Conflict correction method, apparatus and device for content generation result

Abstract

The application provides a conflict correction method, device and equipment of content generation result, and relates to the technical field of safety training content processing. The method comprises the following steps: obtaining a candidate evidence set based on a preset knowledge evidence base and a preset task condition; obtaining target content to be processed, and splitting the target content into multiple content units. According to the candidate evidence set and the multiple content units, the comprehensive conflict degree of each content unit is determined. If it is determined that the target content meets the preset correction trigger condition based on the comprehensive conflict degree, the positioning information is generated according to the comprehensive conflict degree; the target content is corrected according to the positioning information, and the correction result data is output. The method is used to improve the modification accuracy of the training content.

Description

Technical Field

[0001] This application relates to the field of security training content processing technology, and more specifically, to a method, apparatus, and device for conflict correction of content generation results. Background Technology

[0002] Currently, with the development of artificial intelligence technology and digital training platforms, safety training systems can generate training plans, personalized test questions, question analyses, and supplementary explanations targeting weaknesses based on multi-source knowledge resources such as safety procedures, accident cases, management systems, and historical exam question banks. In practical applications, the training content can originate from automatically generated results, template-based results, or manually compiled results, but all must undergo consistency verification and compliance processing before entering the publishing, push, or training execution stages.

[0003] Existing technologies typically focus on content generation, knowledge retrieval, or student assessment itself. If problems are found with the generated training content, the approach is usually to rewrite the entire content or manually modify it section by section.

[0004] However, in existing technologies, modifying training content by completely rewriting it or manually modifying it section by section can easily alter the original correct content, resulting in a lower accuracy of the modified training content. Summary of the Invention

[0005] The purpose of this application is to provide a method, apparatus, and device for correcting conflicts in content generation results, so as to solve the above-mentioned problems existing in the prior art and greatly improve the accuracy of training content modification.

[0006] Firstly, a method for correcting conflicts in content generation results is provided, which may include: Based on a pre-defined knowledge evidence base and pre-defined task conditions, obtain a candidate evidence set; Obtain the target content to be processed, and divide the target content into multiple content units; Based on the candidate evidence set and the multiple content units, determine the overall degree of conflict for each content unit; If the target content is determined to meet the preset correction triggering conditions based on the overall conflict level, then location information is generated according to the overall conflict level; the target content is corrected according to the location information, and the correction result data is output.

[0007] Secondly, a conflict correction device for content generation results is provided, which may include: The acquisition module is used to acquire a set of candidate evidence based on a preset knowledge evidence base and preset task conditions; The splitting module is used to obtain the target content to be processed and split the target content into multiple content units; The determination module is used to determine the overall conflict level of each content unit based on the candidate evidence set and the multiple content units; The correction module is used to generate location information based on the comprehensive conflict level if the target content is determined to meet the preset correction triggering conditions based on the comprehensive conflict level; to perform correction processing on the target content based on the location information; and to output correction result data.

[0008] Thirdly, an electronic device is provided, which includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, used to store computer programs; When a processor executes a program stored in memory, it implements any of the steps described in the first aspect above.

[0009] Fourthly, a computer-readable storage medium is provided, wherein a computer program is stored therein, and when executed by a processor, the computer program implements the steps of any of the methods described in the first aspect above.

[0010] This application provides a method, apparatus, and device for conflict correction of content generation results. Based on a preset knowledge evidence base and preset task conditions, a candidate evidence set is obtained. The target content to be processed is obtained and divided into multiple content units. The overall conflict level of each content unit is determined according to the candidate evidence set and the multiple content units. If the target content meets preset correction triggering conditions based on the overall conflict level, location information is generated according to the overall conflict level. Based on the location information, the target content is corrected, and the correction result data is output. In this solution, the verification basis (knowledge evidence base) and candidate evidence set are first determined, the target object to be verified is determined, and the conflict is triggered, located, and it is determined whether correction is needed based on the overall conflict level. When correction is determined to be needed, the overall conflict level is used as the correction triggering basis, which can limit the correction action to conflict fragments, thereby enabling targeted local correction. This allows for local correction of generated content without expanding the scope of rewriting lesson plans, test questions, or reinforcement content, greatly improving the accuracy of training content modification. Attached Figure Description

[0011] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments of this application will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 A flowchart illustrating a conflict correction method for content generation results provided in an embodiment of this application; Figure 2 A flowchart illustrating another content generation result conflict correction method provided in this application embodiment; Figure 3 A schematic diagram of a conflict correction device for content generation results provided in an embodiment of this application; Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0013] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application. Unless otherwise defined, the technical or scientific terms used in this application should have the ordinary meaning understood by those skilled in the art. The words "first," "second," and similar terms used in this application do not indicate any order, quantity, or importance, but are only used to distinguish different components. The words "comprising" or "including," etc., mean that the element or object preceding the word covers the element or object listed after the word and its equivalents, but do not exclude other elements or objects. The words "connected," "coupled," or "connected," etc., are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. "Up," "down," "left," "right," etc., are only used to indicate relative positional relationships. When the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0014] Currently, with the development of artificial intelligence technology and digital training platforms, safety training systems can generate training plans, personalized test questions, question analyses, and supplementary explanations targeting weaknesses based on multi-source knowledge resources such as safety procedures, accident cases, management systems, and historical exam question banks. In practical applications, the training content can originate from automatically generated results, template-based results, or manually compiled results, but all must undergo consistency verification and compliance processing before entering the publishing, push, or training execution stages.

