Man-machine collaborative auditing and optimizing method for AI (Artificial Intelligence) generated content in field of literature and blog

Abstract

The invention discloses a man-machine collaborative auditing and optimizing method for AI (artificial intelligence) generated content oriented to the field of literature and blog. The man-machine collaborative auditing and optimizing method comprises the following steps: S1, content acquisition and personalized constraint preprocessing: acquiring literature and blog field related content generated by an AI model; if the content is customized by a specific object, inputting an exclusive digital resource library and a style guide of the corresponding object into an AI generation model as constraint conditions; s2, multi-dimensional automatic evaluation: calling a multi-dimensional content evaluation model to carry out automatic analysis on the content, and outputting a quality score vector containing a plurality of evaluation dimension scores; s3, man-machine task dynamic allocation and collaborative auditing: calculating the comprehensive confidence of the content based on the quality score vector, and dynamically allocating the low-confidence content to the matched artificial auditor according to the confidence and the professional portrait of the auditor; and an auditor can trigger traceability verification based on the authoritative literature and blog mapping knowledge domain during auditing. According to the method, the AI generation content in the literature and blog field can be better audited and optimized.

Description

Technical Field

[0001] This invention relates to the field of data processing, and more specifically to a human-machine collaborative review and optimization method for AI-generated content in the cultural heritage field. Background Technology

[0002] Currently, AI-generated content in the cultural heritage field faces challenges in quality control. Existing review methods often lack targeted, multi-dimensional assessments such as historical accuracy and cultural context suitability, leading to biased judgments due to single-dimensional biases. Furthermore, the human-machine collaboration mechanism is imperfect, with task allocation failing to consider content confidence and reviewer professional profiles, resulting in wasted manpower or inaccurate reviews. Customized content lacks specific constraints and suitability assessments, failing to meet the needs of specific audiences. Moreover, the lack of a closed-loop feedback mechanism for review results prevents targeted iterative optimization of AI-generated models, hindering the continuous improvement of the professionalism, reliability, and suitability of AI-generated content in the cultural heritage field. Therefore, an efficient and accurate human-machine collaborative review and optimization solution is urgently needed. This paper proposes a human-machine collaborative review and optimization method for AI-generated content in the cultural heritage field. Summary of the Invention

[0003] The present invention solves the above-mentioned technical problems through the following technical solution, and the present invention includes the following steps: S1. Content Acquisition and Personalization Constraint Pre-processing: Acquire relevant cultural heritage content generated by the AI ​​model; if the content is customized for a specific object, input the object's exclusive digital resource library and style guide as constraints into the AI ​​generation model; S2. Multi-dimensional automatic evaluation: The multi-dimensional content evaluation model is invoked to automatically analyze the content and output a quality score vector containing scores of multiple evaluation dimensions. S3. Dynamic allocation and collaborative review of human and machine tasks: The comprehensive confidence level of the content is calculated based on the quality scoring vector, and low-confidence content is dynamically assigned to the matched human reviewers based on this confidence level and the professional profile of the reviewers; the reviewers can trigger source verification based on authoritative cultural heritage knowledge graph during the review process. S4. Decision Integration and Model Iteration Optimization: The final review conclusion is formed by integrating the results of manual review and automatic evaluation. Based on this conclusion and the review process data, a feedback dataset is constructed to optimize the AI-generated model in a targeted manner, so as to realize the closed-loop feedback of the review results to the generated model.

[0004] 2. The method for human-machine collaborative review and optimization of AI-generated content in the cultural heritage field according to claim 1, characterized in that: the evaluation dimensions of the "multi-dimensional content evaluation model" in step S2 include at least: Historical accuracy dimension: The generated content is semantically compared with authoritative historical document databases to calculate the consistency score Sh of key fact descriptions; Cultural context conformity dimension: Using a cultural context embedding model trained with cultural heritage corpus, we analyze the appropriateness of the generated content in a specific historical context and generate a context deviation score Sc. Logical coherence dimension: Analyzes the self-consistency of events and causal relationships within the content through a pre-trained language model, and outputs a logical coherence score Sl; The quality score vector is .

