is preset. In order to enhance the diversity and expressiveness of the generated content, the present invention designs a multimodal content processing agent. The multimodal processor specializes in processing the charts and picture anchors marked in the context structure skeleton. The multimodal content processing agent first obtains illustrations that match the anchor description through the multimodal capabilities of the retriever; then, it generates tables and charts using the multimodal model. Specifically, when data needs to be displayed, the system will call the multimodal model to draw clear and intuitive charts based on the retrieved reference data; when an illustrative picture needs to be inserted, the system automatically generates the picture through the multimodal generation tool or retrieves pictures that meet the description from existing resources. Through this scheme, the generated content not only has text, but can also be combined with charts and pictures to form a multimodal document with both text and pictures.

[0058] S400, obtaining each quoted sentence in each paragraph content to obtain a quoted sentence list A=(A 1 , A 2 , …, A i , …, A n ), i=1, 2,...,n; among them, A i is the i-th quoted sentence obtained, and n is the number of quoted sentences obtained.

[0059] In this embodiment, the quoted sentences can be identified by a preset quoted computing agent, or can be manually annotated; the quoted sentences in each paragraph content are obtained to obtain A.

[0060] S500, obtaining the similarity between each quoted sentence in A and the corresponding reference content, so as to obtain a similarity list corresponding to A η = (η 1 , η 2 ,…,η i ,…,η n ), where η i A i The similarity between the content and the corresponding reference content.

[0061] In this embodiment, in order to solve the credibility problem of generated content, a citation computing agent is preset in this embodiment. The citation computing agent will identify the sentences that need to be cited in the generated content, and perform similarity calculations with the reference content provided by the retriever to verify its authenticity. The citation computing agent can mark the source of each quoted sentence to ensure that the generated content has clear source support. This solution effectively avoids the problem of the model generating false content, significantly improves the reliability of the article and user trust, and is particularly suitable for writing tasks that require high accuracy, such as academic articles or professional reports.

[0062] Furthermore, η i It can be calculated by referencing a computing agent.

[0063] In this embodiment, the citation calculation is to extract the sentences that need to be added with the citation calculation by LLM. The sentence features of the citation calculation (any of the following conditions can be met): professional terms in a specific field, fact and data citations, opinions of experts, "xxx data", "xxx events", theories first proposed by a scholar, experimental results are recent research, opinions are controversial in the academic community, and contain numbers other than time, etc.; then the correlation between the extracted sentences and the reference content is extracted to obtain citations.

[0064] Furthermore, η i It can also be obtained by following the steps below:

[0065] S510, A is i The corresponding reference content is divided into several sub-sentences to obtain A i The corresponding sub-statement list Y = (Y 1 , Y 2 , …, Y a , …, Y b ), a=1, 2, …, b; Y a To A i The a-th sub-sentence obtained by dividing the corresponding reference content, b is the sub-sentence of A i The number of sub-sentences obtained by dividing the corresponding reference content; among them, the character length of the sliding window is the same as A i The characters are the same length.

[0066] In this embodiment, the grandfather length and A can be preset i The character length of the sliding window is the same as that of A. i The corresponding reference content is divided into several sub-clauses.

[0067] S520, perform word segmentation on each sub-sentence in Y to obtain a word list set CY corresponding to Y = (CY 1 , CY 2,…,CY a ,…,CY b );wherein, CY a Y a Corresponding word list; CY a =(CY a，1 , CY a，2 ,…,CY a，e ,…,CY a，g(a) ), e = 1, 2, ..., g(a); where CY a，e For Y a The e-th word is obtained by word segmentation, and g(a) is the a The number of words obtained by word segmentation.

[0068] S530, for A i Perform word segmentation to obtain A i Corresponding word list CA i =(CA i，1 , C.A. i，2 , …, C.A. i，x , …, C.A. i，y ), x = 1, 2, ..., y; where CA i，x For A i The xth word obtained by word segmentation, y is the i The number of words obtained by word segmentation.

[0069] In this embodiment, the existing word segmentation processing method can be used to perform word segmentation processing on each sub-sentence in Y to obtain CY; and i Perform word segmentation to obtain CA i .

