A Dynamic Governance Retrieval Method Based on Generator-Aware Feedback and List-Based Rewards

Abstract

This specification provides a dynamic governance retrieval method based on generator-aware feedback and list-based rewards, comprising: performing multimodal preprocessing on input multi-format unstructured documents to establish a joint representation of text content and spatial coordinates, and outputting structured data; receiving mixed strings input by users, calling a large language model to perform semantic parsing and intent recognition, and outputting structured query parameters; recalling candidate data based on the structured query parameters, extracting semantic relevance features, timeliness features, data quality features, and multi-source consistency features of each candidate data, and generating a ranking list based on the features; inputting the ranking list into a generation model to generate answers, performing multi-dimensional quality assessment on the answers, and outputting assessment results; converting the assessment results into feedback signals, and dynamically optimizing the parameters and generation strategy of the ranking model based on the feedback signals.

Description

Technical Field

[0001] This document relates to the fields of data retrieval and artificial intelligence, and in particular to a dynamic governance retrieval method based on generator perceptual feedback and list-based rewards. Background Technology

[0002] As enterprises deepen their digital transformation, they accumulate massive amounts of unstructured data, including bidding documents, judgments, scanned copies of various contracts, and internal management documents. This data contains key business entities such as company names, personal names, unified social credit codes, case numbers, and contact information. However, with existing technologies, efficiently governing this mixed input and achieving accurate, unified retrieval remains a significant challenge.

[0003] First, there is a significant disconnect between the data governance process and downstream application logic. Traditional governance methods are typically limited to text extraction during the data entry stage, such as simple OCR text recognition. This approach often overlooks the deep business semantics inherent in the physical layout of documents. Because the governance stage cannot perceive the specific data quality requirements of downstream large language models, the information recalled by the retrieval system, while semantically relevant on the surface, may lack crucial contextual background or structured attributes, thus failing to support LLM in generating high-quality, factually accurate answers.

[0004] Secondly, existing retrieval ranking mechanisms generally lack a global perspective and feedback loop. Traditional algorithms are mainly based on similarity calculations in vector spaces, which essentially evaluate the relevance of individual documents in isolation. However, in complex business decision-making scenarios, large language models require a complete list of results that is logically consistent, factually complementary, and conducive to the stability of the generation model's inference. Existing technologies struggle to dynamically adjust ranking strategies based on the overall quality of the list and by sensing the generator's output feedback.

[0005] Furthermore, the evaluation dimensions for data timeliness are too simplistic. Most retrieval systems employ only simple linear decay or uniform time filtering thresholds, failing to continuously model the differentiated half-life characteristics of data from various business domains such as law, finance, and bidding. The most critical technical pain point lies in the gradient break between the retrieval engine and the generator. Data contradictions, illusion risks, or timeliness conflicts perceived by the generator during the inference phase cannot be backpropagated to the ranking engine through a differentiable path, resulting in a lack of adaptive evolutionary capabilities and output results that often appear as a black box. Summary of the Invention

[0006] This specification provides one or more embodiments of a dynamic governance retrieval method based on generator-aware feedback and list-based rewards, including: S1. Perform multimodal preprocessing on the input multi-format unstructured documents to establish a joint representation of text content and spatial coordinates, and output structured data; S2. Receive the mixed string input by the user, call the large language model to perform semantic parsing and intent recognition, and output structured query parameters; S3. Based on the structured query parameters, recall candidate data, extract semantic relevance features, timeliness features, data quality features and multi-source consistency features of each candidate data, and generate a sorted list based on the features; S4. Input the sorted list into the generation model to generate an answer, perform a multi-dimensional quality assessment on the answer, and output the assessment results; S5. The evaluation results are converted into feedback signals, and the parameters and generation strategy of the ranking model are dynamically optimized based on the feedback signals.

[0007] Furthermore, multimodal preprocessing is performed on the input multi-format unstructured documents to establish a joint representation of text content and spatial coordinates. The output structured data specifically includes: Establish a unified entry point for processing multi-format documents, and automatically distribute them to the corresponding processing channels based on the document's type identifier; The multimodal processing unit, which integrates the OCR recognition model and the document layout analysis model, is invoked to simultaneously extract the text strings in the document and their spatial coordinates on the page. The extracted spatial coordinates are matched with a preset business layout template, and the text string is mapped to the corresponding structured semantic tags based on the matching result. The output includes the text string, the spatial location coordinates, and the structured semantic tags, and generates the OCR recognition confidence index and the structured completeness index of the structured data.