[0015] Because different knowledge sources differ in their formation time, scope of application, and expression, the same training topic often contains heterogeneous information, including current procedural requirements, historical practices in case studies, management requirements in regulations, and existing expressions in question banks. These differences directly impact the training content to be processed, leading to inconsistencies in action suggestions, missing conditions, mismatched target audiences, or inconsistent step sequences between lesson plans, exam questions, explanations, or weak point reinforcement content and the current training guidelines.

[0016] In a typical example, the focus is usually on content generation, knowledge retrieval, or student assessment itself. If problems are found with the generated training content, the approach is often to rewrite the entire content or manually modify it section by section. However, in existing technologies, rewriting the entire content or manually modifying it section by section can easily alter previously correct content, leading to lower accuracy in the revised training content.

[0017] The content generation conflict correction method provided in this application embodiment can be applied to electronic devices, terminal devices, content generation conflict correction devices, or other devices or equipment capable of executing this embodiment, and there are no limitations on this application. In this embodiment, the execution subject is described as an electronic device.

[0018] The terminal can be a user equipment (UE) such as a mobile phone, smartphone, laptop computer, digital broadcast receiver, personal digital assistant (PDA), or tablet computer (PAD), handheld device, in-vehicle device, wearable device, computing device, or other processing device connected to a wireless modem, mobile station (MS), or mobile terminal. This terminal has the ability to communicate with one or more core networks via a radio access network (RAN).

[0019] The preferred embodiments of this application are described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit this application. Furthermore, the embodiments and features in the embodiments of this application can be combined with each other without conflict.

[0020] Figure 1 This is a flowchart illustrating a conflict correction method for content generation results provided in an embodiment of this application. Figure 1 As shown, the method may include: Step S101: Obtain a candidate evidence set based on a preset knowledge evidence base and preset task conditions.

[0021] For example, a preset knowledge evidence base is obtained, which includes a safety training knowledge evidence base SD={sd1,sd2,...,sdn}. SD is a set of knowledge resources that provide verification basis, derived from the training knowledge management base, and includes at least a safety procedure database, a management system database, an accident case database, and a historical exam question bank. Each knowledge resource in SD is segmented, deduplicated, and its source is labeled to obtain an evidence unit set EU={e1,e2,...,eN}. Based on the preset task condition q_k of the current k-th round, an evidence alignment unit EAU_i=(q_k,e_i) is constructed.

[0022] Wherein, sdn represents the nth knowledge resource in SD; eN represents the Nth evidence in EU; e_i represents the i-th evidence unit, e_i includes at least text t_i, training object label a_i, risk topic label r_i, source category type(i) and source credibility cred(i); q_k represents the preset task conditions of the k-th round, q_k includes at least one or more of the training topic, training object label, risk topic label and weakness label.

[0023] Based on the evidence alignment unit EAU_i and the calculation formula, the comprehensive support score S_i of the i-th evidence unit is calculated. Evidence units with scores ranking in the top K positions and whose S_i is not lower than the preset evidence validity threshold θ_s are selected to form a candidate evidence set E_k. The candidate evidence set E_k includes multiple evidence units (i.e., candidate evidence items). E_k = {e_(1), e_(2), ..., e_(K)}. E_k is used for subsequent conflict comparison with the target content T_k to be verified. The calculation formula is as follows:

[0024] Where S_i represents the comprehensive support score of the i-th evidence unit relative to the preset task conditions of the k-th round; sim_v(i) represents the vector similarity between text t_i and preset task conditions q_k; sim_b(i) represents the keyword matching degree; match_a(i) represents the training object matching degree; match_r(i) represents the risk topic matching degree; cred(i) represents the source credibility; θ_s represents the preset evidence validity threshold; α, β, γ, δ, and ε are weight coefficients, satisfying α+β+γ+δ+ε=1.

[0025] Wherein, sim_v(i) is calculated using normalized vector cosine similarity; sim_b(i) = |B(q_k)∩B(e_i)| / |B(q_k)∪B(e_i)|, where B(q_k) represents the set of keywords extracted from the preset task condition q_k, and B(e_i) represents the set of keywords extracted from the evidence unit e_i; match_a(i) is set to 1 when the training object label a_i of the evidence unit matches the training object label in the preset task condition q_k, otherwise it is set to 0; match_r(i) is set to 1 when the risk topic label r_i of the evidence unit matches the risk topic label in the preset task condition q_k, otherwise it is set to 0. Thus, each component of S_i (sim_v(i), match_a(i), match_r(i)) can be calculated from labels, keywords, or vectorized text.

[0026] It should be noted that in this application, the applicable object label information of the evidence unit (training object label a_i, risk topic label r_i, etc.) and the applicable object item description in the preset task conditions (training object label, risk topic label, etc. in preset task conditions q_k) are concepts at different levels. The former are attributes pre-labeled in the knowledge evidence base, used to filter evidence related to the current preset task conditions q_k; the latter are preset attributes included in the preset task conditions q_k, which can be attributes extracted by the user after splitting the target content to be verified. Those skilled in the art should pay attention to distinguishing the data objects to which the two are attached, corresponding to evidence unit e_i and target content, respectively.

