A task research and judgment method, system, device and storage medium

Abstract

This invention discloses a task assessment method, system, device, and storage medium, relating to the field of urban operation and governance technology. The method includes: responding to an assessment task instruction, selecting a subset of evidence matching the assessment task instruction from a pre-constructed evidence library; matching an expected output structure according to the task type corresponding to the assessment task instruction; for each conclusion unit, determining evidence units that meet preset evidence constraints from the subset of evidence, and establishing a mapping relationship between the conclusion unit and the corresponding evidence units; and generating a task assessment result based on the mapping relationship and the content of the evidence units in the subset of evidence, the task assessment result including an assessment conclusion text and an evidence chain list. This ensures that the assessment process is supported by reliable evidence, the assessment conclusion is traceable, and the assessment output meets task requirements, satisfying the standardization and refinement needs of task assessment in various scenarios.

Description

Technical Field

[0001] This invention relates to the field of urban operation and governance technology, specifically to a task assessment method, system, device, and storage medium. Background Technology

[0002] In scenarios requiring the analysis and judgment of complex events, such as public governance and emergency decision-making, the information to be processed typically originates from multiple channels, including hotlines, patrols, the internet, and media, and encompasses various modalities such as text, images, audio, and video. These multi-source, heterogeneous data differ significantly in format, descriptive standards, and reliability, posing a significant challenge to effectively aggregating, correlating, and forming credible conclusions.

[0003] Existing technological solutions are insufficient to support the core requirements of verifiability of conclusions and auditability of processes in the aforementioned scenarios. One type of mainstream retrieval enhancement generation technology aims to generate semantically coherent and highly relevant textual answers, but its architecture typically lacks a rigorous and traceable binding mechanism between conclusions and underlying evidence. Therefore, its outputs are difficult to verify item by item, and it cannot systematically check for evidence conflicts or insufficient coverage during the generation process. Another type of electronic evidence preservation technology focuses on ensuring the integrity and tamper-proof nature of the evidence chain, concentrating on post-event static evidence preservation, and does not address the dynamic and unified governance of multi-source evidence, task-oriented intelligent screening, or real-time correlation with the generated conclusions during the assessment process. Summary of the Invention

[0004] The main objective of this invention is to provide a task assessment method, system, device, and storage medium to address issues such as non-targeted evidence screening, chaotic output structure, disconnect between conclusions and evidence, and poor traceability of assessment results in existing task assessment processes. By accurately screening evidence units that match the assessment task from a pre-built evidence library, determining a standardized output structure based on the task type, establishing a mapping relationship between conclusions and evidence, and generating assessment results containing assessment text and a list of evidence chains, this invention ensures that the assessment process is supported by reliable evidence, that assessment conclusions are traceable, and that assessment outputs meet task requirements, thus satisfying the standardization and refinement needs of task assessment in various scenarios.

[0005] To achieve the above objectives, the embodiments of this application provide the following technical solutions: According to a first aspect of the embodiments of this application, a task assessment method is provided, the method comprising: In response to the assessment task instruction, a subset of evidence matching the assessment task instruction is selected from a pre-built evidence library; the subset of evidence includes several evidence units. The expected output structure is matched according to the task type corresponding to the judgment task instruction; the expected output structure includes several conclusion units predefined by the task type. For each conclusion unit, evidence units that meet the preset evidence constraints are determined from the evidence subset, and a mapping relationship is established between the conclusion unit and the corresponding evidence unit. Based on the mapping relationship and the content of the evidence units in the evidence subset, a task assessment result is generated. The task assessment result includes an assessment conclusion text and an evidence chain list. The evidence chain list is used to identify the evidence units referenced by the assessment conclusion text.

[0006] Optionally, the step of filtering a subset of evidence from a pre-built evidence library that matches the assessment task instruction includes: The judgment task instruction is parsed to obtain retrieval constraints, which include at least one of time range, spatial range, permission range, and task type. Based on the retrieval constraints, conditional filtering is performed on the evidence units in the evidence database to obtain a candidate evidence set; Based on the semantic relevance between each evidence unit in the candidate evidence set and the judgment task instruction, the candidate evidence set is subjected to relevance screening to obtain the evidence subset.

[0007] Optionally, the step of performing relevance screening on the candidate evidence set based on the semantic relevance between each evidence unit in the candidate evidence set and the judgment task instruction to obtain the evidence subset includes: Based on the attribute information of each evidence unit, the corresponding evidence weight is determined; wherein the evidence weight is determined based on at least one of the following: source credibility, content completeness, timeliness, and consistency with other evidence units. The evidence units in the candidate evidence set are sorted according to the semantic relevance and the evidence weight. Based on the sorting results and coverage constraints, evidence units are selected from the candidate evidence set to form the evidence subset. The coverage constraints are used to ensure that the evidence subset covers multiple different information dimensions related to the judgment task instruction.

[0008] Optionally, the step of matching the expected output structure according to the task type corresponding to the assessment task instruction includes: Identify the task type corresponding to the assessment task instruction; Call the corresponding structure template according to the task type; Based on the structural template, several conclusion units are determined, and the several conclusion units are combined to form the expected output structure.

[0009] Optionally, for each conclusion unit, determining the evidence units that meet the preset evidence constraints from the subset of evidence, and establishing the mapping relationship between the conclusion unit and the corresponding evidence units, includes: For each conclusion unit, obtain the preset evidence constraints corresponding to the conclusion unit; Based on the preset evidence constraints, matching evidence units are selected from the evidence subset; If a matching evidence unit is found through screening, the matching evidence unit is used as the cited evidence corresponding to the current conclusion unit, and a mapping relationship is established between the current conclusion unit and the cited evidence. If no matching evidence unit is found through screening, the current conclusion unit is marked as insufficient evidence in the mapping relationship.

[0010] Optionally, generating the task assessment result based on the mapping relationship and the content of the evidence units in the evidence subset includes: Based on the mapping relationship and the content of the cited evidence bound to each conclusion unit, a conclusion text fragment corresponding to each conclusion unit is generated; Multiple conclusion text fragments are combined according to the expected output structure to obtain the judgment conclusion text; The evidence chain list is generated based on the cited evidence corresponding to each conclusion text fragment.

[0011] Optionally, the construction process of the pre-built evidence library includes: Acquire multi-source raw data related to task analysis, and add corresponding source description information to each piece of multi-source raw data; Modal translation processing is performed on the non-text data in the multi-source raw data to obtain the corresponding textual representation; The original text data in the multi-source raw data and the textual representation are semantically isomorphic to extract structured semantic information. Based on the structured semantic information and the corresponding source description information, an evidence unit is constructed; The source credibility is determined based on the type of evidence source of the evidence unit, the content completeness is determined based on the completeness of the content fields, and the timeliness information is determined based on the time interval between the time of evidence occurrence and the time of analysis. The evidentiary weight of the evidentiary unit is determined based on the information regarding the credibility of the source, the completeness of the content, and the timeliness. The evidence units with determined evidence weights are stored, and a retrieval index is created for each evidence unit to form the evidence database.

[0012] Optionally, the step of establishing and storing a retrieval index for the evidence unit includes: The text content of the evidence unit is converted into a vector representation, and a vector index is constructed for semantic similarity retrieval. The structured semantic information of the evidence units is used as a filterable field to construct a structured index for conditional filtering retrieval; Determine temporal associations based on the temporal information contained in each evidence unit, spatial associations based on spatial information, subject associations based on subject information, and semantic associations based on semantic content. Based on the determined temporal, spatial, subject, and semantic associations, event merging processing is performed on multiple evidence units to obtain a unified event identifier; The evidence unit is associated with the corresponding event identifier and stored in the evidence database.

[0013] According to a second aspect of the embodiments of this application, a task assessment system is provided, the system comprising: The evidence subset filtering module is used to filter out a subset of evidence that matches the assessment task instruction from a pre-built evidence library in response to the assessment task instruction; the evidence subset includes several evidence units; The output structure determination module is used to match the expected output structure according to the task type corresponding to the judgment task instruction; the expected output structure includes several conclusion units predefined by the task type. The evidence unit determination module is used to determine, for each conclusion unit, evidence units that meet preset evidence constraints from the evidence subset, and to establish a mapping relationship between the conclusion unit and the corresponding evidence unit. The result generation module is used to generate a task assessment result based on the mapping relationship and the content of the evidence units in the evidence subset. The task assessment result includes an assessment conclusion text and an evidence chain list. The evidence chain list is used to identify the evidence units referenced by the assessment conclusion text.

