A large model-based record review method, device and medium

Abstract

The application discloses a record examination method and device based on a large model and a medium. The method obtains record data, case process data, examination rule version data and associated evidence materials, performs question and answer round division on the record and establishes original text positioning information; fact elements in different question and answer rounds are extracted, a fact chain and a supplementary, correction, denial and confirmation state transition relationship are constructed, current effective facts are determined and conflict records are reserved; target examination rules are dynamically determined according to case time, procedure stage, subject identity and examination matters; examination doubts are generated by using a constrained large model, contents not explicitly recorded in the original text are marked as model inferred facts, and reverse verification is performed through multi-source evidence; the use of model inferred facts is limited according to evidence support states, and a structured examination result containing fact states, rule bases, evidence states and original text positions is output. The application can improve the accuracy, timeliness, adaptability and traceability of record examination.

Description

Technical Field

[0001] This invention relates to the fields of natural language processing, intelligent review of judicial data, and large-scale model application technology, specifically to a method, device, and medium for reviewing transcripts based on a large-scale model. Background Technology

[0002] With the development of natural language processing technology and large language models, using artificial intelligence to automatically review interrogation, inquiry, trial, and investigation records has become an important direction for improving the efficiency of judicial business processing. Current record review methods typically extract factual elements such as time, place, people, actions, and objects based on individual question-and-answer rounds, and then judge whether the record has omissions, contradictions, or procedural risks according to preset rules. However, the facts of the same case are often recorded in multiple question-and-answer rounds and may be supplemented, corrected, or denied in subsequent statements. Existing technology struggles to continuously record the changes in factual elements, easily leading to overlapping statements from different stages or incorrect merging of conflicting content, resulting in inaccurate fact-checking results.

[0003] Furthermore, the rules upon which record review is based may change depending on the rule's effective date, the stage of the case proceedings, the identity of the person being questioned, and the specific matters under review. Existing systems often use fixed rule bases or simply match rules based on time, making it difficult to accurately determine the valid rule version corresponding to a specific review matter. They also lack clear handling mechanisms when rules are missing, versions overlap, or applicable conditions conflict. Simultaneously, while large language models can generate semantically coherent inferences based on context, these inferences may not have direct basis in the original record or other evidentiary materials. Directly using model inferences as review conclusions can easily lead to fact-filling errors and makes it difficult to trace the corresponding original text location, rule basis, and source of evidence. Therefore, it is urgent to address the difficulties in continuously modeling factual states across question-and-answer rounds, dynamically matching review rules, and the lack of evidence verification and result tracing for model-inferred facts. Summary of the Invention

[0004] The purpose of this invention is to provide a method, device and medium for reviewing transcripts based on a large model, which solves the problems of existing transcript review technologies that are difficult to model the state evolution of case facts that change continuously across question and answer rounds, cannot dynamically match review rules according to case time and applicable conditions, and are prone to directly using unrecorded facts supplemented by the large model based on context as the basis for review.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] In a first aspect, the present invention provides a method for reviewing transcripts based on a large model, comprising the following steps:

[0007] Obtain the transcript data, case process data, review rule version data, and related evidence materials of the case to be reviewed; segment the transcript data according to the question-and-answer rounds and establish the original text location information.

[0008] Extract the factual elements explicitly recorded in the original text from each round of question and answer, associate the factual nodes pointing to the same event to be reviewed into a fact chain, distinguish the values ​​explicitly recorded in the original text, the values ​​inherited from the previous state, and the unrecorded values, and form a factual element state matrix across rounds of question and answer.

[0009] Determine the supplementary, corrective, denial, or confirmatory relationships between fact nodes according to different fact elements, update the current valid fact state representing the current statement state, and retain conflicting facts;

[0010] The applicable benchmark time for the rule is determined based on the review item. Candidate rule versions are screened based on the applicable benchmark time, and candidate rule versions are matched according to the procedure stage, subject identity, and review item. When only one candidate rule version meets all applicable conditions, it is determined as the target review rule. When there is no candidate rule version or there are two or more candidate rule versions that meet all applicable conditions, a rule missing mark or rule conflict mark is generated, the model review of the corresponding review item is stopped, and a manual review task is output.

[0011] When the target review rules are determined, the fact chain, the current valid fact state, conflicting facts, original text location information and the target review rules are input into the constrained large model to generate candidate review doubts and mark the model inferred facts that are not explicitly recorded in the original text.

[0012] The system generates reverse verification conditions for the facts inferred by the model, retrieves evidence units based on the reverse verification conditions and determines the evidence support status, restricts the use of the facts inferred by the model based on the evidence support status, and outputs a structured review result containing the fact status, rule basis, evidence status and original text location information.

[0013] Furthermore, the transcript data is segmented according to question-and-answer rounds and original text location information is established, including:

[0014] Multiple transcript data units are generated according to the correspondence between the question text and the answer text. Each transcript data unit includes at least the case identifier, transcript identifier, question and answer round identifier, transcript page number, character start position, character end position, speaker, question text, answer text, and formation time.

[0015] The original text location information is formed by a combination of record identifiers, record page numbers, question-and-answer round identifiers, and character ranges.

[0016] When the transcript originates from a scanned document, the page coordinates of the corresponding text are added to the original text location information; when the transcript is associated with synchronized audio and video recording, the start and end timestamps of the audio or video frames are added to the original text location information.

[0017] Furthermore, fact nodes pointing to the same event under review are linked together as a fact chain, including:

[0018] Perform entity normalization or alias mapping on the names of people, items, locations, and organizations in fact nodes;

[0019] When two fact nodes have the same case identifier, the same fact subject, the same standardized fact behavior, and the same fact object, they are grouped into the same fact chain.

[0020] When two fact nodes have only partial semantic similarity and it cannot be determined whether they point to the same event to be reviewed, the two fact nodes are retained in the candidate fact chain and a review-reviewed association tag is generated.

[0021] The fact chain includes at least a fact chain identifier, a case identifier, a set of fact nodes, the order of nodes, a set of state transition relationships, conflict relationships, and the current valid fact state.

[0022] Further, forming the fact element state matrix and updating the current valid fact state includes:

[0023] The factual element values ​​that are explicitly recorded in the original text and can be located to the current question-and-answer round through the original text location information are marked as explicitly recorded values ​​in the original text.

[0024] The factual element values ​​that are not re-recorded in the current question-and-answer round, originate from the previous question-and-answer round, and have not been negated, modified, or added to make them no longer applicable by the current question-and-answer round, are marked as the previous state inheritance values, and their source question-and-answer round and original text location information are saved.

[0025] When there is no explicitly recorded value in the original text that can be used in the fact chain, the corresponding fact element is marked as unrecorded;

[0026] A supplementary relationship is established when a subsequent fact node adds a fact element; a corrective relationship is established when a subsequent fact node changes the value of an existing fact element; a denial relationship is established when a subsequent fact node negates an existing behavior or result; and a confirmation relationship is established when a subsequent fact node reaffirms the corresponding behavior or result after denial.