[0027] Step S102: Obtain the target content to be processed and split the target content into multiple content units.

[0028] For example, the target content T_k to be verified in the kth round is obtained, and T_k is split into a proposition set P_k={p_k(1),p_k(2),...,p_k(M_k)}; the proposition set P_k includes multiple content units (hereinafter referred to as propositions) p_k(1), p_k(2),..., p_k(M_k), where M_k represents the number of propositions in the kth round; p_k(u) represents the u-th proposition in the kth round, u=1,2,...,M_k; the attribute description information of each proposition p_k(u) includes at least the action item A_u, the object item O_u, the condition item C_u, the time sequence item T_u, and the applicable training object item R_u.

[0029] The sources of T_k include at least: 1) the large model generation entry point, i.e., training plans, personalized test questions, question analysis, or weak point reinforcement content generated by the safety training large model based on training tasks, job tags, risk tags, weak point tags, and related knowledge materials; 2) the template arrangement entry point, i.e., text formed by combining preset training templates; 3) the manual import entry point, i.e., manually written text; and 4) the feedback entry point, i.e., the content returned after the previous round of processing. T_k is provided to this application by the upstream module through the content output interface. The upstream module includes at least the large model generation module, the template arrangement module, the manual import module, or the feedback module from the previous round of processing. Relative to this application, the above modules constitute the input source channels of T_k. This application only performs conflict detection, correction decision-making, and result output on T_k, and does not involve the internal parameters, inference process, or self-correction process of the large model.

[0030] The difference between SD and T_k is as follows: SD is a safety training knowledge evidence base, used to provide verification basis such as procedures, systems, cases, and question banks; T_k is the target content to be verified, which can be the safety training content result generated after the large model is generated, or the content result formed by template arrangement, manual import, or reflow. SD is not used as a correction object, and the internal generation process of the large model is not used as the processing object of this application; T_k is the conflict detection and correction output object of this application.

[0031] Step S103: Determine the overall conflict level of each content unit based on the candidate evidence set and multiple content units.

[0032] For example, attribute description information of each content unit is extracted on multiple preset attribute dimensions. The attribute description information includes action item A_u, object item O_u, condition item C_u, time sequence item T_u, and applicable training object item R_u.

[0033] The preset conflict dimensions include suggested direction conflict dimension, conditional conflict dimension, object adaptation conflict dimension, and temporal conflict dimension. For each proposition p_k(u), the proposition p_k(u) is aligned and compared with the evidence unit e_v in the candidate evidence set E_k. For each preset conflict dimension, the conflict metric between the attribute description information of the proposition and each candidate evidence item is calculated. The conflict metric includes suggested direction conflict Dir_k(u,v) for suggested direction conflict dimension, conditional conflict Cond_k(u,v) for conditional conflict dimension, object adaptation conflict Fit_k(u,v) for object adaptation conflict dimension, and temporal conflict Time_k(u,v) for temporal conflict dimension. The overall conflict degree X_k(u,v) at the proposition level is also calculated, which includes the overall conflict degree between each proposition and each evidence unit.

[0034]

[0035] Where u is the proposition index and v is the evidence index; the values ​​of Dir_k(u,v), Cond_k(u,v), Fit_k(u,v), and Time_k(u,v) are all in the range of [0,1], and the larger the value, the stronger the corresponding type of conflict; Dir_k(u,v) represents the directional conflict between the proposition action suggestion and the evidence requirement; Cond_k(u,v) represents the missing or inconsistent necessary conditions; Fit_k(u,v) represents the mismatch between the training object or the applicable scenario; Time_k(u,v) represents the conflict between the order of steps; X_k(u,v) represents the comprehensive degree of conflict between proposition p_k(u) and evidence e_v; ω1, ω2, ω3, and ω4 are weight coefficients, satisfying ω1+ω2+ω3+ω4=1.

[0036] In one possible implementation, the evidence unit e_v is also parsed into evidence action item A_v, evidence object item O_v, evidence condition item C_v, evidence timing item T_v, and evidence applicable object item R_v. Dir_k(u,v) is 1 when A_u and A_v have opposite directions such as prohibition / permission or should / must not, otherwise it is 0; Cond_k(u,v)=1-|C_u∩C_v| / |C_v|, used to measure the proportion of missing necessary conditions; Fit_k(u,v) is 1 when R_u and R_v are incompatible or O_u and O_v do not match, otherwise it is 0; Time_k(u,v) is 1 when the order of T_u and T_v is inconsistent, otherwise it is 0. The above four types of components can all be calculated through rule tables, label tables, or term mapping tables.

[0037] Step S104: If the target content meets the preset correction triggering conditions based on the comprehensive conflict level, then generate location information according to the comprehensive conflict level; perform correction processing on the target content according to the location information, and output the correction result data.

[0038] For example, if the target content meets the preset correction triggering conditions based on the overall conflict level, then location information is generated based on the overall conflict level. Then, according to the location indicated by the location information, the target content is locally corrected, and the corrected result data is generated and output.