[0014] According to a third aspect of the present application, an electronic device is provided, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the method described in the first aspect above.

[0015] According to a fourth aspect of the embodiments of this application, a computer-readable storage medium is provided having computer-readable instructions stored thereon, the computer-readable instructions being executable by a processor to implement the method described in the first aspect above.

[0016] In summary, this application provides a task assessment method, system, device, and storage medium. Responding to an assessment task instruction, it filters a subset of evidence matching the instruction from a pre-built evidence library. The subset of evidence includes several evidence units. An expected output structure is matched according to the task type corresponding to the assessment task instruction. The expected output structure includes several conclusion units predefined by the task type. For each conclusion unit, evidence units meeting preset evidence constraints are determined from the subset of evidence, and a mapping relationship is established between the conclusion unit and the corresponding evidence unit. Based on the mapping relationship and the content of the evidence units in the subset of evidence, a task assessment result is generated. The task assessment result includes an assessment conclusion text and an evidence chain list, whereby the evidence chain list identifies the evidence units referenced by the assessment conclusion text. By accurately selecting evidence units that match the assessment task from a pre-built evidence library, determining a standardized output structure based on the task type, establishing a mapping relationship between conclusions and evidence, and generating assessment results that include assessment text and a list of evidence chains, the assessment process is supported by reliable evidence, the assessment conclusions are traceable, and the assessment output meets the task requirements, thus satisfying the standardization and refinement needs of task assessment in various scenarios. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or 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 of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0018] The structures, proportions, sizes, etc. illustrated in this specification are only for the purpose of assisting those skilled in the art in understanding and reading the content disclosed herein, and are not intended to limit the conditions under which the present invention can be implemented. Therefore, they have no substantial technical significance. Any modifications to the structure, changes in the proportions, or adjustments to the size, without affecting the effects and objectives that the present invention can produce, should still fall within the scope of the technical content disclosed in the present invention.

[0019] Figure 1 This is a schematic flowchart of a task assessment method provided in an embodiment of this application; Figure 2 This is a schematic diagram of the hierarchical architecture of the task assessment system provided in the embodiments of this application; Figure 3 This is a schematic diagram illustrating the evidence unit construction process provided in an embodiment of this application. Figure 4 This is a schematic diagram of the evidence retrieval and rearrangement process provided in the embodiments of this application; Figure 5 This is a schematic diagram illustrating the binding relationship between the conclusion unit and the evidence unit provided in an embodiment of this application. Figure 6 This is a schematic diagram illustrating the process of verifying the integrity of evidence conflict detection and assessment results in an embodiment of this application. Figure 7 A schematic diagram of a task assessment system provided in an embodiment of this application; Figure 8 This paper shows a structural diagram of an electronic device provided in an embodiment of this application; Figure 9 A diagram of a computer-readable storage medium provided in an embodiment of this application is shown.

[0020] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention 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 the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0022] It should be noted that all directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of the present invention are only used to explain the relative positional relationship and movement of each component in a certain specific posture (as shown in the figure). If the specific posture changes, the directional indication will also change accordingly.

[0023] Furthermore, in this invention, descriptions involving "first," "second," etc., are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0024] In this invention, unless otherwise explicitly specified and limited, the terms "connection," "fixed," etc., should be interpreted broadly. For example, "fixed" can mean a fixed connection, a detachable connection, or an integral part; it can mean a mechanical connection or an electrical connection; it can mean a direct connection or an indirect connection through an intermediate medium; it can mean the internal communication of two components or the interaction between two components, unless otherwise explicitly limited. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0025] Furthermore, the technical solutions of the various embodiments of the present invention can be combined with each other, but only if they are feasible for those skilled in the art. If the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0026] Figure 1 This application illustrates a task assessment method provided by an embodiment of the present application, the method comprising: Step 101: In response to the assessment task instruction, select a subset of evidence that matches the assessment task instruction from the pre-built evidence library; the subset of evidence includes several evidence units; Step 102: Match the expected output structure according to the task type corresponding to the judgment task instruction; the expected output structure includes several conclusion units predefined by the task type; Step 103: For each conclusion unit, determine the evidence units that meet the preset evidence constraints from the evidence subset, and establish a mapping relationship between the conclusion unit and the corresponding evidence unit; Step 104: Based on the mapping relationship and the content of the evidence units in the evidence subset, generate a task assessment result. The task assessment result includes an assessment conclusion text and an evidence chain list. The evidence chain list is used to identify the evidence units referenced by the assessment conclusion text.

[0027] In practical assessment scenarios, traditional methods often suffer from a lack of targeted evidence screening, with a large amount of irrelevant evidence consuming assessment resources, leading to low assessment efficiency. Furthermore, the lack of unified output standards for different types of assessment tasks results in disorganized assessment results that are difficult to review and use later. In addition, the lack of a clear correspondence between conclusions and evidence makes it difficult to verify the rationality of the conclusions, and the assessment process lacks traceability, making it impossible to quickly locate the root cause of problems. This application provides a standardized and refined task assessment method that solves the problems of low efficiency, non-standardized output, insufficient credibility of conclusions, and poor traceability in existing task assessment processes, ensuring that the assessment process is orderly and controllable and the assessment results are accurate and reliable.

[0028] This method first responds to the assessment task instruction and filters out a subset of evidence matching the instruction from a pre-built evidence library. This subset consists of several evidence units; precise filtering eliminates irrelevant evidence, reduces unnecessary workload, and improves assessment efficiency. The pre-built evidence library ensures the standardization and accessibility of evidence sources. Different types of assessment tasks have different requirements for output results; therefore, it is necessary to match the expected output structure according to the task type corresponding to the assessment task instruction. This expected output structure contains several conclusion units predefined by the task type. By pre-setting conclusion units, the content and format of the assessment output can be standardized, ensuring the consistency and relevance of assessment results for different types of tasks and meeting the assessment needs in different scenarios.

[0029] For each predefined conclusion unit, it is necessary to determine the evidence units that meet the preset evidence constraints from the selected evidence subset, and establish a mapping relationship between the conclusion unit and the corresponding evidence unit. The preset evidence constraints can ensure that the evidence used to support the conclusion meets the requirements and avoid unqualified evidence from affecting the credibility of the conclusion. The clear mapping relationship can realize a one-to-one correspondence between the conclusion and the evidence, so that each judgment conclusion has solid evidence support.

[0030] Based on this, and using the established mapping relationship and the content of the evidence units in the evidence subset, a task assessment result is generated, which includes the assessment conclusion text and the evidence chain list. The evidence chain list is used to identify the evidence units cited by the assessment conclusion text. The evidence chain list can clearly present the evidence source of the assessment conclusion, realize the traceability of the assessment process and results, and facilitate subsequent verification, validation and review of the assessment results, thereby improving the efficiency, standardization, credibility and traceability of task assessment.

[0031] In one possible implementation, in step 101, the step of selecting a subset of evidence matching the assessment task instruction from a pre-built evidence library includes: parsing the assessment task instruction to obtain retrieval constraints, the retrieval constraints including at least one of time range, spatial range, permission range, and task type; performing condition filtering on the evidence units in the evidence library based on the retrieval constraints to obtain a candidate evidence set; and performing relevance filtering on the candidate evidence set based on the semantic relevance between each evidence unit in the candidate evidence set and the assessment task instruction to obtain the evidence subset.

[0032] In step 101, a subset of evidence matching the assessment task instructions is selected from a pre-built evidence library. First, the assessment task instructions are parsed to extract retrieval constraints. These constraints can include at least one of time range, spatial range, authority range, and task type. By clearly defining these constraints, the scope of evidence required for assessment can be precisely defined, avoiding the selection of irrelevant evidence units and establishing clear boundaries for subsequent evidence screening. Then, based on the parsed retrieval constraints, a conditional filtering operation is performed on all evidence units in the pre-built evidence library. Evidence units that do not meet the retrieval constraints are eliminated, leaving only those that do meet the conditions to form a candidate evidence set. This quickly narrows the evidence screening scope, reduces the workload of subsequent relevance screening, and improves screening efficiency.

[0033] In one possible implementation, the step of filtering the candidate evidence set based on the semantic relevance between each evidence unit in the candidate evidence set and the judgment task instruction to obtain the evidence subset includes: determining the corresponding evidence weight based on the attribute information of each evidence unit; wherein the evidence weight is determined based on at least one of the following: source credibility, content completeness, timeliness, and consistency with other evidence units; ranking the evidence units in the candidate evidence set according to the semantic relevance and the evidence weight; and selecting evidence units from the candidate evidence set to form the evidence subset according to the ranking result and coverage constraints, wherein the coverage constraints are used to ensure that the evidence subset covers multiple different information dimensions related to the judgment task instruction.

