Intelligent writing method, apparatus and device, and computer program product

Abstract

This application discloses an intelligent writing method, apparatus, device, and computer program product. The method includes: receiving a writing task and identifying its core features; determining a combination of intelligent writing models suitable for the writing task based on the core features of the writing task and a model resource library using a scheduling decision strategy; generating multiple writing ideas using multiple writing models in the intelligent writing model combination and processing them using an idea aggregation processing strategy to obtain a combination of writing ideas; iteratively optimizing the combination of writing ideas using a closed-loop iterative optimization strategy to output the final writing result. The scheduling decision strategy of this application cleverly combines task requirements and model capabilities while considering computational costs, achieving efficient selection of intelligent writing model combinations. The idea aggregation processing strategy fully leverages the characteristics of each model, enriching the diversity and quality of writing ideas. The closed-loop iterative optimization strategy breaks through the bottleneck of traditional iterative rigidity, improving the efficiency and quality of intelligent writing.

Description

Technical Field

[0001] This application relates to the field of intelligent writing technology, and in particular to an intelligent writing method, device and equipment, and computer program product. Background Technology

[0002] In today's era of rapid digital information development, the field of writing is undergoing profound changes. Especially with the continuous advancement of artificial intelligence technology, intelligent writing technology has emerged and gradually become a research hotspot. Intelligent writing aims to utilize advanced algorithms and models to assist or automate various writing tasks, improve writing efficiency and quality, and meet diverse writing needs in different scenarios. However, although existing technologies have promoted the development of intelligent writing to some extent, many key problems still need to be solved in practical applications. These problems seriously restrict the further development and widespread application of intelligent writing technology.

[0003] Currently, there are several technologies related to intelligent writing. The technology with public number CN119167931A uses a large model to provide multi-faceted, fine-grained, and personalized guidance for academic writing; the technology with public number CN117494669A improves the relevance and accuracy of news paragraphs through phased writing; and the technology with public number CN118569216A uses writing templates to guide users in generating documents, reducing the workload of manual review.

[0004] However, these technologies have obvious shortcomings. For example, in terms of multi-model collaboration, the common approach is to use a "one task bound to one model" or "fixed model combination" model, which leads to a lack of creativity and makes it difficult to meet the diverse creative writing needs. In terms of iteration mechanisms, most technologies lack automatic self-iteration functions or only judge the results based on fixed thresholds, which cannot simulate the dynamic process of human creation. In terms of defect improvement and quality control, existing technologies can only identify surface problems and are unable to locate the deep-seated root causes. Summary of the Invention

[0005] This application provides an intelligent writing method, apparatus and device, and computer program product to solve problems such as multi-model collaboration, iteration mechanism, defect improvement and quality control, improve the performance and quality of intelligent writing, meet diverse writing needs, and promote the development of intelligent writing technology.

[0006] The embodiments of this application adopt the following technical solutions:

[0007] In a first aspect, embodiments of this application provide an intelligent writing method, the intelligent writing method comprising:

[0008] Receive a writing task and perform core feature identification on the writing task to obtain the core feature information of the writing task;

[0009] Based on the core feature information of the writing task and the pre-built model resource library, a combination of intelligent writing models adapted to the writing task is determined using a preset scheduling decision strategy.

[0010] The writing ideas of each writing model are generated by using multiple writing models in the intelligent writing model combination and processed by a preset idea aggregation processing strategy to obtain a combination of writing ideas.

[0011] The writing idea combination is iteratively optimized using a pre-defined closed-loop iterative optimization strategy to output the final writing result.

[0012] Optionally, the pre-built model resource library stores capability tags for multiple writing models. The step of determining a combination of intelligent writing models suitable for the writing task based on the core feature information of the writing task and the pre-built model resource library, using a preset scheduling decision strategy, includes:

[0013] Based on the core feature information of the writing task and the capability tags of multiple writing models, calculate the core fit between the writing task and each writing model.

[0014] The core fit between the writing task and each writing model is corrected using a dynamic weight correction strategy to obtain the core fit correction value between the writing task and each writing model.

[0015] The computing cost constraint strategy is used to constrain the core fit correction value of the writing task and each writing model by computing cost, so as to obtain the computing cost constraint value of the writing task and each writing model.

[0016] Based on the core compatibility between the writing task and each writing model, as well as the computational cost constraints between the writing task and each writing model, a hierarchical screening strategy is used to screen multiple writing models and combine intelligent writing models that are compatible with the writing task.

[0017] Optionally, the step of using a hierarchical screening strategy to screen multiple writing models based on the core fit between the writing task and each writing model, and the computational cost constraints between the writing task and each writing model, results in a combination of intelligent writing models that are compatible with the writing task, including:

[0018] Based on the core fit between the writing task and each writing model, multiple writing models are screened using a preset core fit threshold to obtain the first candidate writing model.

[0019] The first candidate writing model is selected based on the type balance screening strategy to obtain the second candidate writing model;

[0020] Based on the writing task and the computational cost constraints of each second candidate writing model, the second candidate writing models are screened to obtain the final intelligent writing model combination.

[0021] Optionally, the step of generating writing ideas for each writing model using multiple writing models in the intelligent writing model combination and processing them using a preset idea aggregation processing strategy to obtain a writing idea combination includes:

[0022] Based on the multiple writing models in the intelligent writing model combination, generate differentiated instructions corresponding to each writing model;

[0023] By using differentiated instructions corresponding to each writing model, we can guide each writing model to generate writing ideas, thereby obtaining the writing ideas and corresponding model feature tags for each writing model.

[0024] Based on the writing ideas of each writing model and the corresponding model feature tags, the preset idea aggregation processing strategy is used to process them to obtain the writing idea combination.

[0025] Optionally, the step of processing the writing ideas based on the writing ideas of each writing model and their corresponding model feature tags using the preset idea aggregation processing strategy to obtain the writing idea combination includes:

[0026] Based on the writing ideas of each writing model and the corresponding model feature tags, the writing ideas of each writing model are classified according to the model feature tags, and the writing task is bound to the core adaptability tag of each writing model.

[0027] The writing ideas of multiple writing models categorized by model feature tags are subjected to multi-stage deduplication to obtain deduplicated writing ideas.

[0028] Based on the deduplicated writing ideas, their corresponding model feature tags, and core fit tags, a hierarchical creative rating strategy is used to rate the deduplicated writing ideas, resulting in a rating result for the writing ideas.

[0029] Based on the rating results of the writing ideas, a multi-dimensional verification and completion strategy is used to verify and complete the writing ideas in order to obtain the final combination of writing ideas.

[0030] Optionally, the step of iteratively optimizing the combination of writing ideas using a preset closed-loop iterative optimization strategy to output the final writing result includes:

[0031] The iterative termination index is dynamically combined based on the task type of the writing task.

[0032] Based on the iteration termination index, determine whether the writing idea combination triggers the preset iteration termination condition;

[0033] If triggered, the final writing result will be output;

[0034] If not triggered, an iterative optimization instruction is generated to iteratively optimize the combination of writing ideas until the preset iteration termination condition is triggered.

[0035] Optionally, the generation of iterative optimization instructions to iteratively optimize the combination of writing ideas includes:

[0036] A multi-dimensional defect scan was performed on the aforementioned writing creative combination to obtain the multi-dimensional defect scan results;

[0037] Based on the multi-dimensional defect scan results, generate executable suggestions;

[0038] Based on the multi-dimensional defect scanning results and the executable suggestions, a multi-path generator is used to initiate multiple optimization paths for concurrent optimization, resulting in optimized writing results.

[0039] Secondly, embodiments of this application also provide an intelligent writing device, the intelligent writing device comprising:

[0040] The feature recognition unit is used to receive the writing task and perform core feature recognition on the writing task to obtain the core feature information of the writing task.

[0041] The scheduling decision unit is used to determine a combination of intelligent writing models that are suitable for the writing task based on the core feature information of the writing task and the pre-built model resource library, using a preset scheduling decision strategy.

