AI-based analysis report quality dynamic optimization method

Abstract

The invention discloses an analysis report quality dynamic optimization method based on AI, and the method comprises the steps: building a user preference vector through balance weight based on a user historical report scoring behavior and user type description; generating a candidate report set in parallel by utilizing a plurality of AI models; an objective score based on multiple quality dimensions is calculated for each candidate report in the candidate report set, candidate report abstracts or key features are displayed to the user through an interactive interface, and subjective original scores of the user for the abstracts are collected; for each candidate report, calculating a similarity value between the user preference vector and the report feature; obtaining an adjustment subjective score in combination with the similarity value and the subjective score of the user; and fusing the objective score and adjusting the subjective score by using a dynamic weight to obtain a fused score. According to the method, triple deployment values can be realized, namely, the manual review cost is greatly reduced, the user satisfaction index is systematically increased, and the weight rule adaptive iteration period is shortened to be within the business demand change rate.

Description

Technical Field

[0001] The present invention relates to the field of AI, and in particular to an AI-based dynamic optimization method for analysis report quality. Background Art

[0002] Existing AI report generation systems rely on uniform standards to assess report quality. This model suffers from the following flaws: First, it ignores individual needs. Financial experts demand rigorous data, while marketers prefer creative analysis. However, existing technology cannot distinguish these differences, resulting in doctors receiving medical reports laden with marketing jargon and pilots seeing complex financial charts. Second, the evaluation mechanism is rigid, making the system unable to adapt based on user feedback. For example, when investors begin to focus on emerging markets, the old system still generates traditional industry analysis based on historical standards. Third, it loses diversity, and over time, the system tends to fall into a fixed routine. Civil aviation marketing reports have gradually become cookie-cutter promotional templates, losing their ability to discern market dynamics and urgently need improvement. Summary of the Invention

[0003] In order to solve the technical problems existing in the background technology, the present invention proposes an AI-based dynamic optimization method for analysis report quality, including: S1. Based on the user's historical report scoring behavior and user type description, a user preference vector is constructed by balancing weights; S2. A candidate report set is generated in parallel using multiple AI models; S3. An objective score based on multiple quality dimensions is calculated for each candidate report in the candidate report set, and a summary or key features of the candidate report is displayed to the user through an interactive interface, and the user's subjective original score for the summary is collected; S4. For each candidate report, the similarity value between the user preference vector and the report feature is calculated; the adjusted subjective score is obtained by combining the similarity value with the user's subjective score; the objective score and the adjusted subjective score are fused using dynamic weights to obtain a fused score; S5. The candidate report with the highest fused score is selected as the final report output.

[0004] Furthermore, it also includes: S6, calculating the arithmetic mean of the objective scores of all candidate reports at the current moment as the average objective score; collecting the user's subjective score for the final report; updating the dynamic weight according to the difference between the subjective score and the average objective score, and applying the updated dynamic weight to the subsequent fusion score calculation.

[0005] Furthermore, S2 specifically includes: S21, configuring difference parameters of multiple AI models, the difference parameters include differences in training data distribution and algorithm architecture differences; S22, responding to report generation instructions, sending parallel trigger signals to all AI models; S23, each AI model independently generates a report entity based on input data, and the report entity includes text content and feature vectors; S24, aggregating all report entities to form a candidate report set.

[0006] Furthermore, S21 specifically includes: S211, obtaining the algorithm architecture descriptor and training data distribution descriptor of each AI model; S212, calculating the architecture difference value based on the algorithm architecture descriptor; S213, calculating the data distribution difference value based on the training data distribution descriptor; S214, integrating the architecture difference value and the data distribution difference value to generate difference parameters.

[0007] Furthermore, S1 includes: S11, extracting the user's historical report scoring behavior characteristics; S12, extracting the user's occupation type related description; S13, normalizing the user's historical report scoring behavior characteristics; S14, vectorizing the user's occupation type related description; S15, weighting and fusing the normalized user's historical report scoring behavior characteristics and the vectorized user's occupation type related description through a balance coefficient; S16. Generate a user preference vector. The formula for constructing the user preference vector is: ; Represents the normalized user historical report rating behavior characteristics; Represents the user's occupation type description after vectorized encoding; is the balance coefficient.