[0027] When the same fact node changes multiple fact elements, the corresponding state transition type is saved according to each fact element.

[0028] The current valid factual state is formed according to the state transition results of each factual element. The current valid factual state is the current statement state of the fact chain and is not used as the result of the case fact determination.

[0029] Further, the target review rules are determined, including:

[0030] Read the rule identifier, rule version, effective start time, effective end time, applicable procedure stage, applicable subject identity, review items, rule applicable base time type, and review conditions from the review rule version data;

[0031] Based on the matters under review and the type of benchmark time for rule application, the time of occurrence of case facts, the time of case filing, the time of interrogation, the time of record formation, or the time of occurrence of corresponding procedural events are selected from the case process data as the benchmark time for rule application.

[0032] Add rule versions whose applicable reference time falls between the valid start time and the valid end time to the candidate rule set, and record the reasons for excluding other rule versions;

[0033] The applicable procedural stage, applicable subject identity, and review matters of the candidate rule version are matched with the procedural stage, subject identity, and review matters of the case to be reviewed, and the matching results of each matching item are saved.

[0034] When only one candidate rule version meets the four applicable conditions of effective time, procedure stage, subject identity, and review matter, it is determined as the target review rule; when there is no candidate rule version or there are two or more candidate rule versions that meet the four applicable conditions, a rule missing mark or rule conflict mark is generated, the model review of the corresponding review matter is stopped, and a manual review task is output.

[0035] Furthermore, candidate review points are generated using the constrained large model, including:

[0036] The data input to the constrained large model is divided into an explicitly recorded fact area, a rule-constrained area, and an inference area to be verified. The explicitly recorded fact area stores the content that can be located to the original transcript or evidence text through the original text location information. The rule-constrained area stores the target review rules and their applicable conditions. The inference area to be verified stores the inference content that is not explicitly recorded in the original text.

[0037] The constrained large model outputs the following review items: fact chain identifier, reference question and answer round, original text location information, target rule version, doubt type, doubt content, fact source type, and model inferred fact markers;

[0038] The output of the constrained large model is subjected to field validation, original source validation, and rule version validation. Content that cannot be located in the original text of the transcript or evidence through the original text location information is marked as the model inferred fact.

[0039] When the output lacks necessary fields, references non-existent original text locations, or uses a version of a non-target rule, the output is determined to be invalid, and candidate review points are regenerated or transferred to manual review.

[0040] Further, determining the state of evidence support includes:

[0041] Extract the subject to be verified, the behavior to be verified, the object of the behavior, the time of occurrence, the location of occurrence, and the relationship to be verified from the facts inferred from the model to form structured reverse verification conditions;

[0042] Evidence units are retrieved from access control records, surveillance videos, witness testimonies, electronic data, inspection records, or original transcripts based on the aforementioned reverse verification conditions.

[0043] The verification relationship between the evidence unit and the facts inferred by the model is determined as direct support, indirect correlation, no direct support, or contrary evidence.

[0044] When there is direct supporting evidence that simultaneously supports the subject to be verified, the behavior to be verified, the object of the behavior, and the time of occurrence, and there is no contrary evidence, the evidence support status is determined to be supported.

[0045] When only indirect evidence exists, or when direct evidence only supports some elements of the subject to be verified, the behavior to be verified, the object of the behavior, and the time of occurrence, the evidence support status is determined to be pending review.

[0046] When there is no direct supporting evidence or indirect related evidence, the evidence support status is determined to be unsupported.

[0047] When there is contrary evidence and no direct supporting evidence that simultaneously supports each of the elements to be verified, the evidence support status is determined to be unsupported.

[0048] When direct supporting evidence and contrary evidence coexist, the evidence support status is determined to be conflicting and the case is transferred to manual review.

[0049] Furthermore, the structured review results also include case identifier, fact chain identifier, historical status, conflict question and answer rounds, rule application benchmark time, rule matching basis, type of doubt, risk level, and recommended handling method;

[0050] The risk level is determined based on the risk mapping conditions pre-configured in the target review rules, as well as the type of doubt, the number of conflicting facts, the factual elements involved in the conflicting facts, and the status of evidence support.

[0051] The structured review results are output to the reviewer's review terminal, enabling the reviewer's review terminal to display the original transcript, factual status, target review rules, and evidence units in a linked manner;

[0052] Receive the results of manual review, save the content before modification, the content after modification, the person making the modification, the modification time, and the basis for modification, and use the results of manual review to update the factual element association conditions, rule matching results, or the output constraints of the constrained large model, without rewriting the original text of the transcript or evidence based on the results of manual review.

[0053] Secondly, the present invention provides a transcript review device based on a large model, including a processor, a memory, a communication interface and a data interface, wherein the processor is connected to the memory, the communication interface and the data interface respectively;

[0054] The data interface is used to obtain transcript data, case process data, review rule version data, and related evidence materials of the case to be reviewed;

[0055] The communication interface is used to interact with the case data domain, the reviewer's terminal, and the structured review result storage terminal.

[0056] The memory is used to store computer programs and the original text location information, fact chains, fact element state matrices, fact node state transition relationships, rule version data, evidence units, reverse verification conditions, and structured review results generated by the computer programs.

[0057] When the computer program is executed by the processor, it causes the processor to implement the large-model-based transcript review method described in the first aspect.

[0058] Thirdly, the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the large-model-based transcript review method described in the first aspect.

[0059] Compared with the prior art, the beneficial effects of the present invention are:

[0060] This invention divides the original transcript into question-and-answer rounds and establishes the association between question-and-answer rounds, fact nodes, and fact chains, enabling fact elements scattered across different question-and-answer rounds to be aggregated and traced according to their original positions. Furthermore, by constructing a fact element state matrix across question-and-answer rounds, the original records, inheritance, corrections, supplements, denials, and conflict states of fact elements are continuously recorded, thus avoiding information omissions or factual misjudgments caused by reviewing based solely on a single question-and-answer round. Further, by dynamically matching a uniquely applicable review rule version based on the rule's applicable benchmark time, procedural stage, subject identity, and review matters, and transferring the review to manual review when rules are missing or conflicting, the invention reduces review bias caused by incorrect rule version application. Simultaneously, by converting candidate facts inferred from a large model into verifiable conditions and using multi-source evidence such as transcripts, access control records, video materials, and witness statements for reverse verification, the invention suppresses conclusions generated by the large model that lack evidentiary support. Finally, by establishing a traceable association between the review results and the original text position, fact chain, rule version, and evidence unit, the invention improves the completeness, accuracy, consistency, and verifiability of transcript review. Attached Figure Description

[0061] Figure 1 This is a schematic diagram of the actual deployment structure and data flow of the transcript review method of the present invention.