[0042] The creative aggregation unit is used to generate writing ideas for each writing model using multiple writing models in the intelligent writing model combination, and process them using a preset creative aggregation processing strategy to obtain a combination of writing ideas.

[0043] The iterative optimization unit is used to iteratively optimize the combination of writing ideas using a preset closed-loop iterative optimization strategy, and output the final writing result.

[0044] Thirdly, embodiments of this application also provide an apparatus, comprising:

[0045] A processor; and a memory arranged to store computer-executable instructions, which, when executed, cause the processor to perform any of the aforementioned intelligent writing methods.

[0046] Fourthly, embodiments of this application also provide a computer program product, including a computer program / instructions, which, when executed by a processor, implement any of the aforementioned intelligent writing methods.

[0047] The above-mentioned technical solutions adopted in the embodiments of this application can achieve the following beneficial effects: The intelligent writing method of the embodiments of this application first receives a writing task and performs core feature identification on the writing task to obtain the core feature information of the writing task; based on the core feature information of the writing task and a pre-built model resource library, a combination of intelligent writing models adapted to the writing task is determined using a preset scheduling decision strategy; multiple writing models in the intelligent writing model combination generate writing ideas for each writing model and process them using a preset idea aggregation processing strategy to obtain a combination of writing ideas; the combination of writing ideas is iteratively optimized using a preset closed-loop iterative optimization strategy to output the final writing result. The intelligent writing method of the embodiments of this application, through comprehensive and accurate task feature identification, can deeply understand the various needs of the writing task, providing a solid foundation for subsequent model scheduling. The preset scheduling decision strategy cleverly combines task requirements and model capabilities, while taking into account computing power costs, realizing efficient screening of intelligent writing model combinations, avoiding resource waste, ensuring matching accuracy, and ensuring efficient task implementation. The idea aggregation processing strategy fully leverages the characteristics of each model, achieving efficient aggregation and selection of multimodal ideas through multiple processing mechanisms, enriching the diversity and quality of writing ideas. The closed-loop iterative optimization strategy establishes a multimodal, deeply collaborative closed-loop iterative system that can automatically optimize based on feedback and evaluation results. This breaks through the bottleneck of traditional rigid iteration and ultimately outputs high-quality, user-responsive, and innovative writing results, significantly improving the efficiency and quality of intelligent writing and providing users with a better and more efficient writing service experience. Attached Figure Description

[0048] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0049] Figure 1 This is a flowchart illustrating an intelligent writing method in an embodiment of this application;

[0050] Figure 2 This is a schematic diagram of an intelligent writing process in an embodiment of this application;

[0051] Figure 3 This is a schematic diagram of a pre-built model resource library in an embodiment of this application;

[0052] Figure 4 This is a schematic diagram of the structure of an intelligent writing device according to an embodiment of this application;

[0053] Figure 5 This is a schematic diagram of the structure of a device according to an embodiment of this application. Detailed Implementation

[0054] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0055] The technical solutions provided by the various embodiments of this application are described in detail below with reference to the accompanying drawings.

[0056] Current status of intelligent writing-related technologies:

[0057] (I) Large-scale model-based academic writing techniques

[0058] Currently, there are publications such as CN119167931A, titled "A Method for Assisted Academic Writing Based on a Large Model and an Intelligent Teaching Assistant Agent." This method utilizes a large language model to generate prompts by inputting academic writing requirements and model essays, target academic papers and fine-grained evaluations, and personalized paper improvement suggestions. It provides students with writing guidance in multiple ways, in a fine-grained and personalized manner, helping them improve their writing skills. While this technology provides strong support for academic writing to a certain extent, it primarily focuses on the academic writing field and has limitations in model collaboration and creative generation.

[0059] (II) News Writing Techniques

[0060] The publication CN117494669A, titled "News Writing Method, News Writing Model Training Method and Device," divides the news writing process into two stages: first, it generates the first paragraph (main idea paragraph) based on the news headline; then, it continues writing subsequent paragraphs one by one based on the existing paragraphs. This method effectively improves the relevance between news paragraphs and the news headline, as well as the relevance between news paragraphs, avoiding problems such as irrelevance and poor coherence between paragraphs, thus improving the accuracy and readability of news writing. However, this technology is only applicable to news writing, has a relatively narrow scope of application, and has limited ability to handle complex writing tasks and diverse creative needs.

[0061] (III) Intelligent Writing Template Technology

[0062] The patent application CN118569216A, titled "Intelligent Writing Method, Device, Electronic Equipment, Medium, and System," describes a method that obtains a writing template for document generation, displays the template and writing guidance controls to the user, responds to user interactions to determine writing materials, and fills in the template to generate the document. This technology solves the problems of low quality and low efficiency in automated writing, improving both document generation efficiency and content quality while reducing the workload of subsequent manual review and refinement. However, this technology focuses on template-guided writing and has shortcomings in dynamic collaboration and deep creative generation within the model.

[0063] The shortcomings of existing technology:

[0064] (i) Multi-model collaboration is inefficient and lacks creative data collection.

[0065] Existing multi-model solutions generally adopt a "one task, one model" or "fixed model combination" approach. This approach cannot dynamically schedule different series and sizes of models to participate collaboratively for the same task. During the writing process, relying solely on the output of a single or fixed model lacks both a mechanism for generating and collecting diverse ideas similar to a "brainstorming" session and a logic for differentiated result interaction between models for the same task. This makes it difficult to aggregate the characteristics of various models to form diverse and high-quality ideas, failing to meet the needs of scenarios such as writing that require diverse creative support. Furthermore, there is the problem of wasting resources by calling high-computing-power models for simple tasks, and the rigid collaboration logic severely impacts writing efficiency and creative quality.

[0066] (ii) The iterative mechanism is rigid and lacks closed-loop improvement capability.

[0067] Existing technologies either lack automatic self-iteration capabilities or only use fixed thresholds to judge whether results meet standards, failing to simulate the dynamic process of "constantly generating new ideas and continuously iterating and updating" in human creation. Furthermore, existing technologies lack the ability to identify defects in intermediate versions; they cannot pinpoint problems in intermediate results, analyze root causes, or propose improvement suggestions, nor can they automatically execute improvements and optimizations. They also fail to incorporate multi-path design to enhance result diversity. This results in iterations remaining largely superficial, making it difficult to obtain better results through continuous optimization, increasing time costs, and limiting optimization effectiveness.

[0068] (iii) Defect improvement is crude, and the path and quality control are insufficient.

[0069] Current technologies can only identify surface-level problems such as logical inconsistencies, failing to accurately pinpoint deeper issues like missing evidence or reversed causality. When proposing improvement suggestions, they lack prioritization and rely on manual screening, resulting in low efficiency. Furthermore, multi-path attempts generate only independent results, leading to fragmented and unsustainable outcomes due to the lack of a fusion mechanism. In addition, existing technologies lack standardized operating procedures (SOPs) for writing, polishing, and typo correction, and do not perform contextual relevance and factual consistency checks, making them prone to content contradictions and compromising output quality.

[0070] In view of the numerous problems existing in the field of intelligent writing, this application aims to propose a novel intelligent writing technology solution to address key issues such as inefficient collaboration among existing multi-model solutions, insufficient creative idea collection, rigid iteration mechanisms, lack of closed-loop improvement capabilities, and inadequate improvement methods and quality control. Through innovative technical means, this solution achieves dynamic collaboration among multiple models, efficient creative idea generation and collection, intelligent iterative improvement, and refined quality control, thereby improving the overall performance and quality of intelligent writing, meeting the diverse needs for high-quality writing in different scenarios, and promoting the development of intelligent writing technology to a higher level.

[0071] Specifically, embodiments of this application provide an intelligent writing method, such as... Figure 1 The diagram shows a flowchart of an intelligent writing method according to an embodiment of this application. The intelligent writing method includes the following steps S110 to S140:

[0072] Step S110: Receive the writing task and perform core feature recognition on the writing task to obtain the core feature information of the writing task.