[0039] The method provided in this application embodiment obtains a candidate evidence set based on a preset knowledge evidence base and preset task conditions. It then obtains the target content to be processed and breaks it down into multiple content units. Based on the candidate evidence set and the multiple content units, it determines the overall conflict level of each content unit. If the overall conflict level determines that the target content meets preset correction triggering conditions, it generates location information based on the overall conflict level. Based on the location information, it performs correction processing on the target content and outputs the correction result data. In this solution, the verification basis (knowledge evidence base) and candidate evidence set are first determined, the target object to be verified is identified, and the conflict is triggered and located based on the overall conflict level. It also determines whether correction is needed. When correction is determined, the overall conflict level is used as the correction triggering basis, which limits the correction action to conflict fragments. This allows for targeted local correction, enabling local correction of generated content without expanding the scope of rewriting lesson plans, test questions, or reinforcement content, thus greatly improving the accuracy of training content modifications.

[0040] Figure 2 A flowchart illustrating a conflict correction method for content generation results provided in this application is shown below. Figure 2 As shown, in this embodiment... Figure 1 Based on the embodiments, the method is described in detail below, and the method includes: Step S201: Based on the preset knowledge evidence base and preset task conditions, obtain the candidate evidence set.

[0041] For example, this step is described in step S101, and will not be repeated here.

[0042] Step S202: Obtain the target content to be processed and split the target content into multiple content units.

[0043] For example, this step is described in step S102, and will not be repeated here.

[0044] Step S203: Determine the overall conflict level of each content unit based on the candidate evidence set and multiple content units.

[0045] In one example, the candidate evidence set includes multiple candidate evidence items; S203 includes: extracting attribute description information of each content unit on multiple preset attribute dimensions; for each content unit, determining the conflict metric between the attribute description information of the content unit and each candidate evidence item on each preset preset conflict dimension; and weighting and summing the conflict metrics on multiple preset conflict dimensions according to preset weight coefficients to determine the comprehensive conflict degree between each content unit and each candidate evidence item.

[0046] For example, this step is described in step S103, and will not be repeated here.

[0047] Step S204: If the target content meets the preset correction triggering conditions based on the comprehensive conflict level, then generate location information according to the comprehensive conflict level.

[0048] In one example, S204 includes: determining the overall quality status of the target content based on the overall conflict level of each content unit; determining whether the overall quality status meets the preset correction triggering conditions; if the overall quality status meets the preset correction triggering conditions, then determining the conflicting content units based on the overall conflict level, and generating the location information of the conflicting content units.

[0049] In one example, the overall quality status of the target content is determined based on the comprehensive conflict level of each content unit. This includes: statistically aggregating the comprehensive conflict level of each content unit to obtain an overall conflict severity index for the target content; determining the number of evidence items for each content unit in the candidate evidence set; the number of evidence items is the number of candidate evidence items; determining the number of content units with non-empty evidence items; determining the proportion of the number of units to the total number of all content units as the basis coverage index for the target content; determining the source credibility of the candidate evidence items that match each content unit; statistically aggregating the source credibility to obtain a candidate evidence credibility index for the target content; and determining the overall quality status of the target content based on the overall conflict severity index, the basis coverage index, and the candidate evidence credibility index.

[0050] In one example, if a preset correction triggering condition is met, the conflicting content units are determined based on the overall conflict level, and the location information of the conflicting content units is generated. This includes: if the preset correction triggering condition is met, generating a set of valid content units based on content units with non-empty evidence; for each content unit in the set of valid content units, obtaining the overall conflict level with each candidate evidence item; determining the pairing relationship where the overall conflict level exceeds a preset conflict threshold as a conflict pair; a conflict pair includes at least a content unit with conflict and a candidate evidence item that conflicts with the content unit; extracting the content units that conflict from each conflict pair and deduplicating identical content units to obtain a set of conflicting content units; for each conflicting content unit in the set of conflicting content units, obtaining all conflict pairs related to that conflicting content unit; and generating the location information of that conflicting content unit based on all conflict pairs.

[0051] For example, based on the overall conflict severity X_k(u,v) at the proposition level, an overall conflict severity index C_k, a coverage index V_k, and a candidate evidence credibility index U_k are aggregated. Then, based on the overall conflict severity index C_k, the coverage index V_k, and the candidate evidence credibility index U_k, a flag R_k is triggered to determine whether to proceed with conflict correction.

[0052] Specifically, the overall conflict severity of each content unit is statistically aggregated to obtain the overall conflict severity index C_k of the target content: ; Determine the number of pieces of evidence for each content unit in the candidate evidence set. The number of evidence items is the number of candidate evidence items; the number of content units with a non-empty number of evidence items is calculated, and the number of these units is calculated as a percentage of the total number of content units (i.e., the number of propositions in the k-th round). The proportion of the target content is used as the coverage index V_k: , Determine the source credibility cred(i) of the candidate evidence items that match each content unit; statistically aggregate the source credibility cred(i) to obtain the candidate evidence credibility index U_k of the target content:

[0053] Based on the overall conflict severity index, coverage index, and candidate evidence credibility index, the overall quality status of the target content is determined. Then, based on the overall conflict severity index C_k, coverage index V_k, and candidate evidence credibility index U_k, a flag R_k is triggered to determine whether to proceed with conflict correction. The formula for calculating flag R_k is as follows:

[0054] Wherein, V_k represents the coverage index, used to indicate the proportion of propositions in T_k that can be supported by evidence; U_k represents the candidate evidence credibility index, used to indicate the average credibility of valid matching evidence; R_k represents the trigger flag; θ_c represents the conflict threshold; θ_v represents the coverage threshold; θ_u represents the reliability threshold; and M_k represents the number of propositions in the k-th round. When N_match,k=0, it means that T_k lacks verifiable evidence, and result A is directly generated and marked for manual review; when R_k=1, it means that the preset correction trigger condition is met, and proceed to the next step; when R_k=0, it means that the preset correction trigger condition is not met, result B is generated, and the process jumps directly to the output correction result data in step S205.