[0070] S540, obtain CA i The number of words in each word list in CY to get the quantity list NUM i =(NUM i，1 , NUM i，2 , …, NUM i，a , …, NUM i，b ), where NUM i，a For CA i Chinese words in CY a The number in .

[0071] In this embodiment, CA can be calculated i The number of words in each word list in CY, and then get NUM i .

[0072] S550, according to NUM i and y, determine ηi =MAX(NUM i ) / y; where MAX() is the preset maximum value function.

[0073] Through the above steps, η can be determined i , the η determined by this method i Compared with the automatic calculation results obtained by large models, the calculation results are based on practical basis and are more reasonable.

[0074] S600, traverse η, if η i >η', then determine A i No exception; otherwise, confirm A i Exception; wherein η' is a preset sentence similarity threshold.

[0075] In this embodiment, if n i >η', then it means A i It has a high similarity with the sub-sentences in the corresponding reference content, so it can be determined that A i is a quoted sentence with a basis, not a sentence automatically generated by the model; and if η i ≤η', then A i The similarity with the sub-sentences in the corresponding reference content is low, so it can be determined that A i It is not a quoted sentence with a basis, but a sentence automatically generated by the model. Therefore, it is determined that A i abnormal.

[0076] Furthermore, after step S600 and before step S700, the method further includes the following steps:

[0077] S610, if A i If an abnormality is found, a preset prompt message is generated; wherein the preset prompt message is used to prompt the user to perform a manual inspection.

[0078] In this embodiment, if A i If there is an exception, a preset prompt message can be generated to prompt the user to check manually to determine the authenticity and accuracy of the quoted sentence.

[0079] S700, if each quoted sentence in A is normal, then the content of each paragraph is verified to generate a multimodal long text corresponding to the writing requirement input by the user.

[0080] Furthermore, the verification of the content of each paragraph includes: basic language error detection and repair, semantic logic coherence check, emotional tone and style consistency adjustment, format standardization check, consistency and completeness check, content simplification and redundant information processing, and content security and compliance detection.

[0081] In this embodiment, a text verification agent is preset. The text verification agent is one of the important modules of the present invention, which aims to comprehensively improve the quality and expression effect of the generated content and provide high reliability and professional guarantee for the final generated article. This module comprehensively optimizes the generated content from language expression, logical structure to format specification through multiple verification and repair mechanisms to ensure high-quality content output. The functions of the text verification agent mainly include the following aspects:

[0082] (1) Basic language error detection and repair

[0083] The text verification agent can automatically detect and repair basic language errors in the generated content, such as typos, spelling errors, grammatical errors, and punctuation problems. Using the language understanding ability of the large model, the verifier can accurately identify grammatical anomalies in the context, such as subject-verb inconsistency, tense errors, or confusing sentences. At the same time, the verifier can also check whether the use of punctuation is in line with language habits, such as the mixing of Chinese punctuation and English punctuation, or the lack of necessary punctuation, and make corresponding corrections. By detecting and repairing language errors, the text verification agent ensures that the generated content is fluent in language, grammatically correct, and with standardized symbols.

[0084] (2) Semantic and logical coherence check

[0085] The generated content may have semantic jumps or context incoherence during the paragraph-by-paragraph writing process. For example, the content of a certain paragraph is not sufficiently relevant to the topic of the previous and next paragraphs, or there are duplications and contradictions in the information logic. The text verification agent uses global semantic analysis technology to check whether the content conforms to the logical progressive relationship. For paragraphs with logical problems, the verifier will provide modification suggestions, or optimize the fluency of the content by automatically adjusting the order of paragraphs and adding transition sentences. In addition, the verifier can also identify possible semantic conflicts in multiple paragraphs, such as contradictory statements, causal logic errors, etc., and repair them by adjusting the text expression.