[0005] Furthermore, the comprehensive confidence level of the calculated content in step S3 is specifically achieved in the following way: The scores of each dimension in the quality score vector S are normalized to obtain a normalized vector. ; Dynamic weight vectors are assigned to each evaluation dimension based on the type of cultural heritage content. ; The calculation process for the overall confidence level is expressed as follows: ; in, Here, b is the sigmoid activation function, and b is the bias term used to map the result to the [0,1] interval. Furthermore, the dynamic allocation based on confidence level and the reviewer's professional profile in step S3 specifically includes: S3.1 Sort all the content to be reviewed according to the comprehensive confidence level C; S3.2 Maintain a dynamic professional competence profile for each human auditor based on historical audit data; S3.3 For content items where C is lower than the dynamic threshold Td1, calculate the matching degree between its quality score vector S and the professional competence profile of each available reviewer, and assign the task to the reviewer with the highest matching degree. S3.4 The dynamic threshold Td1 is periodically adjusted based on the average confidence level of the current task queue, the workload of the review manpower, and the historical review pass rate.

[0006] Furthermore, the specific steps for triggering the source verification based on the authoritative cultural heritage knowledge graph in step S3 are as follows: In response to the reviewer's selection of a specific entity or statement in the generated content, the system uses that fragment as query input; Perform multi-hop association retrieval in the knowledge graph to find attributes, historical events, literature records and academic viewpoints related to the query entity; The search results are presented in the form of evidence chain graphs and text summaries, and the semantic similarity between the generated content fragments and the core evidence is automatically calculated as a reference.

[0007] Furthermore, the S4 step employs a weighted decision fusion strategy to integrate the results of manual review and automatic evaluation: Let the conclusion of the manual review be Vh, and the overall confidence level automatically assessed by the system be C; Set a high confidence threshold Thigh and a low confidence threshold Tlow; The rules for generating the final audit conclusion Vf$ are as follows: when In this case, the main approach is to rely on systematic evaluation and inference, and human conclusions must be accompanied by strong evidence to be overturned; when hour, , where α is the human weighting coefficient that increases with the auditor’s historical accuracy, and sign is the sign function; when In such cases, the conclusion of manual review shall prevail. .

[0008] Furthermore, the construction of the feedback dataset in step S4, used for targeted optimization of the AI-generated model, specifically involves: The final content confirmed as accurate and of high quality, the manually revised version of the content, the corresponding quality score vector, and the source evidence chain are collectively constructed into a feedback dataset. Using the feedback dataset, the AI-generated model is fine-tuned with a reinforcement learning strategy to optimize the model's expected return on the multi-dimensional evaluation metrics. Based on the distribution changes of the overall confidence level $C$ of the content generated by the optimized model during the review process, the threshold Td1 in step S3 is dynamically adjusted.

[0009] Furthermore, in step S1, when performing personalized constraint preprocessing, the personalized compliance evaluation dimension is correspondingly added to the multi-dimensional automatic evaluation in step S2: Calculate the matching score Sp between the generated content and the specific object's dedicated digital resource library in terms of theme, entity, and style; Sp is incorporated as a new dimension into the quality score vector S and participates in the calculation of the comprehensive confidence level C in step S3.

