A mediator performance report generation method based on a large language model

Abstract

The application discloses a mediator performance report generation method based on a large language model, relates to the technical field of performance report generation, and comprises the following steps: obtaining mediation business data and preprocessing the mediation business data into a case performance event sequence; performing stage labeling on the case performance event sequence and converting the case performance event sequence into a stage event vector; configuring a stage index for a DFSMN memory block to form a stage performance memory state; configuring a forward memory coefficient and generating a presequence performance state; configuring a backward memory coefficient to form a bidirectional correction performance state; setting a performance counterargument gate to form a constrained performance state; and calling a large language model to output a performance report. The application improves the matching degree between the performance process of a mediator and performance indicators, reduces the problem that subsequent results are disconnected from previous performance evaluation, and improves the performance report generation efficiency and the reliability of performance conclusions.

Description

Technical Field

[0001] This invention relates to the field of performance report generation technology, and in particular to a method for generating mediator performance reports based on a large language model. Background Technology

[0002] In the field of performance report generation technology, mediator performance reports are typically compiled from handling records, result records, and evaluation records stored within the platform. Existing processing methods mostly revolve around statistical indicators and template texts, first extracting the handling situation within the assessment period from business data, and then filling the statistical results into the report template; some methods utilize intelligent text generation models to summarize and refine the compiled data materials, improving report writing efficiency.

[0003] Existing technologies typically treat mediation data as fragmented indicators, making it difficult to reflect the continuous relationship between different stages of the same case. The mediator's performance in the initial communication, process advancement, and agreement confirmation is simply summarized in the final statistical results, failing to form a sustainable and cumulative memory of performance at each stage, and the report content does not fully reflect the performance process.

[0004] Existing report generation methods often directly draw conclusions based on final results or evaluation content, lacking a mechanism to retrospectively correct prior performance based on subsequent results. When the final case closure, agreement fulfillment, or feedback evaluation is related to the previous performance process, it is difficult to accurately apply these results to the corresponding performance indicators. When intelligent text generation models are involved in report writing, they typically only perform sentence organization functions, lacking performance constraints for abnormal performance information, which easily leads to a disconnect between abnormal performance information and performance conclusions.

[0005] Therefore, how to provide a method for generating mediator performance reports based on a large language model is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a method for generating mediator performance reports based on a large language model. This invention combines DFSMN sequence memory and large language model text generation to collaboratively process the mediator's performance process, subsequent results, and abnormal information, and has the advantages of high report generation efficiency, high performance evaluation accuracy, and strong anomaly constraint capability.

[0007] A method for generating mediator performance reports based on a large language model according to an embodiment of the present invention includes the following steps: Acquire mediation business data within the mediator's assessment period, perform preprocessing, and form a sequence of case performance events; The execution stages of the case performance event sequence are marked, and each event is assigned to the corresponding mediation stage. The event content is transformed into a stage event vector. Configure stage indexes for memory blocks in the DFSMN structure, and write stage event vectors according to the same case and mediation stage to form stage performance memory status; Configure the forward memory coefficient according to the forward processing order of the mediation stage, perform cumulative memory on the performance process characteristics in the stage performance memory state, and generate the previous performance state. Configure backward memory coefficients according to the reverse outcome sequence of the mediation stage, read subsequent outcome information and feedback evaluation information, and apply the read results to the corresponding performance indicators in the previous performance status of the same case to form a two-way correction performance status. Set a performance disproving gate between adjacent stage index memory blocks, read the abnormal information in the stage performance memory state, output the gate value according to the performance index corresponding to the abnormal information, and multiply the vector component corresponding to the same performance index in the bidirectional correction performance state with the gate value to form a constrained performance state. The large language model is invoked to process the constrained performance status into a report text according to the report template, and the mediator performance report is output.

[0008] Optionally, the preprocessing includes performing field unification on the mediation business data, aggregating each event according to the same mediator and the same case, and sorting the aggregated events according to the time of occurrence to form a case performance event sequence.

[0009] Optionally, the stage labeling includes the following steps: According to the mediation stage to which each event belongs, the events in the case performance event sequence are classified into the acceptance stage, communication stage, mediation stage, agreement confirmation stage, case closure stage, and feedback and evaluation stage; Semantic encoding is performed on the event content categorized into each mediation stage; The semantic encoding results are associated with the corresponding mediation stages to form stage event vectors.

[0010] Optionally, the stage index configuration includes the following steps: Set index tags for memory blocks of the DFSMN structure according to the same case and mediation stage; Read the case and mediation stage to which the stage event vector belongs, match the reading result with the index mark, and write the stage event vector into the corresponding stage index memory block when a match is found; According to the order of events corresponding to the stage event vectors in the case performance event sequence, the stage event vectors are stored in the stage index memory block to form the stage performance memory state.

[0011] Optionally, configuring the forward memory coefficients includes the following steps: The memory status of the same case's performance at each stage is arranged in the forward order of acceptance, communication, mediation, agreement confirmation, and case closure. In the DFSMN structure, the forward memory coefficient is set according to the relationship between the reading phase and the preceding mediation phase in the forward processing sequence; Weighted readings are performed on the vector components of the stage performance memory state in the preceding mediation stage according to the forward memory coefficient. The performance process features are extracted from the weighted reading results to form the preceding performance state.