[0062] Figure 2 This is a schematic diagram illustrating the relationship between the original transcript, question-and-answer rounds, fact nodes, and fact chains of this invention.

[0063] Figure 3 This is a schematic diagram of the fact element state matrix across question-and-answer rounds of the present invention.

[0064] Figure 4 This is a schematic diagram illustrating the state transition and conflict retention relationships between fact nodes in this invention.

[0065] Figure 5 This is a schematic diagram of the dynamic matching process for the review rule version of this invention.

[0066] Figure 6 This is a schematic diagram of the multi-source evidence reverse verification process for inferring facts using the model of this invention.

[0067] Figure 7 This is a schematic diagram illustrating the traceable and interconnected review relationship among the personnel responsible for reviewing the written records in this invention. Detailed Implementation

[0068] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0069] Example 1

[0070] like Figure 1 As shown, this embodiment provides a method for reviewing judicial records based on a large model. The method is applied to a judicial record review computing node, which is communicatively connected to a case data domain, a reviewer's terminal, and a structured review result storage terminal.

[0071] The case-handling data domain provides interrogation transcripts, synchronized audio and video recordings, case flow data, review rule version data, and related evidence materials. The judicial transcript review calculation node establishes the original text location relationship, constructs a fact chain across question-and-answer rounds, forms a fact element state matrix, determines the state transitions and conflict relationships between fact nodes, dynamically matches target review rules, generates candidate review doubts using a constrained large model, performs multi-source evidence reverse verification of the model's inferred facts, and forms a structured review result. The reviewer's verification terminal displays the original transcript, factual state, target review rules, and evidence units in a linked manner, and receives the reviewer's verification results.

[0072] The method for reviewing transcripts based on a large model includes the following steps.

[0073] S101. Obtain case-related data and establish original text location relationships.

[0074] Obtain interrogation transcripts, synchronized audio and video recordings, case process data, review rule version data, and related evidence materials for the cases to be reviewed.

[0075] The interrogation transcript shall include at least a case identifier, transcript identifier, transcript creation time, interrogation subject, transcript page number, and question-and-answer text. The synchronized audio and video recording shall include at least an audio timestamp, a video frame timestamp, and the correspondence between the audio and video recording timestamps and the question-and-answer text.

[0076] The case process data includes at least the time of the occurrence of the case facts, the time of case filing, the time of interrogation, the time of record formation, the procedural stage, the case status, the time of occurrence of procedural events, and the identity of the person being interrogated.

[0077] The review rule version data includes at least the rule identifier, rule version, effective start time, effective end time, applicable procedural stage, applicable subject identity, review matters, rule application benchmark time type, and review conditions. The rule application benchmark time type indicates that the corresponding review rule uses the time of occurrence of the case facts, the time of case filing, the time of interrogation, the time of record formation, or the time of occurrence of the corresponding procedural event as the benchmark for judging the validity of the rule version.

[0078] The relevant evidence materials include access control records, surveillance videos, witness testimonies, electronic data, inspection records, and other evidence materials that can verify the content recorded in the transcript.

[0079] The acquired data is subjected to OCR recognition, layout parsing, question-and-answer round segmentation, case identifier alignment, timestamp alignment, entity normalization, subject identity mapping, rule version loading, and evidence unit indexing to form data units with original text location information.

[0080] Based on the correspondence between the question text and the answer text, the interrogation record is divided into multiple record data units. Each record data unit includes at least the case identifier, record identifier, question and answer round identifier, record page number, character start position, character end position, speaker, question text, answer text, and the time of formation.

[0081] This embodiment uses the letter Q followed by two digits to identify the question-and-answer rounds. The letter Q is a fixed prefix, and the two digits following Q indicate the sequence number of the corresponding question-and-answer unit in the original interrogation record. For example, Q01 represents the first question-and-answer round in the original interrogation record, Q04 represents the fourth, Q08 represents the eighth, Q12 represents the twelfth, and Q15 represents the fifteenth. Other question-and-answer rounds not shown in this embodiment are not directly related to the example event under review and do not necessarily indicate the absence of corresponding question-and-answer rounds in the original interrogation record.

[0082] The original text location information is represented by a combination of transcript identifiers, transcript page numbers, question-and-answer round identifiers, and character ranges. When the transcript originates from a scanned document, the original text location information also includes the coordinate area of ​​the corresponding text within the transcript page. When the transcript is associated with synchronized audio and video recordings, the original text location information also includes audio start and end timestamps or video frame start and end timestamps.

[0083] For synchronized audio and video recording, the timestamps of the audio and video recordings are aligned with the corresponding question-and-answer rounds in the transcript. When there is a difference between the time recorded in the transcript and the time captured by the audio and video recording, the recorded value and the captured value in the audio and video recording are saved separately, and the data source of the two time values ​​is recorded, without directly overwriting the other time value with one of them.

[0084] For related evidence materials, they should be divided into multiple evidence units according to the content boundaries that can express a relatively complete subject, behavior, time, place, object, or result. Each evidence unit should include at least an evidence unit identifier, case identifier, evidence source, evidence type, original content, original text location information, time information, subject information, and object information.

[0085] S102. Constructing a fact chain across question-and-answer rounds

[0086] like Figure 2 As shown, the factual elements explicitly recorded in the original transcripts are extracted from each round of question-and-answer sessions. Each factual element includes at least the subject of the fact, the standardized factual act, the object of the fact, the time of occurrence, the place of occurrence, the polarity of the act, and the limiting conditions.

[0087] Each fact node stores the factual elements explicitly recorded in the corresponding question-and-answer round, the question-and-answer round identifier, and the original text location information. For factual elements not explicitly recorded in the original transcript, the corresponding fields will be marked as "not recorded" or "not clear," and no definite values ​​should be automatically added based on the context.

[0088] In this embodiment, the letter F and numbers are used to form the fact chain identifier and the fact node identifier. The fact chain identifier F001 represents the first fact chain in the current case; the fact node identifier F001-1 represents the first fact node in the fact chain F001, the fact node identifier F001-2 represents the second fact node in the fact chain F001, and so on for the remaining fact node identifiers.

[0089] When associating fact nodes in different question-and-answer rounds, first determine whether the corresponding fact nodes belong to the same case, and then perform entity normalization or alias mapping on the names of people, items, locations, and organizations in the fact nodes.

[0090] When two fact nodes have the same case identifier, the same fact subject, the same standardized fact behavior, and the same fact object, they are grouped into the same fact chain.

[0091] When the time or place of occurrence in two fact nodes have an inclusion, overlap, or modification relationship, but the fact subject, standardized fact behavior, and fact object still point to the same event to be examined, the two fact nodes will not be split into different fact chains due to changes in specific time or place values.

[0092] When two fact nodes have only partial semantic similarity and it cannot be determined whether the two fact nodes point to the same event to be reviewed, the two fact nodes are retained in the candidate fact chain respectively, and a correlation tag to be reviewed is generated. The fact nodes are not forcibly merged based on the semantic similarity results output by the large model.

