A call transcription structured knowledge analysis management method and related products

Abstract

The application discloses a call transcription structured knowledge analysis management method and related products. First, the upstream call transcription object is processed by turn normalization to form a unified and standardized turn structure. In the dialogue unit segmentation stage, the semantic change is identified by means of the segmentation line rule, the error segmentation rule is used to avoid damaging the single-turn dialogue semantics, and the overlap determination rule is used to allow the main topic to be accompanied by temporary branches and the business unit turn to overlap and output the overlap correlation unit identifier. Subsequently, the business field distillation is directly carried out based on the original turn text corresponding to the business unit, without relying on the preset standard dialogue library, to generate the unit-level structured knowledge item. Then, the correlation mark is written into the knowledge item through the rule correlation operation, and the interactive correlation information of the call is solidified. Finally, the quality management verification and quality repair processing are used to output the structured knowledge item after management is completed, so that the automatic analysis and sedimentation of the fine-grained unit-level structured knowledge from the original call text are realized.

Description

Technical Field

[0001] This application relates to the field of data processing technology, specifically to a method for managing structured knowledge parsing in call transcription and related products. Background Technology

[0002] In call scenarios such as intelligent customer service, outbound marketing, and complaint follow-up, call recordings can be transcribed into structured text by Automatic Speech Recognition (ASR). This text contains a large amount of implicit business knowledge, such as response paradigms, objection handling strategies, and topic interweaving and response methods. Enterprises need to automatically mine, sort out, and accumulate this kind of implicit knowledge from massive amounts of call data to support multiple downstream intelligent applications such as call review analysis, business quality inspection, agent training, and intelligent response.

[0003] Currently, knowledge parsing and processing of call texts mostly involves segmenting dialogue content and extracting knowledge through conversation turn splitting, semantic clustering matching, or semantic closed-loop slicing distillation. However, these methods have significant drawbacks in real-world call scenarios. Not only do they fail to retain the overlapping relationships between conversation turns after turn splitting, but they also generally adopt a forced semantic closed-loop slicing mode, which easily loses the complex topic interweaving, branching, and parallel interaction details in real calls. Furthermore, the overall processing heavily relies on a pre-set standard dialogue library and cannot autonomously complete fine-grained, unit-level structured knowledge accumulation based on the original call text. This makes it difficult to adapt to the upgraded requirements of high-precision knowledge retrieval, efficient knowledge reuse, and refined business quality inspection under massive call data.

[0004] Therefore, how to achieve fully automated parsing and accumulation of structured knowledge at the fine-grained unit level in original calls is a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] To address the aforementioned issues, this application provides a method and related products for the governance of structured knowledge parsing in call transcription. This method can automatically mine and accumulate unit-level structured business knowledge from the original call transcription text while preserving the true interactive structure of the call, thus meeting the application needs of knowledge retrieval, reuse, and refined quality inspection under massive call data.

[0006] The embodiments of this application disclose the following technical solutions: A method for managing structured knowledge parsing in call transcription, the method comprising: Receive upstream call transcription objects and normalize them into multiple round structures; each round structure includes a round number, a general role tag for this round, a start time for this round, an end time for this round, and round text; the general role tag for this round is obtained by mapping the original role name to a pre-configured role mapping table; The multiple round structures are used as dialogues to be segmented, and the segmentation rules are used as segmentation constraints. These are input into a pre-defined large model to segment dialogue units, resulting in multiple business units. Each business unit includes a round range, business type, topic convergence state, overlapping branch state, and overlapping associated units. The segmentation rules include a dividing line rule, a rule to prohibit incorrect segmentation, and an overlap determination rule. The dividing line rule is used to identify the location where the business semantics change during the call. The rule to prohibit incorrect segmentation is used to constrain the large model from breaking up the single-round dialogue content with the same semantics. The overlap determination rule is used to limit the overlap of round ranges of different business units to only be allowed in call scenarios where the main topic is continuous and accompanied by temporary branches in parallel. It prohibits full-scale stacking overlap without business logic and outputs the corresponding overlapping associated unit identifier when overlap exists. For each business unit, the round range of the business unit and the round text corresponding to the round number covered by the round range, together with the preset distillation constraint prompts, are used as distillation inputs and input into the large model to perform business field distillation, thereby obtaining the structured knowledge entries corresponding to each business unit; the structured knowledge entries are used to describe the business communication information contained in the call dialogue within the corresponding business unit. Based on the time attributes and business field information of the business units corresponding to each structured knowledge entry, rule association operations are performed on each structured knowledge entry to obtain the association tag information corresponding to each structured knowledge entry, and the association tag information is written into the corresponding structured knowledge entry to obtain the tagged structured knowledge entry. Each labeled structured knowledge item undergoes quality governance verification to obtain verification results. Based on each verification result, the corresponding quality repair process is performed on the corresponding labeled structured knowledge item to obtain the remediated structured knowledge item.

[0007] In one possible implementation, the start time of the current round of a business unit is the start time of the current round of the round structure corresponding to the smallest round number covered by the round range of the business unit; the end time of the current round of a business unit is the end time of the current round of the round structure corresponding to the largest round number covered by the round range of the business unit. The process involves performing rule-based association operations on each structured knowledge entry based on the time attributes and business field information of the business unit corresponding to each structured knowledge entry, obtaining association tagging information corresponding to each structured knowledge entry, and writing the association tagging information into the corresponding structured knowledge entry to obtain tagged structured knowledge entries, including: For each structured knowledge entry, the temporal predecessor and successor relationship of the business unit corresponding to the structured knowledge entry is determined based on the start time of the current round, and the temporal predecessor and successor relationship is written into the structured knowledge entry. For each structured knowledge entry, perform interval intersection operation on the round range of the business unit corresponding to the structured knowledge entry to complete the overlapping parallel association verification, obtain the overlapping association unit identifier of the structured knowledge entry, and write the overlapping association unit identifier into the structured knowledge entry. For all structured knowledge entries, perform similar clustering according to the main core topic field within the structured knowledge entry. Structured knowledge entries with the same main core topic are grouped into the same cluster group. Assign a unique cluster group identifier to each cluster group and write the cluster group identifier into the corresponding structured knowledge entry. For all structured knowledge entries, the Jaccard similarity algorithm is used to calculate the text similarity between each pair of structured knowledge entries. When the similarity is greater than or equal to a preset threshold, the corresponding structured knowledge entries are marked as potentially conflicting, and the potentially conflicting marks are written into the corresponding structured knowledge entries. Among them, temporal predecessor-successor relationships, overlapping associated unit identifiers, cluster group identifiers, and suspected conflicts are all association marker information.

[0008] In one possible implementation, the step of determining the temporal predecessor-successor relationship of each structured knowledge entry based on the current round start time of the business unit corresponding to the structured knowledge entry, and writing the temporal predecessor-successor relationship into the structured knowledge entry, includes: Sort all current structured knowledge items in ascending order according to the start time of the current round of the corresponding business unit; For each structured knowledge item, if the topic convergence status of the business unit corresponding to the structured knowledge item is not converged, then the business unit whose start time in this round is later than that business unit and whose overlapping branch status is yes is determined as the successor unit of that business unit; if the topic convergence status of the business unit corresponding to the structured knowledge item is converged, then the business unit whose start time in this round is earlier than the end time of this round of that business unit and whose topic convergence status is not converged is determined as the successor unit of that business unit. Based on the correspondence between successor units recorded in all structured knowledge entries, the predecessor units of each structured knowledge entry are derived in reverse. The information of the predecessor and successor units of the structured knowledge entry is written into the structured knowledge entry as a temporal predecessor-successor relationship.