[0012] Optionally, the extraction of the performance process features includes extracting components corresponding to the continuity of handling actions, the completeness of processing progress, the completeness of material submission, and the completeness of communication feedback records from the weighted reading results according to the source of the vector components in the stage event vector, and writing the extracted components into the previous performance status of the same case.

[0013] Optionally, the formation of the bidirectional correction performance status includes the following steps: The stages of the case's performance memory are arranged in reverse order of results: feedback and evaluation stage, case closure stage, agreement confirmation stage, mediation stage, communication stage, and acceptance stage. The acceptance stage, communication stage, mediation stage, agreement confirmation stage, and case closure stage in the forward processing sequence are used as correction stages. The stages following the correction stages are defined as backward reading stages. Backward memory coefficients are configured according to the correspondence between the correction stages and backward reading stages in the reverse result sequence. The DFSMN structure reads subsequent result information in the backward reading stage according to the backward memory coefficient, and reads feedback evaluation information in the feedback evaluation stage. The retrieved subsequent results and feedback evaluation information are applied to the corresponding performance indicators in the preceding performance status of the same case, forming a two-way performance status correction.

[0014] Optionally, the subsequent result information and feedback evaluation information are applied to the corresponding performance indicators as follows: according to the relationship between the result status and the mediation effect indicator, the agreement performance status and the performance quality indicator, the case closure status and the processing completion indicator, and the evaluation content and complaint record and the service evaluation indicator, the backward reading results are written into the corresponding performance indicators in the preceding performance status to form a two-way correction performance status.

[0015] Optionally, the formation of the constrained performance state includes the following steps: Set a performance disproving gate between adjacent stage index memory blocks, and connect the adjacent stage index memory blocks; The performance verification gate reads abnormal information from the performance memory status during the performance phase and determines the performance indicators corresponding to the abnormal information. Output a gate value according to the performance indicator corresponding to the abnormal information, and map the gate value to the same performance indicator in the two-way correction performance status; Multiply the vector components corresponding to the same performance indicators in the bidirectional performance state with the gate value to form the constrained performance state.

[0016] Optionally, the report textification process includes the following steps: According to the performance indicator field in the report template, read the vector components of the corresponding performance indicator in the constrained performance status; The read vector components are converted into corresponding performance indicator text, and the indicator text is then filled into the corresponding performance indicator field in the report template. The large language model is invoked to organize the sentences in the report template after the indicator text is filled in, and a mediator performance report is generated that includes information on the performance process, subsequent results, feedback and evaluation, abnormal information and performance indicator results.

[0017] The beneficial effects of this invention are: (1) This invention, through stage labeling and stage index configuration, writes the performance information of the same case in different mediation stages into the DFSMN memory block. The performance process is continuously saved and called in stages, avoiding the existing reports that only summarize scattered business data, and improving the completeness of the performance process.

[0018] (2) This invention accumulates prior performance information through forward memory coefficients and reads subsequent result information and feedback evaluation information through backward memory coefficients. The process performance and final result affect the corresponding performance indicators, thereby improving the consistency between performance evaluation and actual mediation effect.

[0019] (3) This invention sets a performance disproving gate between adjacent stage index memory blocks, outputs a gate value based on abnormal information, and constrains the vector components corresponding to the same performance indicators, thereby reducing the problem of weakening abnormal information in report generation and improving the objectivity of performance reports. Attached Figure Description

[0020] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is an overall flowchart of a mediator performance report generation method based on a large language model proposed in this invention; Figure 2 This is a schematic diagram of the stage index configuration of the DFSMN memory block in this invention; Figure 3 This is a schematic diagram illustrating the collaborative processing of forward memory coefficients, backward memory coefficients, and performance counter-proof gates in this invention. Detailed Implementation

[0021] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0022] refer to Figures 1-3 A method for generating mediator performance reports based on a large language model includes the following steps: Acquire mediation business data within the mediator's assessment period, perform preprocessing, and form a sequence of case performance events; The execution stages of the case performance event sequence are marked, and each event is assigned to the corresponding mediation stage. The event content is transformed into a stage event vector. Configure stage indexes for memory blocks in the DFSMN structure, and write stage event vectors according to the same case and mediation stage to form stage performance memory status; Configure the forward memory coefficient according to the forward processing order of the mediation stage, perform cumulative memory on the performance process characteristics in the stage performance memory state, and generate the previous performance state. Configure backward memory coefficients according to the reverse outcome sequence of the mediation stage, read subsequent outcome information and feedback evaluation information, and apply the read results to the corresponding performance indicators in the previous performance status of the same case to form a two-way correction performance status. Set a performance disproving gate between adjacent stage index memory blocks, read the abnormal information in the stage performance memory state, output the gate value according to the performance index corresponding to the abnormal information, and multiply the vector component corresponding to the same performance index in the bidirectional correction performance state with the gate value to form a constrained performance state. The large language model is invoked to process the constrained performance status into a report text according to the report template, and the mediator performance report is output.

[0023] In this embodiment, the preprocessing includes unifying the fields of the mediation business data, aggregating each event according to the same mediator and the same case, and sorting the aggregated events according to the time of occurrence of the events to form a case performance event sequence.

