An intelligent quality inspection method and system based on link division

Abstract

The application provides an intelligent quality inspection method and system based on link division, comprising: obtaining double-recording dialogue audio text and performing semantic coding to obtain dialogue semantic features; separating customer interruption content and seat dialogue flow according to speaker roles; extracting business link keywords in the interruption, matching with a preset keyword library, and marking conflict paragraphs with attribution bias; determining the true attribution link of the conflict paragraphs through semantic association; rearranging the link attribution labels and dividing the boundaries in the order of the double-recording standard process to obtain a target link sequence; performing integrity checking on the target link sequence, detecting missed reading and skipping paragraphs, and correcting attribution and filling in words to update the sequence; extracting compliance rules for each updated link and comparing the actual words with the standard words to output the quality inspection result. The application effectively solves the problems of link attribution confusion and missed reading and skipping difficulty to detect in double-recording quality inspection, and improves the quality inspection accuracy and process integrity.

Description

Technical Field

[0001] This invention relates to the field of information technology, and in particular to an intelligent quality inspection method and system based on process segmentation. Background Technology

[0002] Audio and video recording quality inspection in the financial and insurance industries is a crucial regulatory tool for ensuring compliance in the sales process. Standard dual-recording processes typically include several fixed steps such as identity verification, product introduction, risk disclosure, customer confirmation, and signature confirmation. Each step requires sales personnel to read a designated script in sequence, and the quality inspection system must subsequently review whether each step was fully executed. Existing quality inspection systems generally employ a linear segmentation processing model based on a timeline. This involves scanning along the dialogue timeline, identifying preset keywords in the salesperson's speech to determine which step is currently in, and categorizing all dialogue content after that point into that step until the keyword for the next step is detected. The fundamental premise of this processing model is that both parties strictly follow the preset process sequence, without any jumps in order. However, in actual sales processes, customer interruptions are frequent and unpredictable. Customers may interrupt sales personnel because they cannot hear clearly and request repetition, because they do not understand terminology and request explanations, or because they have doubts about the terms. These interruptions disrupt the linear order of the dialogue, causing sales personnel to pause their current script to respond to the customer, and then return to continue, or even skip parts of the script to proceed to subsequent steps. Current processing methods may forcibly categorize interruptions into the business segment in which they occur. For example, if a customer interrupts to inquire about cancellation procedures while a salesperson is explaining product benefits, the system might mark the cancellation-related dialogue as part of the product introduction segment, causing a significant discrepancy between the interruption's true business nature and its system attribution. If the interruption happens to contain keywords from other segments, the system might misjudge the timeline and jump forward, incorrectly categorizing all subsequent dialogues into later segments, creating a chain reaction of segment sequence misalignments. Another approach is to treat interruptions as noise and filter them directly, retaining only the salesperson's script read according to the procedure. However, filtering out interruptions also discards the salesperson's responses, preventing the system from knowing the complete explanation process. More seriously, after responding to an interruption, the salesperson might skip the remaining script and move directly to the next segment. Because the system checks chronologically, it might incorrectly mark incomplete segments as completed. This exposes two technical problems. First, customer interruptions lead to confusion regarding the attribution of conversation segments. The actual business content of the interruption is inconsistent with the segment at which it occurred, and may trigger incorrect time anchor jumps, causing all subsequent content to be misattributed. Second, this confusion in segment attribution directly leads to the failure of process integrity checks. Because the system cannot correctly determine the start and end boundaries and true attribution of each segment, it cannot accurately detect whether the salesperson has completely read all the required scripts for that segment. Even if the salesperson skips a segment of script due to an interruption, the system cannot recognize it. Incorrect attribution deprives integrity checks of a correct spatiotemporal benchmark, and the lack of integrity further exacerbates the difficulty in determining the attribution of subsequent segments. Ultimately, the quality inspection system determines compliance based on incorrect segment divisions and incomplete conversation content, resulting in compliant conversations being misjudged as non-compliant and non-compliant conversations being missed. Consequently, the quality inspection report cannot accurately reflect the actual compliance status of the sales process. Summary of the Invention

[0003] This invention provides an intelligent quality inspection method based on process segmentation, mainly comprising: The original dialogue recording text is obtained, and semantic encoding processing is performed to form a dialogue semantic feature representation that includes contextual information, semantic associations, speaker identification, and semantic role information. Based on the semantic features of the dialogue, role recognition and content separation are performed to form a set of customer interjection segments and an agent dialogue flow; The set of customer interruption segments is processed to identify conflict segments, extract keywords of business processes, and identify process attribution deviations based on the process thesaurus corresponding to the time of the interruption to form conflict segments. Based on the semantic association between the conflicting paragraph and each segment in the agent's dialogue flow, the process of determining the true attribution stage is performed; Based on the original timestamps of the actual attribution link and conflicting segments, time information adjustment processing is performed to form adjusted time information; Based on the adjusted time information and the standard dual recording process, the process sequence generation and integrity verification are performed. The process attribute tags are rearranged, process boundaries are defined, and a target process sequence is formed. Missed reading nodes and skipped reading paragraphs are identified based on the standard script node list. The process boundaries and sentence fragment positions are adjusted according to the skipped reading paragraphs. Missed node markers are written at the corresponding process positions according to the missed reading nodes, thus forming an updated target process sequence. Compliance checks are performed based on the updated target sequence. Compliance check rules for each step are extracted from the dual-recording quality inspection rule library, and the actual script, standard script, and customer confirmation content are compared to form the quality inspection results.

[0004] Furthermore, the semantic encoding process includes: The speech turns are divided based on speaker switching, sentence pauses, and complete sentence boundaries, and the start and end times and text order of the speech turns are corrected. Extract the question-and-answer sequence between adjacent turns and the referential, negation, conditional, and temporal relationships within the same turn; Encode the relationships of succession, transition, supplementation, and response between objects and statements in the aforementioned relationship.

[0005] Furthermore, the role recognition and content separation process includes: The initial roles of each turn are determined based on the speaker's identity, and the target roles are corrected according to the semantics of inquiry, notification, objection, and confirmation, as well as the question-and-answer relationship between adjacent turns. Extract customer role statements and agent role statements by timestamp, and retain the call boundaries and overlapping speech markers; Merge consecutive customer role statements between adjacent agent turns or overlapping with agent statement times to form the customer interruption segment set, and arrange the agent role statements to form the agent dialogue flow.