[0073] Combination Figure 2 This application provides a schematic diagram of an intelligent writing process in its embodiments. Writing tasks are obtained through user input or interface reception. Task sources can be various terminal devices or applications, and the content can be diverse, covering different themes and text creation needs. Natural language processing technology is used to analyze the writing tasks, automatically identifying the task type. The task type is determined through predefined keyword and semantic pattern matching, such as whether it belongs to the category of year-end summary, project report, daily work report, or promotional copy. Regarding task complexity, the target word count requirement is extracted from the task description, and the depth of logical argumentation is evaluated using text analysis algorithms, such as by analyzing sentence structure, paragraph relationships, and keyword complexity. In terms of creative needs, by analyzing the descriptions of content style and richness of viewpoints in the task, the user's expectations for creativity are identified, such as whether diverse content styles and rich expression of viewpoints are required.

[0074] The extracted core information is integrated according to a predefined format to generate structured task feature tags. For example, under the premise of legality and compliance, for a task that requires "XX research report, 2000 words, multiple viewpoints and styles, and includes bar charts and table visualizations", the tag "XX research + 2000 words + multiple viewpoints and styles + bar charts and table visualizations" is generated to ensure that subsequent steps accurately understand the task requirements.

[0075] Step S120: Based on the core feature information of the writing task and the pre-built model resource library, determine the combination of intelligent writing models that are suitable for the writing task using a preset scheduling decision strategy.

[0076] Collect and organize various intelligent writing models in advance, including large language models and multimodal models. Conduct detailed capability assessments and annotations for each model, generating model capability tags. Tag content can cover, for example, the types of tasks the model excels at, the range of processing complexity, creative generation capabilities, and multimodal support.

[0077] Based on task feature labels and model capability labels, a similarity matching algorithm is used to calculate the degree of matching between task requirements and model capabilities. Simultaneously, considering computational cost factors, cost weights are assigned to different models. While ensuring the matching degree meets a certain threshold, lower-cost model combinations are prioritized. For example, for the "XX Project" task, by analyzing task feature labels and model capability labels, one high-parameter inference model is selected to handle complex logical reasoning, three different series and sizes of text generation models are selected to meet diverse text generation needs, and two multimodal models are selected to generate relevant images and charts, achieving a balance between accurate matching of task requirements and optimization of computational cost.

[0078] Step S130: Use the multiple writing models in the intelligent writing model combination to generate writing ideas for each writing model, and process them using a preset idea aggregation processing strategy to obtain a writing idea combination.

[0079] Task feature information is input into each model in the intelligent writing model suite. Each model generates corresponding writing ideas based on its own algorithm and training data. Then, an idea aggregation processing strategy is used to fully leverage the unique advantages of each model, ensuring that the ideas generated by different models are differentiated. Through differentiated guidance, it is ensured that the ideas complement each other in terms of theme and style, avoiding repetition. In addition, a precise deduplication algorithm can be used to remove highly similar ideas, improving the diversity of ideas. Finally, diversity verification is performed to ensure that the generated writing idea suite meets the task requirements in terms of content, style, and form, achieving efficient aggregation, selection, and retention of multimodal ideas such as text and images.

[0080] Step S140: The writing idea combination is iteratively optimized using a preset closed-loop iterative optimization strategy to output the final writing result.

[0081] We will build a multimodal, deep collaborative closed-loop iterative optimization system to iteratively optimize writing ideas, realize automatic iterative closed-loop optimization, solve the key problems of iterative rigidity and lack of automatic iterative closed-loop improvement, and break through the bottleneck of iterative rigidity.

[0082] The intelligent writing method in this application embodiment, through comprehensive and accurate task feature identification, can deeply understand the various needs of writing tasks, providing a solid foundation for subsequent model scheduling. The preset scheduling decision strategy cleverly combines task requirements with model capabilities while considering computational costs, achieving efficient selection of intelligent writing model combinations. This avoids resource waste, ensures matching accuracy, and guarantees efficient task implementation. The creative aggregation and processing strategy fully leverages the characteristics of each model, achieving efficient aggregation and selection of multimodal creative ideas through multiple processing mechanisms, enriching the diversity and quality of writing ideas. The closed-loop iterative optimization strategy builds a multimodal deep collaborative closed-loop iterative system, which can automatically optimize based on feedback and evaluation results, breaking through the bottleneck of traditional rigid iteration. Ultimately, it outputs high-quality, user-responsive, and innovative writing results, significantly improving the efficiency and quality of intelligent writing and providing users with a better and more efficient writing service experience.

[0083] In some embodiments of this application, the pre-built model resource library stores capability tags for multiple writing models. The step of determining a combination of intelligent writing models suitable for the writing task based on the core feature information of the writing task and the pre-built model resource library, using a preset scheduling decision strategy, includes: calculating the core fit between the writing task and each writing model based on the core feature information of the writing task and the capability tags of the multiple writing models; correcting the core fit between the writing task and each writing model using a dynamic weight correction strategy to obtain a corrected core fit value; applying a computational cost constraint strategy to the corrected core fit value between the writing task and each writing model to obtain a computational cost constraint value; and using a hierarchical screening strategy to screen multiple writing models based on the core fit between the writing task and each writing model and the computational cost constraint value, to obtain a combination of intelligent writing models suitable for the writing task.

[0084] like Figure 3The diagram illustrates a pre-built model resource library as shown in this embodiment. The model resource library pre-stores capability tags for multiple series and sizes of models. The model series cover LLaMA, Qwen, DeepSeek, GLM, GPT, CogVLM, BLIP, Gemini, and many others; model sizes include 0.6B, 2B, 7B, 13B, 70B, and 671B. The capability tags precisely define the advantages of each model, such as "LLaMA7B model excels at concise viewpoint output," "Qwen72B model excels at cross-language professional report writing and high-precision long document structured analysis," "DeepSeekR1 model is suitable for complex industry analysis and logical argumentation scenarios," "GLM series models have significant advantages in Chinese knowledge enhancement and excel at policy interpretation and academic abstract generation," and "CogVLM series is suitable for technical drawing analysis and the creation of reports combining text and graphics," etc.

[0085] Based on the capability tags of the writing models stored in the aforementioned model resource library, the core fit between the writing task and each writing model is calculated. First, the core feature information of the writing task is quantified and scored according to predefined rules. For example, for task type fit, if the task is academic paper writing, each model is scored out of 10 based on its performance on academic paper-related tasks. Similarly, other features such as task complexity, creative requirements, multimodal task fit, and language fit are also quantified and scored. Simultaneously, the capability tags of each model in the model resource library are also converted into corresponding quantified values.

[0086] Based on the weighted cosine similarity formula Let i = [1, n] and j = [1, m]. Here, i represents the i-th task feature dimension, n represents the number of task feature dimensions, j represents the j-th model, and m represents the number of models. The basic weights w are set... ij Furthermore, since the weights sum to 1, the task type fit weight is set relatively large to emphasize the priority of task type matching. The quantized task feature value t... i Model feature values ​​m j And by substituting the basic weights into the formula, the core fit S between the writing task and each writing model is calculated. ij .

[0087] Based on the core fit S calculated above ij By combining the core features of the task with the model's capabilities to adapt to different scenarios, a random perturbation parameter α is introduced. ij This parameter can be determined based on historical data statistical analysis or experience set according to specific scenarios. It is used to dynamically adjust the core adaptability to solve the problem of rigid adaptability with fixed weights.

[0088] Using formula The core fit correction value S′ between the writing task and each writing model was calculated. ij This makes the adaptation calculation more adaptable to the actual needs of different scenarios.

[0089] The core fit adjustment value S′ between the writing task and each writing model is adjusted. ij Furthermore, a computing cost factor is introduced to balance matching degree and resource consumption. Specifically, a global cost weight λ is first determined, which is used to balance the overall weight of matching degree and resource consumption. A computing cost coefficient C is then set for each model. j This coefficient can be determined based on factors such as the computational complexity of the model and the required hardware resources. Simultaneously, a scenario coefficient β is set according to different task scenarios. ij .