[0009] In one possible implementation, the step of deriving the predecessor unit of each structured knowledge entry in reverse order based on the correspondence between successor units recorded in all structured knowledge entries includes: Read the subsequent units of each structured knowledge entry; If the successor unit of a certain structured knowledge entry is a business unit corresponding to another structured knowledge entry, then the business unit corresponding to the structured knowledge entry is determined as the predecessor unit of the other structured knowledge entry. Specifically, for structured knowledge entries that are not pointed to by any successor unit of a structured knowledge entry, their predecessor unit is set to empty.

[0010] In one possible implementation, the quality governance verification includes format verification, segmentation correctness verification, business type correctness verification, and conflict verification. Among them, format verification failure is judged as mechanical problem, segmentation correctness verification failure is judged as large model quality problem, business type correctness verification failure is judged as system problem, and conflict verification failure is classified into mechanical problem, model quality problem or system problem according to the problem category. The quality repair process for the aforementioned mechanical problems is: automatic repair via program execution; The quality repair process for the aforementioned model quality issues is to re-call the large model to perform splitting or distillation processing. The quality repair process for the system-related issues is as follows: modify the configuration and perform reclassification.

[0011] In one possible implementation, the method further includes: The structured knowledge entries, after quality repair processing, are subjected to quality governance verification again. The structured knowledge entries that pass the second verification will be marked as active for external retrieval. The structured knowledge items that fail the re-verification will be stored in the preset rerun database, and a rerun command will be generated.

[0012] A call transcription structured knowledge parsing and governance device, the device comprising: The round normalization unit is used to receive upstream call transcription objects and normalize the upstream call transcription objects into multiple round structures; each round structure includes a round number, a general role tag for this round, a start time for this round, an end time for this round, and round text; the general role tag for this round is obtained by mapping the original role name to a pre-configured role mapping table; The dialogue segmentation unit is used to segment the multiple round structures as dialogues to be segmented, and the segmentation rules as segmentation constraints, inputting them into a preset large model to segment the dialogue units, resulting in multiple business units. Each business unit includes a round range, business type, topic convergence state, overlapping branch state, and overlapping associated units. The segmentation rules include a dividing line rule, a rule to prohibit incorrect segmentation, and an overlap determination rule. The dividing line rule is used to identify the location where the business semantics change during the call. The rule to prohibit incorrect segmentation is used to constrain the large model from breaking up the single-round dialogue content with the same semantics. The overlap determination rule is used to limit the overlap of round ranges of different business units to only be allowed in call scenarios where the main topic is continuous and accompanied by temporary branches in parallel, prohibiting full-scale stacking overlap without business logic, and outputting the corresponding overlapping associated unit identifier when overlap exists. The field distillation unit is used to take the round range of the business unit and the round text corresponding to the round number covered by the round range, together with the preset distillation constraint prompts, as distillation inputs and input them into the large model to perform business field distillation, so as to obtain the structured knowledge entries corresponding to each business unit; the structured knowledge entries are used to describe the business communication information contained in the call dialogue within the corresponding business unit. The rule association operation unit is used to perform rule association operations on each structured knowledge entry based on the time attribute and business field information of the business unit corresponding to each structured knowledge entry, to obtain the association tag information corresponding to each structured knowledge entry, and to write the association tag information into the corresponding structured knowledge entry to obtain the tagged structured knowledge entry. The quality management processing unit is used to perform quality management verification on each marked structured knowledge item to obtain the verification result, and to perform corresponding quality repair processing on the corresponding marked structured knowledge item based on each verification result to obtain the managed structured knowledge item.

[0013] In one possible implementation, the start time of the current round of a business unit is the start time of the current round of the round structure corresponding to the smallest round number covered by the round range of the business unit; the end time of the current round of a business unit is the end time of the current round of the round structure corresponding to the largest round number covered by the round range of the business unit. The rule association operation unit specifically includes: The temporal relationship determination unit is used to determine the temporal predecessor and successor relationship of each structured knowledge entry based on the start time of the current round of the business unit corresponding to the structured knowledge entry, and write the temporal predecessor and successor relationship into the structured knowledge entry. The overlapping association verification unit is used to perform interval intersection operation on the round range of the business unit corresponding to each structured knowledge entry, complete the overlapping parallel association verification, obtain the overlapping association unit identifier of the structured knowledge entry, and write the overlapping association unit identifier into the structured knowledge entry. The topic clustering unit is used to perform similar clustering for all structured knowledge entries according to the main core topic field within the structured knowledge entry. Structured knowledge entries with the same main core topic are grouped into the same cluster group, a unique cluster group identifier is assigned to each cluster group, and the cluster group identifier is written into the corresponding structured knowledge entry. The similarity conflict marking unit is used to calculate the text similarity between any two structured knowledge entries using the Jaccard similarity algorithm. When the similarity is greater than or equal to a preset threshold, the corresponding structured knowledge entry is marked as potentially conflicting, and the potentially conflicting mark is written into the corresponding structured knowledge entry. Among them, temporal predecessor-successor relationships, overlapping associated unit identifiers, cluster group identifiers, and suspected conflicts are all association marker information.

[0014] A call transcription structured knowledge parsing and management device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the call transcription structured knowledge parsing and management method as described above.

[0015] A computer-readable storage medium storing instructions that, when executed on a terminal device, cause the terminal device to perform the call transcription structured knowledge parsing and governance method as described above.

[0016] Compared with the prior art, this application has the following advantages: This application provides a method and related products for the governance of structured knowledge parsing in call transcription. Specifically, when implementing the method provided in this application, the upstream call transcription objects are first normalized by rounds to construct a standardized round structure, providing a unified data foundation for subsequent parsing. In the dialogue unit segmentation stage, semantic changes are identified using dividing line rules, the semantic integrity of a single round is ensured by prohibiting erroneous segmentation rules, and legal round overlaps with temporary branches are allowed by overlap judgment rules. Overlapping associated unit identifiers are output, overcoming the shortcomings of existing solutions that lose round overlap associations after segmentation and lose details of topic interleaving and branch parallel interaction due to forced semantic closure. Next, business field distillation is performed directly based on the original round text corresponding to the business unit, without relying on a preset standard dialogue library, to generate unit-level structured knowledge entries. Then, various association tags are written into the entries through rule association operations to form tagged structured knowledge entries, further solidifying the call interaction association information. Finally, quality governance verification is conducted and corresponding quality repair processes are executed, outputting structured knowledge entries that have undergone governance. This enables the automatic parsing and sedimentation of fine-grained unit-level structured knowledge based on the original call text, adapting to the needs of high-precision knowledge retrieval, efficient knowledge reuse, and refined business quality inspection under massive call data. This application, through a standardized round-of-call normalization process and a multi-dimensional constraint-based intelligent dialogue segmentation mechanism, effectively preserves the real interaction details of overlapping call rounds, topic interweaving, and parallel branching, avoiding the information loss problem caused by traditional forced semantic slicing. Simultaneously, based on the original call transcription data, it automatically completes business segmentation, knowledge distillation, association modeling, and quality governance, autonomously sedimenting high-precision unit-level structured knowledge, effectively adapting to the needs of knowledge retrieval, reuse, and refined quality inspection of massive call data. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in this embodiment or the prior art, the drawings used in the description of the embodiment or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 A flowchart illustrating a method for managing structured knowledge parsing during call transcription, provided in an embodiment of this application; Figure 2 A flowchart illustrating another method for managing structured knowledge parsing in call transcription provided in this application embodiment; Figure 3 This is a schematic diagram of a call transcription structured knowledge parsing and management device provided in an embodiment of this application. Detailed Implementation

[0019] To facilitate understanding of the technical solutions provided in the embodiments of this application, the background technology involved in the embodiments of this application will be described below.