[0006] Furthermore, the conflict paragraph identification process includes: Extract the product name, insurance liability, expense items, risk items, and confirmation items from the customer's interjection paragraphs, and merge synonymous expressions; The agent turn that is being read or has recently ended is determined based on the start time of the interruption, and the target segment vocabulary is determined based on the segment label of the agent turn; Match the target stage's thesaurus with other stage thesauruses, excluding shared words; If the keywords for the remaining business process do not match the target process's keyword library but match other process keyword libraries, mark the corresponding customer interjection paragraph as a conflicting paragraph.

[0007] Furthermore, the process for determining the true attribution includes: Extract the business objects, inquiry content, and referents of the conflict segments, and extract the segment keywords and agent response content of each segment in the agent dialogue flow; Candidate segmentation is formed based on the correspondence of synonyms, the correspondence of referents, and the consistency of business objects; Exclude segments where the business objects are inconsistent or the referents cannot be matched, and arrange them according to the degree of correspondence between questions and answers; Select the segment that directly answers the question. If no direct answer exists, select the segment with the same business object and the corresponding segment keyword, and determine the actual segment to which it belongs based on the segment tag of the segment.

[0008] Furthermore, the time information adjustment process includes: The original duration is determined based on the original start and end times of the conflicting segments, and the start and end times of the actual attribution segment are obtained. The mapping start point is determined by the start time of the agent's statement when the actual attribution is determined, and the mapping end point is determined by the original duration. When the mapping interval exceeds the actual time interval of the assigned link, the mapping start point or mapping end point is corrected according to the corresponding boundary, and the original time, mapping time and actual assigned link are recorded.

[0009] Furthermore, the step sequence generation and integrity verification process includes: Arrange the labels of each stage according to the dual recording standard process, merge the agent statements corresponding to the same stage label and with the adjusted time information, and divide the stage boundaries between adjacent different stage segments. Extract the standard script node list for each stage from the dual-recording script library, and match the actual agent's statements according to the business object, the notification item, and the order of the nodes. Nodes without corresponding statements are marked as missed read nodes, and paragraphs are marked as skipped when there are uncorresponding nodes between adjacent corresponding nodes; Adjust the position of the sentence fragments in the target segment sequence according to the skipped reading paragraph, and merge the agent sentences that belong to the same segment label and have continuous time information after adjustment into segment fragments; Based on the missed reading node, write the missing node mark at the corresponding stage position, and retain the original time and the adjusted time.

[0010] Furthermore, the compliance check process includes: Based on the process tags in the updated target process sequence, mandatory nodes, prohibited expressions, reading order, and customer confirmation requirements are extracted from the dual-recording quality inspection rule base to form a set of compliance inspection rules; The actual agent's statements and standard script nodes are matched according to the business object, the content of the notification, the scope of the rejection, the limiting conditions, and the order of the nodes. The customer's statements are checked against the customer's confirmation requirements to form a rule check record.

[0011] Furthermore, the rule verification records include: Record the corresponding statements of the required reading nodes, the location of the prohibited expressions, the node order correspondence, and the customer confirmation content; If there are missing required reading nodes, prohibited expressions, incorrect node order, or missing customer confirmation, mark the corresponding step as non-compliant; otherwise, mark the corresponding step as compliant. The process labels, judgment results, and corresponding statements are collected to form the quality inspection results.

[0012] This invention provides an intelligent quality inspection system based on process segmentation, mainly comprising: The semantic encoding module is configured to acquire the original dialogue recording text and form a dialogue semantic feature representation; The role recognition and content separation module is configured to form a set of customer interjection segments and an agent dialogue flow based on the dialogue semantic feature representation; The conflict paragraph identification module is configured to generate conflict paragraphs based on the matching relationship between keywords in business processes and process thesaurus; The module for determining the true attribution stage is configured to determine the true attribution stage based on the semantic association between the conflicting paragraph and each stage segment. The time information adjustment module is configured to generate adjusted time information based on the original timestamp and the actual time interval of the affiliated link. The segment sequence generation and integrity verification module is configured to generate a target segment sequence based on the adjusted time information, identify missed reading nodes and skipped reading segments, and update the target segment sequence. The compliance inspection module is configured to generate quality inspection results based on the updated target process sequence and the dual-recording quality inspection rule base.

[0013] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects: This invention discloses an intelligent quality inspection method based on segmentation, providing a systematic solution to the problems of segment misalignment and process confusion caused by customer interruptions in insurance dual recording scenarios. In actual dual recording processes, customers often interrupt sales personnel's standard process explanations due to questions or the need for comprehension, causing the distribution of dialogue content on the timeline to deviate from the standard segment order. Traditional quality inspection methods struggle to accurately identify the true segment to which each dialogue segment belongs. This invention extracts the contextual information and role information of the dialogue through semantic encoding, achieving precise separation of customer and agent content. It then identifies interrupted segments with segment misalignment and maps these conflicting segments back to their true business segments and time intervals through semantic association analysis. This completes the rearrangement and boundary delineation of segment labels, constructing a target segment sequence that conforms to the standard process order. Further, it performs script integrity verification and compliance checks, thereby achieving accurate quality inspection of non-linear dialogue processes. This effectively solves the problem of segment identification confusion caused by customer interruptions, improving the accuracy and automation level of dual recording quality inspection. Attached Figure Description

[0014] Figure 1 This is a flowchart of an intelligent quality inspection method based on process segmentation according to the present invention.

[0015] Figure 2 This is a schematic diagram of an intelligent quality inspection method based on process segmentation according to the present invention.

[0016] Figure 3 This is another schematic diagram of an intelligent quality inspection method based on process segmentation according to the present invention.

[0017] Figure 4 This is a schematic diagram of the structure of an intelligent quality inspection system based on process segmentation according to the present invention. Detailed Implementation

[0018] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. 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.

[0019] like Figures 1-4 This embodiment of an intelligent quality inspection method and system based on process segmentation may specifically include: Step S101: Obtain the original dialogue recording text and perform semantic encoding; extract the contextual information and semantic association features in the dialogue text to obtain the dialogue semantic feature representation.