[0024] After standardizing the fields, mediation data from different sources are organized into a single event format. The event format must at least record the mediator's identifier, case identifier, event source, handling action, event content, event sequence, and result. For records from different sources within the same case but corresponding to the same handling action, the chronological order is preserved. For records lacking an event sequence, the sequence is filled in based on the event source and the position of the handling action within the case flow. Through this process, subsequent stages of annotation are performed based on the unified event format, preventing the same performance content from being repeatedly identified or omitted due to different sources.

[0025] In this embodiment, the stage labeling includes the following steps: According to the mediation stage to which each event belongs, the events in the case performance event sequence are classified into the acceptance stage, communication stage, mediation stage, agreement confirmation stage, case closure stage, and feedback and evaluation stage; The acceptance stage is the initial step after a case enters the mediation process. It records the case receipt, acceptance confirmation, verification of basic information, and the mediator's involvement. The events recorded during the acceptance stage determine whether the case has entered the effective processing stage and provide initial evidence for subsequent communication and mediation.

[0026] After a case is accepted, it enters the communication phase. This phase involves contacting the parties, confirming their claims, understanding the content of the dispute, and exchanging opinions. The content of the case reflects whether the mediator has consistently followed up on feedback from both parties and whether the communication process has remained continuous, providing a basis for subsequent substantive mediation based on the facts of the dispute and the attitudes of both parties.

[0027] During the mediation phase, the mediator sorts out the key points of the dispute based on the results of the previous communication, and promotes the mediation process around the solution, allocation of responsibilities and feedback. This reflects the mediator's actual handling of the case dispute and whether the mediation behavior is consistent with the content of the previous communication, serving as the basis for judging the quality of mediation in the performance indicators.

[0028] The agreement confirmation stage occurs after the mediation opinion is formed. It is used to record the confirmation of the agreement content, the confirmation of the performance matters, the confirmation results of the parties, and the confirmation status of the agreement. It reflects whether the mediation result has been effectively confirmed, avoids judging the mediation effect solely based on the mediation process or the case closing status, and thus provides a clear basis for subsequent result correction.

[0029] The case closure stage is used to record the case handling results, case closure status, result archiving, and case end information. It follows the previous processes of acceptance, communication, mediation, and agreement confirmation, reflects whether the case has been finally handled, and also forms a corresponding relationship with the previous performance of duties, providing result-side information for subsequent performance status correction.

[0030] The feedback evaluation phase occurs after a case is closed, recording the parties' evaluations, feedback content, evaluation levels, complaint records, and follow-up information. Unlike the case closure phase, which reflects whether the case handling was completed, the feedback evaluation phase reflects whether the parties have accepted or evaluated the case handling outcome. The event content generated during the feedback evaluation phase serves as a backward retrieval object, used to correct service evaluation indicators; when feedback anomalies exist, the corresponding vector components in the feedback evaluation phase serve as the source of anomaly information for performance verification.

[0031] Read the source of nodes, handling actions, event types, and event texts in the case performance events. First, classify the events into acceptance, case closure, and feedback evaluation categories according to the source of nodes. Then, classify the events into communication, mediation, and agreement confirmation categories according to the handling actions. When the same event involves several stages, determine the mediation stage in the order of priority of source of nodes, secondly of handling actions, and thirdly of event text for auxiliary verification.

[0032] Semantic encoding is performed on the event content categorized into each mediation stage; After cleaning the text of the event content, standardizing the roles and titles, and standardizing the time expression, the semantic features corresponding to the mediation behavior, dispute content, processing progress, material submission, communication feedback, agreement confirmation, result status and evaluation feedback are extracted. The semantic features are then encoded into vector components that can participate in the DFSMN operation to form the semantic encoding result.

[0033] The semantic encoding results are associated with the corresponding mediation stages to form stage event vectors.

[0034] According to the mediation stage to which the event belongs, the vector components in the semantic coding results are written into the corresponding component areas of the stage event vector; the same type of semantic coding results are written into different component areas in different mediation stages, and the stage event vector reflects both the event content and the mediation stage.

[0035] After the stage event vectors are formed, the event content in the same case is not only included in the report generation in text form, but is also written into the corresponding component areas according to the mediation stage and subsequent processing needs. Content related to mediation behavior is written into the mediation behavior component, content related to case progress is written into the processing progress component, the focus of the dispute and the content of the dispute is written into the dispute content component, the results of supplementary materials and submission of materials are written into the material submission component, the contact between the parties and the exchange of opinions are written into the communication feedback component, the result of agreement confirmation is written into the agreement confirmation component, the result of case closure is written into the result status component, and the evaluation feedback content is written into the evaluation feedback component. The vector component of the acceptance stage records the case reception and initial processing status, the vector component of the communication stage retains the continuous record of communication between the two parties, the vector component of the mediation stage inherits the mediation behavior, the content of the dispute, and the processing progress, the vector component of the agreement confirmation stage records the agreement confirmation status, and the vector components of the case closure stage and the feedback evaluation stage inherit the case result and the evaluation content. The stage event vector simultaneously inherits the event semantics, the mediation stage, and the component area, providing clear input objects for subsequent stage index writing, forward memory, backward correction, and performance counter-evidence reading, avoiding subsequent processing from judging the event's target solely based on text similarity.

