Content-level evidence packaging and multi-source confidence fusion generation method

Abstract

The invention relates to the field of artificial intelligence and natural language processing, and discloses a term-level evidence packaging and multi-source confidence fusion generation method. According to the method, a clause-level evidence packet is constructed through mixed retrieval, ternary confidence coefficients of retrieval margins, coverage rates and generation consistency are calculated, scores are fused, re-retrieval or rewriting is controlled based on an adaptive threshold value, conflict detection and weight reduction are carried out, and finally inline output clause numbers and timeliness labels are generated in a controlled mode. According to the method, the reliability, interpretability and traceability of the generated content are remarkably improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of artificial intelligence and natural language processing, and particularly relates to a clause-level evidence packaging and multi-source confidence fusion generation method. BACKGROUND

[0002] With the rapid development of artificial intelligence and natural language processing technology, retrieval-augmented generation (RAG) based question answering and document generation methods have been widely used in the intelligent interpretation of regulatory, standard and institutional texts. Such methods usually retrieve relevant fragments from a knowledge base and combine them with a generation model to automatically answer input questions and generate content. However, existing technologies still have obvious deficiencies in clause-level evidence organization, evidence credibility assessment and the verifiability of generated content.

[0003] Existing RAG methods usually splice evidence in the form of "text fragment + reference link", which has a coarse granularity and is difficult to accurately correspond to structured elements such as specific clause numbers, table cells or effective times in regulatory documents. This coarse-grained splicing method makes it difficult to review the cited evidence in the generated results, and lacks traceability and consistency at the clause level.

[0004] In terms of evidence credibility assessment, traditional methods mostly rely on a single confidence signal, such as retrieval score or cross-encoder matching score. This single-dimensional evaluation method cannot comprehensively reflect the multi-faceted quality characteristics of evidence, such as the semantic matching degree of evidence and query, the coverage of evidence to query elements, and the consistency relationship between generated content and evidence. As a result, the system cannot accurately judge the reliability of the generated content when faced with insufficient, conflicting or outdated evidence.

[0005] In addition, existing methods usually lack adaptive threshold mechanisms. During the generation process, when the number of available evidence is limited or there are conflicts between evidence, the system cannot actively trigger re-retrieval or rewriting operations, nor can it explicitly warn about potential high-risk outputs. This not only reduces the explainability and stability of the generated results, but also increases the risk of "hallucinations" (i.e. content that does not match the facts).

[0006] On the other hand, current generation models also lack standardization in output format. Most systems only add external reference links in the generated results, without forcing to mark clause numbers and time labels in the output, resulting in a lack of verifiability and reviewability in the generated content, which does not meet the requirements of "clear basis and traceable source" in regulatory or standard documents.

[0007] The existing retrieval enhancement generation-based question answering and generation technology still has the following key problems when facing the structured features and high trust requirements of regulatory documents: (1) The evidence structure is not unified, and clause-level accurate alignment and organization cannot be achieved; (2) The confidence evaluation depends on a single signal, and lacks a multi-dimensional fusion mechanism; (3) There is a lack of adaptive threshold and re-retrieval mechanism, and the system robustness is insufficient; (4) The generated result reference is not standardized, lacks clause ID and time limit identification, and has poor checkability.

[0008] These problems restrict the application of intelligent question answering and document generation systems in high-reliability scenarios such as regulations, policies and standard interpretation, and there is an urgent need for a new technical solution that can achieve clause-level evidence packaging, confidence fusion evaluation and adaptive generation control, to improve the reliability, explainability and review traceability of the generated content. SUMMARY

[0009] The present application aims to solve the following key problems of the existing retrieval enhancement generation technology in the processing of regulations, standards and other documents: (1) the evidence granularity is coarse, and clause-level accurate alignment and traceable reference cannot be achieved; (2) the confidence evaluation depends on a single signal, and lacks a multi-dimensional fusion mechanism; (3) there is a lack of adaptive threshold and closed-loop control, and the system robustness is insufficient; (4) the generated result reference is not standardized, and has poor checkability.

