Content generation method and system based on multi-dimensional attribute balance
By establishing a multi-dimensional attribute indicator system and dynamic optimization algorithm, the problem of uncontrollable results generated by large language models was solved, and the real-time controllability and precise balance of content in multiple dimensions were achieved, thereby improving the balance and value alignment capabilities of the generated content.
Patent Information
- Application Number
- CN202511675622.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-16
- Publication Date
- 2026-02-10
AI Technical Summary
Existing large language models lack multi-dimensional and quantifiable attribute balance constraints in content generation, resulting in uncontrollable generation results with biases and singular tendencies. They fail to meet the needs of personalization and context adaptability, and intervention methods are lagging and singular, lacking calculable balance standards.
A multi-dimensional attribute indicator system is established, a quantitative balance target range is set, and the generation parameters are adjusted through real-time analysis and dynamic optimization algorithms to achieve synchronous balance of content in multiple dimensions. In-depth data from the upstream cognitive system is introduced as a basis for decision-making.
It achieves multi-dimensional real-time controllability and precise balance of generated content, improves the balance and value alignment capabilities of generated content, and provides integrable core technology components.
Smart Images
Figure FT_1 
Figure FT_2 
Figure FT_3
Abstract
Description
Technical Field
[0001] This invention relates to the fields of artificial intelligence, natural language processing, and controllable generative AI, and particularly to a method and system for introducing multi-dimensional attribute constraints and dynamic balance mechanisms in the content generation process. Background Technology
[0002] Current large language models suffer from fundamental technical flaws in content generation: the generated results are uncontrollable, and the models may generate biased, overly inflammatory, or singularly biased content, posing AI security risks. The root cause lies in the disconnect between existing generation systems and upstream perception and cognitive systems, lacking deep, structured contextual constraints. The decision-making process is superficial: mainstream content generation and alignment technologies (such as reinforcement learning based on human feedback) rely on general, coarse feedback signals, failing to integrate and utilize deep user-derived data from upstream cognitive systems (such as structured situational labels with clear business semantics output by mapping methods based on user state and structured situational labels) as context for the generation process. This results in generated content lacking true personalization and contextual adaptability, essentially remaining "undifferentiated" content production, and failing to meet the needs of precise services.
[0003] Intervention is often limited in scope and delayed: Existing intervention methods are mostly focused on single dimensions such as "safety," and are largely post-generation filtering mechanisms with penalties. They lack the ability to simultaneously and precisely guide content in multiple orthogonal attributes such as "action orientation, emotional resonance, and logical rigor" during the generation process. This stems from the lack of a quantifiable, multi-dimensional content attribute indicator system as a benchmark for process constraints, making it impossible to align the intrinsic value during the generation process.
[0004] The concept of balance is not calculable: In existing technologies, there is a lack of clear, calculable, and verifiable quantitative standards for "balance" or "neutrality," making it difficult to integrate into automated content generation pipelines. This results in a lack of core infrastructure for building controllable generation systems, making it impossible to systematically solve the problems of content bias and singular tendencies.
[0005] Therefore, there is an urgent need in this field for a new technological paradigm that can introduce multi-dimensional, quantifiable, and real-time attribute balance constraints in the content generation process and respond to the output of the upstream cognitive system. Summary of the Invention
[0006] (a) Purpose of the invention The purpose of this invention is to overcome the shortcomings of the prior art and provide a content generation method and system based on multi-dimensional attribute balance, so as to solve the technical problem of imbalance in multi-dimensional attributes of the output content caused by the lack of process constraints in the basic large language model.
[0007] (II) Technical Solution To achieve the above-mentioned objectives, the present invention adopts the following technical solution: A content generation method based on multidimensional attribute balance includes: establishing a content attribute index system containing multiple mutually orthogonal dimensions, and setting a quantitative balance target range for each dimension; during the content generation process, analyzing the attribute intensity values of candidate content in each dimension in real time; adjusting the generation parameters in real time based on predefined balance rules and dynamic optimization algorithms to drive the distribution of attribute intensity values in each dimension to approach the target range; and outputting content that meets the requirements of multidimensional attribute balance.
[0008] A content generation system based on multidimensional attribute balancing includes: an attribute indicator system construction module, a content generation and acquisition module, a real-time attribute strength analysis module, a dynamic balancing optimization module, and a balanced content output module.