[0008] Furthermore, the system receives a mixed string input from the user, calls a large language model to perform semantic parsing and intent recognition, and outputs structured query parameters, specifically including: The mixed string input by the user is used as an input prompt, and a large language model is invoked for semantic parsing; The named entity recognition algorithm is invoked in parallel along the decoding path of the large language model to extract entity words from the mixed string and classify their entity types; Based on the semantic feature vector output by the large language model, it is mapped to a predefined query intent category, and the extracted entity words are used to fill the semantic slots corresponding to the query intent category; The output contains structured query parameters including the entity term, the entity type, the query intent category, and the populated semantic slots.

[0009] Furthermore, based on the structured query parameters, candidate data is recalled, and semantic relevance features, timeliness features, data quality features, and multi-source consistency features of each candidate data are extracted. A ranking list is then generated based on these features, specifically including: Calculate the semantic relevance feature value, timeliness feature value, data quality feature value, and multi-source consistency feature value; Weight coefficients are configured for the semantic relevance feature value, timeliness feature value, data quality feature value and multi-source consistency feature value respectively, and a weighted summation operation is performed to generate a comprehensive score for each candidate data. Based on the comprehensive score, a sorted list is generated for all candidate data using a list-based sorting algorithm.

[0010] Furthermore, the sorted list is input into the generation model to generate an answer, and a multi-dimensional quality assessment is performed on the answer. The output assessment results specifically include: The content of each document in the sorted list is combined with the mixed string input by the user according to a predetermined template to form a generation context; The generated context is input into a domain-adaptive large language model, its text generation interface is called, and a structured answer is output. The coverage rate of the structured answer relative to the predefined semantic slots is calculated as the integrity score. The consistency score between the key fact fields in the structured answer and the multi-source authoritative data is calculated. The weighted average of the timeliness feature values ​​of the candidate data on which the structured answer is based is also calculated. Output the structured answer and a quality assessment report including the completeness score, consistency score, and timeliness score.

[0011] Furthermore, converting the evaluation results into feedback signals and dynamically optimizing the parameters and generation strategy of the ranking model based on the feedback signals specifically includes: The evaluation results are analyzed, and the integrity score, consistency score, and timeliness score are weighted and summed to generate a reward signal. The sorted list output by the sorting model is transformed into a probability distribution of candidate documents. The reward signal is used as the return, and the update gradient of the sorting model parameters is calculated using the policy gradient algorithm. Based on the score distribution in the evaluation results, adjust the instruction field and example field in the prompt word template used by the generation model; The updated sorting model parameters and the adjusted prompt word template are deployed to the online inference system.

[0012] Furthermore, the calculation of the semantic relevance feature value, timeliness feature value, data quality feature value, and multi-source consistency feature value is specifically as follows: For each recalled candidate data, the vector cosine similarity between it and the structured query parameters is calculated as a semantic relevance feature value; Obtain the publication timestamp and current system timestamp for each candidate data, and perform exponential decay processing on the time difference based on the half-life factor predefined by the business domain to generate a timeliness feature value; Extract a preset set of core fields from each candidate data, calculate the fill rate of the core field set as a data quality feature value, and read the preset authority weight of its source channel; The multi-source verification unit is invoked to compare the key fields of the same entity from different independent channels, and the field overlap is calculated as the multi-source consistency feature value.

[0013] This specification provides one or more embodiments of a dynamic governance retrieval system based on generator-aware feedback and list-based rewards, including: Multimodal preprocessing module: Used to perform multimodal preprocessing on input multi-format unstructured documents, establish a joint representation of text content and spatial coordinates, and output structured data; Query understanding module: It receives mixed strings input by the user, calls the large language model to perform semantic parsing and intent recognition, and outputs structured query parameters; Feature ranking module: used to recall candidate data based on the structured query parameters, extract semantic relevance features, timeliness features, data quality features and multi-source consistency features of each candidate data, and generate a ranking list based on the features; The evaluation generation module is used to input the sorted list into the generation model to generate answers, perform multi-dimensional quality evaluation on the answers, and output the evaluation results. Feedback optimization module: used to convert the evaluation results into feedback signals, and dynamically optimize the parameters and generation strategy of the ranking model based on the feedback signals.