[0093] In this embodiment, the case is identified as CASE-2026-017, the record is identified as RECORD-01, and the person being questioned is Zhang. Zhang's continuous statements regarding whether he arrived at and entered the warehouse in question are categorized into the fact chain F001. The fact chain F001 includes at least the fact chain identifier, the case identifier, the set of fact nodes, the sequence of nodes, the set of state transition relationships, the conflict relationships, and the current valid fact state.

[0094] In the first question-and-answer round, designated Q01, Zhang stated that he had been to the warehouse that evening, around 8 PM. Based on question-and-answer round Q01, fact node F001-1 is generated and designated as the initial node.

[0095] In the fourth question-and-answer round, designated Q04, Zhang added that he traveled to the warehouse in a black sedan and stated that Li was also present. Fact node F001-2 is generated based on question-and-answer round Q04. Fact node F001-2 adds two fact elements: means of transportation and related personnel, but does not negate the arrival behavior and arrival time explicitly recorded in fact node F001-1.

[0096] In the 8th question-and-answer round, designated Q08, Zhang revised his arrival time at the warehouse from approximately 20:00 to approximately 20:25. Fact node F001-3 is generated based on question-and-answer round Q08. Fact node F001-3 establishes a correction relationship for the time element and simultaneously saves the original time value, the corrected time value, the correction round, and the corresponding original text location information.

[0097] In the 12th question-and-answer round, designated Q12, Zhang stated that he was only outside the warehouse and did not enter it. Fact node F001-4 is generated based on question-and-answer round Q12. Fact node F001-4 denies the act of entering the warehouse, but does not deny the previously explicitly stated elements of the means of transport and related personnel that are not directly mutually exclusive with the act of entering.

[0098] In the 15th question-and-answer round, designated Q15, Zhang stated that he entered the warehouse through a side door at approximately 8:27 PM. Fact node F001-5 is generated based on question-and-answer round Q15. Fact node F001-5 confirms the act of entering the warehouse, modifies the entry time, and supplements the information about the entrance location.

[0099] S103. Form a fact element state matrix across question-and-answer rounds.

[0100] like Figure 3 As shown, a fact element state matrix is ​​established across question-and-answer rounds, with fact elements as rows and question-and-answer rounds and current valid fact states as columns. This fact element state matrix represents the fact element value, fact element source, current round state transition, and current valid fact state after completing the corresponding question-and-answer round.

[0101] Each fact element value in the fact element state matrix is ​​associated with at least the stored value content, value attribute, source question and answer round, and original text location information. The value attribute includes values ​​explicitly recorded in the original text, values ​​inherited from previous states, and unrecorded values.

[0102] The original text clearly states that the corresponding factual elements can be directly located to the original transcript of the current question-and-answer round through the original text location information.

[0103] The inherited value of the previous state indicates that the corresponding factual element has not been re-recorded in the current question-and-answer round, but the factual element is clearly recorded in the original text of the previous question-and-answer round and has not been negated, modified or added by the current question-and-answer round to make the factual element no longer applicable.

[0104] An unrecorded value indicates that there is currently no explicitly recorded original value available in the fact chain.

[0105] In this embodiment, the behaviors "arrival at the warehouse" and the time "around 20:00" in the question-and-answer round Q01 are marked as explicitly recorded values ​​from the original text of the question-and-answer round Q01, while the means of transport and associated personnel are marked as unrecorded values.

[0106] In Q04, the vehicle "black sedan" and the associated person "Li was present" are both marked as original textual records from Q04; the act of arriving at the warehouse and the time of approximately 20:00 are inherited from Q01, and the original textual location information of the inherited factual elements in Q01 is also preserved.

[0107] The time "approximately 20:25" in Q08 is marked as an original textually recorded value from Q08; the act of arriving at the warehouse is inherited from Q01; the vehicle "black sedan" and the associated person "Li was present" are inherited from Q04.

[0108] The action "did not enter the warehouse" in question-and-answer round Q12 is marked as an explicit value from the original text of question-and-answer round Q12. Since question-and-answer round Q12 did not deny the arrival time, means of transportation, or Li's presence, the arrival time is inherited from question-and-answer round Q08, and the means of transportation and related personnel are inherited from question-and-answer round Q04.

[0109] The actions "entering through the side door" and the time "approximately 20:27" in Q15 are both marked as original textual records from Q15. The vehicle "black sedan" and the associated person "Li was present" are inherited from Q04.

[0110] When inheriting factual elements, the inherited value of the previous state is not marked as the original text explicitly recorded value of the current question and answer round. Instead, the source question and answer round and the original text location information of the inherited value of the previous state are saved at the same time, so that the examiner can distinguish the original text recorded content of the current question and answer round from the inherited factual elements.

[0111] S104. Establish state transition and conflict retention relationships between fact nodes.

[0112] like Figure 4 As shown, the state transition types between adjacent fact nodes are determined according to different fact elements. These state transition types include at least initial, supplementary, revised, denial, and confirmation.

[0113] When a subsequent fact node adds fact elements that were not recorded in the previous fact node, a supplementary relationship is established for the newly added fact elements, while other unaffected fact elements are retained.

[0114] When a subsequent fact node explicitly changes the specific value of an existing fact element in a previous fact node, a correction relationship is established for the corresponding fact element. The current value is formed by the corrected fact element value, while retaining the historical value before correction, the original text location information of the fact element before correction, and the correction round.

[0115] When a subsequent factual element explicitly negates the action or result in a preceding factual element, a denial relationship is established for the negated factual element. The denial relationship only applies to the factual elements covered by the negation statement and does not automatically remove unnegotiated time, means of transportation, related persons, or other factual elements.

[0116] When a subsequent factual node reaffirms the corresponding behavior or result after denial, a confirmation relationship is established with the corresponding factual element, while retaining the original textual location information of the denial node, the confirmation node, and both.

[0117] If a factual element is not mentioned in a subsequent round of question and answer, the value explicitly recorded in the previous round of question and answer will be inherited only if the subject and object of the fact remain consistent, and there are no denials, modifications, changes in the scope of application, or limitations that render the factual element inapplicable.

[0118] When the same fact node simultaneously causes different state changes to multiple fact elements, the corresponding state transition type is recorded separately for each fact element, without forcibly limiting the entire fact node to a single state type. Therefore, the same fact node can simultaneously include supplementary relationships, corrective relationships, and confirmatory relationships.

[0119] In this embodiment, fact node F001-1 is the initial node of fact chain F001; fact node F001-2 supplements fact node F001-1 with the means of transport and associated personnel; fact node F001-3 corrects the arrival time relative to the previous state; fact node F001-4 denies the act of entering the warehouse; fact node F001-5 reconfirms the act of entering the warehouse, corrects the entry time and supplements the entrance location. Therefore, fact node F001-5 is a composite transfer node that includes confirmation, correction and supplementation.