[0010] Compared with existing technologies, this invention has the following advantages: This AI-generated content human-machine collaborative review and optimization method for the cultural heritage field evaluates content from multiple dimensions, including historical accuracy, cultural context conformity, and logical coherence. It combines semantic comparison with authoritative historical document databases, analysis of cultural context embedding models trained on cultural heritage corpora, and verification of content self-consistency using pre-trained language models. Simultaneously, it supports reviewers triggering multi-hop association retrieval and evidence chain presentation based on authoritative cultural heritage knowledge graphs, ensuring that the generated content is authentic, accurate, appropriately expressed, and logically consistent. By normalizing the quality score vector and calculating the comprehensive confidence level using a dynamic weight vector and the Sigmoid function, task matching and allocation are performed based on this confidence level and the reviewer's dynamic professional competence profile. This is further enhanced by adjusting a dynamic threshold Td1 based on the average confidence level of the task queue, the reviewer's workload, and historical pass rates. This system achieves precise adaptation of human and machine review tasks, significantly improving review efficiency and avoiding waste of human resources. For customized content needs of specific targets, a dedicated digital resource library and style guide can be used as pre-constraints input into the AI ​​generation model. At the same time, a personalized compliance score (Sp) is added to the multi-dimensional evaluation and integrated into the quality score vector to participate in the confidence calculation, fully meeting the content adaptation requirements in customized scenarios. By integrating the final review conclusion, manually corrected content, quality score vector, and traceability evidence chain to construct a feedback dataset, a reinforcement learning strategy is used to fine-tune the AI ​​generation model. The review threshold is dynamically adjusted according to the changes in the comprehensive confidence distribution of the content generated by the optimized model, continuously improving the quality of AI-generated content and the review pass rate. This provides efficient, reliable, and sustainably optimized technical support for content production in the cultural heritage field, making the system more worthy of widespread use. Attached Figure Description

[0011] Figure 1 This is the overall flowchart of the present invention. Detailed Implementation

[0012] The embodiments of the present invention are described in detail below. These embodiments are implemented based on the technical solution of the present invention, and provide detailed implementation methods and specific operation processes. However, the scope of protection of the present invention is not limited to the following embodiments.

[0013] like Figure 1 As shown, this embodiment provides a technical solution: a method for human-machine collaborative review and optimization of AI-generated content in the cultural heritage field, including the following steps: S1. Content Acquisition and Personalization Constraint Pre-processing: Acquire relevant cultural heritage content generated by the AI ​​model; if the content is customized for a specific object, input the object's exclusive digital resource library and style guide as constraints into the AI ​​generation model; S2. Multi-dimensional automatic evaluation: The multi-dimensional content evaluation model is invoked to automatically analyze the content and output a quality score vector containing scores of multiple evaluation dimensions. S3. Dynamic allocation and collaborative review of human and machine tasks: The comprehensive confidence level of the content is calculated based on the quality scoring vector, and low-confidence content is dynamically assigned to the matched human reviewers based on this confidence level and the professional profile of the reviewers; the reviewers can trigger source verification based on authoritative cultural heritage knowledge graph during the review process. S4. Decision Integration and Model Iteration Optimization: The final review conclusion is formed by integrating the results of manual review and automatic evaluation. Based on this conclusion and the review process data, a feedback dataset is constructed to optimize the AI-generated model in a targeted manner, so as to realize the closed-loop feedback of the review results to the generated model.

[0014] Furthermore, the evaluation dimensions of the "multi-dimensional content evaluation model" in step S2 include at least: Historical accuracy dimension: The generated content is semantically compared with authoritative historical document databases to calculate the consistency score Sh of key fact descriptions; Cultural context conformity dimension: Using a cultural context embedding model trained with cultural heritage corpus, we analyze the appropriateness of the generated content in a specific historical context and generate a context deviation score Sc. Logical coherence dimension: Analyzes the self-consistency of events and causal relationships within the content through a pre-trained language model, and outputs a logical coherence score Sl; The quality score vector is ; By clearly defining core evaluation dimensions such as historical accuracy, cultural context conformity, and logical coherence, and combining professional models with authoritative data verification, we can achieve comprehensive and accurate quality screening of AI-generated cultural heritage content, avoiding the one-sidedness of single-dimensional evaluation. At the same time, we output a structured quality scoring vector, providing a clear and quantifiable basis for subsequent comprehensive confidence calculation and task allocation.