[0020] In business call scenarios such as intelligent customer service, outbound marketing, and complaint follow-up, call recordings are transcribed into text data via Automatic Sentence Responsiveness (ASR). This text contains a wealth of implicit business knowledge, including response paradigms, objection handling strategies, and topic interleaving techniques. Enterprises aim to leverage massive amounts of call data to automatically mine, organize, and accumulate this business knowledge, providing data support for business applications such as call debriefing analysis, business quality inspection, agent training, and intelligent response. Existing call text knowledge parsing methods typically employ conversation turn segmentation, semantic clustering matching, or semantic closed-loop slicing distillation to segment dialogues and extract knowledge.

[0021] However, the above processing methods have obvious shortcomings when adapted to real call scenarios. After splitting the rounds, the overlapping and related information between the conversation rounds will be lost. The forced semantic loop slicing process will cause the loss of interactive details such as topic interleaving and branch parallelism. Furthermore, the processing flow needs to rely on a preset standard dialogue library and cannot directly achieve the autonomous accumulation of fine-grained unit-level structured knowledge based on the original call text. It cannot meet the usage requirements of knowledge retrieval, knowledge reuse, and refined business quality inspection in the context of massive call data.

[0022] To address this issue, this application provides a method and related products for managing structured knowledge parsing in call transcription. First, the upstream call transcription object undergoes round normalization, generating a unified round structure containing round number, role tag, time, and round text, providing a standardized foundation for subsequent parsing. During the dialogue unit segmentation stage, a large model is constrained by splitting rules, rules prohibiting incorrect segmentation, and overlap judgment rules. This not only identifies semantic change locations and ensures the semantic integrity of single-round dialogues but also allows for legitimate round overlaps of the main topic alongside temporary branches, outputting overlapping association unit identifiers. This solves the problems of existing solutions losing conversation round overlap relationships after segmentation and losing details of topic interleaving and branch parallelism due to forced semantic closure. Next, business field distillation is performed based on the original round text corresponding to the business unit, directly generating structured knowledge entries without relying on a pre-set standard dialogue library. Finally, rule-based association operations are used to write overlap, timing, and other association markers into the entries, resulting in tagged structured knowledge entries that fully preserve the association information of the call interaction. Finally, the system performs quality governance verification and corresponding quality repair processing, outputting structured knowledge entries that have been completed. This enables the automatic parsing and accumulation of fine-grained unit-level structured knowledge based on the original call text, meeting the business needs of knowledge retrieval, reuse, and refined quality inspection in massive call scenarios.

[0023] 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 this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0024] See Figure 1 The figure is a flowchart of a method for managing structured knowledge parsing in call transcription provided in an embodiment of this application. Figure 1 As shown, the call transcription structured knowledge parsing and governance method may include steps S101-S105: S101: Receive the upstream call transcription object and normalize the upstream call transcription object into a multiple round structure.

[0025] In the process of knowledge analysis and governance of call transcription text, it is first necessary to receive upstream call transcription objects. To eliminate problems such as inconsistent data formats, chaotic role identification, and missing time sequence information in the speech recognition and transcription output, the upstream call transcription objects can be unified into multiple round structures to complete the standardized preprocessing of the original transcription data. Each round structure is configured with several core fields, including round number, general role label for this round (e.g., customer or agent), start time of this round, end time of this round, and round text. Among them, the general role label for this round can be customer or agent. This label is not directly used as the original role name output by speech recognition, but is obtained by performing a mapping transformation on the original role name in the transcription through a pre-configured role mapping table.

[0026] For example, suppose a customer service call, after round-normalization processing, results in the following two round structures: The structure of the first round is as follows: [Round Number: 1, General Role Tag for this Round: Agent, Start Time for this Round: 00:00:05, End Time for this Round: 00:00:12, Round Text: Hello, how can I help you?] The second round structure is as follows: [Round Number: 2, General Role Tag for this Round: Customer, Round Start Time: 00:00:13, Round End Time: 00:00:22, Round Text: Hello, I would like to inquire about my order refund.] Each round structure fully includes five fields: round number, general role label for this round, start time for this round, end time for this round, and round text. The round number is used to distinguish the order of the conversation, the general role label for this round identifies the speaker, the start and end times record the time range of the speech in this round, and the round text stores the speech transcription content corresponding to this round. The original call transcription is decomposed into structured round data, providing a regular data input for subsequent dialogue unit segmentation.

[0027] It should be noted that the upstream call transcription object is obtained by ASR (Automatic Speech Retrieval System) through voice transcription of the call recording. The transcription result will contain diverse original role identifiers and corresponding dialogue text content.

[0028] In one possible implementation, the method further includes: The upstream call transcription object is validated. If garbled characters are detected in the upstream call transcription object, the speaking role cannot be distinguished, or the role mapping fails when performing role mapping through the pre-configured role mapping table (for example, the identified role identifiers spk0 and spk1 are not configured with corresponding roles in the role mapping table), or the role entries that do not exist in the mapping table are identified, such as third-party (or recording) noise, the call transcription object is directly rejected and will not enter the subsequent processing flow to avoid consuming large model call resources.

[0029] In one possible implementation, the mapping rule for the general role label in this round is as follows: a role mapping configuration table is pre-built and stored. Various original role identifiers in the upstream call transcription object are uniformly mapped. All original role values ​​of agent type, such as Agent 01, Agent_003, customer service personnel, salesperson, and outbound caller, are uniformly mapped to the standardized general role label "Agent". All original role values ​​of customer type, such as 1, caller, policyholder, caller, and user, are uniformly mapped to the standardized general role label "Customer". This ensures that the final round structure retains only the two general role labels "Agent" and "Customer", achieving unified standardization of original role data from different sources and with different naming formats. This avoids subsequent dialogue segmentation and knowledge parsing anomalies caused by messy original role naming, ensuring the standardization of the round structure and the accuracy of subsequent business processing.

[0030] S102: The multiple round structures are used as dialogues to be segmented, and the segmentation rules are used as segmentation constraints. These are then input into a preset large model to segment the dialogue units, resulting in multiple business units.

[0031] To overcome the technical shortcomings of traditional call semantic slicing methods, which force semantic closure and have a single slicing logic, resulting in the loss of interaction details and the inability to retain turn relationships in real-world call scenarios with interleaved topics and parallel branches, a new approach can be developed. After unifying call turns and obtaining a standardized turn structure dataset, the entire normalized turn structure can be used as a complete dialogue to be segmented. Instead of using the traditional single semantic closure slicing logic, a dedicated multi-dimensional segmentation rule is introduced as a strong constraint. The dialogue to be segmented and the constraint rule are input together into a pre-set large model to complete the intelligent and accurate segmentation of dialogue units, ultimately outputting multiple semantically complete and traceable independent business units.

[0032] The segmentation rules configured in this application include three types of constraint logic: dividing line rules, rules prohibiting incorrect segmentation, and overlap determination rules. These three types of rules work together to constrain the segmentation behavior of the large model: the dividing line rules are responsible for accurately identifying semantic nodes where business intent and communication topics have undergone substantial changes during the call dialogue, defining the start and end boundaries of business units with rounds as the smallest granularity, aligning unit boundaries with the start and end of rounds, and not splitting a single round into two different business units; the rules prohibiting incorrect segmentation rigidly constrain the model's segmentation behavior, strictly prohibiting the model from splitting or breaking up the semantically complete dialogue content within a single round, so as to prevent the generation of fragmented units without business meaning and to prevent arbitrarily cutting round boundaries without deviating from the actual interaction logic of the call; the overlap determination rules are specifically adapted to the complex interaction scenarios where the main topic of the real call continues and temporary topic branches are interspersed in parallel, breaking the limitations of traditional solutions that force semantic closure and completely prohibit unit overlap, only allowing the overlap of round ranges of topic branches with actual business logic, strictly prohibiting invalid overlap of all stacked items without business meaning, and ensuring data validity while retaining the characteristics of real call interaction.