[0090] According to the formula Adjust the core adaptation value S′ ij Global cost weight λ, model computing power cost coefficient C j and scene coefficient β ij Substituting these values, we obtain the computational cost constraint value F for the writing task and each writing model. ij This is to ensure a certain degree of matching while considering the cost of computing power.

[0091] Based on the above calculation results, a hierarchical screening mechanism is used to further screen multiple writing models, taking into account accuracy, diversity and cost control, to obtain the final combination of intelligent writing models.

[0092] This application's embodiments utilize a dual-objective scheduling algorithm of "weighted cosine similarity-computational cost balancing" to achieve accurate calculation of the fit between writing tasks and models, as well as effective constraints on computational costs. When calculating core fit, it fully considers various task characteristics and model capabilities, and adapts to different scenario requirements through dynamic weight adjustments, ensuring matching accuracy. The computational cost constraint balances matching accuracy and resource consumption overall, avoiding resource waste. The tiered selection strategy further balances accuracy, diversity, and cost control, efficiently selecting suitable combinations of intelligent writing models. This satisfies the model capability requirements of complex tasks while avoiding resource waste caused by calling high-computational-cost models for simple tasks, ensuring efficient task implementation and significantly improving the efficiency and rationality of model scheduling in intelligent writing tasks, providing strong support for high-quality intelligent writing results.

[0093] In some embodiments of this application, the step of filtering multiple writing models using a hierarchical filtering strategy based on the core fit between the writing task and each writing model, and the computational cost constraint values ​​between the writing task and each writing model, to obtain a combination of intelligent writing models adapted to the writing task includes: filtering multiple writing models using a preset core fit threshold based on the core fit between the writing task and each writing model to obtain a first candidate writing model; filtering the first candidate writing model using a type balance filtering strategy to obtain a second candidate writing model; and filtering the second candidate writing model based on the computational cost constraint values ​​between the writing task and each second candidate writing model to obtain the final combination of intelligent writing models.

[0094] The hierarchical screening mechanism defined in this application embodiment may sequentially include basic admission screening, type balance screening, and computing power cost optimization. When performing basic admission screening, a core matching standard threshold is pre-set, for example, S≥0.8. This threshold is determined based on the analysis and experience of adapting a large number of tasks and models, and is used to quickly distinguish between high and low adaptability.

[0095] The writing task is matched with the core fit S of each writing model. ij The models are compared with a preset threshold S, and those exceeding the threshold are selected to form the first candidate writing model set. This step quickly eliminates models with low suitability for the writing task, reducing the workload of subsequent selections and improving selection efficiency.

[0096] After completing the basic admission screening, a balanced selection rule is formulated based on the actual needs of the intelligent writing task and the characteristics of the model. For example, for general multimodal writing tasks, it is stipulated that there should be no less than one reasoning model to ensure the task's logical reasoning ability; no less than two text models to ensure sufficient text generation and processing capabilities; and no more than two multimodal models (the multimodal model requirement can be exempted for pure text tasks). This avoids excessive multimodal models leading to resource waste and difficulties in collaboration, while ensuring the text-image collaboration capability of multimodal tasks, avoiding excessive concentration of a single type of model, and ensuring creative diversity.

[0097] The models in the first candidate writing model set are then filtered according to the aforementioned type balance selection rules. The type of each model is counted, and models meeting the quantity requirements are retained based on the rules, thus obtaining the second candidate writing model set. This step further optimizes the structure of the model combination, making the model type distribution more reasonable and meeting the needs of the task in different aspects.

[0098] Finally, computational cost optimization and selection are performed. The computational cost constraint value F of the writing task and each second candidate writing model calculated in the aforementioned embodiments comprehensively considers factors such as the core adaptability correction value of the model, the global cost weight, the model computational cost coefficient, and the scenario coefficient, which can reflect the computational cost of the model under certain matching requirements.

[0099] Among the second-tier candidate writing models that meet the type balance screening criteria, their corresponding computational cost constraints F are compared, and the model combination with the lowest F value is prioritized, i.e., the "high matching degree + low computational cost" combination. Redundant models that, while having a high matching degree, have excessively high computational costs are eliminated, ultimately determining the intelligent writing model combination best suited to the writing task. This step effectively controls computational costs and improves resource utilization efficiency while ensuring model-task matching.

[0100] This application employs a tiered screening strategy, progressively refining multiple writing models from basic admission screening and type-balanced screening to computational cost optimization screening. Basic admission screening quickly eliminates low-fitness models, improving subsequent screening efficiency; type-balanced screening ensures the rationality and diversity of model combinations in terms of type, meeting the diverse needs of the task; and computational cost optimization screening effectively controls computational costs while maintaining matching accuracy, avoiding resource waste. Overall, this technical solution can efficiently and accurately screen intelligent writing model combinations suitable for the writing task, balancing accuracy, diversity, and cost control, providing strong support for the high-quality completion of intelligent writing tasks and significantly improving the rationality and efficiency of model scheduling during the intelligent writing process.

[0101] In some embodiments of this application, the step of generating writing ideas for each writing model using multiple writing models in the intelligent writing model combination and processing them using a preset idea aggregation processing strategy to obtain a writing idea combination includes: generating differentiated instructions corresponding to each writing model based on the multiple writing models in the intelligent writing model combination; guiding each writing model to generate writing ideas using the differentiated instructions corresponding to each writing model, thereby obtaining the writing ideas and corresponding model feature tags of each writing model; and processing the writing ideas and corresponding model feature tags of each writing model using the preset idea aggregation processing strategy to obtain the writing idea combination.

[0102] We delve into the unique strengths and areas of expertise of each model in the intelligent writing model portfolio, based on their capability tags within a pre-built model resource library. For example, one model excels at concise viewpoint output, another performs exceptionally well in cross-language professional report writing, and still others are adapted to complex industry analysis and logical argumentation scenarios.

[0103] Continue to refer to Figure 2 Based on the characteristics of each model, targeted and differentiated instructions are constructed. These instructions focus on guiding the model to creatively generate content that leverages its unique strengths. For example, for a model adept at concisely expressing viewpoints, the instruction could be "Summarize the core viewpoint of a given topic in the most concise language." For a model writing cross-lingual professional reports, the instruction could be "Write a report on [a specific professional field] in both Chinese and English, ensuring accurate terminology." This approach establishes a "characteristic tag - differentiated instruction" linkage mechanism, providing guidance for models to output unique creative content from the outset.

[0104] The differentiated instructions are synchronized to the corresponding writing models. Upon receiving the instructions, the models generate writing ideas according to the requirements, based on their own algorithms and training data. For example, a model receiving a concise viewpoint output instruction will generate a brief and refined statement of viewpoint; a cross-language report writing model will output professional report content in two languages. While generating writing ideas, the models also carry model-specific tags related to their own characteristics. These tags, such as "concise viewpoint output type" and "cross-language professional report type," are determined during the model training and capability evaluation phases and are used for subsequent validation and processing of the diversity of the ideas.

[0105] Based on the writing ideas and model feature tags, the writing ideas are processed using a preset idea aggregation processing strategy to ultimately obtain a combination of writing ideas.

[0106] This application's embodiments fully leverage the unique advantages of each model by generating differentiated instructions for each model in the intelligent writing model combination. This enables the models to output unique creative ideas with their own characteristics, effectively avoiding the problem of creative homogenization. By using model-specific tags to annotate and further process the creative ideas, efficient aggregation and selection of a large number of writing ideas are achieved. The final combination of writing ideas is not only rich and diverse in content, covering different styles and perspectives, but also of high quality, meeting the diverse creative needs of complex writing tasks. This provides high-quality creative support for intelligent writing, significantly improving the quality and efficiency of intelligent writing.