[0120] The statement "did not enter the warehouse" recorded in question-and-answer round Q12 and "entered the warehouse through the side door" recorded in question-and-answer round Q15 are opposite in behavioral polarity. Therefore, a conflict relationship is established between fact node F001-4 and fact node F001-5.

[0121] The statements corresponding to Q12 are not deleted because Q15 was formed later than Q12. Instead, the fact nodes, original content, and original location information corresponding to both Q15 and Q12 are saved.

[0122] Based on the confirmation of the entry behavior, the correction of the entry time, and the supplementation of the side door entrance in Q15, "Zhang entered the warehouse through the side door at approximately 20:27" is determined as the current valid fact state of fact chain F001, while the conflicting fact corresponding to Q12 is retained.

[0123] The current valid factual state is the current stated state formed according to the state transition results of each factual element, and is not directly equivalent to the facts of the case as determined through judicial procedures. Without rule review, evidence verification, and manual review, the current valid factual state will not be directly output as the case fact determination result.

[0124] S105, Dynamic Matching Review Rules Version

[0125] like Figure 5 As shown, the applicable benchmark time for the rules is determined based on the matters to be reviewed, candidate rule versions are screened based on the applicable benchmark time, and candidate rule versions are matched according to the procedural stage, subject identity, and matters to be reviewed in the case to be reviewed.

[0126] Read the rule identifier, rule version, effective start time, effective end time, applicable procedure stage, applicable entity identity, review matters, rule applicable base time type, and review conditions from the review rule version data.

[0127] Based on the matters under review and the type of benchmark time for rule application, the time of occurrence of case facts, the time of case filing, the time of interrogation, the time of record formation, or the time of occurrence of corresponding procedural events are selected from the case process data as the benchmark time for rule application.

[0128] When the matter under review concerns the consistency of statements in the written record, the time when the record was created or the corresponding interrogation time shall be determined as the benchmark time for the application of the rules. When the matter under review involves specific procedural acts, the time when the corresponding procedural acts occurred shall be determined as the benchmark time for the application of the rules. When the matter under review involves substantive rules applicable when the facts of the case occurred, the time when the facts of the case occurred shall be determined as the benchmark time for the application of the rules.

[0129] In this embodiment, the matter under review is a consistency review of statements, and the record formation date of June 15, 2026 is used as the base time for the application of the rules; the procedural stage of the case to be reviewed is the first interrogation during the investigation stage, and the subject's identity is a criminal suspect.

[0130] In this embodiment, the rule identifier RA is used to represent the review rule used for conformity review, and the character V followed by a number represents the corresponding rule version. Therefore, the rule version identifier RA-V1 represents the first version of the review rule RA, and the rule version identifier RA-V2 represents the second version of the review rule RA.

[0131] The review rule repository includes rule versions RA-V1 and RA-V2. Rule version RA-V1 is valid from January 1, 2025 to June 30, 2026, and applies to initial interrogations during the investigation phase and consistency reviews of suspects' statements. Rule version RA-V2 will take effect on July 1, 2026.

[0132] The applicable base time of the rules is compared with the effective period of each rule version. Since June 15, 2026, is within the effective period of rule version RA-V1, which is earlier than the effective date of rule version RA-V2, rule version RA-V1 is added to the candidate rule set, rule version RA-V2 is marked as not yet effective, and the reason for the exclusion of rule version RA-V2 is recorded.

[0133] Read candidate rule version RA-V1, compare the procedural stage, subject identity, and review matters of the case to be reviewed with the applicable procedural stage, applicable subject identity, and review matters of rule version RA-V1, and save the matching results of the four applicable conditions: effective time, procedural stage, subject identity, and review matters.

[0134] When only one rule version in the candidate rule set meets all four applicable conditions—validity time, procedural stage, subject identity, and review matter—that rule version is designated as the target review rule, and the corresponding review matter is allowed to enter the subsequent model review process. In this embodiment, only rule version RA-V1 meets all applicable conditions; therefore, rule version RA-V1 is designated as the target review rule.

[0135] When no candidate rule version meets all applicable conditions, a rule missing flag is generated, the applicable conditions that are not met for each candidate rule version are saved, the model review for the corresponding review item is stopped, and a manual review task is output.

[0136] When two or more candidate rule versions simultaneously meet all applicable conditions, a rule conflict flag is generated, the conflicting rule versions and their respective matching criteria are saved, the model review of the corresponding review item is stopped, and a manual review task is output, without the system arbitrarily selecting one of the rule versions.

[0137] The rule matching results should include at least the target rule identifier, target rule version, rule application base time, time matching results, procedure stage matching results, subject identity matching results, review item matching results, and the reason for excluding rules that were not selected.

[0138] S106. Generating candidate review points of contention using a constrained large model.

[0139] After determining the unique target review rule, the fact chain, the current valid fact status, historical conflict facts, original text positioning information, and the target review rule are input into the constrained large model to generate candidate review points of doubt.

[0140] The data input to the constrained large model is divided into a region of explicitly stated facts, a region of rule constraints, and a region of inferences to be verified.

[0141] The explicitly recorded facts area stores content that can be located to the original transcript or evidence text through the original text location information. The rule constraint area stores the target review rules, the target rule version, and the corresponding applicable conditions. The unverified inference area stores inferences generated by the large model based on the context but not explicitly recorded in the original transcript or evidence text.

[0142] Each fact element input into the constrained large model carries a fact chain identifier, value attribute, source question-and-answer round, and original text location information. For inherited values ​​from previous states, the source question-and-answer round and original text location information of the inherited fact element are also carried to prevent the large model from misidentifying inherited values ​​from previous states as explicitly recorded values ​​in the current question-and-answer round.

[0143] Set structured output constraints for the constrained large model, so that each candidate review point of doubt includes at least the review item, fact chain identifier, reference question and answer round, original text location information, target rule version, doubt type, doubt content, fact source type, and model inferred fact label.

[0144] After receiving the output of the constrained large model, perform field validation, original source validation, and rule version validation on the output results.

[0145] When the large model marks a certain content as a fact explicitly stated in the original text, it verifies whether the content has a valid question-and-answer round identifier and original text location information. When the large model references review rules, it verifies whether the version of the referenced rule is consistent with the target review rule. When the output content cannot be located in the original transcript or evidence text through the original text location information, the content is placed in the inference area to be verified and a fact inferred by the model is added.

[0146] When the large model output lacks necessary fields, references non-existent original text locations, or uses a version of a non-target rule, the corresponding output is determined to be invalid, and candidate review points are regenerated. If no valid output is obtained after regenerating a preset number of times, the automatic processing of the corresponding candidate review points is stopped, and the process is transferred to manual review.