[0036] In this embodiment, the stage index configuration includes the following steps: Set index tags for memory blocks of the DFSMN structure according to the same case and mediation stage; Index tags record the same case and mediation stage; index tags are used to define the write objects of the stage event vector.

[0037] Read the case and mediation stage to which the stage event vector belongs, match the reading result with the index mark, and write the stage event vector into the corresponding stage index memory block when a match is found; Read the case and mediation stage to which the stage event vector belongs, compare the case to the same case in the index, and compare the mediation stage to the mediation stage in the index; if the two match, write the stage event vector into the corresponding stage index memory block.

[0038] According to the order of events corresponding to the stage event vectors in the case performance event sequence, the stage event vectors are stored in the stage index memory block to form the stage performance memory state.

[0039] After the stage event vector is written, the stage index memory block stores the stage event vectors within the same mediation stage of the same case. The stage event vectors are kept in the order of the case's performance event sequence. The state of the saved stage index memory block serves as the stage performance memory state.

[0040] When storing stage event vectors in the stage index memory block, the index markers only allow stage event vectors from the same case and the same mediation stage to be written to the current stage index memory block. If the case to which the stage event vector belongs is inconsistent with the case in the index marker, or the mediation stage to which it belongs is inconsistent with the mediation stage in the index marker, it will not be written to the current stage index memory block. Through write control, the data boundaries between different mediation stages of the same case are preserved, and the stage performance content of different cases will not be mixed, providing a stable stage performance memory state for subsequent forward memory and backward correction.

[0041] In this embodiment, the configuration of the forward memory coefficients includes the following steps: The memory status of the same case's performance at each stage is arranged in the forward order of acceptance, communication, mediation, agreement confirmation, and case closure. In the DFSMN structure, the forward memory coefficient is set according to the relationship between the reading phase and the preceding mediation phase in the forward processing sequence; Following the forward processing sequence of acceptance, communication, mediation, agreement confirmation, and case closure, stages are assigned numbers 1 to 5. For the i-th reading stage within the same case, stages with numbers less than i are designated as preceding mediation stages. For the preceding mediation stage with number j, the difference between i and j is calculated, and the reciprocal of the difference plus 1 is used as the initial coefficient. All initial coefficients under the same reading stage are summed, and each initial coefficient is divided by the summation result to obtain the forward memory coefficient of the corresponding preceding mediation stage.

[0042] Weighted readings are performed on the vector components of the stage performance memory state in the preceding mediation stage according to the forward memory coefficient. The performance process features are extracted from the weighted reading results to form the preceding performance state.

[0043] After the forward memory coefficients are formed, the DFSMN structure sequentially reads the stage performance memory states of each preceding mediation stage under the same reading stage, and extracts vector components related to mediation behavior, processing progress, material submission, and communication feedback from the stage performance memory states; the vector components extracted from each preceding mediation stage are multiplied by the forward memory coefficients corresponding to the preceding mediation stage, and the multiplication results of the same component positions are summed to form a weighted reading result.

[0044] In forward memory processing, the set of preceding mediation stages differs for different retrieval stages. The later the retrieval stage, the more preceding mediation stages are invoked; the closer the preceding mediation stage is to the retrieval stage, the more directly it affects the current performance status, and therefore the larger its corresponding forward memory coefficient; preceding mediation stages farther from the retrieval stage still participate in the weighted retrieval, but their corresponding forward memory coefficients are relatively smaller. The preceding performance status retains the early processing steps while highlighting the impact of the nearest stage on the current performance judgment, avoiding the simple averaging of all preceding events.

[0045] In this embodiment, the extraction of the performance process features includes extracting the components corresponding to the continuity of handling actions, the completeness of processing progress, the completeness of material submission, and the completeness of communication feedback records from the weighted reading results according to the source of the vector components in the stage event vector, and writing the extracted components into the previous performance status of the same case.

[0046] The preceding performance status is not a simple superposition of the vectors from the preceding mediation stages, but rather an extraction of key information reflecting the performance process based on weighted readings. Specifically, the continuity of handling actions component describes whether the mediator is continuously advancing the work according to the case procedure; the completeness of processing progress component reflects whether there are any stalls or missing links in the case handling process; the completeness of material submission component records whether relevant materials are continuously supplemented and retained as required; and the completeness of communication and feedback records reflects whether the communication and feedback process between the parties remains consistent. These components collectively form the process-side foundation before backward correction. When subsequent result information and feedback evaluation information enter the processing stage, their relevant impacts are preferentially mapped onto the already formed performance process components, rather than directly affecting the final report text.

[0047] In this embodiment, the formation of the bidirectional correction performance status includes the following steps: The stages of the case's performance memory are arranged in reverse order of results: feedback and evaluation stage, case closure stage, agreement confirmation stage, mediation stage, communication stage, and acceptance stage. The acceptance stage, communication stage, mediation stage, agreement confirmation stage, and case closure stage in the forward processing sequence are used as correction stages. The stages following the correction stages are defined as backward reading stages. Backward memory coefficients are configured according to the correspondence between the correction stages and backward reading stages in the reverse result sequence. When configuring the backward memory coefficient, the stages are assigned numbers 1 to 6 sequentially according to the forward processing order: acceptance stage, communication stage, mediation stage, agreement confirmation stage, case closure stage, and feedback evaluation stage. For the i-th correction stage within the same case, the stage with a number greater than i is designated as the backward reading stage. For the backward reading stage with a number of j, the difference between j and i is calculated, and the reciprocal of the difference plus 1 is used as the initial coefficient. All initial coefficients under the same correction stage are summed, and each initial coefficient is divided by the summation result to obtain the backward memory coefficient for the corresponding backward reading stage.