[0055] M_k^+ represents the number of pieces of evidence with valid matching evidence. Based on this number, a proposition index set can be formed: M_k^+ = {u | there exists e_v ∈ E_k and S_v ≥ θ_s}; N_match,k = |M_k^+|, where N_match,k represents the number of pieces of evidence with valid matching evidence found in the candidate evidence set E_k. A valid match means that proposition p_k(u) has at least one evidence unit e_v satisfying S_v ≥ θ_s, where S_v refers to the comprehensive support score of the v-th evidence unit relative to the preset task conditions of the k-th round; v_u^* represents the optimal evidence unit corresponding to proposition p_k(u), v_u^* = argmax. S_v, and e_v∈E_k, S_v≥θ_s; c_k(u) represents the maximum conflict value of the u-th proposition, when u∈M_k^+ c_k(u)=maxX_k(u,v), and e_v∈E_k, S_v≥θ_s; when u does not belong to M_k^+ c_k(u)=1.

[0056] If the preset correction trigger condition is met, i.e., R_k=1, a set of valid content units is generated based on the number of evidence units M_k^+ which is not empty. For each content unit in the set of valid content units, the comprehensive conflict degree X_k(u,v) with each candidate evidence item is obtained. Pairing relationships with a comprehensive conflict degree exceeding a preset conflict threshold θ_x are identified as conflict pairings, i.e., (u,v) pairings. A conflict pairing includes at least a content unit that conflicts and a candidate evidence item that conflicts with the content unit. Conflicting content units are extracted from each conflict pairing, and duplicate content units are removed to obtain a set of conflicting content units. This leads to a strong conflict location set H_k={(u,v)|u∈M_k^+ and X_k(u,v)≥θ_x}. H_k is used to generate the correction constraint strategy for the next step, which includes a constraint set G_k. For each conflict content unit in the set of conflict content units, obtain all conflict pairs related to that conflict content unit; based on all conflict pairs, generate the location information of that conflict content unit. The location information is a set of conflict pair records between the conflict content unit and each corresponding candidate evidence item. Each conflict pair record includes at least the location identifier of the conflict content unit, the identifier of the candidate evidence item, and the specific value of the comprehensive conflict degree corresponding to the pair.

[0057] Therefore, the quantitative control of the correction trigger condition (label R_k) is achieved through the joint judgment rule of the overall conflict severity index C_k, the coverage index V_k, and the candidate evidence credibility index U_k.

[0058] Step S205: Based on the location information, correct the target content and output the corrected result data.

[0059] In one example, S205 includes: generating a correction constraint strategy based on the location information of the conflicting content unit; performing a targeted correction operation on the conflicting content unit according to the correction constraint strategy to generate a set of candidate correction schemes; the set of candidate correction schemes includes multiple candidate correction schemes; performing a comprehensive performance evaluation on each candidate correction scheme to generate an evaluation result; determining the target correction scheme from the multiple candidate correction schemes based on the evaluation result; and outputting correction result data based on the target correction scheme.

[0060] In one example, the evaluation results include candidate correction scores; based on the evaluation results, the target correction scheme is determined from multiple candidate correction schemes, including: determining the candidate correction scheme with the highest candidate correction score as the preliminary correction scheme based on the candidate correction scores of each candidate correction scheme; determining the correction benefit increment and correction acceptability of the preliminary correction scheme, and determining the acceptance decision flag based on the correction benefit increment and correction acceptability; if the acceptance decision flag indicates that the preliminary correction scheme is adopted, then the preliminary correction scheme is determined as the target correction scheme.

[0061] In one example, the method further includes: if the received judgment mark indicates that the preliminary correction scheme is not adopted, then perform a baseline processing on the target content according to the preset mandatory basis conditions; the mandatory basis conditions are that the source credibility of the candidate evidence item used as the basis for correction is greater than the preset mandatory credibility threshold, and the comprehensive conflict degree between the candidate evidence item used as the basis for correction and the content unit to be processed is greater than the preset mandatory conflict threshold; mark the baseline result data after the baseline processing as pending manual review, and output the marked baseline result data.

[0062] For example, based on the location information of the conflicting content units, a correction constraint strategy is generated. The correction constraint strategy includes a constraint set G_k={g_k(1),g_k(2),...,g_k(Q_k)}, and according to G_k, regularized local correction is performed only on the segment where the conflicting content unit is located, forming a set of candidate correction schemes. R'_k={r_k(1),r_k(2),...,r_k(L_k)}, the candidate correction scheme set includes multiple candidate correction schemes r_k(1),r_k(2),...,r_k(L_k).

[0063]

[0064] Where G_k represents the constraint set of the k-th round; g_k(q) represents the q-th constraint term, Q_k represents the number of constraint terms; R'_k represents the set of candidate correction schemes, r_k(j) represents the j-th candidate correction scheme, and L_k represents the number of candidate schemes. Regularized local correction includes at least: replacing conflicting action terms A_u based on evidence units, supplementing missing condition terms C_u, limiting the applicable training objects to terms R_u, and adjusting the timing terms T_u. The above corrections only apply to the segment containing the conflicting proposition and do not change adjacent content that has not experienced conflict. When L_k=0, it indicates that there are no candidate correction schemes that can be formed, resulting in result C, retaining the original result and outputting a "correction failed" flag.