[0008] Furthermore, S3 includes: S31, extracting information coverage feature vector, coherence feature vector and data severity feature vector for each candidate report; S32, processing information coverage feature vector to generate information coverage score; S33, processing coherence feature vector to generate coherence score; S34, processing data severity feature vector to generate data severity score; S35, weighted fusion of information coverage score, coherence score and data severity score based on preset weight coefficient to generate objective score; S36, collecting users' subjective original scores of candidate reports through an interactive interface; the interactive interface only displays the core feature vector or text summary of the candidate report, and the duration is controlled within the range that users can quickly evaluate, so as to efficiently collect subjective feedback.

[0009] S37. Standardize the subjective original scores to generate user subjective scores.

[0010] Furthermore, S4 includes: S41, calculating the cosine similarity value between the user preference vector and the feature vector of each candidate report; S42, multiplying the cosine similarity value by the user's subjective score to generate an adjusted subjective score; S43, calculating a fusion score using a dynamic weight fusion formula, wherein the fusion score formula is: ; represents an objective rating; indicates adjustment of subjective ratings; Indicates dynamic weight.

[0011] Furthermore, S5 includes: S51, sorting the fusion scores of all reports in the candidate report set in descending order; S52, selecting the candidate report ranked first in the fusion score as the final report; S53, sending the final report to the user terminal through the report output interface.

[0012] Furthermore, S6 includes: S61, calculating the arithmetic mean of the objective scores of all reports in the candidate report set at the current moment to generate an average objective score; S62, collecting the user's subjective original score of the final report; S63, normalizing the subjective original score to generate the user's subjective score; S64, calculating the updated dynamic weight through the weight update formula, where the update formula is: ; Indicates the user's subjective score; represents the average objective rating; Represents the dynamic weight before update; represents the weight adjustment rate, wherein the weight adjustment rate It is a preset constant parameter with a value range set between 0 and 1, and is determined through experimental calibration or system administrator configuration; S65, storing the updated dynamic weight to the weight configuration library; S66, calling the dynamic weight in the weight configuration library in subsequent fusion score calculations.

[0013] The present invention also proposes a non-transitory computer-readable storage medium storing a computer program, which, when executed by a processor, can implement an AI-based dynamic optimization method for analysis report quality proposed by the present invention.

[0014] The present invention proposes an AI-based dynamic optimization method for analysis report quality. By integrating user historical rating data with professional tags to construct a personalized preference vector, the matching accuracy between reports and user needs is significantly enhanced (for example, in aviation marketing scenarios, the adoption rate of route revenue analysis reports has achieved a breakthrough improvement), thereby establishing a precise demand capture mechanism; based on real-time feedback, the evaluation weight is dynamically adjusted. When users continuously give high scores to the rigor of report data, the system automatically increases the weight coefficient of the data rigor, giving the system the ability to continuously evolve; by controlling multi-model difference parameters, the innovation index of a single scenario report is improved while maintaining the accuracy of core data, forming a diversity guarantee mechanism; and ultimately achieving triple deployment value: the cost of manual review is greatly reduced, the user satisfaction index is systematically increased, and the adaptive iteration cycle of weight rules is shortened to within the rate of change of business needs. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 This is the main flow chart of the AI-based dynamic optimization method for analysis report quality proposed by the present invention; Figure 2This is a partial flow chart of an AI-based dynamic optimization method for analysis report quality proposed by the present invention; Figure 3 This is a partial flow chart of an AI-based dynamic optimization method for analysis report quality proposed by the present invention; Figure 4 This is a partial flow chart of an AI-based dynamic optimization method for analysis report quality proposed by the present invention; Figure 5 This is a partial flow chart of an AI-based dynamic optimization method for analysis report quality proposed by the present invention; Figure 6 This is a partial flow chart of an AI-based dynamic optimization method for analysis report quality proposed by the present invention; Figure 7 This is a partial flow chart of an AI-based dynamic optimization method for analysis report quality proposed by the present invention; Figure 8 This is a partial flow chart of the AI-based dynamic optimization method for analysis report quality proposed by the present invention. DETAILED DESCRIPTION

[0016] refer to Figure 1-8 The present invention proposes an AI-based dynamic optimization method for analysis report quality, comprising: S1. Based on the user's historical reporting and scoring behavior and user type description, the user preference vector is constructed by balancing weights. Specifically, it includes: S11. Extract historical user report rating behavior features. This step extracts key behavioral patterns from the user's past rating records, such as frequently viewed report types or rating preferences, to form a structured feature set. This improves the accuracy of subsequent user preference modeling, avoids report recommendation bias caused by ignoring historical behavior, and makes the system more tailored to users' long-term needs.