[0033] Example 1: Suppose a customer service call, when consolidated, contains multiple rounds of dialogue: Round 1 is the agent's greeting and opening, Round 2 is the customer's inquiry about the order refund rules, Round 3 is the agent's answer regarding the refund timeframe, Round 4 is the customer's follow-up question about the shipping reimbursement policy, Round 5 is the agent's response regarding the shipping fee standard, and Round 6 is the confirmation of the refund process by both parties and the end of the dialogue.

[0034] First, semantic change nodes are identified based on the dividing line rules. From round 1 to round 2, the focus shifts from the opening greeting to the core business of refund consultation. From round 3 to round 4, the focus shifts from answering questions about refund timeliness to the branch business of shipping fee consultation. The large model accurately identifies the two semantic change locations as the business unit division boundaries.

[0035] At the same time, the entire process follows the rule of prohibiting incorrect segmentation. Each round of complete semantic dialogue is not split. For example, the customer's complete freight inquiry statement in round 4 and the agent's complete timeliness answer statement in round 3 are kept as a whole and will not be arbitrarily truncated or broken up by the large model, thus ensuring the semantic integrity of each round of dialogue.

[0036] Based on this, complex dialogue scenarios are handled by relying on the overlap judgment rule. In this scenario, the main topic of order refund in rounds 2 to 3 is still ongoing, and the temporary branch topic of shipping reimbursement is inserted in rounds 4 to 5. This is a parallel scenario of a legitimate main topic superimposed on a temporary branch. Therefore, it is allowed for the round range of "refund consultation business unit (rounds 2-3)" and "shipping consultation business unit (rounds 4-5)" to overlap. At the same time, it is strictly prohibited to stack overlaps without logic.

[0037] Ultimately, the large model is divided into two standardized business units. The first business unit has rounds ranging from 2 to 3, the business type is order refund consultation, the topic convergence status is not converged, the overlapping branch status is that there is branch overlap, and the overlapping related unit is the second business unit. The second business unit has rounds ranging from 4 to 5, the business type is freight reimbursement consultation, the topic convergence status is converged, the overlapping branch status is the branch unit, and the overlapping related unit is the first business unit. This fully preserves the interactive details of topic interweaving and parallel primary and secondary topics in real calls, solving the problem of losing complex dialogue logic in traditional single closed-loop slicing.

[0038] After the segmentation based on the above rules, each generated business unit fully carries the core information of round range, business type, topic convergence status, overlapping branch status, and overlapping associated units. It can not only accurately record the call round range and business attributes corresponding to each business unit, but also distinguish whether the topic has closed loop through the topic convergence status, identify the interleaving and parallel attributes of the unit through the overlapping branch status, and output the corresponding overlapping associated unit identifier when there is a legitimate business overlap scenario. It completely retains the unit association relationship under the interleaving interaction of multiple topics, and completely solves the problem of losing round overlap association and missing complex topic interaction details after segmentation in the existing technology. It provides complete and real call structure data support for subsequent refined knowledge distillation and association tagging operation.

[0039] It should be noted that the topic convergence status is used to characterize whether the business topic carried by the corresponding business unit has completed the closed loop. When the dialogue content within the business unit has completed the question answering, the two parties have reached a consensus, or the topic has been actively terminated, the topic convergence status is marked as converged; when the business topic corresponding to the business unit is still ongoing and has not yet received a reply or reached a final conclusion, the topic convergence status is marked as unconverged.

[0040] Referring to Example 1, the business unit corresponding to the order refund consultation has a non-converged topic status because a branch topic on shipping reimbursement was subsequently inserted. The shipping reimbursement consultation business unit completed the Q&A interaction on shipping standards, thus its topic is now considered converged. This status field distinguishes between ongoing and completed business topics, facilitating subsequent categorization, retrieval, and analysis of various business knowledge items.

[0041] In one possible implementation, before step S102, which uses the multiple round structures as dialogues to be segmented and the segmentation rules as segmentation constraints, inputs them into a preset large model for dialogue unit segmentation to obtain multiple business units, the method further includes: performing batch data entry and non-segmentation processing; scanning a batch of call transcription data and estimating the corresponding data volume for each call transcription data; splicing multiple complete calls into processing batches according to preset splicing rules, grouping calls of the same business type into the same batch, separating calls of different business types, and not splitting single calls within or between calls; performing dialogue unit segmentation processing once for each batch, splicing multiple complete calls, marking each call boundary with prompt words, and having the large model mark the call to which each generated business unit belongs. The system identifies the call and performs another round of business field distillation, concatenating multiple business units. The large model labels each business field set with the corresponding unit identifier to prevent issues such as incorrect unit data attribution or data misalignment. If the amount of data transcribed from a single call exceeds a preset threshold, the call is processed as a separate batch using a windowed segmentation method. It is segmented according to the token window and sent to the large model, where deduplication and merging are performed at the segment boundaries, and it is not concatenated with other calls in the batch. Simultaneously, it supports failure retry and breakpoint resumption mechanisms. If a single call processing fails, it is retried several times. If the retry still fails, it is marked as failed, without blocking the processing flow of other calls. After each processing stage is successfully executed, the running accounting information is recorded. When the system restarts, calls that have not been processed are retried according to the processing stage order.

[0042] S103: For each business unit, the round range of the business unit and the round text corresponding to the round number covered by the round range, together with the preset distillation constraint prompts, are used as distillation inputs and input into the large model to perform business field distillation to obtain the structured knowledge entries corresponding to each business unit.

[0043] To accurately, systematically, and finely extract implicit business knowledge hidden in the original call dialogue from the segmented business units, and to achieve an effective transformation from the original dialogue text to standardized structured knowledge, avoiding the problems of traditional methods that rely on fixed script libraries, have coarse extraction granularity, and produce fragmented and non-standardized knowledge, this application, after obtaining multiple business units, performs refined business field distillation processing on each independent business unit separately: First, the round range corresponding to the current business unit is accurately extracted, and the complete original round text corresponding to all round numbers covered by this round range is retrieved. This ensures that the distillation input completely corresponds to the actual dialogue context of this unit, without omission or mixing in irrelevant dialogue content from other units. Simultaneously, it is combined with preset distillation constraint prompts to form a complete distillation input. These prompts standardize the extraction dimensions, output format, and business field standards of the large model, guiding it to parse the dialogue content according to unified business specifications. Then, the integrated input information is fed into the large model for specialized business field distillation, deeply mining the core business information contained in the dialogue of this business unit, such as consultation requests, problem descriptions, solutions, objections, and response strategies. Finally, structured knowledge entries corresponding to each business unit are generated.

[0044] This structured knowledge entry can completely, systematically, and in a standardized manner carry all business communication information within the corresponding business unit. It transforms the originally fragmented, colloquial, and unstructured call dialogue content into searchable, reusable, and quality-inspectable standardized unit-level business knowledge, providing accurate and reliable structured knowledge data support for subsequent knowledge association and labeling, quality governance, and downstream intelligent applications such as call review, business quality inspection, agent training, and intelligent response.

[0045] The distillation constraint prompts are used to define the set of business fields, output format, and field value constraints to guide the large model in extracting corresponding business information from the dialogue text.

[0046] The structured knowledge items include main core topics, interspersed secondary topics, customer needs, underlying customer concerns, agent core actions, key negotiation points, convergence or interruption results, effective highlights, loopholes, and optimizable dialogue. Main core topics record the main business discussion direction of the business unit; interspersed secondary topics record branching topics inserted during the conversation; customer needs clarify the explicitly stated business requirements of the customer; underlying customer concerns uncover the worries and anxieties hidden beneath the surface of the customer's needs; agent core actions summarize the business operations and responses performed by the agent during the conversation; key negotiation points mark the content where the two parties disagree and need to negotiate; convergence or interruption results record whether the business topic ultimately reaches a resolution or the conversation is interrupted; effective highlights extract the agent's excellent responses during the process; loopholes identify business processing defects in the conversation; and optimizable dialogue outputs reference dialogue that can be improved and perfected for this conversation. Through these multi-dimensional fields, the conversational dialogue is transformed into comprehensive, structured business knowledge that can be used for retrospective analysis.