[0107] In some embodiments of this application, the step of processing the writing ideas of each writing model and their corresponding model feature tags using the preset creative aggregation processing strategy to obtain the writing creative combination includes: classifying the writing ideas of each writing model according to their model feature tags, and binding the writing task with the core suitability tags of each writing model; performing multi-stage deduplication processing on the writing ideas of multiple writing models classified by model feature tags to obtain deduplicated writing ideas; rating the deduplicated writing ideas using a hierarchical creative rating strategy based on the deduplicated writing ideas, their corresponding model feature tags, and the core suitability tags to obtain the rating results of the writing ideas; and performing multi-dimensional verification and completion processing of the writing ideas using a diversity verification and completion strategy based on the rating results of the writing ideas to obtain the final writing creative combination.

[0108] Continue to refer to Figure 2 The pre-defined creative aggregation and processing strategy mainly includes the following aspects:

[0109] (1) Enhanced multi-dimensional result reception and feature annotation

[0110] Based on the characteristic tags corresponding to each writing model, the writing ideas for each model are categorized into different categories. For example, if there are characteristic tags such as "concise viewpoint output," "professional report writing," and "creation combining text and images," the corresponding writing ideas are classified into these tag categories respectively.

[0111] The writing task is linked to the core fit tags of each writing model and then to the corresponding writing idea. This step strengthens the correlation between the idea, model features, and task fit, enabling subsequent processing to more clearly understand the source and fit of each idea, and preparing for standardized output of creative materials with distinctive labels.

[0112] (2) Balancing two-stage cross-modal deduplication and diversity preservation

[0113] Single-modal precise deduplication: Based on the creative model's feature tags, writing ideas are grouped according to their features and deduplicated using a single modality. For text or image ideas within the same feature category, appropriate similarity calculation methods are used. For example, text similarity algorithms are used for text ideas, and image feature extraction and comparison algorithms are used for image ideas. Redundant ideas with similarity exceeding a certain threshold are eliminated, while ideas with distinctive features are retained.

[0114] Cross-modal association filtering: After completing single-modal deduplication, cross-modal feature combination filtering is performed. Considering the correlation and complementarity between different modal creatives, such as the coordination between text creatives and image creatives in terms of content expression, cross-modal creative combinations with comprehensive value are selected, and deduplicated logs and retained differentiated creatives are output to provide processing objects for subsequent rating stages.

[0115] (3) Tiered creative rating

[0116] Multi-dimensional scoring: Based on the creative's distinctive tags and quality performance, combined with indicators such as task suitability and distinctive feature prominence, the deduplicated writing creatives are scored from multiple dimensions. For example, task suitability examines the degree to which the creative satisfies the writing task; distinctive feature prominence assesses whether the creative fully reflects the distinctive advantages of the corresponding model, which can be quantified by analyzing the degree of fit between the creative content and the distinctive tags of the corresponding model.

[0117] Human-machine collaborative review: After completing multi-dimensional scoring, the results are reviewed and optimized by humans. Humans can review and adjust the scoring results to ensure the accuracy and rationality of the ratings, and finally output the ranked high-quality creative ideas, providing high-quality creative materials for subsequent diversity verification.

[0118] (4) Diversity verification and completion processing

[0119] Multi-dimensional verification: Based on the feature coverage of the selected creative ideas, verification is conducted from dimensions such as model feature coverage, creative type richness, and perspective differences. This checks for any missing model features, creative types, or perspectives in the creative combinations.

[0120] Model feature coverage can be calculated by comparing the ratio of features already implemented in the model to the total number of features, thus comprehensively evaluating the utilization of model features. Creative type richness can be assessed by identifying the type of each creative idea in a combination, determining its category, and then counting the number of each type and its distribution within the combination, thus comprehensively evaluating creative type richness. Perspective difference can be evaluated by identifying the perspective presented by each writing idea through manual analysis or natural language processing techniques, and then using relevant algorithms to calculate the difference between different creative perspectives. For example, for text-based creative ideas, text similarity algorithms can be used to calculate the similarity in perspective description between different ideas; for image-based creative ideas, image feature extraction and comparison algorithms can be used to calculate the difference in image perspectives. Finally, the perspective difference of all creative perspectives is synthesized to evaluate perspective difference.

[0121] Targeted Completion: If a missing dimension is detected, a targeted model is dispatched to supplement it. For example, if the model's feature coverage is insufficient, a model with the corresponding feature is called to generate more creative ideas; if the richness of creative types is insufficient, a model that can generate specific types of creative ideas is dispatched to supplement it, ultimately outputting a diversity report and a complete creative combination.

[0122] This application's embodiments enhance the correlation and traceability of creative ideas by classifying them according to model feature tags and binding them to core adaptability tags, laying the foundation for subsequent processing. Multi-stage deduplication eliminates redundant ideas while retaining differentiated ones, improving the quality and uniqueness of the ideas. A hierarchical creative rating strategy, combining multi-dimensional scoring with human-machine collaborative review, prioritizes the selection of high-quality ideas with distinctive features and high value. Diversity verification and completion processing ensure the comprehensiveness of creative combinations in terms of model features, types, and perspectives. Overall, it achieves efficient aggregation, selection, and optimization of writing ideas, providing a rich variety of high-quality writing idea combinations to meet the creative needs of intelligent writing tasks, effectively improving the quality and efficiency of creative generation in intelligent writing.

[0123] In some embodiments of this application, the step of using a preset closed-loop iterative optimization strategy to iteratively optimize the writing idea combination and output the final writing result includes: dynamically combining an iteration termination index according to the task type of the writing task; determining whether the writing idea combination triggers a preset iteration termination condition according to the iteration termination index; if triggered, outputting the final writing result; if not triggered, generating an iterative optimization instruction to iteratively optimize the writing idea combination until the preset iteration termination condition is triggered.

[0124] Continue to refer to Figure 2 First, the task type of the writing task is accurately identified. By analyzing the task's characteristic information, such as the task description and objectives, it is determined whether it belongs to different types, such as multimodal tasks, writing + image matching tasks, data visualization tasks, or error correction tasks. Based on the identified task type, appropriate iteration termination indicators are dynamically selected and combined from a pre-set indicator library. For example, if the task is multimodal, the image-text matching indicator "image-text semantic relevance ≥ 85% + image style and text content matching degree ≥ 90%" is selected; if it is a writing + image matching task, "content completeness ≥ 90% + logical coherence ≥ 85% + image-text relevance ≥ 85%" is selected, etc. At the same time, users can customize and supplement thresholds according to specific needs, such as adding indicators such as "industry chart standard compliance ≥ 95%" to meet the requirements of special tasks.

[0125] After determining the iteration termination criteria for the current writing task, various technical means are used to collect the values ​​of each iteration termination criterion in real time. For example, for content completeness, a text integrity algorithm can be used to calculate and analyze whether the text covers all the parts required for the task; the semantic relevance between images and text can be determined using the BLIP-2 model to judge the degree of semantic relevance between images and text; and methods such as edit distance and model refereeing are used to determine whether the content meets the task requirements.

[0126] The collected indicator values ​​are compared with the corresponding iteration termination thresholds, while simultaneously checking whether the maximum number of iterations has been reached. If all indicator values ​​reach or exceed the thresholds, or the maximum number of iterations has been reached, the preset iteration termination condition is triggered, indicating that the writing idea combination has met the task requirements. At this point, the current writing idea combination is output as the final writing result. Otherwise, the preset iteration termination condition is not triggered, and specific iteration optimization instructions are generated based on the indicators that did not meet the requirements. For example, if the content completeness is 82% and the image-text relevance is 78%, instructions such as "add case paragraphs + optimize accompanying image scenarios + polish output text" are generated to guide targeted optimization of the writing idea combination. Then, iteration optimization and condition judgment are performed again until the preset iteration termination condition is triggered.