[0034] In one possible implementation, determining the corresponding evidence weight based on the attribute information of each evidence unit includes: setting preset weight coefficients for source credibility, content completeness, timeliness, and consistency with other evidence units; and performing weighted calculations based on the specific values ​​of each evidence unit in the above four dimensions, combined with the corresponding preset weight coefficients, to obtain the evidence weight of the evidence unit.

[0035] When selecting a subset of evidence based on the semantic relevance between each piece of evidence in the candidate evidence set and the assessment task instructions, the specific process is as follows: First, determine the evidence weight for each piece of evidence based on its own attribute information. The determination of the evidence weight must consider at least one factor among the following: source credibility, content completeness, timeliness, and consistency with other evidence pieces. The evidence weight quantifies the quality and reliability of each piece of evidence, providing a basis for subsequent ranking and selection. Then, combining the semantic relevance between each piece of evidence and the assessment task instructions, as well as the determined evidence weights, comprehensively rank all evidence pieces in the candidate evidence set, prioritizing those with high quality and strong relevance to the task instructions. Finally, based on the ranking results and preset coverage constraints, select suitable evidence pieces from the candidate evidence set to form an evidence subset. The coverage constraint ensures that the selected evidence subset covers multiple different information dimensions related to the assessment task instructions, avoiding the problem of information partiality in the evidence subset.

[0036] In one possible implementation, in step 102, matching the expected output structure according to the task type corresponding to the assessment task instruction includes: identifying the task type corresponding to the assessment task instruction; calling the corresponding structure template according to the task type; determining several conclusion units based on the structure template; and combining the several conclusion units to form the expected output structure.

[0037] Step 102 involves matching the expected output structure based on the task type corresponding to the assessment task instruction. This is achieved through the following steps: First, the assessment task instruction is analyzed to identify the corresponding task type. Different assessment tasks have different core requirements and output requirements. Accurately identifying the task type is the foundation for matching a suitable output structure, ensuring that the subsequent output structure closely matches the task requirements. Then, based on the identified task type, a pre-stored structure template corresponding to that task type is invoked. These templates are pre-defined according to the assessment standards and output specifications for various tasks, ensuring the standardization and uniformity of the assessment output and preventing output structure confusion due to operational differences among different assessors. Finally, based on the invoked structure template, several conclusion units required for this task type are determined. These conclusion units are the core judgment points that the assessment task needs to achieve. The determined conclusion units are combined according to the preset logic of the structure template to form the expected output structure that matches the current assessment task instruction.

[0038] In one possible implementation, in step 103, determining evidence units that meet preset evidence constraints from the evidence subset for each conclusion unit and establishing a mapping relationship between the conclusion unit and the corresponding evidence unit includes: for each conclusion unit, obtaining preset evidence constraints corresponding to the conclusion unit, the preset evidence constraints including at least one of evidence quantity requirements, evidence source type requirements, or evidence content slot requirements; filtering matching evidence units in the evidence subset according to the preset evidence constraints; if a matching evidence unit is found, then the matching evidence unit is used as the cited evidence corresponding to the current conclusion unit, and a mapping relationship is established between the current conclusion unit and the cited evidence; if no matching evidence unit is found, then the current conclusion unit is marked as having insufficient evidence in the mapping relationship.

[0039] In one possible implementation, the evidence content slot requirements in the preset evidence constraints include: for each conclusion unit, predefining the core semantic elements that the evidence required for that conclusion unit should contain, wherein the core semantic elements include at least one of the event subject, time of occurrence, location of occurrence, and core behavior; when filtering matching evidence units, it is necessary to ensure that the evidence unit contains all the predefined core semantic elements.

[0040] In one possible implementation, step 103 involves identifying evidence units that meet preset evidence constraints from the evidence subset for each conclusion unit and establishing a mapping relationship. The specific implementation process is as follows: First, for each conclusion unit, obtain its corresponding preset evidence constraints. These constraints include at least one of the following: evidence quantity requirements, evidence source type requirements, and evidence content slot requirements, ensuring that the selected evidence effectively supports the accuracy of the conclusion. Further, based on the preset evidence constraints, further filtering is performed on the already selected evidence subset to accurately match evidence units suitable for the current conclusion unit, excluding evidence that does not meet the constraints, thus avoiding invalid evidence affecting the judgment results. If a qualified evidence unit can be selected, it is used as the cited evidence corresponding to the current conclusion unit. Simultaneously, a mapping relationship is established between the conclusion unit and the evidence unit, clarifying their correspondence and ensuring that each conclusion has corresponding evidence support. If, after filtering, no evidence unit meeting the preset constraints is found, the current conclusion unit is marked as lacking sufficient evidence in the established mapping relationship for subsequent manual verification and supplementation, ensuring the rigor and reliability of the judgment results and avoiding conclusions without evidence support.

[0041] In one possible implementation, in step 104, generating the task assessment result based on the mapping relationship and the content of the evidence units in the evidence subset includes: generating a conclusion text fragment corresponding to each conclusion unit based on the mapping relationship and the content of the cited evidence bound to each conclusion unit; combining multiple conclusion text fragments according to the expected output structure to obtain the assessment conclusion text; and generating the evidence chain list according to the cited evidence corresponding to each conclusion text fragment.

[0042] In one possible implementation, generating the evidence chain list includes: assigning a unique identifier to each cited evidence unit; associating and recording the unique identifier of each paragraph and conclusion in the judgment conclusion text with the corresponding cited evidence unit in the evidence chain list; and supplementing the evidence chain list with source description information, evidence weight, and core semantic elements of each cited evidence unit to facilitate subsequent verification.

[0043] Step 104 generates the task assessment results based on the mapping relationship and the content of the evidence units in the evidence subset. The specific implementation process is as follows: First, according to the previously established mapping relationship, the cited evidence bound to each conclusion unit is extracted. Combining the specific content of the cited evidence, a conclusion text fragment corresponding to each conclusion unit is generated. This ensures that the conclusion text fragment accurately reflects the information contained in the cited evidence, achieving a one-to-one correspondence between the conclusion and the evidence, logical consistency, and avoiding statements without supporting evidence. Then, according to the previously matched expected output structure, all conclusion text fragments are integrated and combined, clarifying the logical connections between each fragment to form a complete and coherent assessment conclusion text. This ensures that the assessment conclusion text is well-written, clear, and meets the preset output requirements. Finally, based on the cited evidence corresponding to each conclusion text fragment, an evidence chain list is generated. The list clearly records the cited evidence information corresponding to each conclusion text fragment, clearly presenting the connection between the assessment conclusion and the evidence. This facilitates subsequent verification and tracing of the assessment results, ensuring that the entire assessment process is traceable and verifiable, and guaranteeing the rigor and credibility of the assessment results.

[0044] In one possible implementation, the construction process of the pre-built evidence library includes: acquiring multi-source raw data related to task assessment and adding corresponding source description information to each piece of multi-source raw data; performing modal translation processing on the non-text data in the multi-source raw data to obtain the corresponding textual representation; performing semantic isomorphic processing on the original text data in the multi-source raw data and the textual representation to extract structured semantic information; constructing evidence units based on the structured semantic information and the corresponding source description information; determining the source credibility based on the evidence source type of the evidence unit, determining the content completeness based on the completeness of the content fields, and determining the timeliness information based on the time interval between the evidence occurrence time and the assessment time; determining the evidence weight of the evidence unit based on the source credibility, content completeness, and timeliness information; storing the evidence units after determining the evidence weight, and establishing a retrieval index for the evidence units to form the evidence library.

[0045] In one possible implementation, the pre-built evidence base is constructed through the following steps: First, acquire multi-source raw data relevant to the task assessment, and add corresponding source description information to each piece of multi-source raw data to clarify the source channel of the data and ensure subsequent traceability. For non-text data in the multi-source raw data, perform modal translation processing to transform it into a textual representation that can be used for assessment, ensuring that different types of data can participate in subsequent processing in a unified form. Next, perform semantic isomorphism processing on the raw text data in the multi-source raw data and the translated textual representation, and extract the structured semantic information through a unified semantic standard to clarify the core content of each data. Subsequently, based on the extracted structured semantic information and the corresponding source description information, construct standardized evidence units to ensure that each evidence unit has clear core information. On this basis, determine the source credibility of the evidence unit based on the source type, determine its content completeness based on the completeness of the content fields of the evidence unit, and determine its timeliness based on the interval between the occurrence time recorded by the evidence unit and the current assessment time. Combine the information from these three dimensions to determine the evidence weight of each evidence unit. Finally, all evidence units with determined evidence weights are stored uniformly, and a corresponding retrieval index is created for each evidence unit. Through standardized storage and index settings, a complete evidence library is formed.