[0065] g_k(q) includes at least the proposition index u, conflict type type, evidence number v, constraint rule rule, and target value value. Candidate modification schemes r_k(j) are obtained by applying the operation sequence Ω_j to the target content T_k. Ω_j includes at least four types of operations: action item replacement, condition item supplementation, applicable object limitation, and timing item adjustment. If multiple constraints exist for the same proposition, they are first sorted according to the source credibility cred(v) and the overall conflict degree X_k(u,v). For example, the constraint with the highest ranking is determined based on both the source credibility cred(v) and the overall conflict degree X_k(u,v), and then modified; or, the constraint with the highest ranking is determined solely based on either the source credibility cred(v) or the overall conflict degree X_k(u,v), and then modified. If the constraints are mutually exclusive, the constraint with higher source credibility and greater overall conflict degree is retained, and the discarded constraint is written to the conflict log Log_k.

[0066] A comprehensive performance evaluation is performed on each candidate modification scheme r_k(j) in the candidate modification scheme set R'_k. For example, the comprehensive performance evaluation involves calculating the candidate modification score J_k(j) based on the scoring function, and selecting the candidate modification scheme r_k(best) with the highest score as the initial modification scheme. The scoring function is as follows:

[0067]

[0068] Where J_k(j) represents the candidate correction score of the j-th candidate correction scheme; C_k(j), V_k(j), and U_k(j) represent the overall conflict severity index, the basis coverage index, and the candidate evidence credibility index corresponding to the candidate correction scheme r_k(j), respectively; Trace_k(j) represents the traceability score; Edit_k(j) represents the editing cost; Risk_k(j) represents the residual risk cost; λ1 to λ6 are non-negative weights and the sum of the weights is 1. r_k(best) represents the candidate correction scheme with the highest score, and J_k(best) represents the candidate correction score of r_k(best).

[0069] Calculate the incremental benefit ΔJ_k and the acceptableness A_k of the preliminary correction scheme, and determine the acceptance decision flag Acc_k based on the incremental benefit and the acceptableness A_k. Perform gating acceptance or minimum disturbance backoff based on the acceptance decision flag Acc_k.

[0070]

[0071]

[0072]

[0073] Where ΔJ_k represents the score gain of the optimal candidate correction scheme relative to the current result (the current result can be the target content of the kth round); A_k represents the correction acceptability; J_k(cur) represents the original score of the target content T_k determined based on the scoring function; Acc_k represents the acceptance decision flag; ξ represents the gain threshold; θ_a represents the acceptance threshold; θ'_c represents the upper limit of conflict after correction; θ'_v represents the lower limit of coverage after correction; η1 to η4 are non-negative weights and satisfy η1+η2+η3+η4=1; This represents the overall conflict severity index of r_k(best); The basis for representing r_k(best) is the coverage index; The credibility index of candidate evidence for r_k(best); This represents the traceability score of r_k(best).

[0074] If the receiving judgment flag indicates that the preliminary correction scheme is adopted (i.e., when Acc_k=1, r_k(best) is received), the preliminary correction scheme is determined to be the target correction scheme, and result E is formed based on the target correction scheme. If the receiving judgment flag indicates that the preliminary correction scheme is not adopted (i.e., when Acc_k=0), a minimum disturbance backoff is performed, retaining only the necessary repair items that meet the preset mandatory basis conditions, and performing a baseline processing on the target content based on the necessary repair items. The baseline result data after the baseline processing is marked as awaiting manual review, and the marked baseline result data D is output. The mandatory basis conditions are that the source credibility of the candidate evidence items used as the basis for correction is greater than the preset mandatory credibility threshold, and the overall conflict degree between the candidate evidence items used as the basis for correction and the content unit to be processed is greater than the preset mandatory conflict threshold.

[0075] Finally, a unified output structure is adopted to aggregate and output results A, B, C, D, and E, and write them to the conflict log Log_k and the traceable evidence chain. Specifically, result A originates from the branch N_match,k=0 in step S204, indicating a lack of verifiable evidence and requiring manual review; result B originates from the branch R_k=0 in step S204, indicating no correction was triggered and the current T_k is directly output; result C originates from the branch L_k=0 in this step, indicating no candidate correction scheme can be formed and correction failure is output; result D originates from the branch Acc_k=0 in this step, indicating output after performing minimum disturbance backoff and awaiting manual review; and result E originates from the branch Acc_k=1 in this step, indicating output of correction result data after receiving the optimal candidate scheme.

[0076] The output correction results data should include at least the processing status (Status), corresponding training content, a list of conflict locations, conflict type labels, a list of correction actions, a conflict log (Log_k), and a traceable chain of evidence. The Status should include at least one of the following: "Lack of verifiable evidence and awaiting manual review," "Direct output," "Correction failed," "Output after rollback and awaiting manual review," and "Output after correction," corresponding to results A, B, C, D, and E, respectively.