[0047] S104: Based on the time attribute and business field information of the business unit corresponding to each structured knowledge entry, perform rule association operation for each structured knowledge entry to obtain the association tag information corresponding to each structured knowledge entry, and write the association tag information into the corresponding structured knowledge entry to obtain the tagged structured knowledge entry.

[0048] To establish business and temporal relationships between different structured knowledge items, and to overcome the shortcomings of simple field distillation which only extracts single-segment dialogue information and lacks contextual connections between items, thus facilitating subsequent operations such as knowledge retrieval, call debriefing, and business analysis, this application, after obtaining the structured knowledge items corresponding to each business unit, combines the time attribute of the business unit to which each structured knowledge item belongs with the extracted business field information, performs rule-based association operations to obtain the association tag information corresponding to each structured knowledge item, and writes the association tag information into the corresponding structured knowledge item to obtain tagged structured knowledge items. The specific operation process includes: For each structured knowledge entry, the temporal predecessor and successor relationship of the business unit corresponding to the structured knowledge entry is determined according to the start time of the current round, and the temporal predecessor and successor relationship is written into the structured knowledge entry to represent the sequential order of knowledge entries in the call sequence. For each structured knowledge entry, an interval intersection operation is performed on the round range of the business unit corresponding to the structured knowledge entry to complete the overlapping parallel association verification, obtain the overlapping association unit identifier of the structured knowledge entry, and write the overlapping association unit identifier into the structured knowledge entry to record the associated business units with overlapping and parallel rounds. For all structured knowledge entries, perform similar clustering according to the main core topic field within the structured knowledge entry. Structured knowledge entries with the same main core topic are grouped into the same cluster group. Assign a unique cluster group identifier to each cluster group and write the cluster group identifier into the corresponding structured knowledge entry to achieve the aggregation of knowledge entries with the same topic. For all structured knowledge entries, the Jaccard similarity algorithm is used to calculate the text similarity between each pair of structured knowledge entries. When the similarity is greater than or equal to a preset threshold, the corresponding structured knowledge entry is marked as potentially conflicting, and the potentially conflicting mark is written into the corresponding structured knowledge entry to identify knowledge entries with similar content and potential contradictions.

[0049] Among them, temporal predecessor-successor relationships, overlapping associated unit identifiers, cluster grouping identifiers, and suspected conflicts are all association marker information. Through the above multi-dimensional rule association operations, originally independent individual knowledge items can be constructed into a knowledge set with logical connections. This not only preserves the business information of each item, but also reflects the complete dialogue logic of topic flow, branching and interweaving, and topic connection within the conversation, facilitating subsequent association queries, source tracing analysis, and batch business quality inspection processing of knowledge items.

[0050] S105: Perform quality governance verification on each marked structured knowledge item to obtain the verification results, and perform corresponding quality repair processing on the corresponding marked structured knowledge item according to each verification result to obtain the governed structured knowledge item.

[0051] To ensure reliable data quality of the output structured knowledge entries, this application identifies various defects in format, segmentation logic, business affiliation, and content conflicts, preventing problematic knowledge entries from directly flowing into downstream stages such as business analysis and post-mortem quality inspection. After obtaining the marked structured knowledge entries, this application conducts quality governance verification on each marked structured knowledge entry, generating corresponding verification results. Based on the problem type of the verification results, differentiated quality repair processing is performed, ultimately outputting the remediated structured knowledge entries. The quality governance verification in this application covers four dimensions: format verification, segmentation correctness verification, business type correctness verification, and conflict verification. Failure in any verification is categorized into different problem types: format verification failure is classified as a mechanical problem; segmentation correctness verification failure is classified as a large model quality problem; business type correctness verification failure is classified as a systemic problem; and conflict verification failure is categorized into mechanical, model quality, or systemic problems based on the actual cause of the problem. Different remediation strategies are employed for different types of issues. For mechanical issues, the program automatically corrects and repairs errors. For large model quality issues, the large model is re-invoked to perform dialogue unit segmentation or business field distillation, and the corresponding knowledge content is regenerated. For systemic issues, the corresponding business configuration is modified and reclassification is performed. Through classification verification and differentiated remediation mechanisms, issues with different root causes, such as program format errors, large model inference biases, and business configuration system defects, can be identified and remediated accordingly. This improves the overall accuracy and standardization of structured knowledge entries while minimizing manual intervention, outputting structured knowledge entries that can be directly used in business scenarios.

[0052] Based on the content of S101-S105, the upstream call transcription object is first normalized to generate a round structure carrying round number, role, and time information, which can preserve the original call's timing and basic role information. Next, the large model is constrained by splitting rules, rules prohibiting incorrect segmentation, and overlap judgment rules to perform dialogue unit segmentation. The overlap judgment rules output overlapping association unit identifiers for legitimate overlapping scenarios where the main topic is accompanied by temporary branches, solving the problem of losing the overlapping relationships between rounds after round splitting. Simultaneously, the rules prohibiting incorrect segmentation avoid disrupting the semantics of a single round, and do not use forced semantic closed-loop slicing, supporting the generation of cross-round business units and fully preserving the interactive details of call topic interweaving and branch parallelism. Subsequently, business field distillation is performed based on the round text corresponding to the business unit to generate structured knowledge entries for the corresponding business unit. Then, rule-based association operations are performed by combining the time and field attributes of the business unit, and association tags are written to obtain tagged structured knowledge entries, further enriching the association information within the knowledge entries. Finally, the quality governance verification and quality repair process is performed to obtain the structured knowledge entries that have been governed. The entire process does not rely on a preset standard script library. It directly completes the entire process based on the original call transcription data, realizing the autonomous accumulation of fine-grained unit-level structured knowledge. This adapts to the needs of knowledge retrieval, reuse, and refined business quality inspection under massive call data.

[0053] In one possible implementation, the start time of the current round of a business unit is the start time of the current round of the round structure corresponding to the smallest round number covered by the round range of the business unit; the end time of the current round of a business unit is the end time of the current round of the round structure corresponding to the largest round number covered by the round range of the business unit.

[0054] For example, suppose a business unit's round range is round number 2-5, where the start time of the current round structure corresponding to round number 2 is 00:00:13, and the end time of the current round structure corresponding to round number 5 is 00:00:45. Then, the start time of the current round for this business unit is taken as 00:00:13, and the end time is taken as 00:00:45. Using this method, the overall time boundary of the business unit can be calculated based on the round range it encompasses, providing a unified time benchmark for subsequent determination of predecessor and successor relationships and time-dimensional correlation operations.

[0055] In one possible implementation, to completely solve the problems of isolated individual structured knowledge entries, lack of sequential logic, ambiguous topic associations, lack of traceability in parallel branches, and inability to identify potential content conflicts, step S104 performs rule-based association operations on each structured knowledge entry based on the time attributes and business field information of the business unit corresponding to each structured knowledge entry, obtains the association tag information corresponding to each structured knowledge entry, and writes the association tag information into the corresponding structured knowledge entry to obtain a tagged structured knowledge entry. This can be achieved through the following steps: For each structured knowledge entry, the temporal predecessor and successor relationship of the business unit corresponding to the structured knowledge entry is determined based on the start time of the current round, and the temporal predecessor and successor relationship is written into the structured knowledge entry. For each structured knowledge entry, perform interval intersection operation on the round range of the business unit corresponding to the structured knowledge entry to complete the overlapping parallel association verification, obtain the overlapping association unit identifier of the structured knowledge entry, and write the overlapping association unit identifier into the structured knowledge entry. For all structured knowledge entries, perform similar clustering according to the main core topic field within the structured knowledge entry. Structured knowledge entries with the same main core topic are grouped into the same cluster group. Assign a unique cluster group identifier to each cluster group and write the cluster group identifier into the corresponding structured knowledge entry. For all structured knowledge entries, the Jaccard similarity algorithm is used to calculate the text similarity between each pair of structured knowledge entries. When the similarity is greater than or equal to a preset threshold, the corresponding structured knowledge entries are marked as potentially conflicting, and the potentially conflicting marks are written into the corresponding structured knowledge entries. Among them, temporal predecessor-successor relationships, overlapping associated unit identifiers, cluster group identifiers, and suspected conflicts are all association marker information.