[0046] In one possible implementation, the modal translation processing of non-text data in the multi-source raw data includes: if the non-text data is image data, performing optical character recognition and scene feature extraction on the image data to obtain a textual representation containing text information and scene description in the image; if the non-text data is audio data, performing speech transcription and semantic extraction on the audio data to obtain a textual representation corresponding to the audio content; if the non-text data is video data, extracting video keyframes and performing image translation on the keyframes, combining the video and audio transcription content to obtain a complete textual representation.

[0047] In one possible implementation, the semantic isomorphic processing of the original text data in the multi-source original data and the textual representation includes: performing unified word segmentation, synonym normalization and sensitive information desensitization on the original text data and the textual representation; extracting structured semantic information in a unified format from the processed text according to a preset semantic element template to ensure that the structured semantic information format corresponding to different types of original data is consistent.

[0048] In one possible implementation, the step of establishing and storing a retrieval index for the evidence units includes: converting the text content of the evidence units into vector representations to construct a vector index for semantic similarity retrieval; using the structured semantic information of the evidence units as filterable fields to construct a structured index for conditional filtering retrieval; determining temporal associations based on the time information contained in each evidence unit, spatial associations based on the spatial information, subject associations based on the subject information, and semantic associations based on the semantic content; performing event merging processing on multiple evidence units based on the determined temporal associations, spatial associations, subject associations, and semantic associations to obtain a unified event identifier; and storing the evidence units and their corresponding event identifiers in the evidence database.

[0049] In one possible implementation, the event merging process for the evidence units includes: setting a time threshold, a spatial threshold, a subject similarity threshold, and a semantic similarity threshold; if the time association of multiple evidence units satisfies the time threshold, the spatial association satisfies the spatial threshold, the subject association satisfies the subject similarity threshold, and the semantic association satisfies the semantic similarity threshold, then the multiple evidence units are determined to belong to the same event; a unified event identifier is assigned to all evidence units of the same event to achieve event-based classification and storage of evidence units.

[0050] In one possible implementation, the specific process for building and storing retrieval indexes for evidence units is as follows: First, the text content of each evidence unit is converted into a vector representation. This vector representation is then used to construct a vector index for semantic similarity retrieval, enabling rapid identification of the corresponding evidence unit through semantic matching. Simultaneously, the structured semantic information of each evidence unit is extracted as a filterable field, constructing a structured index for conditional filtering retrieval, facilitating the subsequent selection of target evidence units based on specific needs. Based on this, temporal associations are determined by combining time information, spatial associations by combining spatial information, subject associations by combining subject information, and semantic associations by combining semantic content, comprehensively clarifying the relationships between each evidence unit. According to the determined temporal, spatial, subject, and semantic associations, event merging is performed on multiple evidence units, assigning a unified event identifier to all evidence units belonging to the same event, achieving classified management of evidence units. Finally, each evidence unit is associated with its corresponding event identifier and stored in the evidence repository along with the previously constructed vector and structured indexes. Simultaneously, the relevant attribute information of each evidence unit is recorded to ensure rapid location of the required evidence through various retrieval methods, improving the efficiency and accuracy of evidence retrieval.

[0051] In one possible implementation, after generating the task assessment result, the method further includes: performing semantic verification on the assessment conclusion text to determine whether the conclusion text has logical contradictions, ambiguous expressions, or inconsistencies with the cited evidence; if the above situations exist, generating verification prompt information and marking the corresponding conclusion text fragments and cited evidence units for manual review and correction; if the verification passes, associating and archiving the task assessment result, mapping relationship, and related evidence units for subsequent traceability and query.

[0052] In one possible implementation, the semantic verification of the judgment conclusion text includes: semantically comparing the judgment conclusion text with the content of the corresponding cited evidence unit to verify the consistency between the expression of the conclusion text and the evidence content; performing logical association verification on the conclusion text fragments corresponding to different conclusion units in the judgment conclusion text to check for any contradictory expressions; and verifying the standardization of the expression of the conclusion text to ensure that there are no ambiguous, ambiguous, or redundant expressions and that they meet the preset text standardization requirements.

[0053] The present application will now be described in further detail with reference to specific embodiments and accompanying drawings.

[0054] This application provides a method for generating semantically isomorphic and verifiable evidence chains from multiple sources for social risk early warning scenarios. It aims to solve a core engineering problem that has long existed in social risk early warning operations: how to stably organize scattered clues into credible evidence under a unified event caliber in scenarios where multiple sources of data are continuously flowing in, the sources have large differences in caliber, the modal types are highly heterogeneous, the permission boundaries are complex, and the conclusions need to be reviewed afterward. Furthermore, it aims to output risk assessment results that can explain the basis for each item, trace the source of each item, and support audit verification as a whole.

[0055] This method uses evidence units as the smallest governance granularity, forming a closed loop in data access, modality translation, semantic isomorphism, evidence selection and retrieval, constraint generation, integrity verification, and version auditing. This ensures that the final output is not merely a simple risk conclusion text, but a verifiable delivery containing conclusions, evidence numbers, source pointers, hash digests, and version information, thereby improving judgment efficiency, verification capabilities, and credible communication in governance scenarios. This method does not simply feed multi-source data into a large model to directly generate risk conclusions. Instead, it first converts data from different sources, modalities, and permission levels into computable, traceable, and verifiable evidence units. Then, it merges, retrieves, generates constraints, and encapsulates integrity around the same event, ultimately forming a verifiable evidence chain output where the conclusion text corresponds one-to-one with the evidence number, source pointer, and verification information. This embodiment covers both the method flow and the corresponding system device. Modules can be deployed on the same computing node or separately. Figure 2 The diagram shows a breakdown into a data governance layer, an evidence retrieval layer, and a constraint generation and audit traceability layer, which are sequentially connected via internal APIs, message queues, or service buses. The output of each layer serves as the input for the next, and each subsequent layer retains an index pointer and hash fingerprint of the processing results from the previous layer. Therefore, it is not a process of generating data first and then searching for evidence afterward, but rather of generating evidence first and then generating conclusions based on the constraints imposed by that evidence.

[0056] Appendix Figure 2 This is a schematic diagram of the hierarchical architecture of the task assessment system provided in this application embodiment. The architecture is divided into a data governance layer, an evidence retrieval layer, a constraint generation layer, and an audit traceability layer from top to bottom. Data transmission and functional linkage between each layer are achieved through internal interfaces, message queues, or service buses. The processing result of the previous layer serves as the input data for the next layer, and the next layer retains the index pointer and hash fingerprint of the processing result of the previous layer, forming a complete technical processing closed loop.

[0057] The data governance layer includes a data access module, a modality translation module, a semantic isomorphism mapping module, an evidence quality assessment module, and an evidence library construction and maintenance module. It primarily handles the access, standardization, evidence unit construction, and evidence library management of multi-source heterogeneous data. The evidence retrieval layer includes a query parsing module and an evidence selection and retrieval module. These modules parse the judgment task instructions and, based on the parsing results, select and sort evidence from the evidence library. The constraint generation layer includes a conclusion unit construction module, an evidence binding module, and an evidence chain generation module. These modules are responsible for structurally decomposing the judgment output, matching each conclusion unit with evidence that meets the constraint conditions, and establishing corresponding relationships. The audit traceability layer includes an integrity verification and review module, a version management module, and a result output interface module. This enables integrity verification, version retention, and standardized output of the judgment results, ensuring the entire judgment process is traceable and verifiable.

[0058] At the data governance layer, the data access module first receives multi-source data related to social risk events. This multi-source data can include structured records from government business systems, hotline work orders, inspection logs, news texts, social media texts, images, audio, video keyframes, and sensor logs. To prevent subsequent evidence distortion, the system writes source description information for each original record upon access, including at least one or more of the following: source type, collection time, collection channel, collection entity, geolocation marker, permission marker, and confidentiality level marker. It also generates an original reference pointer URI to lay the foundation for subsequent source tracing and access control.