[0017] S12. Extract descriptions of the user's occupational type. This step collects the user's occupational background information, such as financial analyst or medical practitioner labels, and converts it into quantifiable descriptive text. This enhances the comprehensiveness of the user profile, reduces business scenario adaptation errors, and better integrates occupational characteristics into preference analysis.

[0018] S13. Normalize the user's historical reporting and scoring behavior characteristics. This step processes the user's historical scoring characteristics through a standardization algorithm, such as using min-max scaling to eliminate dimensional differences and make the data distribution more even, thereby improving the comparability between different users and reducing the risk of weight imbalance caused by excessively large or small feature values.

[0019] S14. Vectorize and encode the user's occupational type description. This step converts the occupational description into a numerical vector using embedded coding techniques, such as the Word2Vec model, to process text descriptions. This converts abstract attributes into a computable form, thereby increasing the feasibility of feature fusion and preventing text ambiguity from interfering with preference modeling.

[0020] S15. Weighted fusion of the normalized user historical report scoring behavior characteristics and the vectorized description of the user's occupational type using a balance coefficient. This step dynamically adjusts the weights of the normalized characteristics and the vectorized description using the balance coefficient, for example, adjusting the contribution of each to ensure a balanced integration of user behavior and occupational background, thereby improving the representativeness of the preference vector and reducing model bias caused by a single factor dominating.

[0021] S16. Generate a user preference vector. The formula for constructing the user preference vector is: . Represents the normalized user historical report rating behavior characteristics. Represents the user's occupation type description after vectorized encoding. This step generates a final user preference vector by weighting the results. For example, by integrating historical ratings and occupational coding data to form a unique identifier, the vector can be directly used for similarity matching, thereby improving the personalization of subsequent report recommendations and preventing general models from ignoring user uniqueness.

[0022] S2. Generate candidate report sets using multiple AI models in parallel. Specifically: S21. Configure the difference parameters of multiple AI models. The difference parameters include differences in training data distribution and algorithm architecture. Specifically, they include: S211. Obtain the algorithm architecture descriptor and training data distribution descriptor for each AI model. This step obtains the AI ​​model's architecture descriptor (such as Transformer or RNN type) and data distribution descriptor (such as the source distribution of the training set) to quantify and store model characteristics, thereby increasing the basis for differential parameter calculation and reducing the uniformity of reports caused by model duplication.

[0023] S212. Calculate architecture difference values ​​based on the algorithm architecture descriptor. This step calculates difference values ​​based on the algorithm architecture description, such as comparing the number of layers or parameter scale differences, converting architectural characteristics into numerical indicators. This improves the accuracy of model diversity control and reduces the risk of similar reports generated by homogenized models.

[0024] S213. Calculate data distribution difference values ​​based on the training data distribution descriptor. This step calculates difference values ​​based on the training data distribution descriptor. For example, by analyzing the breadth of data domain coverage or sample balance, the impact of data characteristics can be measured, thereby increasing the innovation of the candidate report and avoiding report content limitations caused by data bias.

[0025] S214: Integrate the architecture difference value and the data distribution difference value to generate difference parameters. This step generates comprehensive parameters by integrating the architecture and data difference values. For example, a weighted average method is used to output a single indicator to guide the model selection strategy, thereby improving parallel generation efficiency and reducing resource waste caused by ineffective model calls.

[0026] S22. Respond to the report generation instruction and send a parallel trigger signal to all AI models. This step triggers all AI models simultaneously after receiving the generation instruction, such as waking up a standby model with a broadcast signal. This allows computing tasks to start in parallel, thereby shortening the report generation delay and avoiding response lag caused by sequential execution.

[0027] S23. Each AI model independently generates a report entity based on the input data. The report entity includes text content and feature vectors. This step allows each model to independently process the input data and output a complete report entity, including text and feature vectors. For example, LLM generates financial analysis content with accompanying vector fields. This makes the candidate set diverse and structured, thereby expanding the range of report options and avoiding limiting user choices with a single output.

[0028] S24. Aggregate all report entities to form a candidate report set. This step aggregates the report entities generated by the model into a unified set, such as a list storing all report text and features, so that subsequent evaluation can be processed in batches, thereby improving system throughput and reducing processing delays caused by data dispersion.