[0020] The original dialogue recording text and timestamp are obtained. Turns are divided based on speaker switching, pauses, and complete sentence boundaries. The start and end times and text order of each turn are corrected to obtain a standard dialogue text arranged chronologically. Based on the question-and-answer succession relationships between adjacent turns in the standard dialogue text, and the referential, negation, conditional, and temporal relationships of statements within the same turn, contextual information for each statement is formed. The referential objects, negation ranges, conditional objects, and temporal objects in the contextual information are semantically encoded, and the succession, transition, supplementation, and response relationships between statements in the standard dialogue text are recorded to obtain a semantic feature representation of the dialogue.

[0021] In one implementation, the original dialogue recording text generated by a financial or insurance dual recording service is obtained. The original dialogue recording text includes the content of the sentences transcribed from the recording, the start time and end time of each sentence, the speaker's voiceprint identifier, and the recording channel information. The identity of the customer or salesperson is not determined based on the voiceprint identifier at this time.

[0022] Specifically, the statements are arranged according to their start time, and sentence breaks formed during transcription are corrected. If the pause between adjacent words does not exceed a preset pause threshold, and the preceding word does not form a complete semantic meaning, they are merged into the same statement. If the pause exceeds the preset pause threshold, or the preceding and following words form independent statements, questions, or confirmations, they are divided into different statements. The preset pause threshold is 0.5 seconds, which is based on the common distribution range of pauses between sentences in Mandarin speech speed statistics and is determined in conjunction with the silence segmentation standard commonly used in the field of speech recognition. In practical applications, this threshold can be adjusted within the range of 0.3 seconds to 0.8 seconds according to the recording sampling rate and environmental noise. Furthermore, the turn is divided according to the changes in the speaker's voiceprint identifier. When consecutive statements correspond to the same speaker's voiceprint identifier, they are grouped into the same turn; when the speaker's voiceprint identifier changes, a new turn is established. For overlapping recordings formed by two parties speaking simultaneously, the start and end times of each statement are retained, and overlapping speech marks are added, without forcibly merging them into a single statement.

[0023] In one possible implementation, the text after turn-based segmentation is sequentially calibrated. Statements without overlapping times are arranged according to their start times, while statements with overlapping times are arranged according to their respective turns, with association markers retained at the corresponding time positions. This yields a standard dialogue text containing statement boundaries, turn boundaries, time order, and overlap relationships. Based on this standard dialogue text, for each target statement, its current turn, the previous turn, and the next turn are selected as the context scope. Word segmentation is performed on the statements within the context scope, identifying pronouns, negation words, conditional words, time words, and conjunctions indicating continuation, transition, supplementation, and response. For pronouns such as "this product," "the aforementioned fees," and "this situation," the corresponding object is determined from the preceding statement based on word distance, topic consistency, and grammatical orientation. For negative expressions such as "not liable," "not guaranteed," and "cannot be received in advance," the action or business object modified by the negative word is retrieved from its position, with the sentence end, transition word, or another negative word serving as the boundary of the negation scope. For expressions such as "if," "only," and "after signing," the conditional content, the statements constrained by the condition, and their sequential relationship are recorded to form contextual information corresponding to each target statement.

[0024] For example, when a salesperson responds with "This fee will be collected after the cooling-off period" after a customer's inquiry interrupts the product introduction statement, the context simultaneously includes the preceding product introduction statement, the customer inquiry statement, and the current response statement. The referent of "this fee" is determined based on the fee name in the preceding statement, "after the cooling-off period" is recorded as a time constraint, and "collect" is recorded as a time-constrained business action.

[0025] Specifically, a pre-established term encoding table is used to encode each word in the contextual information. The encoding of each word includes semantic encoding, sentence position encoding, turn-taking position encoding, and temporal sequence encoding. Referential objects, negation scope, conditional objects, and temporal constraints are each marked with relational tags, allowing the same word to generate different encoding results in different contexts. Furthermore, the word codes are compiled according to the order of sentences in the standard dialogue text, and continuation, transition, supplementation, and response relationships are appended between corresponding sentences, forming a dialogue semantic feature representation that simultaneously characterizes sentence content, contextual information, turn-taking position, temporal sequence, and semantic connections.

[0026] Step S102: Based on the speaker identity identifier and semantic role information in the dialogue semantic feature representation, perform speaker role identification and content separation on the original dialogue text, distinguish the customer's speech content from the insurance salesperson's speech content, and form a set of customer interjection paragraphs and agent dialogue flow.

[0027] The initial role of each turn is determined based on the speaker identity identifier in the semantic feature representation of the dialogue, with the agent role corresponding to the insurance salesperson role. If the initial role is inconsistent with the semantics of inquiry, notification, objection, or confirmation, the target role is determined based on the question-and-answer correspondence of adjacent turns, resulting in role-annotated dialogue text. Based on the role-annotated dialogue text, customer role statements and agent role statements are extracted according to timestamps, and the turn boundaries and overlapping speech markers of the statements are retained, resulting in customer speech sequences and agent speech sequences. For the customer speech sequence, consecutive customer role statements located between adjacent agent turns or overlapping with agent role statements in time, and not interrupted by agent role statements, are merged into customer interjection segments, forming a set of customer interjection segments. The agent speech sequence is arranged according to timestamps to form an agent dialogue flow.

[0028] In one implementation, the speaker identification, statement content, start time, end time, and semantic role information corresponding to each turn are extracted from the dialogue semantic feature representation. The speaker identification is derived from the voiceprint consistency relationship of the same speaker in the recording, and the semantic role information includes inquiry, notification, explanation, objection, response, and confirmation, which are used to describe the expressive role of each statement in the dual-recording interaction.

[0029] Specifically, consecutive turn-based conversations are initially assigned roles based on the speaker's identity. At the beginning of the recording, the speaker who provides their identity, organization name, or service number is designated as the agent role; the other speaker who engages in a question-and-answer relationship with the agent role is designated as the customer role. Speakers with the same identity maintain the same initial role throughout the entire recording.

[0030] In one possible implementation, when a complete self-introduction is lacking at the beginning of the recording, the initial role is determined based on semantic role information. Speakers who continuously read out product names, insurance liabilities, fee terms, or risk warnings correspond to the agent role; speakers who raise questions, express objections, or make confirmations regarding the readouts correspond to the customer role, thus forming the initial role markers for each turn. Further, the consistency between the initial role markers and the semantic role information is verified. If a turn marked as an agent role only contains a brief confirmation statement, the role is not directly changed; if the same speaker continuously makes customer identity statements, inquires about terms, and expresses objections, and another speaker in an adjacent turn continuously provides explanations and information, then the initial role is corrected based on the question-and-answer correspondence of adjacent turns to obtain the target role marker.