[0059] For non-textual modalities, the system uses a modality translation module to convert them into manageable textual representations, rather than directly treating the original images or audio as uninterpretable black-box input. (Appendix) Figure 3 This is a schematic diagram of the evidence unit (EU) construction process provided in this application embodiment. The process takes multi-source heterogeneous data as input, and the input data covers various modalities such as text, images, audio, video keyframes, and logs. Each step is connected sequentially to finally complete the construction and storage of standardized evidence units, providing basic support for subsequent risk assessment.

[0060] The process begins with the input stage, receiving various types of raw, multi-source data. This is followed by the data access stage, where the system generates source description information for each piece of raw data, clearly defining the data's source channel, collection time, access permissions, and other traceability attributes, laying the foundation for traceability throughout the entire process.

[0061] Next, the multi-source raw data is processed separately. Raw text data proceeds directly to text normalization, while non-text data such as images, audio, and video undergo modal translation processing sequentially to convert them into computationally computable textual representations. Specifically, image data undergoes at least one or more of the following processes: OCR recognition, object detection, and scene description generation, outputting image-translated text and corresponding confidence scores. Audio data undergoes at least speech recognition and transcription, with speaker separation, emotion recognition, and noise suppression performed as needed. Video data first has keyframes extracted, then undergoes image translation and time segment marking. The purpose of this processing is to uniformly convert visible and audible modal content that is inconvenient to directly compute into normalized text that can be extracted for slots, used in retrieval, and used for evidence binding.

[0062] The modally translated text and the original text enter the semantic isomorphism mapping module. This module first performs text normalization, including removing invalid symbols, standardizing time formats, standardizing place names and aliases, standardizing subject references, normalizing synonyms, and desensitizing sensitive fields. Then, contextual segmentation is performed. To avoid mechanically cutting off an event description, this implementation adopts a sliding window segmentation strategy with overlapping areas. The window length is L, the step size is S, and S is less than L. With this strategy, even if the same factual element falls near paragraph boundaries, sufficient context can be preserved in adjacent segments, thereby reducing evidence fragmentation and extraction bias.

[0063] After segmentation, the semantic isomorphic mapping module extracts slot information such as time, location, subject, object, behavior, demand, emotion, result, and scope of influence from each segment, and encapsulates the original segment, translated text, and structured slots into a unified evidence unit (EU). Each evidence unit contains at least eu_id, event_id, source_type, modality_type, t_collect, t_occurrence, geo_code, actor, target, action, sentiment, privacy_tag, permission_tag, content_hash, and uri. The event_id is not manually filled in statically, but is automatically generated or updated by an event merging algorithm based on spatiotemporal proximity, subject overlap, topic similarity, and propagation chain relationships. This is used to merge multi-source clues under a unified event caliber while retaining relationship markers such as same source, different source, and revised version.

[0064] After the evidence unit is formed, the evidence quality assessment module calculates the evidence weight for each EU. This weight is not a single confidence level, but a result of integrating source confidence, modality translation confidence, timeliness, cross-source consistency, and completeness.

[0065] In one possible implementation, the weight of evidence is calculated according to the following formula:

[0066] in, Indicates the credibility of the source. Indicates the confidence level of modal translation. This indicates cross-source consistency with other evidence related to the same event. Indicates field completeness. Indicates timeliness.

[0067] Timeliness can be calculated using a time decay function, for example:

[0068] Where Δt is the time difference between the time the evidence occurred and the current assessment time, and γ is the attenuation coefficient. This setting allows newer evidence to receive higher weight when other conditions are similar, but if the source is unreliable or there is a severe shortage of slots, the system will not blindly increase its value just because it is new.

[0069] The final stage of the process is event merging and database entry. The evidence database construction and maintenance module writes the aforementioned evidence units into the evidence content storage area, vector index area, structured index area, and source tracing index area, respectively. Standardized text and slot structures enter the evidence content storage area, vector representations enter the vector index area, filterable fields such as time, location, subject, and event type enter the structured index area, and fingerprint hashes, parent version identifiers, change times, and change reasons enter the source tracing index area. To support incremental updates, the system generates a fingerprint hash h_e based on the standardized text and key metadata. When new evidence enters, if the same hash exists, it is determined to be duplicate evidence, and the reference relationship is updated; if the hashes are similar and the slots are highly similar, it is written as a revised version, and the parent version relationship is recorded. This avoids duplicate database entry and preserves the evidence evolution path.

[0070] After the evidence units are constructed and stored, when a risk assessment task is received, the query and parsing module of the evidence retrieval layer first identifies the task type and constraints. For example, the input might be whether there is a trend of escalation of mass incidents in a certain area today, or it might be asking for handling suggestions and explanations of the basis. Based on this, the system extracts the time window, spatial scope, risk type, permission scope, and output template requirements, and sends these conditions to the evidence selection and retrieval module. Evidence selection and retrieval is not simply searching for similar text, but rather performing four actions in sequence: permission filtering, coarse recall, fine reordering, and constraint-based evidence selection, providing reliable evidence support for the subsequent assessment results.

[0071] Appendix Figure 4This is a schematic diagram of the evidence retrieval and rearrangement process provided in an embodiment of this application. After receiving the retrieval conditions related to the assessment task, the system sequentially performs coarse recall and fine rearrangement operations. In the coarse recall stage, the system first encodes the query into a vector and performs an approximate nearest neighbor search in the vector index area to obtain a candidate set. Simultaneously, the structured index area is filtered according to time, location, subject, event type, permission tags, etc., to obtain a collection. The candidate set E is obtained by intersecting or weighting the two sets. During this process, semantic similarity can be calculated using cosine similarity.

[0072] in, For query vector, This is the evidence unit vector. This similarity is mainly used to ensure recall coverage and does not directly determine the final selected evidence.

[0073] After entering the fine rearrangement stage, the system processes each candidate piece of evidence. Calculate the relevance score and the aforementioned weight of evidence By combining the results, a comprehensive score is obtained:

[0074] If sorting is based solely on the score, issues may arise such as evidence originating from the same source, describing only the same point in time, or presenting similar text but conflicting facts. Therefore, this application further incorporates coverage and consistency constraints, requiring that the final selected subset of evidence not only have high scores but also cover key slots and ideally come from different sources, modalities, or time slices.

[0075] In one possible implementation, the certificate selection and retrieval module solves the corresponding objective function while satisfying the constraints:

[0076] in, This indicates consistency within the set of evidence. This represents the conflict penalty term, where λ and μ are configurable parameters. Through this mechanism, the system prevents a single popular piece of information from monopolizing the basis for its conclusions simply because it scores highly, thus ultimately forming a subset of evidence for subsequent analysis.

[0077] In forming a subset of evidence Subsequently, the constraint generation and audit traceability layer does not directly generate the entire report. Instead, it first breaks down the target output into several conclusion units (CUs), such as risk level judgment, main causes, key entities, spatiotemporal evolution trends, and disposal recommendations. Each conclusion unit is pre-configured with evidence slot requirements. For example, risk level judgment requires at least three types of evidence elements—time, location, and entity—to be covered simultaneously, and disposal recommendations require at least one piece of highly credible evidence from an official government source and one piece of mutually corroborating evidence from a publicly available source. The technical significance of this approach is that it first structures the generation task, giving subsequent evidence binding and conflict verification a clear target, rather than making vague judgments about an entire text.

[0078] Appendix Figure 5 This diagram illustrates the binding relationship between conclusion units and evidence units provided in this application embodiment. It shows the complete binding logic from the decomposition of conclusion units to the generation of the evidence chain list. First, multiple independent conclusion units are obtained by decomposing them according to the type of assessment task. In the example, these include conclusion unit CU1 corresponding to risk level judgment, conclusion unit CU2 corresponding to main cause analysis, and conclusion unit CU3 corresponding to diffusion trend and treatment recommendations. Each conclusion unit corresponds to a core judgment content in the assessment result. The middle part is the binding relationship M, which is used to establish the mapping relationship between conclusion units and evidence units. At the same time, coverage calculation, support calculation, and conflict calculation are performed simultaneously during the binding process to provide quantitative basis for the subsequent generation of assessment results.