[0014] This specification provides one or more embodiments of an electronic device, including: Processor; and, A memory is configured to store computer-executable instructions, which, when executed, cause the processor to implement the steps of the dynamic governance retrieval method based on generator-aware feedback and list-based rewards described above.

[0015] This specification provides one or more embodiments of a storage medium for storing computer-executable instructions that, when executed, implement the steps of the above-described dynamic governance retrieval method based on generator-aware feedback and list-based rewards.

[0016] By employing embodiments of this invention, a multimodal preprocessing and text-coordinate joint mechanism is used to achieve immediate governance of unstructured documents upon entry into the database. This significantly improves the accuracy and structuring level of data parsing and reduces the cost of manual cleaning. It overcomes the limitations of traditional static sorting by introducing generator-aware list-style sorting and dynamic timeliness modeling based on half-life, enabling the sorting results to directly optimize the quality of downstream generation tasks. This effectively addresses the pain points of outdated information and relevant but useless data. The constructed end-to-end feedback loop possesses strong self-evolution capabilities, continuously optimizing the sorting strategy based on actual generation results. This enhances the confidence, interpretability, and authority of retrieval results in serious business scenarios such as legal affairs and risk control.

[0017] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and in order to make the above and other objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in one or more embodiments of this specification or in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this specification. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 A flowchart illustrating a dynamic governance retrieval method based on generator-aware feedback and list-based rewards, provided for one or more embodiments of this specification; Figure 2 This diagram illustrates the unstructured data import and governance process of a dynamic governance retrieval method based on generator-aware feedback and list-based rewards, provided for one or more embodiments of this specification. Figure 3 A schematic diagram illustrating the multidimensional feature fusion and ranking of a dynamic governance retrieval method based on generator-aware feedback and list-based rewards, provided for one or more embodiments of this specification; Figure 4 A schematic diagram illustrating the composition of a dynamic governance retrieval system based on generator-aware feedback and list-based rewards, provided for one or more embodiments of this specification; Figure 5 This is a schematic diagram of the structure of an electronic device provided for one or more embodiments of this specification. Detailed Implementation

[0020] To enable those skilled in the art to better understand the technical solutions in one or more embodiments of this specification, the technical solutions in one or more embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this specification, and not all of the embodiments. Based on one or more embodiments of this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of this document.

[0021] Method Implementation Examples According to embodiments of the present invention, a dynamic governance retrieval method based on generator-aware feedback and list-based rewards is provided. Figure 1 A flowchart illustrating a dynamic governance retrieval method based on generator-aware feedback and list-based rewards, provided for one or more embodiments of this specification, is shown below. Figure 1 As shown, the dynamic governance retrieval method based on generator-aware feedback and list-based rewards according to an embodiment of the present invention specifically includes: S1. Perform multimodal preprocessing on the input multi-format unstructured documents to establish a joint representation of text content and spatial coordinates, and output structured data.

[0022] Enterprises need to handle a wide variety of document formats, including but not limited to PDF, Word, JPEG scans, and PNG images. To unify the processing of this heterogeneous data, such as... Figure 2 As shown, a unified entry point for processing multi-format documents is first constructed, with a built-in automatic file type distribution mechanism. Based on the type identifier of the input document, such as the file extension or file header information, the document is automatically distributed to the corresponding processing channel. For editable text documents, their original character stream is directly extracted; for scanned image documents, dynamic resolution optical character recognition processing is initiated; for mixed-type documents, they are automatically split into text and image regions, processed separately, and then merged.

[0023] Next, a multimodal processing unit integrating an optical character recognition model and a document layout analysis model is invoked to extract the string of each character or text block in the document while simultaneously recording its spatial coordinates on the original page, thus establishing a joint text-coordinate input mechanism. Subsequently, the extracted spatial coordinates are matched with pre-defined business layout templates for different business scenarios, and the text string is mapped to the corresponding structured semantic tags based on the matching results. After parsing a preset template and fine-tuning the OCR model, automatic location of key fields can be achieved. For example, entities at the bottom of the page are automatically marked as "signature subject," and 18-digit codes in fixed cells of tables are automatically identified as "Unified Social Credit Code." In this way, automated semantic annotation of unstructured documents is completed, ultimately outputting unified structured data containing the original text string, spatial coordinates, and mapped structured semantic tags. Simultaneously, optical character recognition confidence and structural completeness indices are generated for the output data. The recognition confidence is calculated based on the average predicted probability of each character by the optical character recognition model, while the structural completeness is calculated based on the proportion of successfully filled preset core fields, providing important data quality characteristics for subsequent sorting steps.