[0048] The DFSMN structure reads subsequent result information in the backward reading stage according to the backward memory coefficient, and reads feedback evaluation information in the feedback evaluation stage. When reading subsequent result information, the vector components corresponding to the agreement performance status are extracted from the agreement confirmation stage, and the result status and the vector components corresponding to the case closure status are extracted from the case closure stage. When the feedback evaluation stage is used as a backward reading stage, the vector components corresponding to the evaluation content, evaluation level, and complaint record are extracted from the feedback evaluation stage. The above vector components are multiplied by their corresponding backward memory coefficients and then summed to form the backward reading result.

[0049] After the backward read results are generated, they are written into the preceding performance status according to the corresponding relationship of performance indicators. The result status is used to correct the mediation effect indicator, the agreement performance status is used to correct the performance quality indicator, the case closure status is used to correct the processing completion indicator, and the evaluation content and complaint records are used to correct the service evaluation indicator.

[0050] The retrieved subsequent results and feedback evaluation information are applied to the corresponding performance indicators in the preceding performance status of the same case, forming a two-way performance status correction.

[0051] In this embodiment, the subsequent result information and feedback evaluation information act on the corresponding performance indicators, including the relationship between the result status and the mediation effect indicator, the agreement performance status and the performance quality indicator, the case closure status and the processing completion indicator, and the evaluation content and complaint record and the service evaluation indicator. The backward reading results are written into the corresponding performance indicators in the preceding performance status to form a two-way correction performance status.

[0052] When writing backward read results, the system first searches for the corresponding component position in the preceding performance status according to the performance indicator name. If the corresponding component already exists in the preceding performance status, the backward read result is written to the same indicator position, forming a correction result under the performance indicator together with the original performance process component. If there is no corresponding component in the preceding performance status, but the backward read result already points to a certain performance indicator, the backward read result is written as supplementary content to the performance indicator. If neither the preceding performance status nor the backward read result can correspond to the same performance indicator, no bidirectional correction performance status is written. In this way, subsequent results will not form an evaluation conclusion independently of the preceding performance process, but will instead form a correspondence with the preceding performance content under the same performance indicator.

[0053] In this embodiment, the formation of the constrained performance state includes the following steps: Set a performance disproving gate between adjacent stage index memory blocks, and connect the adjacent stage index memory blocks; In the same case, adjacent stage index memory blocks are determined according to the adjacent relationship of the acceptance stage, communication stage, mediation stage, agreement confirmation stage, case closure stage, and feedback evaluation stage, and a performance counter-evidence gate is set between every two adjacent stage index memory blocks.

[0054] The performance verification gate reads abnormal information from the performance memory status during the performance phase and determines the performance indicators corresponding to the abnormal information. Abnormal information refers to the non-normal performance components recorded in the stage performance memory state, originating from the corresponding vector components of processing time, material submission, communication records, agreement confirmation, and feedback evaluation in the stage event vector. Abnormal processing time results in overdue processing information; missing material submissions result in missing material information; interrupted communication records result in communication interruption information; missing agreement confirmations result in unconfirmed agreement information; and abnormal feedback evaluations result in abnormal feedback information. Overdue processing corresponds to processing efficiency indicators, missing materials correspond to procedural compliance indicators, interrupted communication corresponds to communication quality indicators, unconfirmed agreements correspond to mediation effectiveness indicators, and abnormal feedback corresponds to service evaluation indicators.

[0055] Anomaly component values ​​are obtained from the corresponding components in the stage event vector. Overdue processing information is determined based on the deviation between the processing time component and the interval between adjacent mediation stages; missing materials information is determined based on the proportion of missing records in the materials submission component; communication interruption information is determined based on the gaps in continuous records in the communication feedback component; unconfirmed agreements information is determined based on the missing confirmation status in the agreement confirmation component; and feedback anomaly information is determined based on complaint records, abnormal evaluation content, and abnormal evaluation level in the evaluation feedback component. All the above anomaly component values ​​are uniformly converted to values ​​within the range of 0 to 1, with larger values ​​indicating a higher degree of anomaly. After reading the anomaly component values, the performance verification gate generates corresponding gate values, thereby allowing different degrees of anomaly to exert different levels of constraint on the corresponding performance indicators.

[0056] Output a gate value according to the performance indicator corresponding to the abnormal information, and map the gate value to the same performance indicator in the two-way correction performance status; After reading the vector component corresponding to the abnormal information, the performance counter-verification gate uses this vector component as the abnormal component value. The gate value is formed by subtracting the abnormal component value from 1. When outputting the gate value, the performance indicator corresponding to the abnormal information is recorded simultaneously. The performance indicator is used as the comparison basis to find the vector component corresponding to the performance indicator with the same name in the two-way performance correction status.