[0031] It should be noted that the question-and-answer correspondence is determined by the business object in the statement and the content of the response. When a customer asks "How is the cancellation fee calculated?", if subsequent statements explain the cancellation time, fee deduction, and processing conditions, the preceding statement is marked as an inquiry, and the following statement is marked as a response. The customer role and the agent role are determined by combining the identities of both speakers, avoiding changes in roles based solely on a single question word. Based on the target role markings, the original dialogue text is separated. Statements belonging to the customer role are arranged according to their original timestamps to form the customer speaking sequence, and statements belonging to the agent role are arranged according to their original timestamps to form the agent speaking sequence. Each statement retains its turn, start time, end time, and overlapping speech markers, ensuring that the separated statements still correspond to the original recording position.

[0032] For example, while the agent is reading out the insurance period, a customer raises a payment question before the reading is finished, resulting in an overlap in the time intervals of the two statements. The customer's statement is retained in the customer's speech sequence, and the agent's statement is retained in the agent's speech sequence, with the start and end positions of the overlap recorded separately, without concatenating the two statements into a single turn.

[0033] Specifically, for a customer speaking sequence, consecutive customer statements between adjacent agent turns are merged into a single customer interjection segment. If a customer statement overlaps with an agent statement in time, and subsequent customer supplementary statements occur, the overlapping and supplementary statements are merged. The current customer interjection segment ends upon the reappearance of the agent statement. A set of customer interjection segments is formed based on the start time of each segment, and the agent dialogue flow is then established according to the original chronological order of the agent speaking sequence.

[0034] Step S103: For each interruption segment in the customer interruption segment set, extract the business process keywords and match them with the preset thesaurus. If the process keywords in the interruption segment do not belong to the thesaurus range corresponding to the time of the interruption, it is determined that the interruption segment has a process attribution deviation and is marked as a conflict segment.

[0035] For the set of customer interruption segments, extract product names, insurance liabilities, expense items, risk items, and confirmation items from each segment. Merge synonyms according to a pre-defined synonym table to obtain a set of business process keywords. Based on the start time of each interruption segment, determine the agent's turn currently being read or most recently ended in the agent's dialogue flow, and use the segment label of that turn to determine the target segment thesaurus. Match the set of business process keywords with the target segment thesaurus and other segment thesauruses, excluding shared words that exist in multiple segment thesauruses simultaneously. If the remaining keywords do not match the target segment thesaurus but match other segment thesauruses, the interruption segment is considered to have a segment attribution discrepancy and is marked as a conflicting segment.

[0036] In one implementation, for each interruption segment in the set of customer interruption segments, the segment text, start time, end time, and corresponding turn position are read. The segment text is then segmented into sentences, and business terms are extracted according to the product name, insurance liability, fee items, risk items, term items, cancellation items, and confirmation items in the insurance dual recording business.

[0037] Specifically, a thesaurus corresponding to each step of the dual-recording standard process is pre-established. The identity verification thesaurus records terms related to name, ID, self, and identity confirmation; the product introduction thesaurus records terms related to product name, insurance period, payment period, and coverage; the risk disclosure thesaurus records terms related to uncertain returns, surrender losses, fee deductions, and exclusions; and the customer confirmation thesaurus records terms related to awareness, understanding, agreement, and confirmation. Furthermore, a thesaurus is created for terms with different expressions but consistent business meaning. This thesaurus merges "surrender fee," "contract termination fee," and "surrender deduction" into the "surrender fee" item, and merges "coverage period" and "insurance period" into the "insurance period" item. The merged items correspond to the original terms and their corresponding statements, forming a set of keywords for each business step.

[0038] In one possible implementation, words with only general meanings are not used as standalone keywords for business processes. When words such as "product," "fee," and "confirmation" appear simultaneously in multiple business processes, they must be combined with qualifiers in the same statement to form a complete term. For example, "insurance cancellation fee" and "fee" form the term "insurance cancellation fee," and "identity" and "confirmation" form the term "identity confirmation," to reduce confusion caused by shared terms across processes. Based on the start time of the interrupted segment, the agent's turn corresponding to the time position is retrieved from the agent's dialogue flow. If the interruption start time is between the start and end times of a certain agent's turn, that agent's turn is selected; if the interruption occurs between adjacent agent turns, the agent's turn that most recently ended before the start time is selected to determine the agent's turn corresponding to the time the interruption occurred.

[0039] Specifically, each agent's turn of speech is pre-labeled with a segment tag based on the content they read. When an agent's turn of speech includes their name, ID number, and personal confirmation, it corresponds to an identity verification segment; when it includes product liability, insurance period, and payment details, it corresponds to a product introduction segment; and when it includes information on surrender losses, exclusions, and uncertain returns, it corresponds to a risk disclosure segment. The target segment terminology is then retrieved based on these segment tags.

[0040] For example, when an agent is reading out the product insurance period, a customer requests a refund and asks whether there will be a charge. The terms "refund" and "charge" in the interrupted segment are grouped together as the "refund fee" term. The agent's turn at the moment the interruption occurs is marked as the product introduction segment. Therefore, the product introduction terminology is identified as the target segment terminology, and the "refund fee" term in the risk disclosure terminology is identified as another segment term. The set of keywords for each business segment is compared item by item with the target segment terminology and other segment terminologies, excluding shared words recorded in multiple segment terminologies. If the remaining business segment keywords do not match the target segment terminology and match at least one exclusive term in another segment terminology, it is determined that the corresponding interrupted segment has a segment attribution deviation, and a conflict marker is added to the interrupted segment, forming a conflict segment.

[0041] Step S104: Determine the semantic relationship between each conflict segment and each segment in the agent's dialogue flow, and determine the actual segment to which each conflict segment belongs based on the relationship results.

[0042] For the conflicting segment, the business object, inquiry content, and referent are extracted. Then, segment keywords and agent responses are extracted from each segment of the agent's dialogue flow. Candidate segment segments are obtained based on synonym correspondence, referent correspondence, and business object consistency. For each candidate segment, it is determined whether the agent's response directly answers the inquiry content, and segment segments with inconsistent business objects or uncorrelated referents are excluded. A candidate segment segment sequence is formed based on the degree of question-answer correspondence. Segments that directly answer the inquiry content are selected from the candidate segment segment sequence. If no direct answer exists, segment segments with the same business object and corresponding segment keywords as the conflicting segment are selected, and their true belonging segment is determined based on their segment tags.