[0024] S2. Receives a mixed string input by the user, calls a large language model to perform semantic parsing and intent recognition, and outputs structured query parameters.

[0025] To accurately understand user intent, the received mixed string of user input is used as an input prompt, and a finely tuned large language model is invoked for semantic decoding. This large language model leverages its contextual understanding capabilities to parse out the semantic core and potential relationships within the string, such as... Figure 3 As shown.

[0026] On the decoding path of this large language model, for user input containing one or more mixed strings such as company name, person name, address, unified social credit code or case number, the large language model is called in combination with the named entity recognition algorithm to parse the text, extract entity words from the unstructured text and determine their entity type.

[0027] Based on this, the high-dimensional semantic feature vectors output by the large language model are mapped to a predefined query intent category library. After determining the intent category, the extracted entity words are used to fill the semantic slots in the structured query template corresponding to the intent category. Finally, a structured query parameter containing entity words, entity type, query intent category and all filled semantic slots is output.

[0028] S3. Based on the structured query parameters, recall candidate data, extract semantic relevance features, timeliness features, data quality features and multi-source consistency features of each candidate data, and generate a sorted list based on the features.

[0029] like Figure 4 As shown, after obtaining the structured query parameters, a batch of candidate data blocks is recalled from the generated structured database based on traditional vector retrieval or keyword retrieval methods. For each recalled candidate data, a four-dimensional core feature is calculated, including semantic relevance feature values, timeliness feature values, data quality feature values, and multi-source consistency feature values. Specifically: The semantic relevance feature value is obtained by calculating the cosine similarity between the query parameter vector and the candidate data block vector. This vector is generated by a pre-trained deep semantic embedding model, as shown below: .

[0030] The timeliness feature value is not determined by a simple yes / no threshold, but by an exponential decay model based on the half-life of the business domain. The exponential decay function based on the half-life is as follows: ; in, This indicates the duration of the data since its publication or most recent confirmation. This represents the time half-life factor defined by a specific business domain. Its value is pre-set based on the data turnover frequency of different business domains through experimental testing or historical data distribution statistics. For example, for highly dynamic bidding and tendering change data, its... The value range is typically set to [3, 15] days; for subject qualification data with higher stability, its... The value range is usually set to [180, 720] days. The decay adjustment coefficient is an empirical characteristic parameter determined by the sensitivity of the specific business domain to data timeliness. By introducing a half-life parameter, the time decay process is combined with the specific business scenario to dynamically adjust the timeliness weight for different types of data; timeliness decay factor. Calculated using the exponential decay formula, the input variables include the original time of the data and the current system time.

[0031] Data quality feature value is a composite indicator. First, a set of core fields is extracted from the candidate data, and the proportion of these fields that are successfully filled is calculated as the structured completeness. Then, it is combined with the pre-set authority weight of its data source channel for weighted summation.

[0032] OCR confidence score measures the probability that a visual recognition model accurately reconstructs characters in a document. Its value is directly derived from the Softmax normalized probability value of the output layer of a multimodal parsing model, such as a visual Transformer or a deep convolutional neural network, when classifying characters. The calculation function is: Given a data block to be recognized containing... Each of the identified characters is... The corresponding model prediction confidence level is (in If [0, 1]), then the average recognition confidence of the data block is... Defined as: ; Value definition and logical threshold: High confidence interval [0.85, 1.0] indicates clear text and extremely high recognition reliability; Medium confidence interval [0.60, 0.85) indicates possible presence of rare characters, occlusion, or blurriness, and the feature is marked as requiring manual verification; Low confidence interval [0, 0.60) indicates severe noise interference, and this feature will generate a large negative penalty in the ranking feature vector.