[0057] Multiply the vector components corresponding to the same performance indicators in the bidirectional performance state with the gate value to form the constrained performance state.

[0058] When a match is found, the gating value is multiplied by the vector component of the corresponding performance indicator to obtain the constrained vector component; performance indicators that do not match a gating value retain the original vector component in the two-way corrected performance state. The constrained vector component and the retained original vector component are combined to form the constrained performance state.

[0059] The gating value output by the performance counter-verification gate does not replace the performance indicator itself, but is used to adjust the vector components corresponding to the same performance indicator in the two-way corrected performance state. The higher the abnormal component value, the lower the gating value, and the stronger the constraint on the corresponding performance indicator; the lower the abnormal component value, the closer the gating value is to the original state, and the weaker the constraint on the corresponding performance indicator. For performance indicators without abnormal information, the original vector components in the two-way corrected performance state are retained. Abnormal performance content does not exist independently of the performance indicator and directly affects the final expression of the same performance indicator.

[0060] In this embodiment, the report textification process includes the following steps: According to the performance indicator field in the report template, read the vector components of the corresponding performance indicator in the constrained performance status; By comparing performance indicator names, the fields of processing efficiency, procedural standardization, communication quality, mediation effectiveness, performance quality, and service evaluation in the report template are retrieved. Performance indicators with matching names are found in the constrained performance status, and their corresponding vector components are extracted.

[0061] The read vector components are converted into corresponding performance indicator text, and the indicator text is then filled into the corresponding performance indicator field in the report template. Based on the performance indicator name, the extracted vector components are converted into corresponding indicator text. The indicator text includes the performance indicator name, indicator result description, corresponding content for abnormal information, and suggested content. When abnormal information is associated with the corresponding performance indicator in the constrained performance status, corresponding suggested content is written according to the name of the abnormal information: for overdue processing, write "shorten the processing interval and promptly advance subsequent nodes"; for missing materials, write "replenish material record and verify material submission node"; for interrupted communication, write "supplement communication record and maintain continuous feedback from the parties involved"; for unconfirmed agreements, write "verify agreement confirmation node and supplement confirmation record"; for abnormal feedback, write "review feedback content and follow up on complaint handling record".

[0062] The large language model is invoked to organize the sentences in the report template after the indicator text is filled in, and a mediator performance report is generated that includes information on the performance process, subsequent results, feedback and evaluation, abnormal information and performance indicator results.

[0063] The large language model receives the report template after the indicator text is filled in, the summary of the case performance event sequence, and the performance indicator results corresponding to the constrained performance status. It organizes the statements according to the field order in the report template and generates a mediator performance report.

[0064] During report text processing, the large language model does not directly generate performance conclusions based on the original mediation business data. Instead, it receives the indicator text that has already been transformed from the constrained performance state. The fields of processing efficiency, procedural standardization, communication quality, mediation effect, performance quality, and service evaluation in the report template correspond to the performance indicator vector components with the same names in the constrained performance state. Before generating the report, the large language model first reads the performance indicator vector components with the same names as the template fields from the constrained performance state, and then organizes the reading results into descriptions of indicator results, explanations of the performance process, explanations of subsequent results, and corresponding content for abnormal information. For template fields where the corresponding performance indicator vector components are not read, the large language model does not automatically add performance conclusions. The large language model organizes statements according to the field order in the report template, integrating the indicator text corresponding to each performance indicator into complete report paragraphs. The final generated mediator performance report inherits the processing results of the performance memory state, the two-way correction performance state, and the constrained performance state, reducing the problem of report conclusions deviating from previous calculation results.

[0065] Example 1: To verify the feasibility of this invention in practice, it was applied to the generation of mediator performance reports on an online mediation platform. The platform contains mediators' records of work completed within the assessment period. Existing report generation methods typically extract the number of cases handled, the results, and the evaluation content, then generate a report according to a fixed template. While this method can complete the basic summary, in practice, it easily breaks down the performance process in the same case into scattered statistical items. The report shows the final result but fails to reflect the mediator's continuous performance throughout the case process. Especially when a case involves multiple communications, supplementary materials, agreement confirmation, and subsequent feedback, existing reports often only retain the final conclusion and cannot explain the influence relationship between the preceding performance process and the subsequent results.

[0066] In this embodiment, mediation business data generated by mediators within the assessment period is first acquired. Data from different sources is then standardized in terms of fields. Actions, communication records, document submissions, agreement confirmations, case closure results, and evaluation feedback generated by the same mediator in the same case are grouped into performance events for the same case. These events are then arranged in chronological order to form a sequence of performance events. After forming the event sequence, each event is labeled with a stage, and the event content is converted into a stage event vector. The performance content, originally in text form, is then processed further within a DFSMN structure.

[0067] After generating the stage event vectors, stage indexing is configured for the memory blocks in the DFSMN structure. Index markers are set for the same case and mediation stage, and the stage event vectors are written into the corresponding stage index memory blocks, maintaining the order of the stage event vectors in the case's performance event sequence. Through this process, the performance content of the same case in different mediation stages no longer exists as isolated records, but rather forms a stage performance memory state in the DFSMN structure. When generating reports subsequently, the performance succession relationship between different stages can be retrieved from the stage performance memory state, avoiding reports that only present simple statistical results.