[0127] This application's embodiments, by dynamically combining iterative termination indicators based on task type, can accurately set reasonable termination conditions for the characteristics and requirements of different task types, improving the targeting and effectiveness of iterative optimization. Real-time collection of indicator values ​​and conditional judgments ensure timely monitoring of the optimization status of creative combinations during the iteration process. Iterative optimization or output of the final result based on the judgment results enables continuous improvement of writing creative combinations until the task requirements are met. Overall, it achieves automated and precise iterative optimization of creative combinations during intelligent writing, effectively improving the quality and accuracy of writing results, making the final writing outcome more in line with task requirements and user expectations.

[0128] In some embodiments of this application, the step of generating iterative optimization instructions to iteratively optimize the writing idea combination includes: performing a multi-dimensional defect scan on the writing idea combination to obtain multi-dimensional defect scan results; generating executable suggestions based on the multi-dimensional defect scan results; and using a multi-path generator to initiate multiple optimization paths for concurrent optimization based on the multi-dimensional defect scan results and the executable suggestions to obtain optimized writing results.

[0129] Multi-dimensional defect scanning mainly includes the following aspects:

[0130] (1) Textual defect scanning: The logical relationship extraction algorithm is used to analyze the text content in the creative combination of writing to identify whether there are problems such as "reversal of cause and effect" (such as stating the result first and then explaining the cause with logical confusion) and "lack of evidence" (the argument is not supported by sufficient evidence). At the same time, the keyword matching method is used to check whether there are "information contradictions" in the text, such as the description of the same concept in different paragraphs conflicting with each other.

[0131] (2) Image defect scanning: Use target detection algorithms (such as YOLO) to detect images. Check if there are "missing elements" in the image, such as data charts without coordinate axis labels, which makes the information not accurately conveyed; check if there is "style deviation" in the image, such as a serious report with a cartoon-style picture, which is inconsistent with the overall style.

[0132] (3) Cross-modal defect scanning: The CLIP+BLIP-2 joint algorithm is used to detect the association between text and images. It identifies whether there is a "text-image semantic conflict", such as the text describing "emission reduction effect" but the accompanying image is a "pollution scene image"; it also checks whether there are "labeling errors", such as the data in the chart does not match the text description, or the caption information is inaccurate.

[0133] In-depth analysis of multi-dimensional defect scanning results is conducted to identify the root causes of problems. Based on different root causes, corresponding actionable suggestions are generated and labeled with the suggestion type. For text-based problems, such as missing evidence, the suggestion is "Supplement XX data"; for image-based problems, if the image style is inappropriate, the suggestion is "Adjust the generated keywords to 'Change XX graph to a bar chart, green color scheme, professional business style'"; for cross-modal problems, such as labeling errors, the suggestion is "Synchronously modify text labeling". Figure 1 "Data for XX in 2023".

[0134] Continue to refer to Figure 2 The multi-path generator initiates multiple optimization paths for concurrent optimization. Various optimization paths, such as text, image, and union optimizations, can be pre-stored, and the execution logic between each path is clearly defined. For example, when information inconsistencies are detected in the text and style deviations are detected in the image, the text optimization path and the image optimization path can be initiated simultaneously. Based on the defect type, the multi-path generator initiates multiple adapted optimization paths in parallel. Different paths optimize different problems. During the optimization process, each path runs independently but also collaborates with others to ensure the effectiveness of the overall optimization.

[0135] By comparing the improvement rates of text and image-text adaptation metrics, the optimal result is selected from the results generated in parallel from different improvement directions and used as the output of this iteration optimization.

[0136] This application's embodiments construct a quality control system encompassing full-process SOP execution, multi-dimensional detection, and multi-stage cross-modal consistency verification. Through various advanced algorithms, it achieves precise verification of writing idea combinations in terms of semantics, details, style, and task objective consistency. Coupled with automatic improvement and dual-model cross-validation mechanisms, it effectively ensures the quality of the output results. Simultaneously, the multi-path parallel improvement system can quickly initiate appropriate optimization paths for different defect types, performing concurrent optimization and selecting the optimal result through indicator comparison, greatly improving the efficiency of iterative optimization. Overall, it achieves high-quality and high-efficiency optimization of writing idea combinations, providing users with more compliant and higher-quality writing results.

[0137] In summary, the key points of this application are mainly as follows:

[0138] (1) Design schemes of three engines: “Writing Task Feature Perception and Model Intelligent Scheduling Engine”, “Creative Convergence Engine”, and “Closed-Loop Iterative Optimization Engine”, and the overall design scheme of this application.

[0139] (2) The dual-objective scheduling algorithm of “weighted cosine similarity-computing cost balance” achieves accurate screening and resource optimization of multimodal models by combining the weight matrix and scenario coefficients through the three-level calculation of “basic matching-dynamic adjustment-cost constraint” and the “three-step progressive screening” system.

[0140] (3) The “Model Feature Multidimensional Label - Differentiation Instruction” linkage activation mechanism, combined with the “Group Deduplication by Feature - Cross-modal Association Filtering” two-stage deduplication and the “Model Feature Coverage + Creative Type Richness + Perspective Difference” three-dimensional verification, activates the model feature output of differentiated creative ideas from the source and solves the homogenization problem.

[0141] (4) The “activation-receive-deduplication-rating-verification-storage” full-process collaborative system takes “model characteristics” as the core and runs through each link. It selects high-value ideas through the “core + innovation diversity” rating system and combines the “model characteristics + task type” dual-dimensional idea pool for dynamic iteration to achieve efficient aggregation and accurate reuse of multimodal ideas.

[0142] (5) Multimodal closed-loop iterative system, through dynamic threshold generation, cross-modal defect root cause localization (based on defect sample map) and multi-path parallel improvement, breaks through the bottleneck of iterative rigidity and realizes automatic iterative closed-loop optimization.

[0143] (6) The quality control system of “full-process SOP execution + multi-dimensional detection + cross-modal consistency multi-stage verification” is established by using contextual relevance detection (including image sequence association) and factual consistency detection (including image fact verification), and adopting BLIP-2, CLIP and YOLO to realize the consistency verification of “semantic-detail-style” text and task objectives. Combined with automatic improvement and dual-model cross-validation, the full-process quality control system is established to ensure the quality of output results.

[0144] (7) Multi-path parallel improvement system: pre-store text / image / joint path and associate execution logic, start multiple adaptation paths in parallel for defect type, and select the best result by comparing the improvement of text and image adaptation index, so as to make the iterative optimization process more efficient.

[0145] The main technical effects of this application include:

[0146] (1) Compared with existing technologies that are mostly limited to a single text modality and rely on a single series or fixed-size model output, making it difficult to release the characteristics of different models, this application has better multimodal collaboration capabilities and more fully mobilizes the advantages of models: Through the multi-module collaborative architecture of "dynamic scheduling + creative convergence + closed-loop iteration", the dynamic scheduling module adopts the "weighted cosine similarity - computing power cost balance" algorithm to accurately match the capability tags of multiple series of models such as LLaMA series, Qwen series, and CogVLM series, while adapting to the different size advantages from 0.6B lightweight models to 70B large models - both leveraging the computing power efficiency of lightweight models and releasing the deep adaptation capabilities of large models; then, the "feature tag - differentiated instruction" mechanism of the creative convergence module activates the exclusive capabilities of each model, and finally realizes deep cross-modal fusion of text and image. This approach specifically solves the core problems of existing solutions such as "only text assistance, lack of image-text collaborative creativity", "incomplete release of model advantages" and "insufficient multi-model collaboration", and truly achieves the technical effect of "collective intelligence collaboration and multiplied efficiency".

[0147] (2) Higher creative diversity and accuracy: Existing technologies rely on a single large model output or template filling, which easily leads to the homogenization of creative ideas. This application activates the characteristic output of multiple models to create differentiated creative ideas through the mechanism of "model feature activation + differentiated instructions + two-stage deduplication", and coupled with "three-dimensional diversity verification", it not only ensures the richness of creative ideas but also avoids redundancy.