[0033] Structured completeness is used to evaluate the degree of fit between the parsing results and the business semantic template, reflecting whether the data possesses the necessary information dimensions to support downstream decision-making. For specific business scenarios, a set of core fields that must be extracted is pre-defined. The total number of its fields is Let the number of fields that are successfully matched and filled with non-null values ​​during the actual parsing process be . Then the structured completeness Defined as: ; Value delimitation and weighting logic: Full information state ( = 1): Core fields (such as case number, date, subject, amount, etc.) are all present and have the highest sorting priority; key information is missing ( <0.7): If key fields (such as the signature subject) are missing, even if If the value is very high, the data block will be judged to have a very low contribution to the generation task.

[0034] Finally, the two indicators mentioned above are linearly weighted and fused into a quality input feature vector for the ranking model. : ; in, and The adjustment weights are typically set according to specific business scenarios. > The lack of structured tags has a more severe impact on business logic.

[0035] Multi-source consistency feature values ​​are achieved by calling a multi-source verification unit. Based on a cross-comparison model of multi-dimensional heterogeneous data, the core process includes data feature alignment, similarity measurement calculation, and source authority weighting. Specifically, it retrieves key fields of the same entity, such as registered capital and legal representative, from multiple independent channels, such as business registration system APIs, judicial document databases, and bidding announcement databases. Jaccard similarity or edit distance algorithms are used to calculate the overlap between the source data. If the information from multiple independent sources is completely consistent, a score is awarded. The score approaches 1; if there is a significant conflict, the score decreases.

[0036] After the feature values ​​of the four dimensions mentioned above have been calculated, a weight coefficient that can be dynamically adjusted according to the business scenario is assigned to each feature value, and a comprehensive score for each candidate data is calculated by weighted summation. Finally, based on the comprehensive scores of all candidate data, a list-based sorting algorithm is used to generate a sorted list arranged from highest to lowest score. This sorting layer no longer relies solely on simple semantic similarity, but organically integrates timeliness, data quality, and source credibility, ensuring that the most up-to-date, complete, and reliable high-value information is ranked at the top.

[0037] S4. Input the sorted list into the generation model to generate the answer, perform a multi-dimensional quality assessment on the answer, and output the assessment results.

[0038] After obtaining a high-quality sorted list, the next step is answer generation and evaluation. First, the content of each document in the generated sorted list, along with the obtained structured query parameters and the user's original query string, is assembled according to a predefined prompt template to form a structured generation context.

[0039] Next, the generated context is input into a domain-adaptive large language model, and its text generation interface is invoked. The model then performs understanding, induction, and reasoning based on the provided context, outputting a structured answer in natural language form. In this embodiment, the ranking result list is generated using a generative model LLM. To generate answer A and quantify its quality, as well as to provide a basis for subsequent feedback and optimization, a multi-dimensional automatic evaluation is performed on this answer, specifically including: Calculate a completeness score to assess whether the generated answer covers all necessary semantic slots in the query intent. A predefined set of semantic slots that must be covered is used for different business scenarios. In generating answer A, the number of slots successfully identified and filled is denoted as . Then the integrity score Defined as: ; A consistency score is calculated by extracting key fact fields, such as company name, amount, and date, from the generated answer and cross-referencing them with multi-source verification results from authoritative data sources. The degree of consistency is calculated using string edit distance or numerical error. This embodiment specifically compares the key fact fields involved in the generated answer A, such as entity name, time, amount, and case number, with their corresponding fields from different independent data sources. For each fact field... Calculate its consistency score with multi-source data. ,in The final fact accuracy score is calculated based on string edit distance, numerical error threshold, or semantic similarity model. Defined as a weighted average of the consistency scores for each field: ; in, Indicates the first The importance weights of each fact field in the generation task are used to reflect the degree of influence of different fact fields on the credibility of the generated answer. These weights can be obtained by setting business rules, statistical analysis of historical data, or automatically learned during model training.

[0040] Calculate the timeliness score to assess whether the document block on which the answer is based is the latest valid version. This score is calculated based on the document time decay model. .

[0041] Finally, the structured answer and a quality assessment report containing the completeness score, consistency score, and timeliness score are output simultaneously.

[0042] S5. The evaluation results are converted into feedback signals, and the parameters and generation strategy of the ranking model are dynamically optimized based on the feedback signals.