[0010] To solve the above technical problems, the present application is implemented as follows: The embodiment of the present application provides a generation method of clause-level evidence packaging and multi-source confidence fusion, comprising the following steps: Step 1: evidence collection and granularity alignment, performing hybrid retrieval on the input query, obtaining candidate documents and calculating comprehensive scores through sparse retrieval and dense retrieval fusion; structurally analyzing the candidate documents, and constructing a clause-level evidence package based on clause ID, text span, table cell, source information and time limit interval, wherein the clause-level evidence package comprises clause unique identification, text span set, table cell set, source information and time limit interval, wherein the text span set represents the position interval of the evidence in the document, and the table cell set comprises table number, row key and column key; Step 2: ternary confidence calculation, calculating retrieval margin, coverage rate and generation consistency for each evidence package respectively, and generating a comprehensive confidence score according to a weighted fusion formula; Step 3: fusion scoring and pre-generation, calculating a comprehensive score based on the fusion confidence score, and guiding the generation model to output a draft answer according to the score; Step 4: Adaptive threshold and re-retrieval / rewrite control, set dynamic threshold according to historical confidence distribution, automatically trigger re-retrieval or rewrite operation when fusion score is below threshold or conflict evidence is detected; Step 5: Conflict detection and weight reduction strategy, detect numerical conflict, modal conflict and time-effect conflict between evidence packages, apply weight reduction or exclusion to evidence with conflict; Step 6: Controlled generation and inline output, force output of clause number and time-effect label in pre-set format in the final generation stage, realize traceable structured generation result.

[0011] Further, the mixed retrieval is obtained by proportionally weighting fusion of sparse retrieval scores and dense semantic matching scores, wherein the comprehensive score calculation formula is: ; Wherein, is a weighting coefficient; is the sparse retrieval score of the document ; is the semantic similarity output by the dense encoder.

[0012] Further, the fusion score of the ternary confidence uses the following calculation formula: ; Wherein, is the retrieval margin; is the coverage rate; is the generation consistency; is the time-effect penalty term; to is a learnable weighting parameter.

[0013] Further, the learnable weighting parameter is optimized through end-to-end training, specifically including ranking supervision and calibration supervision.

[0014] Further, the dynamic threshold is adaptively adjusted based on the quantile of the historical fusion score distribution and the query complexity, and the calculation formula is: ; Wherein, is the confidence quantile in the historical distribution; is an adjustment factor; is a query complexity function.

[0015] Further, the conflict detection includes three types of numerical conflict, modal conflict and time-effect conflict, by calculating the conflict degree and applying weight reduction to the evidence with conflict, the weight after weight reduction is used to recalculate the fusion score.

[0016] ​Further, the controlled generation adopts a structured template, and clause numbers and time limit labels are outputted in line after each conclusion sentence, and the output format follows the rules: [§ClauseID | eff=StartDate | exp=EndDate]; ClauseID represents the unique identifier of the clause; eff=StartDate represents the effective date of the clause; StartDate is the date on which the clause comes into effect; exp=EndDate represents the expiration date of the clause; EndDate is the date on which the clause is no longer applicable; To ensure the verifiability and verifiability of the generated results.

[0017] Further, the method supports end-to-end training through manual annotation data, and optimizes the explainability and accuracy of the confidence model through joint optimization of ranking supervision and calibration supervision.

[0018] Further, the method further comprises a multi-task joint training mechanism, which combines natural language entailment and numerical consistency classification auxiliary tasks to jointly optimize the model, so as to improve the cross-modal consistency and generalization ability.

[0019] Further, the method is suitable for text generation scenarios that require high reliability and traceability of regulations, standards and policy explanations, and outputs verifiable answers with clause ID and time limit identification.

[0020] Compared with the prior art, the present application has the following advantages: (1) Traceability is enhanced: through the standardized construction and inline annotation of the clause-level evidence package, the accurate association between the generated content and the source clause is realized, and the verifiability and audit efficiency of the results are significantly improved.