[0068] A forward memory coefficient is configured according to the forward processing sequence of the acceptance stage, communication stage, mediation stage, agreement confirmation stage, and case closure stage to accumulate and remember the performance characteristics of the previous mediation stage. For a case entering the agreement confirmation stage, vector components related to mediation behavior, processing progress, material submission, and communication feedback in previous stages are read and weighted according to the forward memory coefficient to form the previous performance status. The previous performance status reflects the cumulative impact of each stage before the current reading stage on the performance process, and the report illustrates the mediator's continuous performance in the process.

[0069] The report reads subsequent results and feedback evaluation information in reverse order, applying these results to the corresponding performance indicators in the preceding performance status. If a case reaches a clear conclusion at the closing stage, the case closure status is applied to the processing completion indicator; if the agreement performance affects the mediation effect, the agreement performance status is applied to the mediation effect indicator; if there are complaint records or abnormal evaluations in the feedback evaluation, the corresponding content is applied to the service evaluation indicator. The performance evaluation in the report no longer relies solely on preceding performance records but can also incorporate subsequent results and feedback evaluations to adjust performance indicators.

[0070] A performance rebuttal gate is set between adjacent stage index memory blocks. The performance rebuttal gate reads the abnormal performance components from the stage performance memory state and determines the performance indicators corresponding to the abnormal information. If there is overdue processing, it corresponds to the processing efficiency indicator; if there is missing materials, it corresponds to the procedural compliance indicator; if there is communication interruption, it corresponds to the communication quality indicator; if there is unconfirmed agreement, it corresponds to the mediation effect indicator; if there is abnormal feedback, it corresponds to the service evaluation indicator. The performance rebuttal gate outputs a gating value based on the abnormal information, which is multiplied by the vector component corresponding to the same performance indicator in the two-way correction performance state to form a constrained performance state. When generating reports, the large language model does not freely organize conclusions from the previous calculation results, but completes text processing based on the constrained performance state.

[0071] The system reads the vector components of the constrained performance status from the performance indicator fields in the report template, converts these vector components into corresponding indicator text, and fills them into the relevant fields in the report template. The large language model receives the report template with the indicator text filled in, a summary of the case performance event sequence, and the performance indicator results. It then organizes the report statements and outputs a mediator performance report. The output report includes a description of the performance process, as well as subsequent results, feedback evaluations, anomaly information, and performance indicator results, allowing managers to more clearly see the basis for the mediator's performance conclusions.

[0072] To verify the effectiveness of this invention, a comparison was made using the same batch of anonymized mediation business data. Existing generation methods use conventional statistical results and template text to generate reports, while this invention uses stage index DFSMN memory, forward memory, backward correction, and performance counter-evidence gate constraints to generate reports. The comparison results are shown in Table 1.

[0073] Table 1. Comparison of Mediator Performance Report Generation Effects

[0074] As shown in Table 1, this invention outperforms existing methods in terms of report completeness, accuracy of indicator judgment, and generation efficiency. The coverage rate of performance stages increased from 78.6% to 96.4%, indicating that this invention, through stage labeling and stage index configuration, continuously saves performance information of the same case at different stages of handling into the stage performance memory state. Therefore, report generation no longer relies solely on scattered statistical results, resulting in a more complete presentation of the mediator's performance process.

[0075] The accuracy rate of performance indicators improved from 81.2% to 94.7%, indicating that this invention, before generating the report, had already correlated the performance process, subsequent results, and feedback evaluations through forward memory and backward correction. The preceding performance status reflects the mediator's continuous performance during the case process, while subsequent results and feedback evaluations influence the corresponding performance indicators. The conclusions in the report regarding processing efficiency, communication quality, mediation effectiveness, and service evaluation are closer to the actual business process, reducing the problem of inaccurate correspondence between results and indicators in existing methods.

[0076] The accuracy rate of process outcome correlation increased from 73.8% to 92.6%, further demonstrating that this invention does not simply write the case closure result or evaluation content into the report, but rather utilizes a backward memory coefficient to apply subsequent results back to the preceding performance status, forming a traceable link between previous performance and the final result. The manual review and modification rate decreased from 26.3% to 8.9%, indicating a reduction in the need for manual revision of performance conclusions, supplementary process descriptions, or adjustments to indicator correspondences in the generated report. The time required to generate a single report was shortened from 6.8 minutes to 1.9 minutes, demonstrating that this invention improves the accuracy and interpretability of reports while reducing the workload of manual compilation, review, and modification, thus increasing the efficiency of mediator performance report generation.