[0056] Specifically, this application completes comprehensive correlation operations from four dimensions: temporal logic, spatial round overlap, topic clustering and aggregation, and text conflict verification. First, based on the standardized start time of each business unit, the temporal relationship of the business units corresponding to all knowledge items is deduced, the predecessor and successor units of each business unit in the whole call are accurately located, the temporal predecessor and successor relationship is generated and written into the corresponding structured knowledge items, and the sequential flow and progression logic of the call topics are fully restored. Secondly, mathematical intersection operations are performed on the round range of each business unit to automatically verify whether there is a legal branch overlap relationship between different business units, accurately match the related business units that are parallel branches, generate a unique overlapping related unit identifier and complete the entry backfilling to realize the branch tracing of complex interleaved dialogues. Furthermore, relying on the main core topic fields output by the large model distillation, the full set of structured knowledge items are clustered and aggregated according to the same category. Knowledge items with the same topic theme and the same business scenario are uniformly collected into the same cluster group, and a unique cluster group identifier is assigned to each group, so as to realize the batch collection, unified management and subsequent batch statistical analysis of similar business knowledge. Finally, the Jaccard similarity algorithm is introduced to perform full-text similarity quantification on any two structured knowledge items. When the similarity between the two items reaches a preset threshold, it is determined that the two knowledge items have suspected conflict issues such as overlapping content, contradictory expressions, and repeated parsing. The corresponding items are automatically marked with suspected conflict tags to complete the screening of implicit conflicts in knowledge content.

[0057] In summary, the temporal predecessor and successor relationships, overlapping associated unit identifiers, cluster grouping identifiers, and suspected conflict markers generated by this application together constitute complete association marker information. This upgrades the originally fragmented and independent single business knowledge into a structured knowledge network with temporal chains, branch associations, similar groupings, and conflict identification, significantly improving the comprehensiveness and accuracy of subsequent knowledge quality inspection, review analysis, demand mining, and wording optimization.

[0058] See Figure 2 , Figure 2 The flowchart of another call transcription structured knowledge parsing and governance method provided in this application embodiment shows that, for each structured knowledge entry, the temporal predecessor-successor relationship of the business unit corresponding to the structured knowledge entry is determined according to the current round start time of the business unit, and the temporal predecessor-successor relationship is written into the structured knowledge entry. Specifically, this can be achieved through steps S201-S204: S201: Sort all current structured knowledge items in ascending order according to the start time of the current round of the corresponding business unit.

[0059] To establish a unified time benchmark and facilitate accurate identification of the temporal relationships between various business units, this step acquires all structured knowledge entries and extracts the start time of the current round for each business unit corresponding to each structured knowledge entry. All structured knowledge entries are then sorted in ascending order according to their start times. After sorting, the arrangement of all structured knowledge entries in the list clearly reflects the order in which the corresponding business units are initiated throughout the entire call, eliminating the sequential confusion caused by business unit segmentation and parallel branching. This provides a reliable temporal basis for subsequently selecting successor units and deducing predecessor units based on time conditions, ensuring that the logic for determining temporal predecessor-successor relationships has a unified and reproducible time reference.

[0060] S202: For each structured knowledge item, if the topic convergence status of the business unit corresponding to the structured knowledge item is not converged, then the business unit whose start time in this round is later than that business unit and whose overlapping branch status is yes is determined as the successor unit of that business unit; if the topic convergence status of the business unit corresponding to the structured knowledge item is converged, then the business unit whose start time in this round is earlier than the end time of this round of that business unit and whose topic convergence status is not converged is determined as the successor unit of that business unit.

[0061] To adapt to two different dialogue modes in real-world calls—one where the main topic continues to advance and the other where the topic has completed its loop—and to achieve differentiated identification of successor units, this step traverses all structured knowledge entries after chronological sorting. It uses two different judgment rules to filter successor units based on the topic convergence status of each business unit. When a business unit's topic convergence status is not converged, it means the current main topic has not yet been processed, and parallel branch topics may appear interspersed during the call. In this case, business units whose start time is later than the current business unit and whose overlapping branch status is "yes" are identified as successor units, thus capturing newly appearing parallel branch business units during the main topic's existence. When a business unit's topic convergence status is converged, it means the current topic has completed its loop. In this case, instead of searching for subsequent topics based on time sequence, business units whose start time is earlier than the end time of the current business unit's current round and whose topic convergence status is not converged are retrieved as successor units. This is used to identify interspersed branch topics that started before the end time of the current main topic but have not yet been processed. By distinguishing whether a topic converges and executes different filtering logics, it is possible to identify parallel branches derived from the main topic's existence stage, as well as capture unfinished interspersed dialogues before and after the topic's closure. This avoids the shortcomings of a single time-based judgment rule that cannot adapt to complex call branch scenarios, and provides accurate subsequent relationship basis for subsequent reverse derivation of precursor units and construction of complete temporal correlation links.

[0062] S203: Based on the correspondence between successor units recorded in all structured knowledge entries, the predecessor units of each structured knowledge entry are derived in reverse.

[0063] In order to construct a complete bidirectional temporal correlation link, after obtaining the successor unit information corresponding to all structured knowledge entries, this step performs reverse derivation based on the already recorded successor unit correspondence to obtain the predecessor unit of each structured knowledge entry.

[0064] Specifically, iterate through all structured knowledge entries. For each entry's successor unit, if entry A marks the business unit corresponding to entry B as its own successor unit, it means that entry B is sequentially inherited from entry A. Thus, it can be determined that entry A is the predecessor unit of entry B. For structured knowledge entries that no other entry marks as their successor unit, it means that the entry does not have a preceding business unit, and its predecessor unit is set to empty.

[0065] This reverse derivation method eliminates the need for complex time condition traversal and matching. Relying solely on the already generated successor unit mapping relationship, it efficiently obtains the predecessor unit corresponding to each knowledge item, establishing a two-way temporal relationship between predecessor and successor. This allows each structured knowledge item to simultaneously possess upstream source and downstream inheritance information, fully restoring the sequential dependency relationship between topics within the conversation, and providing complete data support for subsequently writing temporal predecessor and successor relationships into structured knowledge items.

[0066] S204: The information of the predecessor and successor units of the structured knowledge entry is written into the structured knowledge entry as a temporal predecessor-successor relationship.

[0067] To solidify the derived topic temporal dependencies into structured knowledge entries and achieve persistent storage of temporal association information between knowledge entries, this step integrates the predecessor and successor unit information corresponding to each structured knowledge entry to form a temporal predecessor-successor relationship, and writes it into the corresponding structured knowledge entry. After the write operation, each tagged structured knowledge entry not only stores its own business field content, but also carries the upstream preceding topic unit and downstream succeeding topic unit information of that business unit in the entire call.

[0068] When conducting call debriefing, topic tracing, and business statistical analysis, the temporal predecessor and successor relationships within the entries can be directly read to quickly locate the source topic and subsequent branch topics, reconstructing the complete topic flow chain of the entire call. This facilitates subsequent business operations such as topic link backtracking, branch dialogue location, and business problem tracing, allowing discrete knowledge entries to form an interconnected temporal knowledge network.