[0077] Simultaneously, a conflict log (Log_k) is written. Log_k includes at least the task ID (TaskID), round number (k), proposition index (u), overall conflict level (X_k(u,v)), triggering reason (C_k, V_k, U_k, R_k), candidate correction score (J_k(j)), acceptance decision (ΔJ_k, A_k, Acc_k), final status (Status), and corresponding evidence unit number.

[0078] Therefore, by controlling the correction boundary through constraint sets, training basis, applicable objects, and temporal conditions can be transformed into correction rules. Limiting the correction scope through constraint sets ensures that only conflicting segments are corrected locally, avoiding unbounded overall rewriting. Ranking candidate correction schemes using a scoring function J_k(j) guarantees that the selected scheme simultaneously considers conflict reduction, training basis coverage, evidence reliability, and traceability. Using both incremental benefit and acceptability indicators to determine whether to accept the correction result avoids replacing invalid corrections; when conditions are not met, a minimal disturbance rollback is performed, maintaining the stability of non-conflicting content. The process involves determining candidate correction schemes r_k(j) and the optimal candidate correction scheme r_k(best), deciding whether to accept the correction, and completing the final output and audit documentation after correction, facilitating content auditing and process traceability.

[0079] The method provided in this application, based on a preset knowledge evidence base and preset task conditions, obtains a candidate evidence set. It then acquires the target content to be processed and breaks it down into multiple content units. Based on the candidate evidence set and the multiple content units, it determines the overall conflict level of each content unit. If the overall conflict level determines that the target content meets preset correction trigger conditions, it generates location information based on the overall conflict level. Based on the location information, it performs correction processing on the target content and outputs the correction result data. This solution implements a complete scheme for target content, including "evidence source construction—target content breakdown—proposition-level conflict calculation—correction trigger judgment—constrained local correction—candidate correction scheme scoring—gating acceptance or rollback—unified output audit." When correction is determined to be necessary, the overall conflict level is used as the correction trigger, limiting the correction action to conflict fragments. This allows for targeted local correction, enabling local correction of generated content without expanding the scope of rewriting lesson plans, test questions, or reinforcement content, significantly improving the accuracy of training content modifications.

[0080] Corresponding to the above method, embodiments of this application also provide a conflict correction device for content generation results, such as... Figure 3 As shown, the device includes: The acquisition module 41 is used to acquire a candidate evidence set based on a preset knowledge evidence base and preset task conditions; The splitting module 42 is used to obtain the target content to be processed and split the target content into multiple content units; Module 43 is used to determine the overall degree of conflict of each content unit based on the candidate evidence set and multiple content units; The correction module 44 is used to generate location information based on the comprehensive conflict level if the target content meets the preset correction triggering conditions; and to perform correction processing on the target content based on the location information and output the correction result data.

[0081] The functions of each functional unit in the content generation result conflict correction device provided in the above embodiments of this application can be implemented through the above method steps. Therefore, the specific working process and beneficial effects of each unit in the content generation result conflict correction device provided in the embodiments of this application will not be repeated here.

[0082] This application also provides an electronic device, such as... Figure 4 As shown, it includes a processor 510, a communication interface 520, a memory 530, and a communication bus 540, wherein the processor 510, the communication interface 520, and the memory 530 communicate with each other through the communication bus 540.

[0083] Memory 530 is used to store computer programs; The processor 510 performs the above steps when executing the program stored in the memory 530.

[0084] The communication bus mentioned above can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus can be divided into address bus, data bus, control bus, etc. For ease of illustration, only one thick line is used to represent it in the diagram, but this does not mean that there is only one bus or one type of bus.

[0085] The communication interface is used for communication between the aforementioned electronic devices and other devices.

[0086] The memory may include random access memory (RAM) or non-volatile memory (NVM), such as at least one disk storage device. Optionally, the memory may also be at least one storage device located remotely from the aforementioned processor.

[0087] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0088] The implementation methods and beneficial effects of the various components of the electronic device in the above embodiments for solving the problem can be found in [reference needed]. Figure 1 The steps in the illustrated embodiments are used to implement the electronic device. Therefore, the specific working process and beneficial effects of the electronic device provided in this application will not be repeated here.

[0089] In another embodiment provided in this application, a computer-readable storage medium is also provided, which stores instructions that, when executed on a computer, cause the computer to perform the conflict correction method for the content generation result described in any of the above embodiments.

[0090] In another embodiment provided in this application, a computer program product containing instructions is also provided, which, when run on a computer, causes the computer to execute the conflict correction method for the content generation result described in any of the above embodiments.

[0091] Those skilled in the art will understand that the embodiments in this application can be provided as methods, systems, or computer program products. Therefore, the embodiments in this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the embodiments in this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0092] This application describes embodiments of methods, apparatus (systems), and computer program products according to embodiments of this application with reference to flowchart illustrations and / or block diagrams. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0093] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0094] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0095] Although preferred embodiments have been described in this application, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the embodiments of this application.

[0096] Obviously, those skilled in the art can make various modifications and variations to the embodiments of this application without departing from the spirit and scope of the embodiments of this application. Therefore, if these modifications and variations to the embodiments of this application fall within the scope of the claims in this application and their equivalents, then this application also intends to include these modifications and variations.

Claims

1. A method for correcting conflicts in content generation results, characterized in that, include: Based on a pre-defined knowledge evidence base and pre-defined task conditions, obtain a candidate evidence set; Obtain the target content to be processed, and divide the target content into multiple content units; Based on the candidate evidence set and the multiple content units, determine the overall degree of conflict for each content unit; If the target content is determined to meet the preset correction triggering conditions based on the overall conflict level, then location information is generated according to the overall conflict level. Based on the location information, the target content is corrected, and the corrected result data is output.