[0021] (2) Reliability is improved: the ternary confidence fusion mechanism comprehensively considers multiple-dimensional signals such as retrieval, coverage and consistency, so that the evidence evaluation is more comprehensive and objective, and the "illusion" risk is effectively reduced.

[0022] (3) System robustness is optimized: the adaptive threshold and closed-loop control mechanism enable the system to automatically cope with complex situations such as insufficient evidence, conflict or expiration, improving the stability and reliability of the output.

[0023] (4) Automation and standardization: the method realizes the full-process automation and output standardization from evidence retrieval, evaluation to generation, and is particularly suitable for text generation scenarios such as regulations, standards and policy explanations that require high reliability. BRIEF DESCRIPTION OF DRAWINGS

[0024] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiments description. Obviously, the drawings in the following description only some of the embodiments of the present application, and for those skilled in the art, other drawings can be obtained without creative labor based on these drawings. Figure 1 The flow chart of the clause-level evidence packaging and multi-source confidence fusion generation method provided by the present application is shown in the figure. Figure 2 The figure is one of the hardware structure schematic diagrams of the electronic device provided by the embodiments of the present application. Figure 3 The figure is another hardware structure schematic diagram of the electronic device provided by the embodiments of the present application. DETAILED DESCRIPTION

[0025] The technical solutions in the embodiments of the present application will be described clearly and completely in the following with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of the present application.

[0026] The terms "first", "second", etc. in the specification of the present application are used to distinguish similar objects, and are not used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged under appropriate circumstances, so that the embodiments of the present application can be implemented in an order other than those illustrated or described herein, and the objects distinguished by "first", "second", etc. are generally of a kind and do not limit the number of objects, for example, the first object can be one or more. In addition, "and / or" in the specification means at least one of the connected objects, and the character " / ", generally represents that the front and rear associated objects are in an "or" relationship.

[0027] Please refer to Figure 1 The embodiments of the present application provide a clause-level evidence packaging and multi-source confidence fusion generation method, which includes the following steps: Step S101: evidence collection and granularity alignment, performing hybrid retrieval on the input query, obtaining candidate documents and calculating comprehensive scores through sparse retrieval and dense retrieval fusion calculation; structurally analyzing the candidate documents, constructing a clause-level evidence package based on clause ID, text span, table cell, source information and time interval, wherein the clause-level evidence package includes clause unique identifier, text span set, table cell set, source information and time interval, and the text span set represents the position interval of the evidence in the document, and the table cell set includes table number, row key and column key; Step S102: Ternary confidence calculation, calculate retrieval margin, coverage and generation consistency for each evidence package respectively, and generate comprehensive confidence score according to the weighted fusion formula; Step S103: Fusion score and pre-generation, calculate comprehensive score based on fusion confidence score, and guide the generation of model output draft answer with the score; Step S104: Adaptive threshold and re-search / re-write control, set dynamic threshold according to historical confidence distribution, and automatically trigger re-search or re-write operation when fusion score is lower than threshold or conflict evidence is detected; Step S105: Conflict detection and weight reduction strategy, detect numerical conflict, modal conflict and time conflict between evidence packages, and apply weight reduction or rejection to evidence with conflict; Step S106: Controlled generation and inline output, force to output clause number and time label according to preset format in the final generation stage, and realize traceable structured generation result.

[0028] In step S101, the system first performs hybrid retrieval process to collect preliminary evidence for the input query. Assuming that the input query is , the corresponding candidate document set is , , where represents the i-th document in the candidate document set . The comprehensive matching score is obtained by proportionally weighting the sparse retrieval score and the dense semantic matching score : ; wherein is the weighted coefficient for balancing the two signals; is the sparse retrieval score of document ; and is the semantic similarity output by the dense encoder.

[0029] It should be further pointed out that the sparse retrieval adopts BM25 algorithm, and the dense retrieval adopts a double encoder model based on BERT.

[0030] The system rearranges the candidate set, and the top two scores are , which are used for calculating retrieval margin. This refers to the ranking of the document scores in the candidate document set after hybrid retrieval. represents the document with the highest score, represents the second highest score.