[0069] Steps S201-S204 combine the time attributes of business units with the convergence and overlapping branch states of topics to distinguish between two different dialogue scenarios: one where the main topic is ongoing and the other where the topic has closed its loop. This allows for the systematic organization of the sequential relationships between various knowledge items. First, all items are sequentially sorted based on the start time of the current round. Then, successor units are selected using differentiated rules based on whether the topic has converged. Next, the predecessor units are derived in reverse. Finally, the information of the predecessor and successor units is backfilled into the structured knowledge items. This reconstructs the temporal progression of topics and the interweaving of branches within the call, achieving a digital model of the call's topic flow logic. This provides a temporal-level correlation basis for subsequent call debriefing, topic tracing, and business analysis.

[0070] In one possible implementation, step S203, based on the successor unit correspondence recorded in all structured knowledge entries, reverse-engineers the predecessor units of each structured knowledge entry, including: Read the successor unit of each structured knowledge entry; if the successor unit of a structured knowledge entry is a business unit corresponding to another structured knowledge entry, then determine the business unit corresponding to the structured knowledge entry as the predecessor unit of the other structured knowledge entry; wherein, for structured knowledge entries that are not pointed to by any successor unit of a structured knowledge entry, set their predecessor unit to null.

[0071] Specifically, to efficiently construct a bidirectional temporal correlation mapping, this step derives the predecessor unit based on the already generated successor unit mapping relationship. It iterates through all structured knowledge entries, reads the successor unit information stored in each entry, and establishes the pointing relationships between units. When entry A marks the business unit corresponding to entry B as a successor unit, it means that unit B follows unit A in temporal sequence; therefore, unit A is set as the predecessor unit of unit B. For structured knowledge entries that are not pointed to by any other entry as a successor unit, it indicates that the business unit does not have a preceding successor unit in the call topic chain; therefore, its predecessor unit is directly set to empty.

[0072] By using this reverse mapping derivation method, the determination of all predecessor units can be completed in batches without repeating the time condition filtering, ensuring the logical consistency of the predecessor and successor relationship, avoiding the disorder of the timing relationship, and providing accurate data for subsequent writing of the predecessor and successor relationship.

[0073] In one possible implementation, the quality governance verification includes format verification, segmentation correctness verification, business type correctness verification, and conflict verification; wherein, a format verification failure is determined to be a mechanical problem, a segmentation correctness verification failure is determined to be a large model quality problem, a business type correctness verification failure is determined to be a system problem, and a conflict verification failure is classified into a mechanical problem, a model quality problem, or a system problem according to the problem's affiliation. The quality repair process for the aforementioned mechanical problems is: automatic repair via program execution; The quality repair process for the aforementioned model quality issues is to re-call the large model to perform splitting or distillation processing. The quality repair process for the system-related issues is as follows: modify the configuration and perform reclassification.

[0074] Specifically, to conduct hierarchical and categorized quality checks on labeled structured knowledge items, distinguish the root causes of problems, and implement differentiated remediation strategies, this application sets up four quality governance verification dimensions. Format verification mainly checks for procedural issues such as the completeness of knowledge item fields, data format, and the standardization of field values; failure to pass this verification is classified as a mechanical problem. Segmentation correctness verification checks whether the segmentation results of business units conform to the segmentation rules, determining whether there are deviations caused by large model inference, such as incorrect unit segmentation or boundary errors; failure to pass this verification is classified as a large model quality problem. Business type correctness verification checks whether the business type to which the business unit belongs matches the business configuration rules; failure to pass this verification is classified as a systemic problem. Conflict verification identifies content contradictions, logical conflicts, and other problems between items; these problems need to be classified into mechanical, model quality, or systemic problems based on their actual causes.

[0075] Different types of problems are addressed with corresponding remediation methods. Mechanical problems can be automatically corrected directly through program logic; large model quality problems require re-invoking the large model, re-executing dialogue unit segmentation or business field distillation, and regenerating the corresponding knowledge content; system problems require adjusting business configuration parameters and then performing reclassification. By classifying problems according to their sources and adopting corresponding remediation methods, different root causes of problems such as format errors, large model inference biases, and business configuration defects can be addressed in a targeted manner, thereby improving the overall quality of structured knowledge items.

[0076] In one possible implementation, the method further includes: After the quality repair process is completed, the structured knowledge entries are re-verified for quality management. The structured knowledge entries that pass the re-verification are marked as active for external retrieval and access. The structured knowledge entries that fail the re-verification are stored in a preset rerun database and a rerun command is generated.

[0077] Specifically, to ensure the output structured knowledge items are of reliable and usable quality, and to prevent data with residual defects after the first round of repair from flowing into business processes, a complete quality governance verification process needs to be re-executed on the repaired structured knowledge items after the first round of quality repair. Items that pass the second verification are marked as active, indicating that the knowledge item verification meets the standards and can be used for business calls such as retrieval and retrospective analysis. Items that fail the second verification indicate that a single repair measure cannot solve the existing problems, and they are no longer processed on-site. These items are stored in a preset rerun database, and a corresponding rerun command is automatically generated, awaiting subsequent scheduling to re-execute the complete parsing and governance process.

[0078] The filtering mechanism of secondary verification distinguishes between knowledge items that can be used directly and those that need to be reprocessed. On the one hand, it ensures the quality of data provided to external parties, and on the other hand, it unifies and collects abnormal items that cannot be repaired, making it easier to process them in batches and preventing unqualified knowledge items from being misused by business systems.

[0079] In one possible implementation, the method further includes: Configure and implement a business type system iteration governance mechanism, supporting four types of iteration capabilities: business type alias mapping, abandoned type reclassification, manual unit correction, and model addition classification control. This enables smooth updates and standardized governance of the business type system, avoiding problems such as large-scale data reruns, classification chaos, classification explosion, and retrieval failure caused by type iteration.

[0080] Specifically, this application supports alias mapping. A list of corresponding aliases is pre-configured for each business type in the type set configuration. During system operation, a preset mapping function automatically maps historical old type names to the latest new type names. The structured index continuously retains the original type data. Type mapping is only dynamically completed in business processes such as business queries and whitelist verification. This allows for the classification and updating of old type business units simply by modifying the configuration, without needing to reprocess historical data. Simultaneously, it supports a type deletion and reclassification mechanism. When a business type is deleted in the iterative iteration of the business type system, the full reclassification task automatically detects all business units using that obsolete type, determines that the unit is not in the valid business whitelist, and automatically includes its associated call transcription data in the rerun list. A manual forced rerun task is then initiated, and the business type determination is re-completed through a large model to adapt to the updated type system. Meanwhile, this application supports refined, unit-level manual type modification. For a small number of independent business units with misclassification issues, it allows manual correction of their business type and front-end display tags. After correction, the unit is directly set to an active and available state without triggering a complete rerun or changing the item's flat storage path, achieving lightweight and accurate error correction for individual abnormal units. In addition, this application sets up a model-added classification constraint mechanism, strictly prohibiting large models from arbitrarily adding business categories. For newly identified business scenarios, they are uniformly classified into a pre-defined pending edge type and stored in a rerun database. After system governance and statistics, and manual sorting and iterative improvement of classification rules, formal classification is completed. This fundamentally avoids problems such as disorderly expansion of business categories, category explosion, and retrieval failure and data chaos caused by the lack of rule control over new types, ensuring the long-term stability and controllable iteration of the business type system.