[0079] Further, multiple evidence units are bound to the conclusion unit. In the example, evidence units e1, e2, e3, and e4 are included. Each evidence unit carries a unique identifier eu_id, a content hash, and corresponding core attribute information. For example, evidence unit e1 contains basic factual elements such as time, location, subject, and behavior; evidence unit e2 contains supplementary information such as demands, emotions, and source tags; evidence unit e3 is marked with a high credibility attribute from government sources; and evidence unit e4 is marked with a supplementary attribute of mutual verification from publicly available sources. Through the binding relationship M, each conclusion unit corresponds to a set of evidence units that meet the preset evidence constraints. Finally, the evidence chain list on the right is generated based on the binding relationship. The evidence chain list includes at least the statement ID, the text of the conclusion fragment, the binding eu_id set, the support score, the conflict score, the source hash list, and metadata of the generation process such as model version, timestamp, and parameters. This ensures a one-to-one correspondence between each judgment conclusion and the corresponding evidence, guaranteeing that the conclusions are traceable and verifiable.

[0080] The evidence binding module selects a set of evidence units that satisfy the slot constraints from E* for each conclusion unit CU and establishes a mapping relationship. The system also calculates evidence coverage:

[0081] in, This represents the number of conclusion units that have been successfully bound. This represents the total number of conclusion units. If the binding evidence for a conclusion unit is insufficient, or if key slots are still missing after binding, the system will not allow the model to operate freely. Instead, it will directly output a standardized statement indicating insufficient evidence requiring verification and generate supplementary evidence suggestions. For conclusion units that have already been bound, the system will also perform conflict detection. Conflict detection can be based on structured slot rules, such as whether there are contradictions between time, location, number of people, and subject identities, or it can be based on natural language inference models to calculate the probability of support and the probability of contradiction.

[0082] Appendix Figure 6 This is a schematic diagram of the evidence conflict detection and judgment result integrity verification process provided in the embodiments of this application. After the binding of the conclusion unit and the evidence unit is completed, the system performs conflict detection on the evidence set corresponding to each conclusion unit. The conflict detection methods include rule comparison based on structured slots and semantic consistency judgment based on natural language reasoning, which identifies contradictory content between evidence in terms of time, place, subject, behavior, etc.

[0083] This process takes the bound evidence set E_CU for a single conclusion unit as input. First, it performs a conflict detection step, identifying factual contradictions and semantic conflicts between bound evidence through element consistency comparison and natural language inference implicature judgment. After conflict detection, a conflict score is calculated and compared with a preset threshold τ. If the conflict score is greater than the threshold τ, the evidence is considered highly conflicting, and the process enters the manual review branch, marking it as requiring manual review and outputting supplementary evidence collection suggestions, without directly generating a definitive conclusion. If the conflict score is not greater than the threshold τ, the process enters the coverage check step. In the coverage check step, the evidence coverage of the current conclusion unit is calculated and compared with a preset compliance threshold η. If the coverage is less than the threshold η, the evidence is considered insufficient, and the process enters the fallback output branch, outputting a standardized statement indicating insufficient evidence requiring verification and generating a supplementary evidence collection list. If the coverage is not less than the threshold η, the evidence meets the requirements, and the process enters the normal output step, outputting the conclusion and the corresponding evidence chain eu_id list. After normal output is completed, the integrity verification and encapsulation steps are performed. The specific process is as follows: recalculate the fingerprint hash of each evidence unit (EU), generate a Merkle tree based on the hash value and obtain the root hash Merkle Root, perform signature and version encapsulation on the root hash, and finally output the verifiable final result to ensure that the evidence chain is not tampered with during transmission and archiving, and realize the full auditability and traceability of the judgment results.

[0084] When there is a high degree of conflict among the evidence corresponding to a certain conclusion unit, the system marks the conclusion as requiring manual review and explicitly lists the conflicting evidence, rather than silently discarding it in the background. For generative models, this constraint means that every factual judgment in its output must reference the already bound eu_id, and factual extensions of unbound evidence are prohibited from being output, thereby suppressing the generation of illusions.

[0085] For evidence sets without obvious conflicts, the system generates a conclusion text and a corresponding list of evidence chains based on the binding relationships. The evidence chain generation module then generates the final conclusion text and simultaneously forms the evidence chain list. The list includes at least the conclusion fragment identifier, the conclusion fragment text, the corresponding set of evidence identifiers, support score, conflict score, evidence weight summary, source hash list, and metadata about the generation process. The metadata preferably includes the model version, prompt template version, evidence selection parameters, call timestamp, and operator identifier, ensuring that the final deliverable is a structured assessment result that can be traced sentence by sentence and verified line by line.

[0086] To prevent tampering with the chain of evidence during output, transmission, or archiving, the integrity verification and review module performs verifiable encapsulation of the evidence chain list. Specifically, it first recalculates the fingerprint hash of each EU in the evidence chain list, then serializes the evidence chain list and constructs a Merkle tree to obtain the root hash. ; and then if necessary Perform digital signature to obtain During third-party review, only the evidence content, hash path, and signature information are needed to verify whether the evidence chain has been tampered with, who produced it, and which version it corresponds to.

[0087] The version management module further records the version number of the evidence set, the root hash of the evidence chain, and the version number of the conclusion used in each assessment. When new evidence is added, corrective evidence is added, or permissions are changed, resulting in an updated conclusion, the system does not overwrite the old results but retains the historical version. This allows reviewers to answer why they made that judgment at the time and why they changed their judgment later. This is of practical significance for accountability, review, and cross-departmental collaboration in social risk governance scenarios.

[0088] The following example illustrates the situation. In a certain administrative region, multiple complaints about gatherings and online dissemination of information occurred within 24 consecutive hours. The data access module received 18 work orders from the 12345 hotline, 6 grid patrol records, 4 public news articles, 27 social media texts, 11 images, and 9 sets of short video keyframes. The system first writes source description information and permission tags for each original record; then, the modal translation module transcribes the banner text, location markers, and gathering information in the images into text, extracts the video keyframes into combined fragments of time points, scene descriptions, and OCR text, and uniformly translates the audio and video content into manageable text.

[0089] The semantic isomorphic mapping module further segments and extracts slots from the above content to form a batch of evidence units. For example, a work order records that residents of a certain street reported an abnormal water supply at 09:10 on March 10; an image was extracted by OCR to reveal the text of a banner requesting the restoration of water supply and the location of the photo was identified; a news report states that relevant departments have initiated emergency response; and a social media post reflects the expansion of the crowd gathering at the scene. Based on spatiotemporal proximity and subject overlap, the system merges these pieces of evidence under the same event_id.

[0090] Next, the evidence quality assessment module calculates the weight of each piece of evidence, lowering the score for evidence with unclear sources, serious content deficiencies, or significant conflicts with other evidence. After the assessors input the task of whether the event has an escalation trend, what the main triggers are, and whether immediate joint action is needed, the evidence selection and retrieval module selects the final subset of evidence from the candidate set, provided that it contains at least one piece of evidence from an official source and one piece of evidence from a publicly available source.

[0091] The constraint generation layer breaks down the output into four conclusion units: risk level judgment, cause analysis, diffusion trend, and disposal recommendations, and binds each to a corresponding eu_id. If the diffusion trend unit is only supported by online posts without on-site or government evidence to corroborate it, the system directly marks the unit as lacking sufficient evidence; if the cause analysis unit shows two competing causes, namely abnormal water supply and construction disturbance, the system outputs a conflict warning and requires manual review.

[0092] Ultimately, the output interface module does not produce a regular report, but rather an analytical text with evidence numbers, along with an evidence chain list, root hash, signature value, and version number. Analysts can click on any conclusion fragment in the interface to directly view its associated EU content, source pointer, and original record location. They can also export a verifiable list with one click for archiving, transfer, or auditing. Thus, from multi-source data access, evidence unit construction, evidence selection generation to integrity verification, a closed-loop, engineering-ready complete technical solution has been formed.

[0093] It should be noted that the above embodiments are only one of the preferred embodiments of the present invention. Any equivalent substitutions or modifications made by those skilled in the art to the evidence slot set, weight calculation method, retrieval model, conflict detection model, Merkle tree implementation method, signature algorithm, deployment architecture, and conclusion unit granularity without departing from the concept of the present invention should fall within the protection scope of the present invention.

[0094] In summary, this application provides a multi-source evidence semantic isomorphism mechanism for social risk early warning scenarios. Specifically, it refers to mapping different modal data such as text, images, audio, video keyframes, and log records into a unified evidence unit EU with a consistent field structure after modal translation, text normalization, and slot extraction. This evidence unit also retains the original source pointer, content hash, permission tag, and event attribution information.

[0095] This application also provides an evidence merging, indexing, and version maintenance mechanism around a unified event caliber. Specifically, it refers to merging multi-source evidence at the event level based on event_id, spatiotemporal proximity, subject overlap, topic similarity, and propagation relationship, and simultaneously constructing vector indexes, structured indexes, and source indexes. This allows evidence to be retrieved semantically and filtered by business conditions such as time, location, subject, and permissions. It also supports duplicate evidence identification and revision version tracking.