[0043] In one implementation, the agent's response content is fully preserved and used as a key basis for semantic association analysis with the conflicting segments. Then, the conflicting segments with attached conflict tags and the agent's dialogue flow are obtained. The agent's dialogue flow is divided into multiple segment segments according to the dual-recording business segment tags. Each segment segment contains continuous agent statements, start time, end time, segment tag, and the standard business item corresponding to that segment.

[0044] Specifically, the business object, inquiry content, and referent are extracted from the conflicting segments. The business object refers to the product name, insurance liability, payment matters, policy cancellation matters, fee matters, or risk matters involved in the customer's statement; the inquiry content consists of the question object and the questioning action; the referent is the preceding business object to which "this fee," "the aforementioned liability," or "this method" refers. Further, the agent's statements in each segment are broken down into business content, and segment keywords and agent responses are extracted. Segment keywords are words that indicate the scope of business in the segment, and agent responses are explanations, limitations, confirmations, or risk disclosures made to the business object. Each agent response retains its original statement and corresponding segment tag.

[0045] In one possible implementation, preliminary matching is first performed based on synonym correspondences. "Cancellation deduction" in the conflicting paragraph and "contract termination fee" in the segment are identified as the same fee item; "coverage duration" and "insurance period" are identified as the same time period item. Segments where the business objects do not have synonym correspondences are not included in the candidate scope.

[0046] Specifically, the initially matched segment continues to verify the referents and question-and-answer correspondences. If the conflicting segment contains a clear business object, it is determined whether the agent's response revolves around that business object. If the conflicting segment uses pronouns, the customer's and agent's statements are read backward from the point of interruption until the business object corresponding to the pronoun in terms of singular / plural, matter category, and modifiers is obtained. Then, it is determined whether the agent's response answers the question in the inquiry. When a customer asks "whether there will be a deduction for canceling the policy," an agent's statement that includes the conditions for collecting or the scope of deductions for canceling the policy is a direct answer; an agent's statement that only mentions the cancellation application method is an example of having the same business object but a different answer. After verification, segment segments with consistent business objects and valid referents are retained to obtain candidate segment segments.

[0047] It should be noted that when a customer interrupts and the agent does not respond immediately, the question-and-answer relationship is not limited to adjacent turns. The agent can search for segments following the conflicting segment along the conversation flow. If an explanatory statement corresponding to the inquiry is found within such a segment, it is included as a candidate segment. The search scope ends when the same business object is fully explained or the agent moves to another unrelated business object. For multiple candidate segments, they are arranged according to the degree of correspondence between the agent's response and the inquiry. The degree of correspondence decreases sequentially as follows: the first level is when the agent's response directly answers the question; the second level is when the agent's response shares the same business object as the inquiry and contains corresponding explanatory keywords; the third level is when the agent's response only shares the same business object as the inquiry but does not answer the question. Within the same level, segments with clearly corresponding referents and whose segment keywords match the keywords of the conflicting segment are prioritized, forming a candidate segment sequence.

[0048] For example, when a customer asks "How much will be deducted if I don't want to buy it anymore?" during the product introduction, the business object in the conflicting paragraph is identified as the cancellation fee, and the question is identified as the fee deduction upon cancellation. The product introduction only includes the premium amount, the risk disclosure includes an explanation of the losses incurred upon contract termination and the cancellation amount, and the customer confirmation only includes confirmation of awareness. The risk disclosure segment satisfies both the consistency of the business object and the direct answer relationship, and is therefore ranked first in the candidate segment sequence. A segment that directly answers the question is selected from the candidate segment sequence. If no direct answer exists, a segment with consistent business object, corresponding referent, and whose segment keywords match the keywords of the conflicting paragraph is selected. The segment tag of the selected segment is read, and this segment tag is determined as the true belonging segment of the conflicting paragraph.

[0049] Step S105: Based on the true attribution information, extract the original timestamp of the conflicting segment, map the conflicting segment to the time interval corresponding to the true attribution segment, and obtain the adjusted time information.

[0050] Based on the true attribution information, the original start and end times of the conflicting segments are obtained. The original duration is determined by the time difference between the two. The start and end times of the corresponding segment fragment of the true attribution are extracted to obtain the target segment time interval. Based on the target segment time interval and the original sequence of the conflicting segments, the start time of the agent's statement corresponding to the true attribution is used as the mapping starting point. The mapping ending point is determined according to the formula "mapping ending point = mapping starting point + original duration" to obtain the mapping time interval. If the mapping ending point is later than the end time of the segment fragment, the mapping ending point is determined by the end time, and the mapping starting point is determined backward. If the mapping starting point is earlier than the start time of the segment fragment, the target segment time interval is used as the mapping time interval. The original time, mapping time, and true attribution segment are recorded to obtain the adjusted time information.

[0051] In one implementation, the true attribution point, original start time, and original end time corresponding to each conflicting segment are obtained. The original start time represents the recording moment when the customer begins to interrupt, and the original end time represents the recording moment when the continuous customer speech ends. The time difference between the two is recorded as the original duration, and the original time is not overwritten due to attribution adjustments.

[0052] Specifically, the start and end times of the segment corresponding to the true attribution stage are read from the agent dialogue flow. If the same true attribution stage contains multiple discontinuous segment segments, the segment containing the agent's statement is locked based on the agent's statement selected when determining the true attribution stage, and the time range of that segment is determined as the target stage time interval. Further, the start time of the agent's statement is obtained and used as the initial mapping anchor point for the conflicting segment. The agent's statement is either a direct answer to the customer's inquiry or a statement with the same business object and corresponding stage keyword as the conflicting segment, thus ensuring that the mapping position uses the same agent's content as the true attribution determination.

[0053] In one possible implementation, when the agent statement corresponding to the conflicting paragraph is a complete response statement, the start time of the agent statement is used as the mapping starting point, and the mapping ending point is determined backward according to the original duration. When the corresponding content consists of multiple consecutive agent statements, the start time of the first agent statement is used as the mapping starting point, so that the mapping time covers the beginning of the group of response content. If the mapping ending point does not exceed the end time of the target stage time interval, the mapping starting point and the mapping ending point are retained. If the mapping ending point is later than the end time, the end time of the target stage time interval is first used as the corrected mapping ending point, and then the corrected mapping starting point is determined backward according to the original duration, so that the mapping range of the conflicting paragraph is within the actual belonging stage.