[0015] Suppose the AI-generated content is: "Tang Dynasty tri-color pottery mainly uses yellow, green, and white colors, and was prevalent from the early to the high Tang Dynasty. Because it was often used as burial goods in aristocratic tombs, its shapes mostly imitated objects and animals in real life. After the craftsmanship matured, it became an important representative of Tang Dynasty handicrafts." Historical accuracy dimension (Sh): The content is semantically compared with authoritative literature databases such as "History of Chinese Ceramics". If the key facts (prevailing period, use) are consistent, Sh=0.95. Cultural context fit dimension (Sc): Through model analysis trained on cultural relics corpus, expressions such as "burial goods in noble tombs" and "representative of Tang Dynasty handicrafts" fit the social and cultural background of the Tang Dynasty, and Sc=0.92. Logical coherence dimension (Sl): The pre-trained language model verifies the self-consistency of the causal relationship "mature technology → rich styling → becoming an important representative", and obtains Sl=0.9; The final quality score vector is .

[0016] The overall confidence level of the content calculated in step S3 is specifically achieved in the following way: The scores of each dimension in the quality score vector S are normalized to obtain a normalized vector. ; Dynamic weight vectors are assigned to each evaluation dimension based on the type of cultural heritage content. ; The calculation process for the overall confidence level is expressed as follows: ; in, is the Sigmoid activation function, and b is the bias term used to map the result to the [0,1] interval; By normalizing the multi-dimensional scores to eliminate differences in dimensional scales, dynamically assigning weights based on the type of cultural heritage content, and then mapping the results to the [0,1] interval using the Sigmoid function, the overall confidence level is quantified and standardized. This approach takes into account the differences in importance of different dimensions and can intuitively reflect the overall quality and credibility of AI-generated content, providing a precise and unified decision-making basis for the dynamic allocation of subsequent human-machine tasks.

[0017] For example, in the case of AI-generated content of Tang Dynasty tri-colored pottery, the known quality score vector is S=[S_h,S_c,S_l]=[0.95,0.92,0.9]: Normalization: Since the scores of each dimension are already in the range of [0,1], after normalization, Snorm = [0.95, 0.92, 0.9]; Dynamic weight allocation: If the content is related to cultural heritage and popular science (historical accuracy has the highest priority), assign a weight vector. ; Set the bias term $b=0.1$, and substitute it into the overall confidence formula: ; Calculation process: First calculate the inner product , then calculate The final overall confidence level is C≈0.737.

[0018] The dynamic allocation based on confidence level and the professional profile of the reviewer in step S3 specifically includes: S3.1 Sort all the content to be reviewed according to the comprehensive confidence level C; S3.2. Maintain a dynamic professional ability profile for each human reviewer based on historical review data; S3.3. For content items with C lower than the dynamic threshold Td1, calculate the matching degree between its quality score vector S and the professional ability profiles of each idle reviewer, and assign the task to the reviewer with the highest matching degree; S3.4. The dynamic threshold Td1 is periodically adjusted according to the average confidence of the current task queue, the review manpower load, and the historical review passing rate; By sorting the content to be reviewed according to the comprehensive confidence, and maintaining the dynamic professional profiles of reviewers, the accurate matching of low-confidence content and professional reviewers is achieved, ensuring the review quality; at the same time, the dynamic threshold Td1 is periodically adjusted according to the task queue status, manpower load, and historical data, flexibly adapting to changes in the review scenario, avoiding manpower idleness or overload, and maximizing the overall efficiency and resource utilization of human-machine collaborative review.

[0019] For example, for the AI-generated content of Tang Dynasty tri-colored glazed pottery (comprehensive confidence C≈0.737), assuming there are 3 idle reviewers: Confidence ranking and threshold setting: The average confidence of the current task queue is 0.75, the review manpower load is moderate, and the historical review passing rate is 80%. The dynamic threshold Td1 = 0.8. Since C = 0.737 < Td1, this content needs to be manually reviewed; Reviewer professional profile and matching degree calculation: Reviewer A (majoring in ceramic cultural relics, the historical review accuracy rate of ceramic content is 95%, matching degree 0.92), Reviewer B (majoring in bronze wares, matching degree 0.65), Reviewer C (majoring in calligraphy and painting, matching degree 0.58). After calculating the matching degree between the quality score vector S = [0.95, 0.92, 0.9] and the profiles of the three, the matching degree of A is the highest; Task assignment and dynamic threshold adjustment: Assign the tri-colored glazed pottery-related content to Reviewer A; if the average confidence of the task queue drops to 0.68 and the manpower load increases later, Td1 is adjusted to 0.75. At this time, C = 0.737 is still lower than Td1, and it continues to be assigned to the corresponding professional reviewer according to the matching degree.