[0031] In order to realize clause-level accurate alignment, the candidate document is parsed and decomposed into clause set , wherein represents the text of the i-th clause in the set The smallest evidence span set most relevant to the query is extracted by attention backtracking, integral gradient or contrast mask, etc. , that is, the position interval of the evidence in the document. and represent the left boundary and the right boundary of the evidence text, that is, the position interval of the evidence in the document.

[0032] It should be further explained that the structured analysis adopts a segmentation model based on an attention mechanism, and a rule engine is combined to extract clauses and table structures.

[0033] If the clause contains table information, the associated table number , row key and column key are detected and uniformly mapped to: ; To ensure consistency of multiple sources, the system normalizes the units and precision of numerical content. For example, the voltage "220V" and "0.22kV" are mapped to a unified dimension. Through the above process, the evidence package structure containing the unique identification of the clause , the text span set , the table cell set , the source information and the time interval can be formed: ; ; Each evidence package records the source credibility , for example, the weight of an official source is higher than that of a third-party source.

[0034] In step S102, in order to objectively evaluate the credibility of the evidence, the invention designs a ternary confidence index: retrieval margin , coverage and generation consistency . The fusion of the three constitutes the final confidence score.

[0035] (1) Retrieval margin

[0036] The retrieval margin reflects the discrimination between correct evidence and suboptimal evidence, and is defined as: ; Wherein, is the set of selected effective evidence packages; is the candidate evidence package that is not selected; is the rearrangement score corresponding to the evidence package; represents the i-th evidence selected as the valid evidence package; denotes the evidence package selected as the valid evidence package.

[0037] It can be seen that, the greater the value, the clearer the system's discrimination of correct evidence, and the higher the confidence level.

[0038] (2) Coverage C Coverage measures the degree to which evidence supports the content of the query and the generated results. First, the query is decomposed into a set of elements denotes the element in the query.

[0039] Query alignment rate is defined as: ; Subsequently, the temporary answer is obtained in the generation phase, and its semantic element set is extracted, and the coverage of the generated evidence is calculated : ; wherein, denotes the candidate answer in the generated answer. For content involving numerical values, numerical consistency also needs to be verified. Let the set of numerical value pairs in the evidence package be , wherein vi denotes a numerical value in the evidence package , usually referring to a certain numerical information mentioned in the evidence, denotes the related unit in the generated result, and , then the numerical consistency coverage is: ; Comprehensive coverage is defined as a weighted average: ; wherein, is an empirical weight, and satisfies .

[0040] (3) Generation consistency U Generation consistency measures the semantic matching, self-consistency and reference accuracy between the generated answer and the evidence.

[0041] 1. Semantic entailment degree : the entailment probability of the evidence to the answer is calculated using a natural language reasoning model .

[0042] 2. Consistency : The consistency of different sampling temperatures under the same evidence input ; , where is the th generated candidate answer. The main answer is determined by the voting result, and the consistency is defined as: ; , where is the th generated candidate answer.

[0043] 3. Alignment rate : Calculate the proportion of tokens in the answer covered by the evidence : ; Synthetic consistency is: ; , where .

[0044] , where is a constraint condition of the weighted coefficient, indicating that the sum of the weights of the three components of the generated consistency should be 1.

[0045] Step 3: Fusion score and pre-generation After obtaining the ternary confidence, the evidence is weighted and fused to obtain the final fusion score : ; , where is a learnable weighting parameter optimized through end-to-end training, including ranking supervision and calibration supervision. It is supervised and learned through artificially labeled query-evidence pairs, and the optimization goal is the ranking loss function; is a time penalty term. If the query time is not within the evidence time interval , set , otherwise 0.

[0046] The system uses the fusion score as the index of adoptability and generates a draft answer accordingly. The generation process can use template-based controlled decoding to make the output initially have a clause reference format.

[0047] The model parameters can be learned by minimizing the negative log-likelihood or maximizing the AUC / NDCG metrics. The probabilistic calibration of the fusion scores can be done by Platt scaling or temperature scaling.