[0147] In this embodiment, the large model generates a candidate point of contention for review based on "not entering the warehouse" in question-and-answer round Q12 and "entering the warehouse through the side door" in question-and-answer round Q15. This candidate point of contention references the fact chain F001, question-and-answer rounds Q12 and Q15, and uses rule version RA-V1 as the rule basis.

[0148] When the large model generates "Li was also present" based on the question-and-answer round Q04, "Zhang entered through the side door" in question-and-answer round Q15, and other contextual information, this content is marked as a fact inferred by the model because the original transcript does not explicitly record Li's actions of opening the door or providing assistance.

[0149] In this embodiment, the model-inferred fact identifier is composed of the letter 'I' and a number, and the aforementioned model-inferred fact is marked as I001. The model-inferred fact I001 is only used as an object to be verified, not as a fact explicitly recorded in the original text, nor as the basis for a definitive review conclusion.

[0150] S107. Perform multi-source evidence reverse verification of the facts inferred from the model.

[0151] like Figure 6As shown, a reverse verification condition is generated for the model-inferred fact I001. In this embodiment, the reverse verification condition identifier is composed of the letter C and numbers, and the reverse verification condition generated for the model-inferred fact I001 is marked as C001.

[0152] The reverse verification condition C001 is used to retrieve independent evidence units that can support, limit, or refute the model's inferred fact I001, and is not used to further suggest that the large model expands the model's inferred fact I001.

[0153] Extract the subject to be verified, the behavior to be verified, the object of the behavior, the time of occurrence, the location of occurrence, and the relationship to be verified from the facts inferred from the model. Then, standardize the subject name, behavior name, object name, and time expression to form a structured reverse verification condition.

[0154] In this embodiment, the model inferred fact I001 is "Li provided assistance to Zhang in entering the warehouse". The reverse verification condition C001 generated from the model inferred fact I001 includes: the subject to be verified is Li; the behavior to be verified is opening the door or providing assistance; the object of the behavior is the side door of the warehouse in question; the time range is 20:20 to 20:30; the location is the area of ​​the side door of the warehouse in question; and the relationship to be verified is that there is a connection between Li's behavior and Zhang's entry into the warehouse.

[0155] For access control records, search criteria are constructed based on subject identification, access control point identification, and time range. For surveillance video, search criteria are constructed based on monitoring location, video time period, and target personnel trajectory. For witness statements and transcripts, search criteria are constructed based on standardized subject, behavior, object, and time elements. For electronic data, search criteria are constructed based on account identification, device identification, communication object, and time range.

[0156] In this embodiment, the letter E and numbers are used to form the evidence unit identifier. Evidence units E001, E002, E003, and E004 are retrieved based on the reverse verification condition C001.

[0157] Evidence unit E001 originates from the access control record, which states that the side door of the warehouse in question was opened at 20:26, but the identity of the person who opened the door was not recorded. Therefore, evidence unit E001 can prove that the side door was opened at the relevant time, but it cannot prove that the opening was carried out by Li. The verification relationship between evidence unit E001 and the model-inferred fact I001 is determined to be an indirect association.

[0158] Evidence unit E002 originates from surveillance video, showing two people approaching the side door at a relevant time, but the second person cannot be identified. Therefore, evidence unit E002 proves the presence of a second person at the scene, but it cannot confirm that this person is Li, nor can it confirm that this person provided assistance. The verification relationship between evidence unit E002 and the model-inferred fact I001 is determined to be an indirect association.

[0159] Evidence unit E003 originates from witness testimony, which states that another male was present at the scene, but the witness could not confirm that the male was Li. Therefore, the verification relationship between evidence unit E003 and the model-inferred fact I001 is determined to be an indirect association.

[0160] Evidence unit E004 is derived from the original transcript, which only records that Li was present, but does not record that Li opened the door or provided any other assistance. Therefore, the verification relationship between evidence unit E004 and the model-inferred fact I001 is determined to be without direct support.

[0161] The verification relationship between evidence units and model-inferred facts includes direct support, indirect correlation, no direct support, and contrary evidence.

[0162] Direct support indicates that the evidence unit simultaneously supports the subject to be verified, the act to be verified, the object of the act, and the time of occurrence. Indirect association indicates that the evidence unit only supports some elements or peripheral facts of the subject to be verified, the act to be verified, the object of the act, and the time of occurrence. No direct support indicates that the evidence unit does not contain any content that can prove the act to be verified. Contrary evidence indicates that the evidence unit explicitly denies the subject to be verified, the act to be verified, the object of the act, the time of occurrence, or the relationship to be verified.

[0163] When there is direct supporting evidence that simultaneously supports the subject, act, object, and time of the act to be verified, and no contrary evidence exists, the evidence support status is determined to be "supported." When the evidence support status is "supported," a review prompt with an evidence source identifier is allowed to be generated, but the system does not directly make a factual determination of the case.

[0164] When only indirect evidence exists, or when direct supporting evidence only supports some of the elements to be verified, the evidence support status is determined to be pending review, and it is prohibited to output the corresponding model-inferred facts as a definitive review conclusion.

[0165] When there is no direct supporting evidence or indirect related evidence, the evidence support status is determined to be unsupported, and the certain conclusions formed by relying on the corresponding model to infer the facts are deleted or downgraded.

[0166] When there is contrary evidence and no direct supporting evidence that simultaneously supports each element to be verified, the evidence support status is determined to be unsupported, and the source and original text location information of the contrary evidence are saved.

[0167] When both supporting and contradictory evidence exist, the evidence support status is determined to be conflicting, and the case is transferred to manual review. The system does not automatically select one side based on the number of supporting and contradictory evidence.

[0168] In this embodiment, the verification results corresponding to evidence units E001 to E004 are: zero direct support items, three indirect association items, one item without direct support, and zero contrary evidence items. Since there is no direct supporting evidence that simultaneously supports the subject to be verified, the behavior to be verified, the object of the behavior, and the time of occurrence, the evidence support status of model-inferred fact I001 is determined to be pending review. It is prohibited to use model-inferred fact I001 as the basis for independently forming a definitive review conclusion, and the definitive conclusion formed based on model-inferred fact I001 is downgraded to a manual review prompt.

[0169] S108. Generate structured review results and receive manual review feedback.

[0170] like Figure 7 As shown, the current valid factual status, historical conflict facts, target review rules, and evidence support status are jointly processed to generate a structured review result.

[0171] The structured review results include at least the case identifier, fact chain identifier, current valid fact status, historical status, historical conflict facts, conflict question and answer rounds, original text location information, target rule version, rule application benchmark time, rule matching basis, model inferred facts, evidence support status, doubt type, risk level, and recommended handling method.

[0172] The risk level is determined based on the risk mapping conditions pre-configured in the target review rules. These risk mapping conditions include at least the type of doubt, the number of conflicting facts, the factual elements involved in the conflicting facts, the status of evidence support, and whether manual review is required.