[0086] (3) Adjustment of emotional tone and style consistency

[0087] According to the needs of different writing scenarios and target audiences, the text verification agent can adjust the emotional tone and style of the content to make the overall expression of the article more in line with expectations. For example, for academic papers that require a rigorous style, the verifier can remove unnecessary exclamatory tones or overly colloquial expressions in the content; and for marketing copy that requires a friendly style, the verifier will appropriately introduce more relaxed and attractive expressions. In addition, through contextual analysis, the verifier ensures that the tone of the content is consistent throughout the text, and there will be no problem of sudden changes in style between paragraphs.

[0088] (4) Format compliance check

[0089] The text validation agent supports the formatting requirements of the article, ensuring that the content meets standardized standards in layout design and visual presentation. For example, for academic writing, the validator will check whether the paragraph indentation, citation mark format and title hierarchy meet the requirements; for marketing or business content, the validator will optimize the paragraph distribution and layout to make the content more readable. At the same time, the validator can identify tables and image anchors in paragraphs to ensure that these multimodal elements are neatly and logically laid out after insertion. In addition, the validator will unify the font, font size, line spacing, etc. in the article, so that the final output content has good aesthetics both on the screen and in print.

[0090] (5)Consistency and integrity check

[0091] The text verification agent can check whether the content in the article is consistent, for example, whether the entity information such as names of people, places, and time is consistent, whether the referenced terms are standardized, or whether there are repeated content expressions. At the same time, the verifier can also check the overall integrity of the article, such as whether some necessary paragraphs, data or image anchors are missing, to ensure that the final generated article covers all the information points required by the user. For multimodal content, the verifier can also verify whether the inserted charts and pictures match the text content to avoid the problem of inconsistency between the description and the displayed content.

[0092] (6) Content simplification and redundant information processing

[0093] During the text generation process, the model may introduce redundant information or unnecessary elaboration, especially in long articles, which are prone to repeated expressions or low information density. The text verification intelligent body will optimize the compactness of the article content through information condensation and redundancy removal. For example, for paragraphs that are too lengthy, the verifier will extract the core information and reorganize the language expression; for repeated statements, the verifier will merge or delete them to ensure that the article content is concise and the information is clear.

[0094] (7) Content security and compliance testing

[0095] The text verification agent can also perform content security and compliance checks on the generated content, identifying sensitive information or potential inappropriate expressions. For example, when writing legal, medical or financial content, the verifier will check the accuracy of professional terminology and identify possible legal risks or misleading information, and issue warnings or suggest modifications. This feature makes the generated content more reliable and professional when applied to formal scenarios.

[0096] 7. Finally generate credible multimodal long text

[0097] Finally, the system of the present invention can generate credible, well-illustrated, and clearly structured long-form content. The searcher and reader work together to ensure that the generated content has the latest and accurate knowledge support; the context structure skeleton clarifies the logic of the full text and the distribution of multimodal content; the paragraph writing method solves the limitation problem of the model output window; the multimodal content processing agent enriches the content form; the reference calculation agent and the text verification agent ensure the credibility and quality of the content. In the end, the content generated by the system not only meets the user's demand for long-form text, but also provides richer forms of expression through charts and pictures, comprehensively improving the efficiency and effectiveness of content creation.

[0098] Through this invention scheme, this patent realizes the comprehensive optimization of existing AI writing technology, which can meet the needs of high-quality content creation in multiple fields and scenarios, and provide users with a new intelligent writing tool; the specific process can be referred to Figure 2 shown.

[0099] The key points of this embodiment are as follows:

[0100] 1. Dynamic Knowledge Acquisition Agent: Introduction of Retriever and Reader

[0101] By designing a retriever module, the present invention realizes the ability to dynamically acquire knowledge from search engines, private domain knowledge bases and external databases. The retriever not only supports text information retrieval, but also has a multimodal retrieval function of searching images with text. In conjunction with a reader driven by a large model, the retrieved text, web pages, and file contents can be extracted into condensed core content fragments, and the image information can be processed, thereby making up for the problems of insufficient model knowledge and real-time defects.