[0048] In step S104, in order to enhance the robustness of the system, the application introduces a dynamic threshold mechanism. Let the historical quantile function of the global confidence distribution be The dynamic threshold is defined as: ; wherein is the th quantile of the historical distribution; is an adjustment factor; is a query complexity function, which can be estimated by query length, entity number, etc.

[0049] The decision rule is as follows: If , the system automatically triggers re-retrieval or rewriting; If there is conflicting evidence (see step S105), perform purification and weight reduction; When , enter the final generation stage.

[0050] Re-retrieval includes synonym expansion, time / version relaxation, and query rewriting, and rewriting is achieved by adjusting the control parameters (such as temperature, sampling range) to achieve more conservative or multi-reference output.

[0051] In step S105, the conflict detection includes three types of numerical conflict, modal conflict and time conflict, by calculating the conflict degree and applying weight reduction to the evidence with conflict, the weight after weight reduction is used to recalculate the fusion score, which specifically includes: (1) Numerical conflict detection Let two evidence bags contain numerical items , when and the units are consistent, is a threshold defined in numerical conflict detection. The numerical conflict degree is: ; (2) Modal conflict detection If the obligation modal set in the evidence has an opposite relationship (such as "should" and "prohibit") with , the modal conflict degree , otherwise 0.

[0052] (3) Time conflict detection If the evidence bag effective interval With query time incompatible, and Each represents an evidence package The effective start time and effective end time. Then there is a time limit conflict. for: ; Overall Conflict Level for: ; Apply a weight reduction to conflicting evidence packages: ,in Indicates evidence package The weight is used to measure the importance of the evidence in the final fusion result.

[0053] The reduced weights are used to recalculate the fusion score. This ensures that the output is based on consistent and valid evidence.

[0054] In step S106, when the fusion is complete... At this point, the system enters the controlled generation phase. It employs a structured prompt template, mandating the appending of clause numbers and expiration information after each concluding sentence.

[0055] Inline annotation rules are defined as follows (using Extended Backus Normal Form EBNF): [§ClauseID|eff=StartDate|exp=EndDate]; The above rules define the format of inline annotation rules, requiring that clause numbers and expiration information be appended to the generated text to ensure the structured and traceable nature of the generated content. Here, ClauseID represents the unique identifier of the clause. Each clause in a regulation, standard, or policy document will have a unique clause number (e.g., Article 12 of a regulation) to identify that clause; eff=StartDate indicates the effective date of the clause; StartDate is the date the clause begins to take effect; exp=EndDate indicates the expiration date of the clause; EndDate is the date the clause is no longer applicable.

[0056] For example: "According to Article 12 of the 'Safety Management Regulations,' rectification should be completed within 24 hours." [§12|eff=2023-07-01|exp=2026-06-30]".

[0057] When the answer contains table references, append the table location identifier after the numerical value, such as "(Table 3: Row 2)". (Column 4).

[0058] After generation, the system performs a coverage self-check, if the coverage or consistency of the final answer is lower than the draft phase result, then it reverts to the re-search branch to regenerate.

[0059] Further, the method supports end-to-end training through artificial annotation data, and optimizes the explainability and accuracy of the confidence model through ranking supervision and calibration supervision joint optimization, specifically including: First, based on the artificially annotated adoptability score, a training set is constructed, and the fusion function is executed for ranking supervision, and the objective function is: ; Wherein, represents a pair of high and low confidence samples; and are the fusion scores of two samples and respectively; represents the Sigmoid function.

[0060] Secondly, in the calibration supervision stage, the confidence threshold is fitted with the artificial "adoptability / need to check" label, so that the confidence output is monotonically related to the true adoptability probability.

[0061] In addition, the system can jointly execute natural language inference (NLI) and numerical consistency discrimination auxiliary tasks to improve the generalization ability of the model under different modalities.

[0062] Further, the method further includes a multi-task joint training mechanism, which combines natural language implication and numerical consistency classification auxiliary tasks to jointly optimize the model to improve cross-modal consistency and generalization ability.

[0063] Further, the method is suitable for text generation scenarios that require high reliability and traceability for the interpretation of regulations, standards and policies, and outputs verifiable answers with clause ID and time limit identification.