[0148] (3) Improved closed-loop iteration and quality control: Existing technology improvements are mostly single-suggestion outputs without closed-loop optimization. This application constructs a closed loop of "defect analysis - multi-path improvement - effect verification" and combines "semantic-detail-style" multi-stage graphic verification to solve the problem of "extensive improvement and insufficient quality control" in existing solutions, thereby improving the stability of output quality.

[0149] (4) Better resource adaptation and reuse efficiency: Existing technologies do not consider computing power cost and creative reuse. This application optimizes resource allocation through the "weighted cosine similarity-computing power cost balance" scheduling algorithm, combined with dynamic iteration of the "creative pool", persistent storage, and result reuse, which reduces computing power consumption and improves the creative reuse efficiency of subsequent tasks.

[0150] This application also provides an intelligent writing device 400, such as... Figure 4 As shown, a structural schematic diagram of an intelligent writing device according to an embodiment of this application is provided. The intelligent writing device 400 includes:

[0151] The feature recognition unit 410 is used to receive the writing task and perform core feature recognition on the writing task to obtain the core feature information of the writing task.

[0152] The scheduling decision unit 420 is used to determine a combination of intelligent writing models that are suitable for the writing task based on the core feature information of the writing task and the pre-built model resource library, using a preset scheduling decision strategy.

[0153] The creative aggregation unit 430 is used to generate writing ideas for each writing model using multiple writing models in the intelligent writing model combination and process them using a preset creative aggregation processing strategy to obtain a combination of writing ideas.

[0154] The iterative optimization unit 440 is used to iteratively optimize the combination of writing ideas using a preset closed-loop iterative optimization strategy, and output the final writing result.

[0155] In some embodiments of this application, the pre-built model resource library stores capability tags for multiple writing models. The scheduling decision unit 420 is specifically used for: calculating the core fit between the writing task and each writing model based on the core feature information of the writing task and the capability tags of the multiple writing models; correcting the core fit between the writing task and each writing model using a dynamic weight correction strategy to obtain a corrected core fit value; constraining the core fit value between the writing task and each writing model using a computing cost constraint strategy to obtain a computing cost constraint value between the writing task and each writing model; and selecting multiple writing models based on the core fit between the writing task and each writing model and the computing cost constraint value between the writing task and each writing model using a hierarchical screening strategy to combine intelligent writing models that are compatible with the writing task.

[0156] In some embodiments of this application, the scheduling decision unit 420 is specifically used to: filter multiple writing models based on the core fit between the writing task and each writing model using a preset core fit threshold to obtain a first candidate writing model; filter the first candidate writing model according to a type balance filtering strategy to obtain a second candidate writing model; and filter the second candidate writing models according to the computing power cost constraints between the writing task and each second candidate writing model to obtain the final intelligent writing model combination.

[0157] In some embodiments of this application, the creative aggregation unit 430 is specifically used to: generate differentiated instructions corresponding to each writing model according to the multiple writing models in the intelligent writing model combination; guide each writing model to generate writing ideas using the differentiated instructions corresponding to each writing model, thereby obtaining the writing ideas and corresponding model feature tags of each writing model; and process the writing ideas and corresponding model feature tags of each writing model using the preset creative aggregation processing strategy to obtain the writing creative combination.

[0158] In some embodiments of this application, the creative aggregation unit 430 is specifically used for: classifying the writing ideas of each writing model according to the model feature tags based on the writing ideas of each writing model and the corresponding model feature tags, and binding the writing task with the core adaptability tags of each writing model; performing multi-stage deduplication processing on the writing ideas of multiple writing models after classification by model feature tags to obtain deduplicated writing ideas; rating the deduplicated writing ideas using a hierarchical creative rating strategy based on the deduplicated writing ideas, the corresponding model feature tags, and the core adaptability tags to obtain the rating results of the writing ideas; and performing multi-dimensional verification and completion processing of the writing ideas using a diversity verification and completion strategy based on the rating results of the writing ideas to obtain the final combination of writing ideas.

[0159] In some embodiments of this application, the iterative optimization unit 440 is specifically used to: dynamically combine iterative termination indicators according to the task type of the writing task; determine whether the combination of writing ideas triggers a preset iterative termination condition according to the iterative termination indicators; if triggered, output the final writing result; if not triggered, generate iterative optimization instructions to iteratively optimize the combination of writing ideas until the preset iterative termination condition is triggered.

[0160] In some embodiments of this application, the iterative optimization unit 440 is specifically used to: perform multi-dimensional defect scanning on the writing idea combination to obtain multi-dimensional defect scanning results; generate executable suggestions based on the multi-dimensional defect scanning results; and, based on the multi-dimensional defect scanning results and the executable suggestions, use a multi-path generator to initiate multiple optimization paths for concurrent optimization to obtain optimized writing results.

[0161] It is understood that the above-mentioned intelligent writing device can realize each step of the intelligent writing method provided in the foregoing embodiments. The relevant explanations of the intelligent writing method are applicable to the intelligent writing device and will not be repeated here.

[0162] Figure 5 This is a schematic diagram of the structure of a device according to an embodiment of this application. For example... Figure 5 As shown, the device includes one or more processors (or processing units), and may also include one or more memories coupled to the processors, and may also include a communication module coupled to the processors.

[0163] A communication module can be used to communicate with other devices or apparatuses, such as sending or receiving data and / or signals. A communication module may have at least one communication module for communication. A communication module may include any interface necessary for communicating with other devices. Exemplarily, a communication module may be a transceiver, circuit, bus, module, or other type of communication module.

[0164] The processor may include, but is not limited to, one or more of the following: a general-purpose computer, a special-purpose computer, a microcontroller, a digital signal processor (DSP), or a controller-based multi-core controller architecture. The device may have multiple processors, such as application-specific integrated circuit (ASIC) chips, which are time-dependent on a clock synchronized with the main processor.

[0165] The memory may include one or more non-volatile memories and one or more volatile memories. Examples of non-volatile memories include, but are not limited to, at least one of the following: read-only memory (ROM), electrically programmable read-only memory (EPROM), flash memory, hard disk, compact disc (CD), digital video disc (DVD), or other magnetic and / or optical storage. Examples of volatile memories include, but are not limited to, at least one of the following: random access memory (RAM), or other volatile memories that do not persist during the duration of a power outage.

[0166] A computer program consists of computer-executable instructions that are executed by an associated processor. Programs can be stored in ROM. A processor can perform any appropriate action and processing by loading the program into RAM.

[0167] Possible implementations of this application can be achieved through a program, enabling the communication device to execute any of the processes discussed in the foregoing embodiments. Possible implementations of this application can also be achieved through hardware or a combination of software and hardware.

[0168] In some implementations, the program may be tangibly contained in a computer-readable storage medium, which may include in a device (such as in memory) or other storage device accessible by the device. The program may be loaded from the computer-readable storage medium into RAM for execution. The computer-readable storage medium may include any type of tangible non-volatile memory, such as ROM, EPROM, flash memory, hard disk, CD, DVD, etc.

[0169] This application also provides a computer-readable storage medium storing computer instructions or program code thereon, which, when executed by a processor, causes the processor to perform the methods and functions involved in any of the above embodiments. A computer-readable medium can be any tangible medium that contains or stores a program for or relating to an instruction execution system, apparatus, or device. A computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. A computer-readable medium can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination thereof. The computer-readable storage medium can be any available medium accessible to a computer or a data storage device such as a server or data center that integrates one or more available media. More detailed examples of computer-readable storage media include electrical connections with one or more wires, magnetic media (e.g., disks, floppy disks, hard disks, magnetic tapes, magnetic storage devices), optical media (e.g., optical storage devices, DVDs), semiconductor media (e.g., solid-state drives), random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), or any suitable combination thereof.

[0170] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. Embodiments of this application also provide at least one computer program product tangibly stored on a non-transitory computer-readable storage medium. This computer program product includes one or more computer-executable instructions, such as instructions included in a program module, which execute in a device on a target real or virtual processor to perform the processes, methods, and functions involved in any of the above embodiments. When the computer program instructions are loaded and executed on a computer, all or part of the flow or function according to the embodiments of this application is generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line) or wireless (e.g., infrared, wireless, microwave, etc.) means.