[0029] S3. Calculate an objective score based on multiple quality dimensions for each candidate report in the candidate report set, present a summary or key features of the candidate report to the user through an interactive interface, and collect the user's original subjective score for the summary. Specifically, this includes: S31. Extract feature vectors for information coverage, coherence, and data rigor for each candidate report. This step extracts feature vectors covering information completeness, logical coherence, and data rigor. For example, NLP tools can analyze keyword coverage to quantify multi-dimensional quality, thereby increasing objectivity and preventing subjective judgments from overlooking critical flaws.

[0030] S32. Process the information coverage feature vector to generate an information coverage score. This step generates a score by processing coverage features, such as calculating the proportion of missing key information and converting it into a 0-1 score. This makes the information completeness quantitatively comparable, thereby improving the practicality of the report and reducing decision-making errors caused by missing data.

[0031] S33. Process the coherence feature vector to generate a coherence score. This step generates a score based on coherence features, such as analyzing paragraph cohesion and fluency, and assigning a score based on logical chain integrity. This makes the report readability measurable, thereby increasing user comprehension efficiency and avoiding cognitive burden caused by structural confusion.

[0032] S34. Process the data severity feature vector to generate a data severity score. This step generates a score based on the data severity features, such as verifying the reliability of data citations and scoring based on the strength of evidence support. This makes the report's rigor more concrete, thereby reducing the probability of erroneous inferences and preventing unfounded statements from affecting user trust.

[0033] S35. The information coverage score, the consistency score, and the data severity score are weighted and integrated based on preset weight coefficients to generate an objective score. This step weights the three-dimensional scores using preset weights, such as assigning a higher weight to coverage, to output a comprehensive objective score. This allows the quality assessment to take multiple factors into account, thereby improving the fairness of the judgment and preventing a single dimension from dominating and misjudging high-quality reports.

[0034] S36. Collect users' raw subjective ratings of candidate reports through an interactive interface. The interactive interface displays only the core feature vectors or text summaries of the candidate reports, keeping the duration of the display within a user-friendly range to efficiently collect subjective feedback. This step collects raw user ratings through the interactive interface, such as using a five-level scale to collect accuracy feedback. This allows subjective perceptions to be directly input into the system, thereby increasing user engagement and reducing optimization lag caused by feedback delays.

[0035] S37: Standardize the original subjective scores to generate user subjective scores. This step standardizes the original scores, for example, normalizing them to a uniform dimension, eliminating individual scoring biases and making the subjective data comparable and reliable, thereby improving the accuracy of the fused score and preventing outliers from distorting the weight adjustment.

[0036] S4. For each candidate report, calculate the similarity between the user preference vector and the report features. Combine this similarity with the user's subjective score to obtain an adjusted subjective score. Use dynamic weights to fuse the objective score and the adjusted subjective score to obtain a fused score. This specifically includes: S41. Calculate the cosine similarity between the user preference vector and each candidate report feature vector. This step calculates the cosine similarity between the user preference vector and the report feature vector, for example, by dividing the vector dot product by the modulus length, and outputs a 0-1 matching value. This quantifies the degree of personalized fit, thereby increasing the relevance of recommendations and reducing the impact of irrelevant reports on user decision-making.

[0037] S42. Multiply the cosine similarity value by the user's subjective score to generate an adjusted subjective score. This step generates an adjusted score by multiplying the similarity value by the user's subjective score. For example, high similarity amplifies the weight of the subjective score, dynamically strengthening the influence of preferences, thereby improving report satisfaction and avoiding static methods that ignore immediate user feedback.

[0038] S43. Calculate the fusion score using a dynamic weight fusion formula, where the fusion score formula is: . Indicates objective rating. Indicates adjustment of subjective ratings. Represents a dynamic weight. This step fuses objective scores and adjusts subjective scores using a dynamic weight formula, such as λ, to balance the contributions of both. The final fusion value is then output, allowing the evaluation to balance both quality and individuality, thereby reducing biased judgments and increasing report acceptance rates.

[0039] S5. Select the candidate report with the highest fusion score as the final report output. Specifically including: S51. Sort the fusion scores of all reports in the candidate report set in descending order. This step sorts the fusion scores of all reports in descending order, for example, by using a quick sort algorithm to process the score list to generate a ranking of merits and demerits. This makes the selection process orderly and efficient, thereby shortening decision-making time and avoiding random selection that reduces output quality.