[0009] (III) Beneficial Effects Compared with the prior art, the present invention has the following beneficial effects: 1. By incorporating balance control into the content generation process, the generated content is intrinsically guided, enhancing the real-time controllability of the content's multi-dimensional attributes; 2. Through a quantifiable attribute indicator system and dynamic optimization algorithms, the content is synchronized and precisely balanced across multiple dimensions; 3. It provides a core technology component that can be integrated into various AIGC applications, providing standardized capabilities for the balance and value alignment of generated content. Attached Figure Description
[0010] Figure 1 This is a schematic diagram of the system module composition provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the method flow provided in an embodiment of the present invention; Figure 3 This is a flowchart illustrating the working logic of the dynamic balance optimization module in this embodiment of the invention. Detailed Implementation
[0011] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be described in further detail below with reference to the accompanying drawings and embodiments.
[0012] Example 1: Basic Flow and Implementation of the Method See Figure 2 The content generation method provided in this embodiment includes the following steps: Step S301: Establish a multi-dimensional attribute indicator system.
[0013] Based on business scenarios and value objectives, a set of mutually orthogonal and quantifiable content attribute dimensions are defined, and a balanced target range is set for each dimension.
[0014] Step S302: Obtain candidate content.
[0015] Connect to a basic large language model to obtain its candidate content generated based on user input.
[0016] Step S303: Analyze attribute strength in real time.
[0017] A pre-trained multi-label attribute classifier is used to analyze candidate content in real time and calculate the attribute strength value of each attribute dimension.
[0018] Step S304: Dynamic balance optimization.
[0019] The dynamic balancing optimization module receives attribute strength values for each dimension. Its optimization objective is to minimize the absolute deviation of each dimension's strength value from the target interval's median. Optimization methods include: dynamically adjusting the sampling parameters of the underlying large language model; applying lexical probability biases based on attribute strength during decoding; and executing balancing rules that define coupling constraints.
[0020] Step S305: Output balanced content.
[0021] When the intensity values of the generated content fall within the balanced target range of all attribute dimensions, or the deviations of the content are all less than the threshold, the content is output as the final result.
[0022] Example 2: System Performance Verification To verify the technical effectiveness of the present invention, a comparative experiment was designed.
[0023] Experimental setup: Task: Generate mental health recommendations; Comparison System: The system of this invention compared to a baseline model that has not undergone multidimensional attribute balancing optimization; Evaluation metric: Attribute balance, which is the proportion of each dimension's intensity value falling within the target range.
[0024] Experimental results: On a test set of 1000 generated content items, the attribute balance of the system of this invention was 91.7%, while the attribute balance of the baseline model was 52.1%.
[0025] Experimental conclusion: Experimental data show that by introducing multi-dimensional attribute balance constraints, this invention can significantly improve the balance of generated content across preset dimensions.
[0026] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims
1. A content generation method based on multidimensional attribute balancing, characterized in that, include: Step S301: Establish a content attribute indicator system that includes multiple dimensions, and set a quantitative balance target range for each dimension; Step S302: Obtain candidate content generated by the basic large language model; Step S303: Analyze the attribute strength values of the candidate content in each dimension of the indicator system in real time; Step S304: Based on predefined balancing rules and dynamic optimization algorithms, intervene and optimize the generation process of the candidate content to drive the distribution of attribute intensity values in each dimension to approach the balancing target range; Step S305: Output the optimized content that meets the requirements for multidimensional attribute balance.
2. The method according to claim 1, characterized in that, The content attribute index system described in step S301 includes multiple mutually orthogonal attribute dimensions.
3. The method according to claim 2, characterized in that, The attribute dimensions include at least two of the following: action orientation, emotional resonance, logical rigor, innovation inspiration, and cognitive certainty.
4. The method according to claim 1, characterized in that, The real-time analysis described in step S303 employs a multi-label classification method based on a pre-trained attribute classifier to perform streaming and parallel evaluation of the token sequence during the generation process.
5. The method according to claim 1, characterized in that, The intervention and optimization of the generation process described in step S304 includes adjusting the sampling parameters of the basic large language model, applying a word probability bias based on attribute strength during the decoding process, or using model predictive control theory to perform rolling optimization of the generation trajectory.
6. The method according to claim 5, characterized in that, The balancing rule in step S304 defines the coupling constraint relationship between attribute dimensions. When the attribute strength of a certain dimension deviates from the target range, the generation weight of other dimensions that have a preset constraint relationship with it will be adjusted in conjunction.
7. The method according to claim 1, characterized in that, The method constructs a real-time technical closed loop of analysis-optimization-generation, and the attribute balance of its output content can be used as a feedback signal for iterative updates of the optimization model.
8. A content generation system based on multidimensional attribute balancing, used to implement the method as described in any one of claims 1 to 7, characterized in that, include: The attribute indicator system construction module is used to execute step S301; The content generation and acquisition module is used to execute step S302; The attribute strength real-time analysis module is used to execute step S303; The dynamic balance optimization module is used to execute step S304; The balanced content output module is used to perform step S305.
9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1 to 7.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1 to 7.