[0102] 2. Autonomous planning of the network structure skeleton intelligent body

[0103] This paper proposes the concept of "context structure skeleton" to solve the shortcomings of the model in long-form content planning. The skeleton not only contains the outline of the article, but also defines the core theme of each paragraph of content, search question planning, and the setting of multimodal anchor points (including the location of image and chart insertion). The skeleton design makes the logic of generating content clearer and provides comprehensive support for the introduction of multimodal content.

[0104] 3. Write in paragraphs to break through the length limit

[0105] The present invention breaks through the length limit of the model generation window by generating content one by one in paragraphs, ensuring that each paragraph of content has high quality and maintains logical coherence. The step-by-step generation method also enables the system to create very long content (such as documents of tens of thousands of words) without affecting the overall logic and consistency.

[0106] 4. Multimodal Content Processing Agent

[0107] The system designs a multimodal content processing agent to generate and process the multimodal content anchors marked in the context skeleton. This module uses the multimodal retrieval capability of the retriever to obtain relevant illustrations, and generates the required charts through the multimodal model. This function enriches the expression of content, so that the final generated content is no longer limited to text, but forms a multimodal document with both text and pictures.

[0108] 5. Use computational agents to ensure content credibility

[0109] In order to improve the credibility of generated content, the present invention introduces a citation computing agent module. This module verifies the factuality of the content and marks the source by identifying sentences that need to be cited in the content and matching them with the retrieved references. This effectively solves the problem that the model-generated content may have hallucinations or false information, making the generated results more reliable.

[0110] 6. Text verification agent fully optimizes generated content

[0111] The text verification agent is an important part of the present invention. It improves the overall quality of the generated content through multiple verification and repair functions. The verifier is responsible for checking typos, grammatical errors, semantic coherence, emotional style consistency, and format norms, and simplifies redundant content. At the same time, it also supports content security and compliance testing to ensure that the generated articles meet professional standards in various application scenarios.

[0112] The effects of the present invention are as follows:

[0113] 1. Dynamic knowledge support to enhance writing intelligence

[0114] By introducing a retriever and a reader, the present invention can dynamically acquire real-time knowledge from search engines, private domain knowledge bases, and external databases, making up for the problem of knowledge blind spots and lack of timeliness caused by fixed training data in AI models. At the same time, the reader performs structured extraction and concentration on the retrieved content, making the content generated by the model more accurate and authoritative. This dynamic knowledge support greatly expands the intelligence and applicability of AI writing, making the generated article content more in line with user needs in terms of depth and breadth.

[0115] 2. Break through generation limitations and achieve super long text creation

[0116] The present invention achieves structured planning of the full text content by designing a "context structure skeleton", and clarifies the logical framework, outline, paragraph theme and multimodal anchor points of the article. The writing method of paragraph-by-paragraph generation solves the problem of the limitation of the traditional model on the generation length, allowing the system to create high-quality content with a length of tens of thousands of words. At the same time, the paragraph-by-paragraph writing method ensures the coherence of the content logic and the integrity of each paragraph, breaking through the constraints of the model's single generation window, and greatly improving the writing efficiency and operability.

[0117] 3. Realize multimodal content generation and enrich the forms of expression

[0118] The multimodal content processing agent of the present invention supports the generation and insertion of multimodal content such as pictures and charts. Through the multimodal retrieval function, the system can obtain illustration content from external resources and draw charts using multimodal models. This function effectively solves the problem of single content form in traditional AI writing, making the final generated article more vivid and intuitive, with multimodal display capabilities of both pictures and texts, meeting the needs of diverse scenarios such as academic, reporting, and publicity.

[0119] 4. Ensure the credibility and reliability of the content

[0120] In order to solve the problem of low credibility of content in AI writing, the present invention designs a citation computing agent. The citation computing agent can automatically verify the basis of factual statements in the generated content, mark its source, and ensure the authenticity of the content by matching the similarity with the reference material. This mechanism effectively reduces the problems of fiction and illusion in AI-generated content, significantly improves the reliability and professionalism of the generated results, and provides users with a credible writing solution.