[0054] It should be noted that if the forward-determined mapping starting point is earlier than the start time of the target segment's time interval, it indicates that the original duration exceeds the acceptable duration of the target segment's time interval. In this case, the start time of the target segment's time interval is used as the mapping starting point, and the end time of the target segment's time interval is used as the mapping ending point, while recording the duration truncation mark. For cases where multiple conflicting segments are mapped to the same true segment, the mapping order is arranged according to the original start time of each conflicting segment. When the initial mapping starting point of a later conflicting segment is earlier than the mapping ending point of a previous conflicting segment, the previous mapping ending point is used as the starting point for the later mapping, and the mapping ending point is re-determined based on the original duration of the later conflicting segment; if the duration exceeds the segment's end time, boundary correction is performed.

[0055] Specifically, for each conflict segment, the original start time, original end time, original duration, target segment time interval, mapping start point, mapping end point, actual belonging segment, and boundary correction mark are recorded to form adjusted time information corresponding to each conflict segment.

[0056] Step S106: Based on the adjusted time information, rearrange the segment attribution tags of the entire dialogue according to the segment order of the dual recording standard process and divide the segment boundaries to obtain a target segment sequence that conforms to the standard process order. Perform integrity verification on the target segment sequence, extract the standard script node list that the insurance salesperson should read in each segment, compare it with the actual script content to detect missed or skipped segments, perform attribution correction and script supplementation, and update the target segment sequence.

[0057] Based on the adjusted time information, the labels of each stage are arranged according to the standard dual-recording process. Agent statements with consecutive adjusted time information corresponding to the same stage label are merged into stage segments. The adjusted customer interjection segments are included within the time range corresponding to their actual stage. Stage boundaries are defined between adjacent stage segments to obtain the target stage sequence. For the target stage sequence, a standard script node list for each stage is obtained from a preset dual-recording script library. These nodes are matched item by item with the actual agent statements according to the business object, notification item, and node order. Nodes without corresponding statements are marked as missed read nodes. When there are uncorresponding nodes between adjacent corresponding nodes, a skipped reading segment is determined. Based on the skipped reading segment, the actual agent statements incorrectly assigned to adjacent stages are corrected, and a missing node mark is written at the corresponding stage position. The original time and adjusted time are retained to obtain the updated target stage sequence.

[0058] Specifically, the process involves obtaining the stage labels, original time, adjusted time, and true stage attribution for each agent's statements throughout the entire dialogue. Following the standard dual-recording process of identity verification, product introduction, risk disclosure, customer confirmation, and signature confirmation, a sequence identifier is assigned to each stage, and the original stage labels with attribution discrepancies are replaced with the true stage attribution.

[0059] Specifically, agent statements are arranged according to their sequence identifiers. Agent statements belonging to the same segment and whose adjusted times are consecutive are merged into one segment segment; agent statements belonging to the same segment but separated by other segment statements are re-integrated into the corresponding segment segment based on their actual belonging segment, while retaining their original times, ensuring that the rearrangement does not change the actual occurrence position in the original recording. Furthermore, segment boundaries are defined between adjacent segment segments. The end time of the last agent statement in the preceding segment segment is recorded as the front boundary, and the start time of the first agent statement in the following segment segment is recorded as the back boundary. When there is a customer interruption segment between the two segments, its adjusted time is incorporated into the time range of the corresponding segment segment based on the actual belonging segment of the customer interruption segment, resulting in the target segment sequence.

[0060] In one possible implementation, when the same segment is read multiple times in the original recording, the agent statements with the same business target and consistent standard process position are grouped into the same segment segment. If the salesperson responds to a customer's inquiry about surrender losses during the product introduction, the response statement is included in the risk disclosure segment, but the product introduction segment can continue to record subsequent product liability statements after the response ends. For the target segment sequence, a list of standard script nodes corresponding to each segment is read from a preset dual-recording script library. Each standard script node records at least the node name, the segment it belongs to, the business target, the matters to be disclosed, and the node order. The standard script nodes for the risk disclosure segment may include disclosure of uncertain returns, disclosure of surrender losses, disclosure of expense deductions, and disclosure of liability exclusions, with each node corresponding to independent matters to be disclosed.

[0061] Specifically, following the node sequence, each standard script node is matched item by item with the actual agent statements in its corresponding segment. During the matching process, the business target, action content, limiting conditions, and negation scope are extracted from the actual agent statements. When the standard script node is for informing about policy cancellation losses, the actual agent statement must simultaneously include information about contract termination or policy cancellation, as well as notification of amount reduction, fee deduction, or loss liability. If only "policy cancellation is possible" appears without any loss content, the actual agent statement is not considered to have matched the aforementioned standard script node. When the standard script node is for informing about uncertain returns, the actual agent statement must express that returns are not fixed, the demonstrated amount does not represent actual returns, or the actual result is affected by agreed conditions. After completing the item-by-item matching, the matched or unmatched status of each standard script node is recorded.

[0062] For example, the standard script node list sequentially includes statements about insurance liability, insurance period, and payment method. If the target segment contains statements about insurance liability and payment method, but not about the insurance period, then the statement about the insurance period is marked as uncorresponding. When both the preceding and following standard script nodes are corresponding, the uncorresponding node between them is identified as a skip node, and its corresponding flow position is identified as a skip segment.

[0063] It should be noted that when there are consecutive uncorresponding nodes at the end of a segment, the uncorresponding nodes are marked as missed read nodes; uncorresponding nodes located between two corresponding nodes in a segment are marked as skipped read nodes. If multiple consecutive uncorresponding nodes are within the same interval of read nodes, they are merged and recorded into a skipped read segment according to the standard script node order, and the node names of each uncorresponding node are retained.

[0064] In one embodiment, the actual agent's statement corresponds to a standard script node. However, when this statement is reclassified into an adjacent segment, its segment affiliation is corrected based on the business object, the notification item, and the node order. The corrected agent's statement is removed from the original segment and written into the corresponding segment segment, while the start time, end time, and adjacent segment boundaries of the affected segment are redefined. Further, for missed or skipped nodes, a missing node marker is written at the corresponding node position in the target segment sequence. The missing node marker includes the segment it belongs to, node name, node order, missing category, and adjacent corresponding nodes; it does not include the script text not actually read by the salesperson. The corrected actual agent's statement, missing node markers, original time, adjusted time, and redefined segment boundaries are combined to obtain the updated target segment sequence.