[0096] This application also provides an evidence slot constraint and binding mechanism for conclusion units. Specifically, before generating risk assessment results, the output task is decomposed into several conclusion units (CUs), and minimum evidence number, source type, time consistency, and key slot coverage requirements are pre-set for each conclusion unit. Only evidence unit sets that meet the requirements are allowed to be bound to the corresponding conclusion unit and participate in subsequent conclusion generation.

[0097] This application also provides a constrained evidence chain generation and conflict verification mechanism. Specifically, each factual judgment output by the generation model must explicitly reference the bound eu_id; if the evidence is insufficient, it outputs "insufficient evidence to be verified" rather than "unfounded inference"; if there is a conflict between the evidence, it outputs a conflict prompt and triggers a manual review mark, thereby truly implementing evidence constraints within the generation process.

[0098] This application also provides an evidence chain encapsulation and integrity verification mechanism for audit review. Specifically, it refers to serializing and encapsulating an evidence chain list consisting of conclusion fragments, evidence identifier sets, support, conflict degree, source hash, and generation process metadata, and using Merkle root hash, digital signature, and version number management to achieve tamper-proof verification of results, playback of historical versions, and cross-departmental review.

[0099] Compared with existing technologies, the beneficial effects of the embodiments of this application are mainly reflected in the following aspects. First, the embodiments of this application elevate social risk assessment from a readable answer to a verifiable conclusion. Many existing solutions can provide a complete statement, but it is difficult to explain which evidence each judgment is based on; the embodiments of this application, through conclusion unit decomposition, evidence binding, coverage calculation, and conflict prompts, enable the final output to naturally possess the ability to trace back sentence by sentence and review item by item, making it more suitable for auditing, accountability, and collaborative consultation in governance scenarios. Second, the embodiments of this application significantly enhance the unified utilization capability of multimodal evidence. Through modal translation and semantic isomorphic mapping, content such as images, audio, and video, which were originally difficult to directly participate in retrieval and alignment, are transformed into computable, indexable, and comparable evidence units, thereby reducing the possible omission of key evidence when judging solely based on textual information.

[0100] Third, this application's embodiments do not place access control and compliance processing on the periphery of the system. Instead, they embed access tags, anonymization rules, and process logging into the evidence access, retrieval, generation, and output stages, enabling the system to meet business real-time requirements while also taking into account data security and compliance requirements. Fourth, this application's embodiments improve the stability of the conclusion formation process through coverage constraints, consistency constraints, and conflict verification mechanisms. In cases with a single source, insufficient evidence, or contradictory facts, the system will not blindly output a definitive conclusion but will clearly indicate insufficient evidence or the need for manual review. This is particularly important for high-responsibility scenarios such as risk assessment. Fifth, this application's embodiments establish evidence chain version management and integrity verification mechanisms, which can retain historical evidence when adding new evidence, revising evidence, or reassessing the case. This supports reviewing why the evidence was used at the time and why it changed later, thus possessing both engineering practical value and long-term traceability value in governance scenarios.

[0101] Without departing from the core concept of this invention, there are various alternative implementation methods in the embodiments of this application. In the modal translation stage, image processing can either use a unified visual language model to generate scene descriptions, or use object detection plus OCR plus rule templates to extract elements step by step; audio processing can either perform only speech recognition and transcription, or further superimpose speaker separation, voiceprint recognition, emotion recognition, or noise suppression modules; video processing can either be based on keyframe extraction, or based on shot segmentation and separate translation.

[0102] In the semantic isomorphism and slot extraction stages, either a combination of dedicated models such as named entity recognition, event extraction, and relation extraction can be used, or a large model can be used to directly output structured fields, supplemented by rule validation. Context segmentation can use either a fixed-length overlapping window or an adaptive segmentation strategy based on syntactic boundaries, event boundaries, or topic switching points.

[0103] In the retrieval and evidence selection stages, candidate recall can employ a hybrid approach of sparse and dense retrieval, or only one of these methods. The re-ranking mechanism can use a cross-encoder, a learned ranking model, a dual-tower fine ranking model, or a ranking method that combines rules and models. Evidence subset selection can be achieved through objective function optimization, or through greedy strategies, threshold filtering strategies, or rule priority strategies.

[0104] In the conflict detection and evidence binding stage, the probability of support and the probability of contradiction can be calculated based on the natural language reasoning model, or rule comparison can be performed based on structured slots such as time, location, subject, number of people, and event results. The granularity of the conclusion unit (CU) can be set to sentence level, paragraph level, indicator level, or template field level to adapt to the report formats of different departments.

[0105] In the integrity verification and evidence preservation stage, in addition to Merkle trees and digital signatures, timestamp services, blockchain anchoring, trusted execution environment certification, or other equivalent integrity verification methods can also be used; the evidence content can be fully encapsulated, or only the digest, fingerprint hash, and key metadata can be encapsulated. As long as a verifiable binding relationship between the conclusion and the evidence can be achieved, it can be regarded as an alternative implementation of the embodiments of this application.

[0106] In summary, this application provides a task assessment method that, in response to an assessment task instruction, filters a subset of evidence matching the assessment task instruction from a pre-built evidence library; the subset of evidence includes several evidence units; matches an expected output structure according to the task type corresponding to the assessment task instruction; the expected output structure includes several conclusion units predefined by the task type; for each conclusion unit, determines evidence units that meet preset evidence constraints from the subset of evidence and establishes a mapping relationship between the conclusion unit and the corresponding evidence unit; based on the mapping relationship and the content of the evidence units in the subset of evidence, generates a task assessment result, which includes an assessment conclusion text and an evidence chain list, the evidence chain list being used to identify the evidence units referenced by the assessment conclusion text. By accurately selecting evidence units that match the assessment task from a pre-built evidence library, determining a standardized output structure based on the task type, establishing a mapping relationship between conclusions and evidence, and generating assessment results that include assessment text and a list of evidence chains, the assessment process is supported by reliable evidence, the assessment conclusions are traceable, and the assessment output meets the task requirements, thus satisfying the standardization and refinement needs of task assessment in various scenarios.

[0107] Based on the same technical concept, embodiments of this application also provide a task assessment system, such as... Figure 7 As shown, the system includes: The evidence subset filtering module 701 is used to filter out a subset of evidence that matches the judgment task instruction from a pre-built evidence library in response to the judgment task instruction; the evidence subset includes several evidence units; The output structure determination module 702 is used to match the expected output structure according to the task type corresponding to the judgment task instruction; the expected output structure includes several conclusion units predefined by the task type. The evidence unit determination module 703 is used to determine, for each conclusion unit, evidence units that meet the preset evidence constraints from the evidence subset, and to establish a mapping relationship between the conclusion unit and the corresponding evidence unit; The result generation module 704 is used to generate a task assessment result based on the mapping relationship and the content of the evidence units in the evidence subset. The task assessment result includes an assessment conclusion text and an evidence chain list. The evidence chain list is used to identify the evidence units referenced by the assessment conclusion text.

[0108] This application also provides an electronic device corresponding to the method provided in the foregoing embodiments. Please refer to... Figure 8 The diagram illustrates an electronic device provided by some embodiments of this application. The electronic device 20 may include: a processor 200, a memory 201, a bus 202, and a communication interface 203, wherein the processor 200, the communication interface 203, and the memory 201 are connected via the bus 202; the memory 201 stores a computer program that can run on the processor 200, and when the processor 200 runs the computer program, it executes the method provided by any of the foregoing embodiments of this application.

[0109] The memory 201 may include high-speed random access memory (RAM) or non-volatile memory, such as at least one disk storage device. Communication between this system network element and at least one other network element is achieved through at least one physical port (which can be wired or wireless), such as the Internet, wide area network, local area network, or metropolitan area network.

[0110] Bus 202 can be an ISA bus, PCI bus, or EISA bus, etc. The bus can be divided into an address bus, a data bus, a control bus, etc. The memory 201 is used to store programs. After receiving an execution instruction, the processor 200 executes the program. The method disclosed in any of the foregoing embodiments of this application can be applied to the processor 200, or implemented by the processor 200.