[0121] 5. Comprehensively optimize content quality

[0122] The text verification agent of the present invention optimizes the quality of the generated content in an all-round way. By checking and repairing typos, grammatical errors, logical coherence, emotional consistency and format specifications, the generated article content has a high degree of language fluency, logical clarity and expression accuracy. At the same time, the verifier ensures that the article can meet high standards in different scenarios through content redundancy removal, style adjustment and compliance inspection. The content finally generated is not only of excellent quality, but also more professional and reliable.

[0123] 6. Efficiently implement credible long articles with pictures and texts

[0124] The present invention realizes an efficient process from knowledge acquisition, structure planning to content generation through the collaborative work of multiple modules. The final output content is logically rigorous and fluent in language, and shows a higher level of visual and information expression through the richness of multimodal content. At the same time, through the dual guarantee of reference computing agents and text verification agents, the generated results are credible and reliable, without the need for additional modification or verification by users, which greatly improves writing efficiency and user experience.

[0125] The present invention significantly improves the intelligence, reliability and diversity of AI writing by innovatively combining dynamic knowledge acquisition agents, multimodal generation, long-form writing planning and content credibility assurance. Compared with the prior art, the present invention not only solves the key problems of traditional AI writing in terms of knowledge coverage, real-time, multimodal support and generation quality, but also provides users with an efficient, reliable, and illustrated intelligent content creation tool. The realization of this system marks a major advancement in AI writing technology and has broad application value in academia, business, creativity and other fields.

[0126] In an exemplary embodiment, an agent autonomously plans a multimodal large model content creation device, the device comprising:

[0127] The context structure skeleton generation agent is used to generate the context structure skeleton of the content to be created according to the writing requirements and reference files input by the user; wherein the context structure skeleton includes a number of paragraph information and a number of multimodal anchor point information corresponding to the content to be created; the multimodal anchor point information is used to represent the corresponding chart or image position; each paragraph information includes a number of questions to be retrieved.

[0128] The paragraph traversal writing agent is used to traverse each paragraph information and generate the paragraph content corresponding to each paragraph information according to a number of questions to be searched corresponding to each paragraph information; wherein, the paragraph content corresponding to each paragraph information is obtained by retrieving the reference content obtained by searching the questions to be searched.

[0129] The multimodal content processing agent is used to traverse each anchor point information and generate a chart or picture corresponding to each anchor point information according to each anchor point information.

[0130] The quoted sentence acquirer is used to acquire each quoted sentence in each paragraph content to obtain a quoted sentence list A=(A 1 , A 2 , …, A i , …, A n ), i=1, 2,...,n; among them, A i is the i-th quoted sentence obtained, and n is the number of quoted sentences obtained.

[0131] The citation calculation agent is used to obtain the similarity between each quoted sentence in A and the corresponding reference content, so as to obtain the similarity list corresponding to A η = (η 1 , η 2 ,…,η i ,…,η n ), where η i A i The similarity between the content and the corresponding reference content.

[0132] Exception determiner, traverse η, if η i >η', then determine A i No exception; otherwise, confirm A i Exception; wherein η' is a preset sentence similarity threshold.

[0133] The text verification agent is used to verify the content of each paragraph if there is no abnormality in each quoted sentence in A, so as to generate a multimodal long text corresponding to the writing requirements input by the user.

[0134] In an exemplary embodiment, an Agent autonomously planned multimodal large model content creation system is also provided, the system comprising: a processor and a storage medium; the storage medium stores at least one instruction or at least one program, and the at least one instruction or the at least one program is loaded and executed by the processor to implement the Agent autonomously planned multimodal large model content creation method as described in any one of the above embodiments.

[0135] In addition, although the steps of the method in the present disclosure are described in a specific order in the drawings, this does not require or imply that the steps must be performed in this specific order, or that all the steps shown must be performed to achieve the desired results. Additionally or alternatively, some steps may be omitted, multiple steps may be combined into one step, and / or one step may be decomposed into multiple steps, etc.

[0136] Although some specific embodiments of the present invention have been described in detail by way of example, it should be understood by those skilled in the art that the above examples are only for illustration, not for limiting the scope of the present invention. It should also be understood by those skilled in the art that various modifications may be made to the embodiments without departing from the scope and spirit of the present invention.