[0043] First, the generated quality assessment report is parsed. A predefined evaluation function is used to weight and sum the completeness score, consistency score, and timeliness score in the report, calculating a unique reward signal, Reward. This signal reflects the contribution of the current ranking list to generating high-quality answers, as detailed below: ; in, It generates rewards; For completeness score, For the sake of factual accuracy, Score for timeliness; , , The weighting coefficients for each dimension of the reward function are used to reflect the importance of completeness, factual consistency, and timeliness to the generation quality in different business scenarios. These weighting coefficients can be determined through one or a combination of the following methods: (1) Based on historical labeled data, offline calibration is performed using grid search or Bayesian optimization. (2) During end-to-end training, it is updated along with the ranking model as a learnable parameter; (3) Initial values ​​are set by business experts based on the application scenario and dynamically fine-tuned during actual operation.

[0044] Since the process from sorting to generation involves discrete sampling, this embodiment employs a policy gradient-based reinforcement learning method to feed the reward signal back to the sorting model. This is coupled into a list-based loss function. Specifically, the raw score of each candidate document in the ranking model's output is transformed into a probability distribution using a softmax function, representing the probability of each document being selected. Let the ranking model evaluate the documents... The predicted score is Then the sampling probability P of the document in the list is... Defined through the Softmax function: ; Overall list-based reward-loss function Defined as: ; in To reward incremental It is the baseline mean of the reward, used to reduce the variance of the gradient. If the quality of the answers generated by the current list is higher than the average level, this value is positive; otherwise, it prompts the model to suppress the ranking logic. For log-likelihood terms, it ensures that the update direction of the ranking model parameters moves towards the list distribution that yields higher rewards.

[0045] To preserve the semantic foundation of the ranking model, the final loss function... This typically combines the traditional list sorting loss with the generative-aware reward loss: ; in, [0, 1] are moderating factors used to balance the ratio between semantic relevance and generation utility.

[0046] This reward signal can comprehensively consider the completeness, consistency, fact matching degree, and timeliness coverage of the answers, and is used to measure the actual contribution of the ranking results to the generation task.

[0047] The ranking model parameters are updated via backpropagation based on the aforementioned gradients. A policy gradient-based reinforcement learning optimization algorithm (such as the REINFORCE algorithm) can be employed, using generated rewards as the feedback signal to weight the log probability of the ranking model on the current sampled list. This is then combined with an Adam or stochastic gradient descent optimizer to complete the parameter update. This update mechanism can guide the ranking model to gradually favor ranking strategies that produce high-quality generated answers, without relying on the explicit differentiability of the generation process.

[0048] Meanwhile, based on the score distribution in the evaluation results, we analyze which types of queries or documents led to low scores, and dynamically adjust the instruction fields and example fields in the prompt word template used in the generation stage accordingly. Finally, we deploy the updated sorting model parameters and the adjusted prompt word template to the online inference system to complete the optimization.

[0049] The beneficial effects of this invention are as follows: By employing multimodal preprocessing and a text-coordinate joint mechanism, unstructured documents are processed immediately upon entry into the database, significantly improving the accuracy and structuring of data parsing while reducing manual cleaning costs. Breaking through the limitations of traditional static sorting, the introduction of generator-aware list-based sorting and dynamic timeliness modeling based on half-life enables the sorting results to directly optimize the quality of downstream generation tasks, effectively addressing the pain points of outdated information and relevant but useless data. The constructed end-to-end feedback loop possesses strong self-evolution capabilities, continuously optimizing the sorting strategy based on actual generation results, enhancing the confidence, interpretability, and authority of search results in serious business scenarios such as legal affairs and risk control.

[0050] System Implementation Examples According to embodiments of the present invention, a dynamic governance retrieval system based on generator-aware feedback and list-based rewards is provided. Figure 4 This specification provides a schematic diagram illustrating the composition of a dynamic governance retrieval system based on generator-aware feedback and list-based rewards, as shown in one or more embodiments. Figure 4 As shown, the dynamic governance retrieval system based on generator-aware feedback and list-based rewards according to an embodiment of the present invention specifically includes: Multimodal preprocessing module 40: Used to perform multimodal preprocessing on the input multi-format unstructured documents, establish a joint representation of text content and spatial coordinates, and output structured data; Query understanding module 42: It is used to receive mixed strings input by users, call the large language model to perform semantic parsing and intent recognition, and output structured query parameters; Feature ranking module 44: used to recall candidate data based on the structured query parameters, extract semantic relevance features, timeliness features, data quality features and multi-source consistency features of each candidate data, and generate a ranking list based on the features; Evaluation generation module 46: used to input the sorted list into the generation model to generate answers, perform multi-dimensional quality evaluation on the answers, and output the evaluation results; Feedback optimization module 48: used to convert the evaluation results into feedback signals, and dynamically optimize the parameters and generation strategy of the sorting model based on the feedback signals.