[0065] Step S107: For each step in the updated target step sequence, extract the corresponding compliance check rules from the preset dual-recording quality inspection rule library, compare the actual scripts of each step with the standard scripts, determine whether each step is compliant, and obtain the quality inspection results.

[0066] For the updated target sequence, based on the tags of each stage, mandatory reading nodes, prohibited expressions, reading order, and customer confirmation requirements are extracted from the preset dual-recording quality inspection rule library to form a set of compliance inspection rules corresponding to each stage. According to the set of compliance inspection rules, the actual agent statements for each stage are matched with standard script nodes according to business object, notification content, scope of rejection, restrictive conditions, and node order. The corresponding customer statements are then checked against the customer confirmation requirements to obtain a rule verification record. Based on the rule verification record, if mandatory reading nodes are missing, prohibited expressions appear, node order is incorrect, or customer confirmation is missing, the corresponding stage is determined to be non-compliant; otherwise, the corresponding stage is determined to be compliant. The stage tags, judgment results, and corresponding statements are then compiled to obtain the quality inspection result.

[0067] In one implementation, an updated target sequence is obtained. The target sequence includes each sequence label, actual agent statements, corresponding customer statements, standard script node status, missing node markers, original time, adjusted time, and sequence boundaries, arranged in the order of the dual-recording standard process.

[0068] Specifically, based on the stage label of each stage, the corresponding compliance inspection rules are extracted from the preset dual-recording quality inspection rule library. The compliance inspection rules include mandatory node rules, prohibited expression rules, node order rules, and customer confirmation rules. Each rule records the stage it belongs to, the object to be checked, the items to be checked, and the non-compliance conditions.

[0069] For example, the mandatory reading rules for the identity verification stage may include the customer's name, identification information, and personal identity confirmation; the product introduction stage may include the product name, insurance period, payment period, and insurance liability; and the risk disclosure stage may include uncertain returns, surrender losses, fee deductions, and exclusions. Furthermore, for the mandatory reading rules, the corresponding standard script nodes and actual agent statements in the target stage sequence are read. Verification is performed according to the business target, the notification action, the limiting conditions, and the scope of negation. If the actual agent statement only mentions the business target name but does not express the corresponding notification, it is not recorded as a completed mandatory reading node.

[0070] In one possible implementation, for prohibited expression rules, promise-making words and their modified objects are extracted from the actual agent statements. If an agent statement makes a definite promise regarding the amount of benefit, the outcome of protection, or the amount of surrender value, and the corresponding rule prohibits such statements, then the result of the prohibited expression is recorded, and the original time and the process in which the statement was executed are associated with it.

[0071] Specifically, the node order rule checks the actual agent statements against their positions in the standard script node list. If the agent statement corresponding to a later standard script node precedes the previous standard script node, and there is no reasonable interruption due to a customer's interruption, the node order is recorded as incorrect; missing node markers are not considered read nodes in the order check. For customer confirmation rules, the agent's inquiry statement and subsequent customer statements are read from the corresponding stage. When the customer's statement clearly expresses awareness, understanding, agreement, or personal confirmation, customer confirmation is recorded as complete; when the customer only raises questions, changes the topic, or does not respond, customer confirmation is recorded as missing. The customer confirmation content retains the speaker's role and time position. Based on the results of each rule check, if a corresponding stage has missing mandatory nodes, violated prohibited expressions, incorrect node order, or missing customer confirmation, the stage is deemed non-compliant; other stages are deemed compliant. The quality inspection results are obtained by compiling stage tags, compliance status, violated rules, corresponding statements, missing nodes, and time information.

[0072] This invention provides an intelligent quality inspection system based on process segmentation, mainly comprising: The semantic encoding module is used to acquire the original dialogue recording text and perform semantic encoding, extracting contextual information and semantic association features from the dialogue text to obtain a dialogue semantic feature representation. The role recognition and content separation module is used to perform speaker role recognition and content separation on the original dialogue text based on the speaker identity identifier and semantic role information in the dialogue semantic feature representation, distinguishing the customer's speech content from the insurance salesperson's speech content, forming a set of customer interjections and agent dialogue flow; The conflict paragraph identification module is used to extract business process keywords from each interruption paragraph in the customer interruption paragraph set and match them with a preset thesaurus. If the process keyword in the interruption paragraph does not belong to the thesaurus range corresponding to the time of the interruption, it is determined that the interruption paragraph has a process attribution deviation and is marked as a conflict paragraph. The True Attribution Determination Module is used to determine the semantic association between each conflict segment and each segment in the agent's dialogue flow, and to determine the true attribution of each conflict segment based on the association results. The time adjustment module is used to extract the original timestamp of the conflict segment based on the information of the actual attribution link, map the conflict segment to the time interval corresponding to the actual attribution link, and obtain the adjusted time information. The segment sequence generation and update module is used to rearrange the segment attribution tags of the entire dialogue according to the segment order of the dual recording standard process based on the adjusted time information and divide the segment boundaries to obtain a target segment sequence that conforms to the standard process order. The module performs integrity verification on the target segment sequence, extracts the list of standard script nodes that the insurance salesperson should read in each segment, compares it with the actual script content to detect missed and skipped segments, adjusts the segment boundaries and sentence fragment positions according to the skipped segments, and writes missing node markers at the corresponding segment positions according to the missed nodes, and updates the target segment sequence. The compliance check module is used to extract the corresponding compliance check rules from the preset dual-recording quality inspection rule library for each step in the updated target step sequence, compare the actual scripts of each step with the standard scripts, determine whether each step is compliant, and obtain the quality inspection results.