Claims

1. An agent-autonomous planning multimodal large model content creation method, characterized in that: The method comprises the following steps: S100, generating a context structure skeleton of the content to be created according to the writing requirements and reference files input by the user; wherein the context structure skeleton includes a plurality of paragraph information and a plurality of multimodal anchor information corresponding to the content to be created; the multimodal anchor information is used to represent the position of the corresponding chart or picture; each paragraph information includes a plurality of questions to be searched; S200, traversing each paragraph information, inputting the reference content corresponding to each paragraph information, the paragraph content of the previous paragraph and the context structure skeleton of the whole text into the preset large language model LLM to create paragraph content, so as to obtain the paragraph content corresponding to each paragraph information; wherein, the reference content corresponding to each paragraph information is obtained by searching the corresponding to-be-searched question; S300, traversing each anchor point information, and generating a chart or picture corresponding to each anchor point information according to each anchor point information; S400, obtaining each sentence in each paragraph that needs to be cited, to obtain a cited sentence list A=(A1, A2, ..., A i , …, A n ), i=1, 2,...,n; where, A i is the i-th quoted sentence obtained, and n is the number of quoted sentences obtained; S500, obtaining the similarity between each quoted sentence in A and the corresponding reference content, so as to obtain a similarity list corresponding to A η = (η1, η2, ..., η i ,…,η n ), where η i A i Similarity with the corresponding reference content; S600, traverse η, if η i >η', then determine A i No exception; otherwise, confirm A i Abnormal; where η' is the preset sentence similarity threshold; S700, if each quoted sentence in A is normal, then the content of each paragraph is verified to generate a multimodal long text corresponding to the writing requirement input by the user.

2. According to claim 1, the method for creating content of a multi-modal large model by autonomously planning by an agent is characterized in that: The reference content corresponding to each paragraph information is obtained by the following steps: S210, obtaining a number of to-be-searched questions corresponding to each paragraph of information, so as to obtain a to-be-searched question list set B = (B1, B2, ..., B j , …, B m ), j = 1, 2, ..., m; where B j is the list of questions to be retrieved corresponding to the jth paragraph information, and m is the number of paragraph information; B j =(B j，1 , B j，2 , …, B j，r , …, B j，f(j) ), r = 1, 2, ..., f(j); where B j，r is the rth question to be retrieved corresponding to the jth paragraph information, and f(j) is the number of questions to be retrieved corresponding to the jth paragraph information; S220, traverse B j , if B j，r For the first type of problem, the default local large model is used to calculate B j，r Search; the results corresponding to the first type of questions are not timely; S230, if B j，r For the preset second type of question, the network is used to j，r Search to obtain the reference content list set DB corresponding to B = (DB1, DB2, ..., DB j , …, DB m ), where DB j For B j Corresponding reference content list; DB j =(DB j，1 , DB j，2 , …, DB j，r , …, DB j，f(j) );DB j，r For B j，r The retrieved reference content.

3. According to claim 1, the method for creating content of a multi-modal large model by autonomously planning by an agent is characterized in that: Step S300 includes the following steps: S310, obtaining each anchor point information to obtain an anchor point information list C = (C1, C2, ..., C p , …, C q ), p = 1, 2, ..., q; where C p is the pth anchor point information obtained, q is the number of anchor point information obtained; S320, traverse C, if C p For the preset data display anchor information, the preset multimodal model is used to convert C p The corresponding data are plotted into graphs; S330, if C p For the preset image display anchor information, the preset multimodal model is used. p Generate images or retrieve images matching the description from reference files and the Internet.

4. The Agent autonomous planning multimodal large model content creation method according to claim 1 is characterized in that: η i Calculated by the preset reference calculation agent.