[0064] As Figure 2 shown, the embodiment of the present application also provides an electronic device 600, which includes a processor 601, a memory 602, a program or instruction stored in the memory 602 and executable on the processor 601, which implements each process of the above-mentioned clause-level evidence packaging and multi-source confidence fusion generation method embodiment when executed by the processor 601, and can achieve the same technical effect. To avoid repetition, it will not be repeated here.

[0065] It should be noted that the first electronic device in the embodiments of the present application includes the mobile electronic device and the non-mobile electronic device described above.

[0066] Figure 3 A hardware structure schematic diagram of an electronic device according to an embodiment of the present application.

[0067] The electronic device 700 includes, but is not limited to, a radio frequency unit 701, a network module 702, an audio output unit 703, an input unit 704, a sensor 705, a display unit 706, a user input unit 707, an interface unit 708, a memory 709, and a processor 710, etc.

[0068] Those skilled in the art can understand that the electronic device 700 can further include a power supply (such as a battery) for supplying power to each component, and the power supply can be logically connected to the processor 710 through a power management system, so as to realize the functions of power management, such as charging, discharging, and power consumption management, through the power management system. Figure 3 The electronic device structure shown in the figure does not constitute a limitation on the electronic device, and the electronic device can include more or fewer components than the figure, or combine certain components, or different component arrangements, which are not described here.

[0069] It should be understood that in the embodiments of the present application, the input unit 704 can include a graphics processing unit (GPU) 7041 and a microphone 7042, and the graphics processing unit 7041 processes image data of a still image or a video obtained by an image capture device (such as a camera) in a video capture mode or an image capture mode. The display unit 706 can include a display panel 7061, which can be configured in the form of a liquid crystal display, an organic light-emitting diode, etc. The user input unit 707 includes a touch panel 7071 and other input devices 7072. The touch panel 7071 is also called a touch screen. The touch panel 7071 can include two parts of a touch detection device and a touch controller. The other input devices 7072 can include, but are not limited to, a physical keyboard, function keys (such as volume control keys, on-off keys, etc.), trackballs, mice, joysticks, etc., which are not described here. The memory 709 can be used to store software programs and various data, including but not limited to application programs and operating systems. The processor 710 can integrate an application processor and a modem processor, wherein the application processor mainly processes operating systems, user interfaces, and application programs, etc., and the modem processor mainly processes wireless communication. It can be understood that the above-mentioned modem processor can also not be integrated into the processor 710.

[0070] The embodiment of the present application further provides a readable storage medium, which stores programs or instructions, and the programs or instructions are executed by a processor to realize the processes of the method for generating clause-level evidence packaging and multi-source confidence fusion and achieve the same technical effects. To avoid repetition, details are not described herein.

[0071] The processor is the processor in the electronic device in the above embodiment. The readable storage medium includes a computer readable storage medium, such as a computer read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0072] The embodiment of the present application further provides a chip, which includes a processor and a communication interface, the communication interface is coupled with the processor, and the processor is used to run programs or instructions to realize the processes of the method for generating clause-level evidence packaging and multi-source confidence fusion and achieve the same technical effects. To avoid repetition, details are not described herein.

[0073] It should be understood that the chip mentioned in the embodiment of the present application can also be referred to as a system-level chip, a system chip, a chip system or a system-on-chip chip, etc.

[0074] It should be noted that, in this document, the term "comprising" or "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that the process, method, article or device including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such process, method, article or device. Without more limitations, the element defined by the statement "comprising a" does not exclude the presence of other identical elements in the process, method, article or device including the element.

[0075] In addition, it should be noted that the scope of the method and system in the embodiment of the present application is not limited to performing functions in the order shown or discussed, but can also include performing functions in a substantially simultaneous manner or in a reverse order according to the functions involved, for example, the described method can be performed in an order different from the described order, and various steps can also be added, omitted or combined. In addition, the features described with reference to some examples can be combined in other examples.