[0173] The risk level is not generated freely by a large model, but is determined by matching the types of doubts, the number of conflicting facts, the factual elements involved in the conflicting facts, and the status of evidence support in the structured review results with the risk mapping conditions in the target review rules.

[0174] The structured review results generated in this embodiment include: the fact chain identifier is F001; the current valid fact status is that Zhang entered the warehouse through the side door at approximately 20:27; the historical conflicting facts are that question-and-answer round Q12 records that he did not enter the warehouse and question-and-answer round Q15 records that he entered the warehouse through the side door; the target review rule is rule version RA-V1; the doubt type is a contradiction between statements; the risk level is medium; the evidence support status of the model-inferred fact I001 is pending review; the recommended handling method is to manually review the factual conflict between question-and-answer rounds Q12 and Q15, and to prohibit the output of model-inferred fact I001 as a definitive review fact.

[0175] The reviewers used the review terminal to view the original transcript, the status of the facts, the target review rules, and the evidence units in a linked manner. After the reviewers selected the conflict relationship between question-and-answer rounds Q12 and Q15, the original text location window displayed the original text content of question-and-answer rounds Q12 on page 7 and Q15 on page 9, respectively, and showed the corresponding transcript identifier, page number, question-and-answer round identifier, and character range.

[0176] For the model-inferred fact I001, the review terminal shows that evidence units E001 to E003 are indirectly related to the model-inferred fact I001, and evidence unit E004 does not directly support the model-inferred fact I001. The reviewer is allowed to check each evidence unit through the original text location information.

[0177] For review conclusions supported by sufficient original text and evidence, the review terminal retains the corresponding review conclusion. For review conclusions based on model inferences derived from evidence that is pending review or unsupported, the review terminal downgrades the corresponding review conclusion to a manual review prompt. For review conclusions where both directly supporting evidence and contrary evidence coexist, the review terminal displays an evidence conflict status and prohibits the automatic formation of a definitive conclusion.

[0178] When reviewers modify the current valid factual status, conflict relationships, target review rules, or evidence verification relationships, they should save the content before modification, the content after modification, the person making the modification, the modification time, and the basis for the modification.

[0179] The results of manual review are used to update the factual element association conditions, rule matching results, or output constraints of the constrained large model, but the original transcript or evidence text is not rewritten based on the manual review results. The updated factual element association conditions or large model output constraints are only used for subsequent review tasks, and the corresponding version identifier and effective time are saved to maintain the traceability of the review process.

[0180] Example 2

[0181] This embodiment provides a transcript review device based on a large model, including a processor, a memory, a communication interface, and a data interface. The processor is connected to the memory, the communication interface, and the data interface, respectively.

[0182] The data interface is used to obtain transcript data, case process data, review rule version data, and related evidence materials of the case to be reviewed.

[0183] The communication interface is used to interact with the case data domain, the reviewer's terminal, and the structured review result storage terminal.

[0184] The memory is used to store computer programs and the original text location information, fact chains, fact element state matrices, fact node state transition relationships, rule version data, evidence units, reverse verification conditions, and structured review results generated by the computer programs.

[0185] When the processor executes the computer program, it establishes a correspondence between the question-and-answer rounds and the original text position, associates fact nodes pointing to the same event to be reviewed as a fact chain across question-and-answer rounds, distinguishes between explicitly recorded values ​​in the original text, inherited values ​​from previous states, and unrecorded values, and forms a fact element state matrix.

[0186] The processor identifies supplementary, corrective, denial, and confirmative relationships according to different factual elements, establishes fact node state transition relationships, updates the current valid factual state, and retains conflicting facts.

[0187] The processor determines the applicable benchmark time for the rule based on the review item, and determines the target review rule based on the applicable benchmark time, procedure stage, subject identity and review item; when a rule missing mark or rule conflict mark is generated, the model review of the corresponding review item is stopped and a manual review task is output.

[0188] After determining the unique target review rule, the processor uses a constrained large model to generate candidate review doubts, distinguishes between facts explicitly stated in the original text and facts inferred by the model, generates reverse verification conditions for the facts inferred by the model, retrieves multi-source evidence units and determines the evidence support status, retains, denies or downgrades the corresponding review results according to the evidence support status, and outputs a structured review result with original text location information, rule matching basis and evidence source information to the reviewer's review terminal.

[0189] Example 3

[0190] This embodiment also provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it enables the large-model-based transcript review method described in Embodiment 1 or Embodiment 2. The computer-readable storage medium can be a read-only memory, random access memory, solid-state memory, disk, optical disk, or other non-transitory storage medium capable of storing a computer program.

[0191] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for reviewing written records based on a large model, characterized in that, Includes the following steps: Obtain the transcript data, case process data, review rule version data, and related evidence materials of the case to be reviewed; segment the transcript data according to the question-and-answer rounds and establish the original text location information. Extract the factual elements explicitly recorded in the original text from each round of question and answer, associate the factual nodes pointing to the same event to be reviewed into a fact chain, distinguish the values ​​explicitly recorded in the original text, the values ​​inherited from the previous state, and the unrecorded values, and form a factual element state matrix across rounds of question and answer. Determine the supplementary, corrective, denial, or confirmatory relationships between fact nodes according to different fact elements, update the current valid fact state representing the current statement state, and retain conflicting facts; The applicable benchmark time for the rules is determined based on the review items, candidate rule versions are screened based on the applicable benchmark time, and candidate rule versions are matched according to the procedure stage, subject identity and review items; When only one candidate rule version meets all applicable conditions, it is determined as the target review rule; when there is no candidate rule version or there are two or more candidate rule versions that meet all applicable conditions, a rule missing mark or rule conflict mark is generated, the model review of the corresponding review item is stopped, and a manual review task is output. When the target review rules are determined, the fact chain, the current valid fact state, conflicting facts, original text location information and the target review rules are input into the constrained large model to generate candidate review doubts and mark the model inferred facts that are not explicitly recorded in the original text. The system generates reverse verification conditions for the facts inferred by the model, retrieves evidence units based on the reverse verification conditions and determines the evidence support status, restricts the use of the facts inferred by the model based on the evidence support status, and outputs a structured review result containing the fact status, rule basis, evidence status and original text location information.

2. The method for reviewing transcripts based on a large model according to claim 1, characterized in that, The transcript data is segmented by question-and-answer rounds and original text location information is established, including: Multiple transcript data units are generated according to the correspondence between the question text and the answer text. Each transcript data unit includes at least the case identifier, transcript identifier, question and answer round identifier, transcript page number, character start position, character end position, speaker, question text, answer text, and formation time. The original text location information is formed by a combination of record identifiers, record page numbers, question-and-answer round identifiers, and character ranges. When the transcript originates from a scanned document, the page coordinates of the corresponding text are added to the original text location information; when the transcript is associated with synchronized audio and video recording, the start and end timestamps of the audio or video frames are added to the original text location information.