5. The Agent autonomous planning multimodal large model content creation method according to claim 1 is characterized in that: η i Obtained through the following steps: S510, A is i The corresponding reference content is divided into several sub-sentences to obtain A i The corresponding sub-statement list Y = (Y1, Y2, ..., Y a , …, Y b ), a=1, 2, …, b; Y a To A i The a-th sub-sentence obtained by dividing the corresponding reference content, b is the sub-sentence of A i The number of sub-sentences obtained by dividing the corresponding reference content; among them, the character length of the sliding window is the same as A i The characters are the same length; S520, perform word segmentation on each sub-sentence in Y to obtain a word list set CY corresponding to Y = (CY1, CY2, ..., CY a ,…,CY b );wherein, CY a Y a Corresponding word list; CY a =(CY a，1 , CY a，2 ,…,CY a，e ,…,CY a，g(a) ), e = 1, 2, ..., g(a); where CY a，e For Y a The e-th word is obtained by word segmentation, and g(a) is the a The number of words obtained by word segmentation; S530, for A i Perform word segmentation to obtain A i Corresponding word list CA i =(CA i，1 , C.A. i，2 , …, C.A. i，x , …, C.A. i，y ), x = 1, 2, ..., y; where CA i，x For A i The xth word obtained by word segmentation, y is the i The number of words obtained by word segmentation; S540, obtain CA i The number of words in each word list in CY to get the quantity list NUM i =(NUM i，1 , NUM i，2 , …, NUM i，a , …, NUM i，b ), where NUM i，a For CA i Chinese words in CY a The number of S550, according to NUM i and y, determine η i =MAX(NUM i ) / y; where MAX() is the preset maximum value function.

6. The Agent autonomous planning multimodal large model content creation method according to claim 1 is characterized in that: After step S600 and before step S700, the method further includes the following steps: S610, if A i If an abnormality is found, a preset prompt message is generated; wherein the preset prompt message is used to prompt the user to perform a manual inspection.

7. The Agent autonomous planning multimodal large model content creation method according to claim 1 is characterized in that: The verification of the content of each paragraph includes: basic language error detection and repair, semantic logic coherence check, emotional tone and style consistency adjustment, format standardization check, consistency and completeness check, content simplification and redundant information processing, and content security and compliance detection.

8. An Agent autonomously planned multi-modal large model content creation device, characterized in that: The device comprises: The context structure skeleton generating agent is used to generate the context structure skeleton of the content to be created according to the writing requirements and reference files input by the user; wherein the context structure skeleton includes a number of paragraph information and a number of multimodal anchor information corresponding to the content to be created; the multimodal anchor information is used to represent the corresponding chart or picture position; each paragraph information includes a number of questions to be retrieved; The paragraph traversal writing agent traverses each paragraph information, and inputs the reference content corresponding to each paragraph information, the paragraph content of the previous paragraph, and the context structure skeleton of the whole text into the preset large language model LLM for paragraph content creation, so as to obtain the paragraph content corresponding to each paragraph information; wherein, the reference content corresponding to each paragraph information is obtained by retrieving the corresponding to-be-searched question; A multimodal content processing agent, used for traversing each anchor point information and generating a chart or picture corresponding to each anchor point information according to each anchor point information; The quoted sentence acquirer is used to acquire each quoted sentence in each paragraph content to obtain a quoted sentence list A = (A1, A2, ..., A i , …, A n ), i=1, 2,...,n; where, A i is the i-th quoted sentence obtained, and n is the number of quoted sentences obtained; The citation calculation agent is used to obtain the similarity between each quoted sentence in A and the corresponding reference content, so as to obtain the similarity list η corresponding to A = (η1, η2, ..., η i ,…,η n ), where η i A i Similarity with the corresponding reference content; Exception determiner, traverse η, if η i >η', then determine A i No exception; otherwise, confirm A i Abnormal; where η' is the preset sentence similarity threshold; The text verification agent is used to verify the content of each paragraph if there is no abnormality in each quoted sentence in A, so as to generate a multimodal long text corresponding to the writing requirements input by the user.

9. An Agent autonomously planned multimodal large model content creation system, characterized in that: The system includes: a processor and a storage medium; the storage medium stores at least one instruction or at least one program, and the at least one instruction or the at least one program is loaded and executed by the processor to implement the Agent autonomous planning multimodal large model content creation method as described in any one of claims 1-7.

Images