[0051] The embodiments of the present invention are system embodiments corresponding to the above method embodiments. The specific operation of each module can be understood by referring to the description of the method embodiments, and will not be repeated here.

[0052] Device Example 1 This invention provides an electronic device, such as... Figure 5 As shown, it includes: a memory 50, a processor 52, and a computer program stored in the memory 50 and executable on the processor 52. When the computer program is executed by the processor 52, it performs the following method steps: S1. Perform multimodal preprocessing on the input multi-format unstructured documents to establish a joint representation of text content and spatial coordinates, and output structured data; S2. Receive the mixed string input by the user, call the large language model to perform semantic parsing and intent recognition, and output structured query parameters; S3. Based on the structured query parameters, recall candidate data, extract semantic relevance features, timeliness features, data quality features and multi-source consistency features of each candidate data, and generate a sorted list based on the features; S4. Input the sorted list into the generation model to generate an answer, perform a multi-dimensional quality assessment on the answer, and output the assessment results; S5. The evaluation results are converted into feedback signals, and the parameters and generation strategy of the ranking model are dynamically optimized based on the feedback signals.

[0053] Device Example 2 This invention provides a computer-readable storage medium storing an information transmission implementation program. When executed by a processor 52, the program performs the following method steps: S1. Perform multimodal preprocessing on the input multi-format unstructured documents to establish a joint representation of text content and spatial coordinates, and output structured data; S2. Receive the mixed string input by the user, call the large language model to perform semantic parsing and intent recognition, and output structured query parameters; S3. Based on the structured query parameters, recall candidate data, extract semantic relevance features, timeliness features, data quality features and multi-source consistency features of each candidate data, and generate a sorted list based on the features; S4. Input the sorted list into the generation model to generate an answer, perform a multi-dimensional quality assessment on the answer, and output the assessment results; S5. The evaluation results are converted into feedback signals, and the parameters and generation strategy of the ranking model are dynamically optimized based on the feedback signals.

[0054] The computer-readable storage media described in this embodiment include, but are not limited to, ROM, RAM, disk, or optical disk.

[0055] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; 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 or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A dynamic governance retrieval method based on generator-aware feedback and list-based rewards, characterized in that, include: S1. Perform multimodal preprocessing on the input multi-format unstructured documents to establish a joint representation of text content and spatial coordinates, and output structured data; S2. Receive the mixed string input by the user, call the large language model to perform semantic parsing and intent recognition, and output structured query parameters; S3. Based on the structured query parameters, recall candidate data, extract semantic relevance features, timeliness features, data quality features and multi-source consistency features of each candidate data, and generate a sorted list based on the features; S4. Input the sorted list into the generation model to generate an answer, perform a multi-dimensional quality assessment on the answer, and output the assessment results; S5. The evaluation results are converted into feedback signals, and the parameters and generation strategy of the ranking model are dynamically optimized based on the feedback signals.

2. The method according to claim 1, characterized in that, Multimodal preprocessing is performed on the input multi-format unstructured documents to establish a joint representation of text content and spatial coordinates. The output structured data specifically includes: Establish a unified entry point for processing multi-format documents, and automatically distribute them to the corresponding processing channels based on the document's type identifier; The multimodal processing unit, which integrates the OCR recognition model and the document layout analysis model, is invoked to simultaneously extract the text strings in the document and their spatial coordinates on the page. The extracted spatial coordinates are matched with a preset business layout template, and the text string is mapped to the corresponding structured semantic tags based on the matching result. The output includes the text string, the spatial location coordinates, and the structured semantic tags, and generates the OCR recognition confidence index and the structured completeness index of the structured data.

3. The method according to claim 1, characterized in that, It receives a mixed string input from the user, calls a large language model to perform semantic parsing and intent recognition, and outputs structured query parameters, including: The mixed string input by the user is used as an input prompt, and a large language model is invoked for semantic parsing; The named entity recognition algorithm is invoked in parallel along the decoding path of the large language model to extract entity words from the mixed string and classify their entity types; Based on the semantic feature vector output by the large language model, it is mapped to a predefined query intent category, and the extracted entity words are used to fill the semantic slots corresponding to the query intent category; The output contains structured query parameters including the entity term, the entity type, the query intent category, and the populated semantic slots.