[0111] The processor 200 may be an integrated circuit chip with signal processing capabilities. In implementation, each step of the above method can be completed by the integrated logic circuitry in the hardware of the processor 200 or by instructions in software form. The processor 200 may be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it may also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), an off-the-shelf programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this application. The general-purpose processor may be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this application can be directly embodied in the execution of a hardware decoding processor, or executed by a combination of hardware and software modules in the decoding processor. The software modules may reside in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. The storage medium is located in memory 201. The processor 200 reads the information in memory 201 and, in conjunction with its hardware, completes the steps of the above method.

[0112] The electronic devices and methods provided in the embodiments of this application are based on the same inventive concept and have the same beneficial effects as the methods they employ, operate, or implement.

[0113] This application also provides a computer-readable storage medium corresponding to the method provided in the foregoing embodiments. Please refer to... Figure 9 The computer-readable storage medium shown is an optical disc 30, on which a computer program (i.e., a program product) is stored, which, when run by a processor, executes the methods provided in any of the foregoing embodiments.

[0114] It should be noted that examples of the computer-readable storage medium may also include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other optical and magnetic storage media, which will not be elaborated here.

[0115] The computer-readable storage medium provided in the above embodiments of this application and the method provided in the embodiments of this application are based on the same inventive concept and have the same beneficial effects as the methods adopted, run or implemented by the applications stored therein.

[0116] The above description is only a preferred embodiment of the present invention and does not limit the patent scope of the present invention. All equivalent structural transformations made under the concept of the present invention using the contents of the present invention specification and drawings, or direct / indirect applications in other related technical fields, are included within the patent protection scope of the present invention.

Claims

1. A task judgment method, characterized in that, The method includes: In response to the assessment task instruction, a subset of evidence matching the assessment task instruction is selected from a pre-built evidence library; the subset of evidence includes several evidence units. The expected output structure is matched according to the task type corresponding to the judgment task instruction; the expected output structure includes several conclusion units predefined by the task type. For each conclusion unit, evidence units that meet the preset evidence constraints are determined from the evidence subset, and a mapping relationship is established between the conclusion unit and the corresponding evidence unit. Based on the mapping relationship and the content of the evidence units in the evidence subset, a task assessment result is generated. The task assessment result includes an assessment conclusion text and an evidence chain list. The evidence chain list is used to identify the evidence units referenced by the assessment conclusion text. The step of matching the expected output structure according to the task type corresponding to the assessment task instruction includes: identifying the task type corresponding to the assessment task instruction; calling the corresponding structure template according to the task type; determining a number of conclusion units based on the structure template; and combining the number of conclusion units to form the expected output structure. The step of generating task assessment results based on the mapping relationship and the content of evidence units in the evidence subset includes: generating conclusion text fragments corresponding to each conclusion unit based on the mapping relationship and the content of the cited evidence bound to each conclusion unit; combining multiple conclusion text fragments according to the expected output structure to obtain the assessment conclusion text; and generating the evidence chain list according to the cited evidence corresponding to each conclusion text fragment. The construction process of the pre-built evidence library includes: acquiring multi-source raw data related to task analysis and adding corresponding source description information to each piece of multi-source raw data; performing modal translation processing on the non-text data in the multi-source raw data to obtain the corresponding textual representation; performing semantic isomorphic processing on the original text data in the multi-source raw data and the textual representation to extract structured semantic information; constructing evidence units based on the structured semantic information and the corresponding source description information; determining the source credibility based on the evidence source type of the evidence unit, determining the content completeness based on the completeness of the content fields, and determining the timeliness information based on the time interval between the evidence occurrence time and the analysis time; determining the evidence weight of the evidence unit based on the source credibility, content completeness, and timeliness information; storing the evidence units after determining the evidence weight, and establishing a retrieval index for the evidence units to form the evidence library.

2. The method of claim 1, wherein, The step of selecting a subset of evidence that matches the assessment task instructions from a pre-built evidence library includes: The judgment task instruction is parsed to obtain retrieval constraints, which include at least one of time range, spatial range, permission range, and task type. Based on the retrieval constraints, conditional filtering is performed on the evidence units in the evidence database to obtain a candidate evidence set; Based on the semantic relevance between each evidence unit in the candidate evidence set and the judgment task instruction, the candidate evidence set is subjected to relevance screening to obtain the evidence subset.

3. The method of claim 2, wherein, Based on the semantic relevance between each evidence unit in the candidate evidence set and the assessment task instruction, the candidate evidence set is subjected to relevance screening to obtain the evidence subset, including: Based on the attribute information of each evidence unit, the corresponding evidence weight is determined; wherein the evidence weight is determined based on at least one of the following: source credibility, content completeness, timeliness, and consistency with other evidence units. The evidence units in the candidate evidence set are sorted according to the semantic relevance and the evidence weight. Based on the sorting results and coverage constraints, evidence units are selected from the candidate evidence set to form the evidence subset. The coverage constraints are used to ensure that the evidence subset covers multiple different information dimensions related to the judgment task instruction.

4. The method of claim 1, wherein, For each conclusion unit, determining the evidence units that meet the preset evidence constraints from the subset of evidence, and establishing the mapping relationship between the conclusion unit and the corresponding evidence unit, includes: For each conclusion unit, obtain the preset evidence constraints corresponding to the conclusion unit; Based on the preset evidence constraints, matching evidence units are selected from the evidence subset; If a matching evidence unit is found through screening, the matching evidence unit is used as the cited evidence corresponding to the current conclusion unit, and a mapping relationship is established between the current conclusion unit and the cited evidence. If no matching evidence unit is found through screening, the current conclusion unit is marked as insufficient evidence in the mapping relationship.

5. The method of claim 1, wherein, The process of establishing and storing a retrieval index for the evidence unit includes: The text content of the evidence unit is converted into a vector representation, and a vector index is constructed for semantic similarity retrieval. The structured semantic information of the evidence units is used as a filterable field to construct a structured index for conditional filtering retrieval; Determine temporal associations based on the temporal information contained in each evidence unit, spatial associations based on spatial information, subject associations based on subject information, and semantic associations based on semantic content. Based on the determined temporal, spatial, subject, and semantic associations, event merging processing is performed on multiple evidence units to obtain a unified event identifier; The evidence unit is associated with the corresponding event identifier and stored in the evidence database.

6. A task assessment system, characterized in that, The system includes: The evidence subset filtering module is used to filter out a subset of evidence that matches the assessment task instruction from a pre-built evidence library in response to the assessment task instruction; the evidence subset includes several evidence units; An output structure determination module is used to match an expected output structure according to the task type corresponding to the assessment task instruction; the expected output structure includes several conclusion units predefined by the task type; the matching of the expected output structure according to the task type corresponding to the assessment task instruction includes: identifying the task type corresponding to the assessment task instruction; calling the corresponding structure template according to the task type; determining several conclusion units based on the structure template, and combining the several conclusion units to form the expected output structure. The evidence unit determination module is used to determine, for each conclusion unit, evidence units that meet preset evidence constraints from the evidence subset, and to establish a mapping relationship between the conclusion unit and the corresponding evidence unit; The result generation module is used to generate task assessment results based on the mapping relationship and the content of the evidence units in the evidence subset. The task assessment results include assessment conclusion text and an evidence chain list, whereby the evidence chain list identifies the evidence units referenced by the assessment conclusion text. Generating the task assessment results based on the mapping relationship and the content of the evidence units in the evidence subset includes: generating conclusion text fragments corresponding to each conclusion unit based on the mapping relationship and the content of the referenced evidence bound to each conclusion unit; combining multiple conclusion text fragments according to the expected output structure to obtain the assessment conclusion text; and generating the evidence chain list based on the referenced evidence corresponding to each conclusion text fragment. The construction process of the pre-built evidence library includes: acquiring multi-source raw data related to the task assessment and preparing... Each piece of multi-source raw data is given a corresponding source description; the non-text data in the multi-source raw data is modally translated to obtain a corresponding textual representation; the original text data in the multi-source raw data and the textual representation are semantically isomorphically processed to extract structured semantic information; based on the structured semantic information and the corresponding source description information, an evidence unit is constructed; the source credibility is determined based on the evidence source type of the evidence unit, the content completeness is determined based on the completeness of the content fields, and the timeliness information is determined based on the time interval between the evidence occurrence time and the judgment time; the evidence weight of the evidence unit is determined based on the source credibility, content completeness, and timeliness information; the evidence units with determined evidence weights are stored, and a retrieval index is established for the evidence units to form the evidence library.

7. An electronic device, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the method as claimed in any one of claims 1-5.

8. A computer-readable storage medium, characterized in that, It stores computer-readable instructions that can be executed by a processor to implement the method as described in any one of claims 1-5.

Images