3. The method for reviewing transcripts based on a large model according to claim 1, characterized in that, Fact nodes pointing to the same event under review are linked together as a fact chain, including: Perform entity normalization or alias mapping on the names of people, items, locations, and organizations in fact nodes; When two fact nodes have the same case identifier, the same fact subject, the same standardized fact behavior, and the same fact object, they are grouped into the same fact chain. When two fact nodes have only partial semantic similarity and it cannot be determined whether they point to the same event to be reviewed, the two fact nodes are retained in the candidate fact chain and a review-reviewed association tag is generated. The fact chain includes at least a fact chain identifier, a case identifier, a set of fact nodes, the order of nodes, a set of state transition relationships, conflict relationships, and the current valid fact state.

4. The method for reviewing transcripts based on a large model according to claim 1, characterized in that, Forming the fact element state matrix and updating the current valid fact state includes: The factual element values ​​that are explicitly recorded in the original text and can be located to the current question-and-answer round through the original text location information are marked as explicitly recorded values ​​in the original text. The factual element values ​​that are not re-recorded in the current question-and-answer round, originate from the previous question-and-answer round, and have not been negated, modified, or added to make them no longer applicable by the current question-and-answer round, are marked as the previous state inheritance values, and their source question-and-answer round and original text location information are saved. When there is no explicitly recorded value in the original text that can be used in the fact chain, the corresponding fact element is marked as unrecorded; A supplementary relationship is established when a subsequent fact node adds a fact element; a corrective relationship is established when a subsequent fact node changes the value of an existing fact element; a denial relationship is established when a subsequent fact node negates an existing behavior or result; and a confirmation relationship is established when a subsequent fact node reaffirms the corresponding behavior or result after denial. When the same fact node changes multiple fact elements, the corresponding state transition type is saved according to each fact element. The current valid factual state is formed according to the state transition results of each factual element. The current valid factual state is the current statement state of the fact chain and is not used as the result of the case fact determination.

5. The method for reviewing transcripts based on a large model according to claim 1, characterized in that, Determining the target review rules includes: Read the rule identifier, rule version, effective start time, effective end time, applicable procedure stage, applicable subject identity, review items, rule applicable base time type, and review conditions from the review rule version data; Based on the matters under review and the type of benchmark time for rule application, the time of occurrence of case facts, the time of case filing, the time of interrogation, the time of record formation, or the time of occurrence of corresponding procedural events are selected from the case process data as the benchmark time for rule application. Add rule versions whose applicable reference time falls between the valid start time and the valid end time to the candidate rule set, and record the reasons for excluding other rule versions; The applicable procedural stage, applicable subject identity, and review matters of the candidate rule version are matched with the procedural stage, subject identity, and review matters of the case to be reviewed, and the matching results of each matching item are saved. When only one candidate rule version meets the four applicable conditions of effective time, procedure stage, subject identity, and review matter, it is determined as the target review rule; when there is no candidate rule version or there are two or more candidate rule versions that meet the four applicable conditions, a rule missing mark or rule conflict mark is generated, the model review of the corresponding review matter is stopped, and a manual review task is output.

6. The method for reviewing transcripts based on a large model according to claim 1, characterized in that, The constrained large model is used to generate candidate review points of contention, including: The data input to the constrained large model is divided into an explicitly recorded fact area, a rule-constrained area, and an inference area to be verified. The explicitly recorded fact area stores the content that can be located to the original transcript or evidence text through the original text location information. The rule-constrained area stores the target review rules and their applicable conditions. The inference area to be verified stores the inference content that is not explicitly recorded in the original text. The constrained large model outputs the following review items: fact chain identifier, reference question and answer round, original text location information, target rule version, doubt type, doubt content, fact source type, and model inferred fact markers; The output of the constrained large model is subjected to field validation, original source validation, and rule version validation. Content that cannot be located in the original text of the transcript or evidence through the original text location information is marked as the model inferred fact. When the output lacks necessary fields, references non-existent original text locations, or uses a version of a non-target rule, the output is determined to be invalid, and candidate review points are regenerated or transferred to manual review.

7. The method for reviewing transcripts based on a large model according to claim 1, characterized in that, Determining the state of evidence support includes: Extract the subject to be verified, the behavior to be verified, the object of the behavior, the time of occurrence, the location of occurrence, and the relationship to be verified from the facts inferred from the model to form structured reverse verification conditions; Evidence units are retrieved from access control records, surveillance videos, witness testimonies, electronic data, inspection records, or original transcripts based on the aforementioned reverse verification conditions. The verification relationship between the evidence unit and the facts inferred by the model is determined as direct support, indirect correlation, no direct support, or contrary evidence. When there is direct supporting evidence that simultaneously supports the subject to be verified, the behavior to be verified, the object of the behavior, and the time of occurrence, and there is no contrary evidence, the evidence support status is determined to be supported. When only indirect evidence exists, or when direct evidence only supports some elements of the subject to be verified, the behavior to be verified, the object of the behavior, and the time of occurrence, the evidence support status is determined to be pending review. When there is no direct supporting evidence or indirect related evidence, the evidence support status is determined to be unsupported. When there is contrary evidence and no direct supporting evidence that simultaneously supports each element to be verified, the evidence support status is determined to be unsupported. When direct supporting evidence and contrary evidence coexist, the evidence support status is determined to be conflicting and the case is transferred to manual review.

8. The method for reviewing transcripts based on a large model according to claim 1, characterized in that, The structured review results also include case identifier, fact chain identifier, historical status, conflict question and answer rounds, rule application benchmark time, rule matching basis, type of doubt, risk level, and recommended handling method; The risk level is determined based on the risk mapping conditions pre-configured in the target review rules, as well as the type of doubt, the number of conflicting facts, the factual elements involved in the conflicting facts, and the status of evidence support. The structured review results are output to the reviewer's review terminal, enabling the reviewer's review terminal to display the original transcript, factual status, target review rules, and evidence units in a linked manner; Receive the results of manual review, save the content before modification, the content after modification, the person making the modification, the modification time, and the basis for modification, and use the results of manual review to update the factual element association conditions, rule matching results, or the output constraints of the constrained large model, without rewriting the original text of the transcript or evidence based on the results of manual review.

9. A transcript review device based on a large model, characterized in that, It includes a processor, a memory, a communication interface, and a data interface, wherein the processor is connected to the memory, the communication interface, and the data interface respectively; The data interface is used to obtain transcript data, case process data, review rule version data, and related evidence materials of the case to be reviewed; The communication interface is used to interact with the case data domain, the reviewer's terminal, and the structured review result storage terminal. The memory is used to store computer programs and the original text location information, fact chains, fact element state matrices, fact node state transition relationships, rule version data, evidence units, reverse verification conditions, and structured review results generated by the computer programs. When the computer program is executed by the processor, the processor performs the transcript review method based on a large model as described in any one of claims 1-8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the transcript review method based on a large model as described in any one of claims 1-8.

Images