[0171] This application also proposes a computer program product, including a computer program or instructions that, when run on a computer, cause the computer to perform the processes, methods, and functions described in the above embodiments. Typically, program modules include routines, programs, libraries, objects, classes, components, data structures, etc., that perform specific tasks or implement specific abstract data types. In various embodiments, the functionality of program modules can be combined or divided as needed. The machine-executable instructions for the program modules can be executed locally or in a distributed device. In a distributed device, the program modules can reside in both local and remote storage media.

[0172] Generally, the various embodiments of this application can be implemented in hardware or dedicated circuitry, software, logic, or any combination thereof. Some aspects can be implemented in hardware, while others can be implemented in firmware or software, which can be executed by a controller, microprocessor, or other computing device. Although various aspects of the embodiments of this disclosure are shown and described as block diagrams, flowcharts, or represented using some other illustration, it should be understood that the blocks, apparatuses, systems, techniques, or methods described herein can be implemented as, as non-limiting examples, in hardware, software, firmware, dedicated circuitry or logic, general-purpose hardware or controllers or other computing devices, or some combination thereof.

[0173] It should be noted that although embodiments of this application have been described above with reference to the accompanying drawings, these embodiments are not independent of each other, and they can be combined to obtain other embodiments. The methods, situations, categories, and classifications of embodiments in this application are only for the convenience of description and should not constitute a special limitation. Various methods, categories, situations, and features in embodiments can be combined with each other if logically consistent. The various embodiments of this application can be arbitrarily combined to achieve different technical effects. The embodiments of this application will not list various combinations.

[0174] Furthermore, although the operation of the methods of this disclosure is described in a specific order in the accompanying drawings, this does not require or imply that these operations must be performed in that specific order, or that all of the operations shown must be performed to achieve the desired result. Rather, the steps depicted in the flowcharts may be performed in a different order. Additionally or alternatively, certain steps may be omitted, multiple steps may be combined into one step, and / or one step may be broken down into multiple steps. It should also be noted that the features and functions of two or more devices according to this disclosure may be embodied in one device. Conversely, the features and functions of one device described above may be further divided and embodied by multiple devices.

[0175] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0176] The above description is merely an embodiment of this application and is not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. An intelligent writing method, characterized in that, The intelligent writing method includes: Receive a writing task and perform core feature identification on the writing task to obtain the core feature information of the writing task; Based on the core feature information of the writing task and the pre-built model resource library, a combination of intelligent writing models adapted to the writing task is determined using a preset scheduling decision strategy. The writing ideas of each writing model are generated by using multiple writing models in the intelligent writing model combination and processed by a preset idea aggregation processing strategy to obtain a combination of writing ideas. The writing idea combination is iteratively optimized using a pre-defined closed-loop iterative optimization strategy to output the final writing result.

2. The intelligent writing method according to claim 1, characterized in that, The pre-built model resource library stores capability tags for multiple writing models. The step of determining a combination of intelligent writing models suitable for the writing task based on the core feature information of the writing task and the pre-built model resource library, using a preset scheduling decision strategy, includes: Based on the core feature information of the writing task and the capability tags of multiple writing models, calculate the core fit between the writing task and each writing model. The core fit between the writing task and each writing model is corrected using a dynamic weight correction strategy to obtain the core fit correction value between the writing task and each writing model. The computing cost constraint strategy is used to constrain the core fit correction value of the writing task and each writing model by computing cost, so as to obtain the computing cost constraint value of the writing task and each writing model. Based on the core compatibility between the writing task and each writing model, as well as the computational cost constraints between the writing task and each writing model, a hierarchical screening strategy is used to screen multiple writing models and combine intelligent writing models that are compatible with the writing task.

3. The intelligent writing method according to claim 2, characterized in that, The step involves using a hierarchical screening strategy to select multiple writing models based on the core compatibility between the writing task and each writing model, as well as the computational cost constraints of the writing task and each writing model. The combination of intelligent writing models adapted to the writing task includes: Based on the core fit between the writing task and each writing model, multiple writing models are screened using a preset core fit threshold to obtain the first candidate writing model. The first candidate writing model is selected based on the type balance screening strategy to obtain the second candidate writing model; Based on the writing task and the computational cost constraints of each second candidate writing model, the second candidate writing models are screened to obtain the final intelligent writing model combination.

4. The intelligent writing method according to claim 1, characterized in that, The process of generating writing ideas for each writing model using multiple writing models in the intelligent writing model combination and processing them using a preset idea aggregation and processing strategy to obtain a writing idea combination includes: Based on the multiple writing models in the intelligent writing model combination, generate differentiated instructions corresponding to each writing model; By using differentiated instructions corresponding to each writing model, we can guide each writing model to generate writing ideas, thereby obtaining the writing ideas and corresponding model feature tags for each writing model. Based on the writing ideas of each writing model and the corresponding model feature tags, the preset idea aggregation processing strategy is used to process them to obtain the writing idea combination.

5. The intelligent writing method according to claim 4, characterized in that, The process of processing the writing ideas and corresponding model feature tags of each writing model using the preset idea aggregation processing strategy to obtain the writing idea combination includes: Based on the writing ideas of each writing model and the corresponding model feature tags, the writing ideas of each writing model are classified according to the model feature tags, and the writing task is bound to the core adaptability tag of each writing model. The writing ideas of multiple writing models categorized by model feature tags are subjected to multi-stage deduplication to obtain deduplicated writing ideas. Based on the deduplicated writing ideas, their corresponding model feature tags, and core fit tags, a hierarchical creative rating strategy is used to rate the deduplicated writing ideas, resulting in a rating result for the writing ideas. Based on the rating results of the writing ideas, a multi-dimensional verification and completion strategy is used to verify and complete the writing ideas in order to obtain the final combination of writing ideas.

6. The intelligent writing method according to claim 1, characterized in that, The step of iteratively optimizing the combination of writing ideas using a pre-defined closed-loop iterative optimization strategy to output the final writing result includes: The iterative termination index is dynamically combined based on the task type of the writing task. Based on the iteration termination index, determine whether the writing idea combination triggers the preset iteration termination condition; If triggered, the final writing result will be output; If not triggered, an iterative optimization instruction is generated to iteratively optimize the combination of writing ideas until the preset iteration termination condition is triggered.

7. The intelligent writing method according to claim 6, characterized in that, The generation of iterative optimization instructions for iteratively optimizing the combination of writing ideas includes: A multi-dimensional defect scan was performed on the aforementioned writing creative combination to obtain the multi-dimensional defect scan results; Based on the multi-dimensional defect scan results, generate executable suggestions; Based on the multi-dimensional defect scanning results and the executable suggestions, a multi-path generator is used to initiate multiple optimization paths for concurrent optimization, resulting in optimized writing results.

8. An intelligent writing device, characterized in that, The intelligent writing device includes: The feature recognition unit is used to receive the writing task and perform core feature recognition on the writing task to obtain the core feature information of the writing task. The scheduling decision unit is used to determine a combination of intelligent writing models that are suitable for the writing task based on the core feature information of the writing task and the pre-built model resource library, using a preset scheduling decision strategy. The creative aggregation unit is used to generate writing ideas for each writing model using multiple writing models in the intelligent writing model combination, and process them using a preset creative aggregation processing strategy to obtain a combination of writing ideas. The iterative optimization unit is used to iteratively optimize the combination of writing ideas using a preset closed-loop iterative optimization strategy, and output the final writing result.

9. An apparatus comprising: processor; And a memory arranged to store computer-executable instructions, which, when executed, cause the processor to perform any of the intelligent writing methods of claims 1 to 7.

10. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instruction is executed by the processor, it implements any one of the intelligent writing methods described in claims 1 to 7.

Images