4. The method according to claim 1, characterized in that, Based on the structured query parameters, candidate data is recalled, and semantic relevance features, timeliness features, data quality features, and multi-source consistency features of each candidate data are extracted. A ranking list is then generated based on these features, specifically including: Calculate the semantic relevance feature value, timeliness feature value, data quality feature value, and multi-source consistency feature value; Weight coefficients are configured for the semantic relevance feature value, timeliness feature value, data quality feature value and multi-source consistency feature value respectively, and a weighted summation operation is performed to generate a comprehensive score for each candidate data. Based on the comprehensive score, a sorted list is generated for all candidate data using a list-based sorting algorithm.

5. The method according to claim 1, characterized in that, The sorted list is input into the generative model to generate answers, and a multi-dimensional quality assessment is performed on the answers. The output assessment results specifically include: The content of each document in the sorted list is combined with the mixed string input by the user according to a predetermined template to form a generation context; The generated context is input into a domain-adaptive large language model, its text generation interface is called, and a structured answer is output. The coverage rate of the structured answer relative to the predefined semantic slots is calculated as the integrity score. The consistency score between the key fact fields in the structured answer and the multi-source authoritative data is calculated. The weighted average of the timeliness feature values ​​of the candidate data on which the structured answer is based is also calculated. Output the structured answer and a quality assessment report including the completeness score, consistency score, and timeliness score.

6. The method according to claim 1, characterized in that, The evaluation results are transformed into feedback signals, and the parameters and generation strategy of the ranking model are dynamically optimized based on the feedback signals, specifically including: The evaluation results are analyzed, and the integrity score, consistency score, and timeliness score are weighted and summed to generate a reward signal. The sorted list output by the sorting model is transformed into a probability distribution of candidate documents. The reward signal is used as the return, and the update gradient of the sorting model parameters is calculated using the policy gradient algorithm. Based on the score distribution in the evaluation results, adjust the instruction field and example field in the prompt word template used by the generation model; The updated sorting model parameters and the adjusted prompt word template are deployed to the online inference system.

7. The method according to claim 4, characterized in that, The calculation of the semantic relevance feature value, timeliness feature value, data quality feature value, and multi-source consistency feature value is specifically as follows: For each recalled candidate data, the vector cosine similarity between it and the structured query parameters is calculated as a semantic relevance feature value; Obtain the publication timestamp and current system timestamp for each candidate data, and perform exponential decay processing on the time difference based on the half-life factor predefined by the business domain to generate a timeliness feature value; Extract a preset set of core fields from each candidate data, calculate the fill rate of the core field set as a data quality feature value, and read the preset authority weight of its source channel; The multi-source verification unit is invoked to compare the key fields of the same entity from different independent channels, and the field overlap is calculated as the multi-source consistency feature value.

8. A dynamic governance retrieval system based on generator-aware feedback and list-based rewards, characterized in that, include: Multimodal preprocessing module: Used to perform multimodal preprocessing on input multi-format unstructured documents, establish a joint representation of text content and spatial coordinates, and output structured data; Query understanding module: It receives mixed strings input by the user, calls the large language model to perform semantic parsing and intent recognition, and outputs structured query parameters; Feature ranking module: used to recall candidate data based on the structured query parameters, extract semantic relevance features, timeliness features, data quality features and multi-source consistency features of each candidate data, and generate a ranking list based on the features; The evaluation generation module is used to input the sorted list into the generation model to generate answers, perform multi-dimensional quality evaluation on the answers, and output the evaluation results. Feedback optimization module: used to convert the evaluation results into feedback signals, and dynamically optimize the parameters and generation strategy of the ranking model based on the feedback signals.

9. An electronic device, characterized in that, include: processor; as well as, A memory configured to store computer-executable instructions, which, when executed, cause the processor to implement the steps of the dynamic governance retrieval method based on generator-aware feedback and list-based rewards as described in any one of claims 1 to 7.

10. A storage medium, characterized in that, Used to store computer-executable instructions, which, when executed, implement the steps of the dynamic governance retrieval method based on generator-aware feedback and list-based rewards as described in any one of claims 1 to 7.

Images