2. The method as described in claim 1, characterized in that, The candidate evidence set includes multiple candidate evidence items; based on the candidate evidence set and the multiple content units, the overall conflict level is determined, including: Extract the attribute description information of each content unit across multiple preset attribute dimensions; For each content unit, the conflict metric between the attribute description information of the content unit and each candidate evidence item is determined in each preset conflict dimension. The conflict metrics on multiple preset conflict dimensions are weighted and summed according to preset weighting coefficients to determine the overall degree of conflict between each content unit and each candidate evidence item.

3. The method as described in claim 1, characterized in that, If the target content is determined to meet the preset correction triggering conditions based on the overall conflict level, then location information is generated according to the overall conflict level, including: The overall quality status of the target content is determined based on the degree of comprehensive conflict of each content unit. Determine whether the overall quality status meets the preset correction trigger conditions; If the overall quality status is determined to meet the preset correction triggering conditions, then the conflict content unit is determined according to the overall conflict degree, and the location information of the conflict content unit is generated.

4. The method as described in claim 3, characterized in that, Based on the overall conflict level of each content unit, the overall quality status of the target content is determined, including: The overall conflict severity index of the target content is obtained by statistically aggregating the comprehensive conflict degree of each content unit. Determine the number of evidence items for each content unit in the candidate evidence set; the number of evidence items is the number of candidate evidence items. Determine the number of content units for which the number of evidences is not empty; determine the proportion of the number of these units to the total number of all content units, as the basis coverage index for the target content; Determine the source credibility of candidate evidence items that match each of the content units; statistically aggregate the source credibility to obtain the candidate evidence credibility index of the target content; The overall quality status of the target content is determined based on the overall conflict severity index, the basis coverage index, and the candidate evidence credibility index.

5. The method as described in claim 4, characterized in that, If the preset correction triggering conditions are met, then based on the overall conflict level, the conflict content unit is determined, and the location information of the conflict content unit is generated, including: If the preset correction triggering conditions are met, a set of valid content units is generated based on the content units where the number of evidence units is not empty. For each content unit in the set of valid content units, obtain the overall degree of conflict with each candidate evidence item; The pairing relationship where the overall conflict level exceeds a preset conflict threshold is determined as a conflict pairing; the conflict pairing includes at least the content unit that has a conflict and the candidate evidence item that conflicts with the content unit; Extract the content units that conflict from each of the conflict pairs, and remove duplicates from the same content units to obtain a set of conflicting content units; For each conflicting content unit in the set of conflicting content units, obtain all conflict pairs related to that conflicting content unit; and generate the location information of that conflicting content unit based on all conflict pairs.

6. The method as described in claim 3, characterized in that, Based on the location information, the target content is corrected, and the corrected result data is output, including: Based on the location information of the conflicting content unit, a correction and limitation strategy is generated; According to the aforementioned correction and limitation strategy, a targeted correction operation is performed on the conflicting content unit to generate a candidate correction scheme set; the candidate correction scheme set includes multiple candidate correction schemes. A comprehensive performance evaluation is performed on each of the candidate modification schemes to generate an evaluation result; based on the evaluation result, a target modification scheme is determined from the multiple candidate modification schemes. Based on the target correction scheme, the correction result data is output.

7. The method as described in claim 6, characterized in that, The evaluation results include candidate correction scores; Based on the evaluation results, a target correction scheme is determined from among the multiple candidate correction schemes, including: Based on the candidate correction scores of each candidate correction scheme, the candidate correction scheme with the highest candidate correction score is determined as the preliminary correction scheme; Determine the incremental benefit and acceptability of the preliminary revised plan, and determine the acceptance judgment flag based on the incremental benefit and acceptability of the revised plan; If the receiving determination flag indicates that the preliminary correction scheme is adopted, then the preliminary correction scheme is determined to be the target correction scheme.

8. The method as described in claim 7, characterized in that, The method further includes: If the receiving determination mark indicates that the preliminary correction scheme is not adopted, then the target content is subjected to a minimum protection process according to the preset mandatory basis conditions; the mandatory basis conditions are that the source credibility of the candidate evidence item used as the basis for correction is greater than the preset mandatory credibility threshold, and the comprehensive conflict degree between the candidate evidence item used as the basis for correction and the content unit to be processed is greater than the preset mandatory conflict threshold. Mark the guaranteed result data after the guaranteed result processing as pending manual review, and output the marked guaranteed result data.

9. A conflict correction device for content generation results, characterized in that, The device includes: The acquisition module is used to acquire a set of candidate evidence based on a preset knowledge evidence base and preset task conditions; The splitting module is used to obtain the target content to be processed and split the target content into multiple content units; The determination module is used to determine the overall conflict level of each content unit based on the candidate evidence set and the multiple content units; The correction module is used to generate location information based on the comprehensive conflict level if the target content is determined to meet the preset correction triggering conditions based on the comprehensive conflict level; to perform correction processing on the target content based on the location information; and to output correction result data.

10. An electronic device, characterized in that, The electronic device includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, used to store computer programs; A processor, when executing a program stored in memory, implements the method of any one of claims 1-8.

Images