[0020] The triggering of the traceability verification based on the authoritative cultural relics knowledge graph in the S3 step is specifically as follows: In response to the reviewer's selection operation on a specific entity or statement in the generated content, the system uses this fragment as the query input; Perform multi-hop association retrieval in the knowledge graph to find the attributes, historical events, literature records, and academic views related to the query entity; Present the retrieval results in the form of an evidence chain graph and a text summary, and automatically calculate the semantic similarity between the generated content fragment and the core evidence as a reference; It provides authoritative and traceable verification support for manual review. Through multi-hop association retrieval of authoritative cultural heritage knowledge graph, it comprehensively covers the relevant attributes, historical events, document records and academic viewpoints of specific entities / statements, avoiding reliance on subjective experience in review. It presents the source tracing results intuitively with evidence chain graph and text summary, and with the help of automatically calculated semantic similarity reference, it helps reviewers quickly verify the authenticity of content. This not only improves the accuracy of review, but also saves time in manually consulting authoritative materials, and enhances the professionalism and efficiency of human-machine collaborative review.

[0021] For example, AI-generated content of Tang Dynasty tri-colored pottery (including the statement "mainly popular from the early Tang to the high Tang period"): The reviewer had doubts about the statement and selected that excerpt as the query input. The system performs multi-hop association search in the authoritative cultural heritage knowledge graph: the first hop locates the attribute "Tang Sancai - Flourishing Period", the second hop associates "Early Tang to High Tang - Characteristics of Ceramic Development", and the third hop searches for relevant records in "History of Chinese Ceramics", archaeological reports of Tang Dynasty tombs in Xi'an, and academic papers in the field of cultural heritage. Search results present: a chain of evidence (Tang Sancai → period of popularity → early Tang to high Tang → corroborated by the literature "History of Chinese Ceramics" → supported by archaeological discoveries in Xi'an tombs) + text summary ("According to the "History of Chinese Ceramics", Tang Sancai originated in the early Tang Dynasty, flourished in the high Tang Dynasty, and gradually declined after the mid-Tang Dynasty. It was a common funerary object in the tombs of Tang Dynasty nobles"). Semantic similarity calculation: The semantic similarity between the generated content fragment and the core evidence (recorded in "A History of Chinese Ceramics") was automatically compared, and Sim=0.94 was obtained, which provides a direct reference for the reviewer to judge the accuracy of the statement.

[0022] The S4 step, which integrates the results of manual review and automatic evaluation, employs a weighted decision-making fusion strategy. Let the conclusion of the manual review be Vh, and the overall confidence level automatically assessed by the system be C; Set a high confidence threshold Thigh and a low confidence threshold Tlow; The rules for generating the final audit conclusion Vf$ are as follows: when In this case, the main approach is to rely on systematic evaluation and inference, and human conclusions must be accompanied by strong evidence to be overturned; when hour, , where α is the human weighting coefficient that increases with the auditor’s historical accuracy, and sign is the sign function; when In such cases, the conclusion of manual review shall prevail. ; Adopt a weighted decision fusion strategy for different intervals, which not only respects the reliability of high-confidence AI evaluations to improve the review efficiency, but also highlights the authority of manual review in low-confidence scenarios to ensure quality. In the middle interval, the manual weight coefficient α that dynamically adjusts according to the historical accuracy rate of reviewers is used to balance the human-machine decision weights, achieving the precision, flexibility, and maximization of the credibility of the review conclusion, avoiding the one-sidedness of a single decision-making entity. At the same time, clarify the decision rules for different confidence intervals to make the review process more standardized and operable.