[0040] S52: Select the candidate report ranked first in the fusion score as the final report. This step selects the top-ranked report as the final output, for example, by indexing the highest-scoring item, to ensure optimal content delivery, thereby increasing user satisfaction and avoiding wasting resources with suboptimal reports.

[0041] S53. Send the final report to the user terminal through the report output interface. This step sends the final report to the user terminal through the report interface, such as API transmission of file data, so that the results are immediately available, thereby improving the interaction fluency and reducing the impact of delivery delays on user experience.

[0042] S6. Calculate the arithmetic mean of the objective scores of all candidate reports at the current moment as the average objective score. Collect the user's subjective score for the final report. Update the dynamic weight based on the difference between the subjective score and the average objective score, and apply the updated dynamic weight to the subsequent fusion score calculation. Specifically include: S61. Calculate the arithmetic mean of the objective scores of all reports in the current candidate report set to generate an average objective score. This step calculates the arithmetic mean of the objective scores of all reports, for example by summing and dividing by the number, to generate a baseline quality value. This makes the system status monitorable, thereby increasing the basis for weight adjustment and preventing isolated data from misleading the optimization direction.

[0043] S62. Collect the user's subjective original score for the final report. This step collects the user's subjective original score for the final report, such as collecting feedback through pop-up windows on the interface, so that real-time feedback is entered into the system, thereby reducing the feedback loss rate and ensuring the integrity of the closed-loop optimization data.

[0044] S63: Standardize the original subjective scores to generate user subjective scores. This step standardizes the original scores, for example, scaling them to a fixed range, to generate comparable subjective scores, adapting the data to the updated formula, thereby reducing noise interference and improving the accuracy of weight calculation.

[0045] S64. Calculate the updated dynamic weight using a weight update formula, where the update formula is: . Indicates the user's subjective rating. represents the average objective rating. Represents the dynamic weight before update. represents the weight adjustment rate, wherein the weight adjustment rate It is a preset constant parameter with a value range of 0 to 1 and is determined by experimental calibration or system administrator configuration; for example, in the initial deployment phase, Set to 0.05 to balance weight update speed and stability. Weight adjustment rate As a hyperparameter, it is preset when the system is initialized and does not need to be calculated in real time. The typical value is determined based on business scenario testing, such as in the aviation marketing report. , to quickly respond to changes in user preferences. This step calculates new dynamic weights using a weight update formula. For example, based on the difference between the subjective score and the average objective score and the old weight adjustment, the parameters evolve dynamically, increasing system adaptability and preventing rigid rules from lagging behind changing user needs.

[0046] S65: Store the updated dynamic weights in the weight configuration library. This step makes the parameters persistent by storing the updated weights in the configuration library, such as writing new lambda values ​​to the database, thereby ensuring consistency in subsequent calls and reducing the risk of data loss.

[0047] S66. Invoke the dynamic weights in the weight configuration library in subsequent fusion scoring calculations. This step enables optimization results to take effect in real time by invoking the configuration library weights in subsequent fusion scoring, such as reading the latest values ​​and applying them to the formula. This increases system responsiveness and prevents historical parameters from degrading decision quality.

[0048] The present invention also proposes a non-transitory computer-readable storage medium storing a computer program, which, when executed by a processor, implements the AI-based dynamic optimization method for analysis report quality proposed by the present invention.

[0049] The above are only preferred specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with this technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solutions and inventive concepts of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A dynamic optimization method for analysis report quality based on AI, characterized in that: include: S1. Based on the user's historical reporting and rating behavior and user type description, the user preference vector is constructed by balancing the weights. S2. Generate a set of candidate reports using multiple AI models in parallel; S3. Calculate an objective score based on multiple quality dimensions for each candidate report in the candidate report set, present a summary or key features of the candidate report to the user through an interactive interface, and collect the user's original subjective score for the summary; S4. For each candidate report, calculate the similarity between the user preference vector and the report features; combine the similarity with the user's subjective score to obtain an adjusted subjective score; Use dynamic weights to fuse objective scores and adjust subjective scores to obtain a fused score; S5. Select the candidate report with the highest fusion score as the final report output.