[0077] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for generating mediator performance reports based on a large language model, characterized in that, The steps include the following: Acquire mediation business data within the mediator's assessment period, perform preprocessing, and form a sequence of case performance events; The execution stages of the case performance event sequence are marked, and each event is assigned to the corresponding mediation stage. The event content is transformed into a stage event vector. Configure stage indexes for memory blocks in the DFSMN structure, and write stage event vectors according to the same case and mediation stage to form stage performance memory status; Configure the forward memory coefficient according to the forward processing order of the mediation stage, perform cumulative memory on the performance process characteristics in the stage performance memory state, and generate the previous performance state. Configure backward memory coefficients according to the reverse outcome sequence of the mediation stage, read subsequent outcome information and feedback evaluation information, and apply the read results to the corresponding performance indicators in the previous performance status of the same case to form a two-way correction performance status. Set a performance disproving gate between adjacent stage index memory blocks, read the abnormal information in the stage performance memory state, output the gate value according to the performance index corresponding to the abnormal information, and multiply the vector component corresponding to the same performance index in the bidirectional correction performance state with the gate value to form a constrained performance state. The large language model is invoked to process the constrained performance status into a report text according to the report template, and the mediator performance report is output.

2. The method for generating mediator performance reports based on a large language model according to claim 1, characterized in that, The preprocessing includes standardizing the fields of the mediation business data, grouping each event according to the same mediator and the same case, and sorting the grouped events according to the time of occurrence to form a case performance event sequence.

3. The method for generating mediator performance reports based on a large language model according to claim 1, characterized in that, The stage labeling includes the following steps: According to the mediation stage to which each event belongs, the events in the case performance event sequence are classified into the acceptance stage, communication stage, mediation stage, agreement confirmation stage, case closure stage, and feedback and evaluation stage; Semantic encoding is performed on the event content categorized into each mediation stage; The semantic encoding results are associated with the corresponding mediation stages to form stage event vectors.

4. The method for generating mediator performance reports based on a large language model according to claim 1, characterized in that, The stage index configuration includes the following steps: Set index tags for memory blocks of the DFSMN structure according to the same case and mediation stage; Read the case and mediation stage to which the stage event vector belongs, match the reading result with the index mark, and write the stage event vector into the corresponding stage index memory block when a match is found; According to the order of events corresponding to the stage event vectors in the case performance event sequence, the stage event vectors are stored in the stage index memory block to form the stage performance memory state.

5. The method for generating mediator performance reports based on a large language model according to claim 1, characterized in that, The configuration of the forward memory coefficients includes the following steps: The memory status of the same case's performance at each stage is arranged in the forward order of acceptance, communication, mediation, agreement confirmation, and case closure. In the DFSMN structure, the forward memory coefficient is set according to the relationship between the reading phase and the preceding mediation phase in the forward processing sequence; Weighted readings are performed on the vector components of the stage performance memory state in the preceding mediation stage according to the forward memory coefficient. The performance process features are extracted from the weighted reading results to form the preceding performance state.

6. The method for generating mediator performance reports based on a large language model according to claim 5, characterized in that, The extraction of the performance process features includes extracting components corresponding to the continuity of handling actions, the completeness of processing progress, the completeness of material submission, and the completeness of communication feedback records from the weighted reading results according to the source of the vector components in the stage event vector, and writing the extracted components into the previous performance status of the same case.

7. The method for generating mediator performance reports based on a large language model according to claim 1, characterized in that, The formation of the bidirectional correction performance status includes the following steps: The stages of the case's performance memory are arranged in reverse order of results: feedback and evaluation stage, case closure stage, agreement confirmation stage, mediation stage, communication stage, and acceptance stage. The acceptance stage, communication stage, mediation stage, agreement confirmation stage, and case closure stage in the forward processing sequence are used as correction stages. The stages following the correction stages are defined as backward reading stages. Backward memory coefficients are configured according to the correspondence between the correction stages and backward reading stages in the reverse result sequence. The DFSMN structure reads subsequent result information in the backward reading stage according to the backward memory coefficient, and reads feedback evaluation information in the feedback evaluation stage. The retrieved subsequent results and feedback evaluation information are applied to the corresponding performance indicators in the preceding performance status of the same case, forming a two-way performance status correction.

8. The method for generating mediator performance reports based on a large language model according to claim 7, characterized in that, The subsequent results and feedback evaluation information are applied to the corresponding performance indicators as follows: according to the relationship between the result status and the mediation effect indicator, the agreement performance status and the performance quality indicator, the case closure status and the processing completion indicator, and the evaluation content and complaint record and the service evaluation indicator, the backward reading results are written into the corresponding performance indicators in the preceding performance status to form a two-way correction performance status.

9. The method for generating mediator performance reports based on a large language model according to claim 1, characterized in that, The formation of the constrained performance state includes the following steps: Set a performance disproving gate between adjacent stage index memory blocks, and connect the adjacent stage index memory blocks; The performance verification gate reads abnormal information from the performance memory status during the performance phase and determines the performance indicators corresponding to the abnormal information. Output a gate value according to the performance indicator corresponding to the abnormal information, and map the gate value to the same performance indicator in the two-way correction performance status; Multiply the vector components corresponding to the same performance indicators in the bidirectional performance state with the gate value to form the constrained performance state.

10. The method for generating mediator performance reports based on a large language model according to claim 1, characterized in that, The report textification process includes the following steps: According to the performance indicator field in the report template, read the vector components of the corresponding performance indicator in the constrained performance status; The read vector components are converted into corresponding performance indicator text, and the indicator text is then filled into the corresponding performance indicator field in the report template. The large language model is invoked to organize the sentences in the report template after the indicator text is filled in, and a mediator performance report is generated that includes information on the performance process, subsequent results, feedback and evaluation, abnormal information and performance indicator results.

Images