[0023] For example, for the AI-generated content of Tang Dynasty tri-colored glazed pottery (the comprehensive confidence level C≈0.737), set the high-confidence threshold Thigh = 0.85, the low-confidence threshold Tlow = 0.7, and the manual weight coefficient α = 0.6 for reviewer A (historical accuracy rate 95%). The manual review conclusion Vh = 1 (indicating that the content is qualified): Threshold judgment: Since (0.7 ≤ 0.737 < 0.85), the weighted fusion formula ; Sign function calculation: ; Final review conclusion: , that is, it is determined that the AI-generated content is qualified; Supplementary scenario: If C = 0.88 ≥ Thigh = 0.85, then the system evaluation shall prevail. If the manual review wants to overturn it, strong evidence such as "the wrong period of the prevalence of tri-colored glazed pottery" needs to be provided; if C = 0.65 < Tlow = 0.7, then directly take the manual review conclusion Vh as the final conclusion.

[0024] In the S4 step, constructing a feedback data set for targeted optimization of the AI generation model is specifically as follows: Construct the content that is finally reviewed as accurate and high-quality, the content version after manual correction, the corresponding quality scoring vector, and the traceability evidence chain together into a feedback data set; Use the feedback data set to fine-tune the AI generation model with a reinforcement learning strategy to optimize the expected return of the model on the multi-dimensional evaluation metrics; According to the distribution change of the comprehensive confidence level $C$ of the content generated by the optimized model in the review link, dynamically adjust the threshold Td1 in the S3 step; By integrating high-quality content, manually revised versions, quality scoring vectors, and traceability evidence chains to construct a complete feedback dataset, a reinforcement learning strategy is used to fine-tune the AI ​​generation model, achieving **two-way closed-loop optimization of the model and the review system**. This allows the model to specifically improve the performance of multi-dimensional evaluation indicators in the cultural heritage field, continuously reducing content error rates and improving generation quality. At the same time, the review threshold can be dynamically adjusted based on the comprehensive confidence distribution after model optimization, ensuring that human-machine collaborative review always adapts to the model's evolution rhythm, avoiding ineffective manual review or quality omissions, and improving the efficiency and reliability of AI content production in the cultural heritage field in the long term.

[0025] For example, the AI-generated content of Tang Dynasty tri-colored pottery (original quality score vector S=[0.95,0.92,0.9], original comprehensive confidence level C≈0.737, review conclusion: qualified): Feedback dataset construction: The content confirmed to be accurate by the review, the version without manual correction (content is qualified and does not need to be modified), the quality score vector S=[0.95,0.92,0.9], and the source evidence chain (recorded in "History of Chinese Ceramics" + archaeological report of Tang Dynasty tombs in Xi'an) are all included in the feedback dataset; Model-oriented optimization: With the goal of maximizing the expected returns of historical accuracy (Sh), cultural context conformity (Sc), and logical coherence (Sl), a reinforcement learning strategy is used to fine-tune the AI-generated model; The optimized model generates new content: the multi-dimensional scores of the new content are improved to [percentage missing]. Since the scores of each dimension are already in the [0,1] interval, after normalization... Using the weight vector W=[0.5,0.3,0.2] and bias term b=0.1, we substitute them into the comprehensive confidence formula. ; Step-by-step calculation of the new overall confidence level: Step 1: Calculate the inner product: ; The second step is to add the bias term: 0.968 + 0.1 = 1.068; The third step is to calculate the Sigmoid function value: The final new overall confidence level ; Dynamic adjustment of review threshold: The average confidence level of all content in the task queue after optimization has increased from 0.75 to 0.78, while maintaining a moderate workload for review personnel. The original dynamic threshold Td1=0.8 has been adjusted to Td1=0.82, focusing on manual review of higher-risk content and adapting to the quality distribution after model optimization.