[0073] 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. An intelligent quality inspection method based on process segmentation, characterized in that, The method includes: The original dialogue recording text is obtained, and semantic encoding processing is performed to form a dialogue semantic feature representation that includes contextual information, semantic associations, speaker identification, and semantic role information. Based on the semantic features of the dialogue, role recognition and content separation are performed to form a set of customer interjection segments and an agent dialogue flow; The set of customer interruption segments is processed to identify conflict segments, extract keywords of business processes, and identify process attribution deviations based on the process thesaurus corresponding to the time of the interruption to form conflict segments. Based on the semantic association between the conflicting paragraph and each segment in the agent's dialogue flow, the process of determining the true attribution stage is performed; Based on the original timestamps of the actual attribution link and the conflicting segment, time information adjustment processing is performed to form the adjusted time information; Based on the adjusted time information and the standard dual recording process, the process sequence generation and integrity verification are performed. The process attribute tags are rearranged, process boundaries are defined, and a target process sequence is formed. Missed reading nodes and skipped reading paragraphs are identified based on the standard script node list. The process boundaries and sentence fragment positions are adjusted according to the skipped reading paragraphs. Missed node markers are written at the corresponding process positions according to the missed reading nodes, thus forming an updated target process sequence. Compliance checks are performed based on the updated target sequence. Compliance check rules for each step are extracted from the dual-recording quality inspection rule library, and the actual script, standard script, and customer confirmation content are compared to form the quality inspection results.

2. The intelligent quality inspection method based on process segmentation according to claim 1, characterized in that, The semantic encoding process includes: The speech turns are divided based on speaker switching, sentence pauses, and complete sentence boundaries, and the start and end times and text order of the speech turns are corrected. Extract the question-and-answer sequence between adjacent turns and the referential, negation, conditional, and temporal relationships within the same turn; Encode the sequential, transitional, supplementary, and responsive relationships between objects and statements in the aforementioned relationship.

3. The intelligent quality inspection method based on process segmentation according to claim 1, characterized in that, The role recognition and content separation process includes: The initial roles of each turn are determined based on the speaker's identity, and the target roles are corrected according to the semantics of inquiry, notification, objection, and confirmation, as well as the question-and-answer relationship between adjacent turns. Extract customer role statements and agent role statements by timestamp, and retain the call boundaries and overlapping speech markers; Merge consecutive customer role statements between adjacent agent turns or overlapping with agent statement times to form the customer interruption segment set, and arrange the agent role statements to form the agent dialogue flow.

4. The intelligent quality inspection method based on process segmentation according to claim 1, characterized in that, The conflict paragraph identification process includes: Extract the product name, insurance liability, expense items, risk items, and confirmation items from the customer's interjection paragraphs, and merge synonymous expressions; The agent turn that is being read or has recently ended is determined based on the start time of the interruption, and the target segment vocabulary is determined based on the segment label of the agent turn; Match the target stage's thesaurus with other stage thesauruses, excluding shared words; If the keywords for the remaining business process do not match the target process's keyword library but match other process keyword libraries, mark the corresponding customer interjection paragraph as a conflicting paragraph.

5. The intelligent quality inspection method based on process segmentation according to claim 1, characterized in that, The process for determining the true ownership includes: Extract the business objects, inquiry content, and referents of the conflict segments, and extract the segment keywords and agent response content of each segment in the agent dialogue flow; Candidate segmentation is formed based on the correspondence of synonyms, the correspondence of referents, and the consistency of business objects; Exclude segments where the business objects are inconsistent or the referents cannot be matched, and arrange them according to the degree of correspondence between questions and answers; Select the segment that directly answers the question. If no direct answer exists, select the segment with the same business object and the corresponding segment keyword, and determine the actual segment to which it belongs based on the segment tag of the segment.

6. The intelligent quality inspection method based on process segmentation according to claim 1, characterized in that, The time information adjustment process includes: The original duration is determined based on the original start and end times of the conflicting segments, and the start and end times of the actual attribution segment are obtained. The mapping start point is determined by the start time of the agent's statement when the actual attribution is determined, and the mapping end point is determined by the original duration. When the mapping interval exceeds the actual time interval of the assigned link, the mapping start point or mapping end point is corrected according to the corresponding boundary, and the original time, mapping time and actual assigned link are recorded.

7. The intelligent quality inspection method based on process segmentation according to claim 1, characterized in that, The process of generating and verifying the integrity of the sequence includes: Arrange the labels of each stage according to the dual recording standard process, merge the agent statements corresponding to the same stage label and with the adjusted time information, and divide the stage boundaries between adjacent different stage segments. Extract the standard script node list for each stage from the dual-recording script library, and match the actual agent's statements according to the business object, the notification item, and the order of the nodes. Nodes without corresponding statements are marked as missed read nodes, and paragraphs are marked as skipped when there are uncorresponding nodes between adjacent corresponding nodes; Adjust the position of the sentence fragments in the target segment sequence according to the skipped reading paragraph, and merge the agent sentences that belong to the same segment label and have continuous time information after adjustment into segment fragments; Based on the missed reading node, write the missing node mark at the corresponding stage position, and retain the original time and the adjusted time.

8. The intelligent quality inspection method based on process segmentation according to claim 1, characterized in that, The compliance check process includes: Based on the process tags in the updated target process sequence, mandatory nodes, prohibited expressions, reading order, and customer confirmation requirements are extracted from the dual-recording quality inspection rule base to form a set of compliance inspection rules; The actual agent's statements and standard script nodes are matched according to the business object, the content of the notification, the scope of the rejection, the limiting conditions, and the order of the nodes. The customer's statements are checked against the customer's confirmation requirements to form a rule check record.

9. The intelligent quality inspection method based on process segmentation according to claim 8, characterized in that, The rule-based record verification includes: Record the corresponding statements of the required reading nodes, the location of the prohibited expressions, the node order correspondence, and the customer confirmation content; If there are missing required reading nodes, prohibited expressions, incorrect node order, or missing customer confirmation, mark the corresponding step as non-compliant; otherwise, mark the corresponding step as compliant. The process labels, judgment results, and corresponding statements are collected to form the quality inspection results.

10. An intelligent quality inspection system based on process segmentation, characterized in that, include: The semantic encoding module is configured to acquire the original dialogue recording text and form a dialogue semantic feature representation; The role recognition and content separation module is configured to form a set of customer interjection segments and an agent dialogue flow based on the dialogue semantic feature representation; The conflict paragraph identification module is configured to generate conflict paragraphs based on the matching relationship between keywords in business processes and process thesaurus; The module for determining the true attribution stage is configured to determine the true attribution stage based on the semantic association between the conflicting paragraph and each stage segment. The time information adjustment module is configured to generate adjusted time information based on the original timestamp and the actual time interval of the affiliated link. The segment sequence generation and integrity verification module is configured to generate a target segment sequence based on the adjusted time information, identify missed reading nodes and skipped reading segments, and update the target segment sequence. The compliance inspection module is configured to generate quality inspection results based on the updated target process sequence and the dual-recording quality inspection rule base.

Images