2. The AI-based dynamic optimization method for analysis report quality according to claim 1, characterized in that: Also includes: S6. Calculate the arithmetic mean of the objective scores of all candidate reports at the current moment as the average objective score; Collect users' subjective scores on the final report; The dynamic weight is updated according to the difference between the subjective score and the average objective score, and the updated dynamic weight is applied to the subsequent fusion score calculation.

3. The AI-based dynamic optimization method for analysis report quality according to claim 1 is characterized in that ,S2 specifically includes: S21. Configure the difference parameters of multiple AI models, including differences in training data distribution and algorithm architecture; S22. In response to the report generation instruction, a parallel trigger signal is sent to all AI models; S23. Each AI model independently generates a report entity based on input data, and the report entity includes text content and feature vectors; S24. Aggregate all reporting entities to form a candidate reporting set.

4. The AI-based dynamic optimization method for analysis report quality according to claim 3, characterized in that: S21 specifically includes: S211. Obtain an algorithm architecture descriptor and a training data distribution descriptor for each AI model; S212, calculating an architecture difference value based on the algorithm architecture descriptor; S213, calculating a data distribution difference value based on the training data distribution descriptor; S214 , integrating the architecture difference value and the data distribution difference value to generate a difference parameter.

5. The AI-based dynamic optimization method for analysis report quality according to claim 1, characterized in that: S1 includes: S11, extracting the user's historical reporting and scoring behavior characteristics; S12, extracting descriptions related to the user's occupation type; S13, normalizing the user's historical reporting and scoring behavior characteristics; S14. Vectorize and encode the description of the user's occupation type; S15, the user's historical report scoring behavior characteristics after weighted fusion and normalization processing by the balance coefficient and the description of the user's occupation type after vector coding; S16. Generate a user preference vector; Among them, the construction formula of the user preference vector is: ; Represents the normalized user historical report rating behavior characteristics; Represents the user's occupation type description after vectorized encoding; is the balance coefficient.

6. The AI-based dynamic optimization method for analysis report quality according to claim 1, characterized in that S3 include: S31. Extract the information coverage feature vector, coherence feature vector, and data severity feature vector for each candidate report; S32, processing the information coverage feature vector to generate an information coverage score; S33, processing the coherence feature vector to generate a coherence score; S34, processing the data severity evaluation feature vector to generate a data severity evaluation score; S35. Weighting and fusing the information coverage score, the coherence score, and the data severity score based on a preset weight coefficient to generate an objective score; S36. Collecting the user's original subjective ratings of the candidate reports through an interactive interface; the interactive interface only displays the core feature vectors or text summaries of the candidate reports, and the duration is controlled within a range that the user can quickly evaluate, so as to efficiently collect subjective feedback; S37. Standardize the subjective original scores to generate user subjective scores.

7. The AI-based dynamic optimization method for analysis report quality according to claim 1, characterized in that S4 include: S41, calculating the cosine similarity value between the user preference vector and the feature vector of each candidate report; S42, multiplying the cosine similarity value by the user's subjective score to generate an adjusted subjective score; S43. Calculate the fusion score using a dynamic weight fusion formula, where the fusion score formula is: ; represents an objective rating; indicates adjustment of subjective ratings; Indicates dynamic weight.

8. The AI-based dynamic optimization method for analysis report quality according to claim 1, characterized in that S5 include: S51, sorting the fusion scores of all reports in the candidate report set in descending order; S52. Select the candidate report ranked first in the fusion score as the final report; S53. Send the final report to the user terminal through the report output interface.

9. The AI-based dynamic optimization method for analysis report quality according to claim 2, characterized in that S6 include: S61. Calculate the arithmetic mean of the objective scores of all reports in the candidate report set at the current moment to generate an average objective score; S62. Collecting users' subjective original scores on the final report; S63, standardizing the subjective original scores to generate user subjective scores; S64. Calculate the updated dynamic weight using a weight update formula, where the update formula is: ; Indicates the user's subjective score; represents the average objective rating; Represents the dynamic weight before update; represents the weight adjustment rate, wherein the weight adjustment rate It is a preset constant parameter with a value range set between 0 and 1 and determined by experimental calibration or system administrator configuration; S65. Storing the updated dynamic weight in the weight configuration library; S66. In subsequent fusion score calculations, the dynamic weights in the weight configuration library are called.

10. A non-transitory computer-readable storage medium, characterized in that A computer program is stored, and when the program is executed by a processor, an AI-based dynamic optimization method for analysis report quality as described in any one of claims 1 to 9 is implemented.

Images