[0026] Furthermore, in step S1, when performing personalized constraint preprocessing, the personalized compliance evaluation dimension is correspondingly added to the multi-dimensional automatic evaluation in step S2: Calculate the matching score Sp between the generated content and the specific object's dedicated digital resource library in terms of theme, entity, and style; Sp is incorporated as a new dimension into the quality score vector S and used in the calculation of the comprehensive confidence C in step S3. To address the customized content needs in the cultural heritage sector, a new personalized conformity assessment dimension, $S_p$, has been added. This dimension incorporates the matching degree between the generated content and the specific resources and style constraints of the target audience into the quality assessment system, making the quality scoring vector more aligned with the customized scenario and improving the accuracy and adaptability of customized content assessment. Simultaneously, $S_p$ participates in the comprehensive confidence calculation, ensuring that the confidence score not only reflects the core quality of the content but also the degree of personalized fit. This expands the applicability of the method in specific cultural heritage content generation scenarios, guaranteeing that customized content is both professional and compliant, and meets specific needs.

[0027] For example, AI-generated content of Tang Dynasty tri-colored pottery (new scenario: customizing popular science texts for a museum, which must conform to its collection resources and style specifications): Personalized constraints are implemented upfront: the museum's exclusive digital resource database (information on its collection of Tang tri-colored pottery artifacts and exhibition themes) and style guidelines (popular language and highlighting the characteristics of its collection) are used as input constraints to the AI ​​model; Added personalized compatibility assessment: Calculate the matching degree between the generated content and the exclusive resource library in terms of theme (Tang Sancai popular science), entity (description of museum collection cultural relics), and style (popularization), and get Sp=0.94; Update the quality score vector: (Scores for each dimension are all in the range [0,1], after normalization) ); Dynamic weight allocation: Customized content needs to be personalized, and weight vectors should be allocated accordingly. (The total weight is 1); Calculate the overall confidence level: Set the bias term b = 0.1, and substitute it into the formula. : Inner product calculation: ; Sigmoid function calculation: The final overall confidence level is C≈0.738; With the addition of the Sp dimension, the confidence level reflects not only the historical accuracy and logical coherence of the content, but also its fit with the museum's specific needs, ensuring that customized science popularization copy is both professional and tailored to the characteristics of the collection.

[0028] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0029] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples, without contradiction. Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of the present invention.

Claims

1. A method for human-machine collaborative review and optimization of AI-generated content in the cultural heritage field, characterized in that, Includes the following steps: S1. Content Acquisition and Personalization Constraint Pre-processing: Acquire relevant cultural heritage content generated by the AI ​​model; if the content is customized for a specific object, input the object's exclusive digital resource library and style guide as constraints into the AI ​​generation model. S2. Multi-dimensional automatic evaluation: Calls a multi-dimensional content evaluation model to automatically analyze the content and outputs a quality score vector containing scores of multiple evaluation dimensions; S3. Dynamic allocation and collaborative review of human and machine tasks: The comprehensive confidence level of the content is calculated based on the quality score vector, and the low confidence level content is dynamically assigned to the matching human reviewers based on this confidence level and the professional profile of the reviewers; the reviewers can trigger source verification based on authoritative cultural heritage knowledge graph during the review process. S4. Decision Integration and Model Iteration Optimization: The final review conclusion is formed by integrating the results of manual review and automatic evaluation. Based on this conclusion and the review process data, a feedback dataset is constructed to optimize the AI-generated model in a targeted manner, so as to realize the closed-loop feedback of the review results to the generated model.

2. A method for human-machine collaborative review and optimization of AI-generated content in the cultural heritage field according to claim 1, characterized in that: The evaluation dimensions of the "multi-dimensional content evaluation model" in step S2 should include at least: Historical accuracy dimension: The generated content is semantically compared with authoritative historical document databases to calculate the consistency score Sh of key fact descriptions; Cultural context conformity dimension: Using a cultural context embedding model trained with cultural heritage corpus, we analyze the appropriateness of the generated content in a specific historical context and generate a context deviation score Sc. Logical coherence dimension: Analyzes the self-consistency of events and causal relationships within the content through a pre-trained language model, and outputs a logical coherence score Sl; The quality score vector is .