[0076] The embodiments of the present application are described above with reference to the accompanying drawings, but the present application is not limited to the above-described specific embodiments, and the above-described specific embodiments are merely illustrative, but not restrictive, and a person of ordinary skill in the art can make many forms under the inspiration of the present application without departing from the scope of the present application, and all of them belong to the protection of the present application.

Claims

1. A method for generating clause-level evidence through packaging and multi-source confidence fusion, characterized in that, Includes the following steps: Step 1: Evidence collection and granularity alignment. Perform hybrid retrieval on the input query to obtain candidate documents and calculate the comprehensive score by fusing sparse retrieval and dense retrieval. Candidate documents are structured and parsed. Clause-level evidence packages are constructed based on clause ID, text span, table cells, source information, and time limit. The clause-level evidence package includes a unique clause identifier, a text span set, a table cell set, source information, and time limit. The text span set represents the position range of the evidence in the document, and the table cell set includes the table number, row key, and column key. Step 2: Calculate the ternary confidence score. For each evidence package, calculate the retrieval margin, coverage, and generation consistency, and generate a comprehensive confidence score based on the weighted fusion formula. Step 3: Fusion scoring and pre-generation. Calculate the comprehensive score based on the fusion confidence score, and use this score to guide the generation model to output the draft answer; Step 4: Adaptive threshold and re-retrieval / rewrite control. Set a dynamic threshold based on the historical confidence distribution. When the fusion score is lower than the threshold or conflicting evidence is detected, the re-retrieval or rewrite operation is automatically triggered. Step 5: Conflict detection and deweighting strategy. Detect numerical conflicts, modal conflicts, and timeliness conflicts between evidence packets, and deweight or remove conflicting evidence. Step 6: Controlled generation and inline output. In the final generation stage, the clause number and time label are forced to be output in a preset format to achieve a traceable structured generation result.

2. The method according to claim 1, characterized in that, The hybrid retrieval is obtained by weighting and fusing sparse retrieval scores and dense semantic matching scores proportionally, wherein the formula for calculating the comprehensive score is: ； in, These are weighting coefficients; For document Sparse retrieval score; This represents the semantic similarity output by the dense encoder.

3. The method according to claim 1, characterized in that, The fusion score of the ternary confidence level is calculated using the following formula: ； in, To retrieve the margins; For coverage; To generate consistency; This is a time-limited penalty item; to These are learnable weighted parameters.

4. The method according to claim 3, characterized in that, The learnable weighted parameters are optimized through end-to-end training, specifically including ranking supervision and calibration supervision.

5. The method according to claim 1, characterized in that, The dynamic threshold is adaptively adjusted based on the quantiles of the historical fusion distribution and the query complexity, and its calculation formula is as follows: ； in, The confidence level in the historical distribution quantiles; For adjustment factors; This is a query complexity function.

6. The method according to claim 1, characterized in that, The conflict detection includes three categories: numerical conflict, modal conflict, and time-related conflict. The conflict degree is calculated and the evidence with conflict is weighted down. The weighted down weight is used to recalculate the fusion score.

7. The method according to claim 1, characterized in that, The controlled generation uses a structured template, with the clause number and expiration tag output inline after each concluding sentence. The output format follows the following rules: [§ClauseID|eff=StartDate|exp=EndDate]; Wherein, ClauseID represents the unique identifier of the clause; eff=StartDate represents the effective date of the clause; StartDate is the date on which the clause begins to take effect; exp=EndDate represents the expiration date of the clause; EndDate is the date on which the clause ceases to apply; To ensure the verifiability and verifiability of the generated results.

8. The method according to claim 1, characterized in that, The method described supports end-to-end training using manually labeled data and optimizes the interpretability and accuracy of the confidence model through a combination of ranking supervision and calibration supervision.

9. The method according to claim 1, characterized in that, The method further includes a multi-task joint training mechanism, which combines natural language entailment and numerical consistency classification auxiliary tasks to jointly optimize the model, thereby improving cross-modal consistency and generalization ability.

10. The method according to claim 1, characterized in that, The method described is applicable to text generation scenarios where the interpretation of regulations, standards, and policies requires high reliability and traceability, and outputs verifiable answers with clause IDs and timeliness indicators.

Images