3. A method for human-machine collaborative review and optimization of AI-generated content in the cultural heritage field according to claim 2, characterized in that: The overall confidence level calculated in step S3 is achieved in the following way: The scores of each dimension in the quality rating vector S are normalized to obtain the normalized vector. ; Dynamic weight vectors are assigned to each evaluation dimension based on the type of cultural heritage content. ; The calculation process for the overall confidence level is expressed as follows: ； in, is the Sigmoid activation function, and b is the bias term used to map the result to the [0,1] interval.

4. A method for human-machine collaborative review and optimization of AI-generated content in the cultural heritage field according to claim 3, characterized in that: The S3 step, which dynamically allocates resources based on confidence level and the professional profile of the reviewers, specifically includes: S3.1 Sort all items to be reviewed according to the overall confidence level C; S3.2 Maintain a dynamic professional competence profile for each human auditor based on historical audit data; S3.3 For content items where C is lower than the dynamic threshold Td1, calculate the matching degree between its quality score vector S and the professional competence profile of each available reviewer, and assign the task to the reviewer with the highest matching degree. S3.4 The dynamic threshold Td1 is periodically adjusted based on the average confidence level of the current task queue, the workload of the review manpower, and the historical review pass rate.

5. A method for human-machine collaborative review and optimization of AI-generated content in the cultural heritage field according to claim 4, characterized in that: The specific steps in step S3 that trigger the source verification based on the authoritative cultural heritage knowledge graph are as follows: In response to the reviewer's selection of a specific entity or statement in the generated content, the system uses that fragment as query input; Perform multi-hop association retrieval in the knowledge graph to find attributes, historical events, literature records and academic viewpoints related to the query entity; The search results are presented in the form of evidence chain graphs and text summaries, and the semantic similarity between the generated content fragments and the core evidence is automatically calculated as a reference.

6. A method for human-machine collaborative review and optimization of AI-generated content in the cultural heritage field according to claim 1, characterized in that: The "integration of manual review results and automatic evaluation results" step S4 adopts a weighted decision integration strategy: Let the conclusion of the manual review be Vh, and the overall confidence level automatically assessed by the system be C; Set a high confidence threshold Thigh and a low confidence threshold Tlow; The rules for generating the final audit conclusion Vf$ are as follows: when In this case, the main approach is to rely on systematic evaluation and inference, and human conclusions must be accompanied by strong evidence to be overturned; when hour, , where α is the human weighting coefficient that increases with the auditor’s historical accuracy, and sign is the sign function; when In such cases, the conclusion of manual review shall prevail. .

7. A method for human-machine collaborative review and optimization of AI-generated content in the cultural heritage field according to claim 6, characterized in that: Step S4 involves constructing a feedback dataset for targeted optimization of the AI-generated model. The final content confirmed as accurate and of high quality, the manually revised version of the content, the corresponding quality score vector, and the source evidence chain are collectively constructed into a feedback dataset. Using the feedback dataset, the AI-generated model is fine-tuned with a reinforcement learning strategy to optimize the model’s expected returns on multi-dimensional evaluation metrics. Based on the distribution changes of the overall confidence level $C$ of the content generated by the optimized model during the review process, the threshold Td1 in step S3 is dynamically adjusted.

8. A method for human-machine collaborative review and optimization of AI-generated content in the cultural heritage field according to claim 7, characterized in that: In step S1, when performing personalized constraint preprocessing, the personalized compliance assessment dimension is added to the multi-dimensional automatic assessment in step S2 accordingly: Calculate the matching score Sp between the generated content and the specific object's dedicated digital resource library in terms of theme, entity, and style; Sp is incorporated as a new dimension into the quality score vector S and participates in the calculation of the comprehensive confidence level C in step S3.

Images