[0081] See Figure 3 , Figure 3 This is a schematic diagram of a call transcription structured knowledge parsing and management device provided in an embodiment of this application. Figure 3 As shown, the call transcription structured knowledge parsing and governance device includes: Round normalization unit 301 is used to receive upstream call transcription objects and normalize the upstream call transcription objects into multiple round structures; each round structure includes a round number, a general role tag for this round, a start time for this round, an end time for this round, and round text; the general role tag for this round is obtained by mapping the original role name to a pre-configured role mapping table; Dialogue segmentation unit 302 is used to take the multiple round structures as dialogues to be segmented, and the segmentation rules as segmentation constraints, and input them into a preset large model to segment dialogue units, resulting in multiple business units. Each business unit includes a round range, business type, topic convergence state, overlapping branch state, and overlapping associated units. The segmentation rules include a dividing line rule, a rule to prohibit incorrect segmentation, and an overlap determination rule. The dividing line rule is used to identify the location where the business semantics change during the call. The rule to prohibit incorrect segmentation is used to constrain the large model from breaking up the single-round dialogue content with the same semantics. The overlap determination rule is used to limit the round ranges of different business units to intersect only in call scenarios where the main topic is continuous and accompanied by temporary branches in parallel, prohibiting full-scale stacking overlap without business logic, and outputting the corresponding overlapping associated unit identifier when overlap exists. Field distillation unit 303 is used to, for each business unit, take the round range of the business unit and the round text corresponding to the round number covered by the round range, together with the preset distillation constraint prompts, as distillation inputs and input them into the large model to perform business field distillation, so as to obtain the structured knowledge entries corresponding to each business unit; the structured knowledge entries are used to describe the business communication information contained in the call dialogue within the corresponding business unit. The rule association operation unit 304 is used to perform rule association operation on each structured knowledge entry based on the time attribute and business field information of the business unit corresponding to each structured knowledge entry, to obtain the association mark information corresponding to each structured knowledge entry, and to write the association mark information into the corresponding structured knowledge entry to obtain the marked structured knowledge entry. The quality management processing unit 305 is used to perform quality management verification on each marked structured knowledge item to obtain the verification result, and to perform corresponding quality repair processing on the corresponding marked structured knowledge item according to each verification result to obtain the managed structured knowledge item.

[0082] In one possible implementation, the start time of the current round of a business unit is the start time of the current round of the round structure corresponding to the smallest round number covered by the round range of the business unit; the end time of the current round of a business unit is the end time of the current round of the round structure corresponding to the largest round number covered by the round range of the business unit.

[0083] In one possible implementation, the rule association operation unit 304 specifically includes: The temporal relationship determination unit is used to determine the temporal predecessor and successor relationship of each structured knowledge entry based on the start time of the current round of the business unit corresponding to the structured knowledge entry, and write the temporal predecessor and successor relationship into the structured knowledge entry. The overlapping association verification unit is used to perform interval intersection operation on the round range of the business unit corresponding to each structured knowledge entry, complete the overlapping parallel association verification, obtain the overlapping association unit identifier of the structured knowledge entry, and write the overlapping association unit identifier into the structured knowledge entry. The topic clustering unit is used to perform similar clustering for all structured knowledge entries according to the main core topic field within the structured knowledge entry. Structured knowledge entries with the same main core topic are grouped into the same cluster group, a unique cluster group identifier is assigned to each cluster group, and the cluster group identifier is written into the corresponding structured knowledge entry. The similarity conflict marking unit is used to calculate the text similarity between any two structured knowledge entries using the Jaccard similarity algorithm. When the similarity is greater than or equal to a preset threshold, the corresponding structured knowledge entry is marked as potentially conflicting, and the potentially conflicting mark is written into the corresponding structured knowledge entry. Among them, temporal predecessor-successor relationships, overlapping associated unit identifiers, cluster group identifiers, and suspected conflicts are all association marker information.

[0084] In one possible implementation, the timing relationship determination unit specifically includes: The sorting unit is used to sort all current structured knowledge items in ascending order according to the start time of the current round of the corresponding business unit; The successor unit selection unit is used to determine the successor unit of each structured knowledge item if the topic convergence status of the business unit corresponding to the structured knowledge item is not converged. If the topic convergence status of the business unit corresponding to the structured knowledge item is converged, the business unit whose topic convergence status is not converged is determined to be the successor unit of the business unit. The precursor unit derivation unit is used to deduce the precursor unit of each structured knowledge item in reverse based on the correspondence between the successor units recorded in all structured knowledge items. The temporal relationship writing unit is used to write the information of the predecessor and successor units of the structured knowledge entry as a temporal predecessor-successor relationship into the structured knowledge entry.

[0085] In one possible implementation, the precursor unit derivation unit is specifically used for: Read the subsequent units of each structured knowledge entry; If the successor unit of a certain structured knowledge entry is a business unit corresponding to another structured knowledge entry, then the business unit corresponding to the structured knowledge entry is determined as the predecessor unit of the other structured knowledge entry. Specifically, for structured knowledge entries that are not pointed to by any successor unit of a structured knowledge entry, their predecessor unit is set to empty.

[0086] In one possible implementation, the quality governance verification includes format verification, segmentation correctness verification, business type correctness verification, and conflict verification. Among them, format verification failure is judged as mechanical problem, segmentation correctness verification failure is judged as large model quality problem, business type correctness verification failure is judged as system problem, and conflict verification failure is classified into mechanical problem, model quality problem or system problem according to the problem category. The quality repair process for the aforementioned mechanical problems is: automatic repair via program execution; The quality repair process for the aforementioned model quality issues is to re-call the large model to perform splitting or distillation processing. The quality repair process for the system-related issues is as follows: modify the configuration and perform reclassification.

[0087] In one possible implementation, the device further includes: The secondary verification unit is used to perform a second quality management verification on the structured knowledge items after the quality repair process has been completed. The status marking unit is used to mark structured knowledge entries that pass the second verification as active, so that they can be retrieved and called from outside. The rerun processing unit is used to store the structured knowledge items that fail the re-verification into a preset rerun database and generate a rerun command.

[0088] In addition, this application embodiment also provides a call transcription structured knowledge parsing and management device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the call transcription structured knowledge parsing and management method as described above.

[0089] In addition, this application embodiment also provides a computer-readable storage medium storing instructions that, when executed on a terminal device, cause the terminal device to perform the call transcription structured knowledge parsing and governance method as described above.

[0090] This application embodiment performs round-based normalization on the upstream call transcription object, and completes business unit segmentation by combining segmentation constraints including splitting rules, rules to prevent incorrect segmentation, and rules to determine overlap. In scenarios where legal round overlap is allowed with the main topic and temporary branches running in parallel, it outputs overlapping and related unit identifiers, thus preserving the overlapping relationships between conversation rounds and preventing the loss of interactive information after round splitting. Simultaneously, it does not employ a forced semantic loop slicing mode, but relies on rules to prevent incorrect segmentation to ensure the semantic integrity of a single round, supporting cross-round construction of business units and fully preserving the real interactive details of topic interweaving and branch parallelism within the call. Furthermore, this application does not rely on a preset standard dialogue library; it directly performs business unit segmentation, business field distillation, and rule association operations based on the original call transcription data, automatically generating structured knowledge entries with association tags. After quality governance verification and repair processing, it outputs the treated structured knowledge entries, achieving the autonomous accumulation of fine-grained unit-level structured knowledge from the original call text. This adapts to the needs of high-precision knowledge retrieval, efficient knowledge reuse, and refined business quality inspection in environments with massive call data.

[0091] The foregoing has provided a detailed description of the call transcription structured knowledge parsing and governance method and related products provided in this application. The various embodiments in the specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to in the method section. It should be noted that those skilled in the art can make several improvements and modifications to this application without departing from the principles of this application, and these improvements and modifications also fall within the protection scope of the claims of this application.

[0092] It should be understood that in this application, "and / or" is used to describe the relationship between related objects, indicating that there can be three relationships. For example, "A and / or B" can mean: only A exists, only B exists, and both A and B exist simultaneously, where A and B can be singular or plural. The character " / " generally indicates that the related objects before and after are in an "or" relationship. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one of a, b, or c can mean: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple. Moreover, the terms "comprising", "including", or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. In the absence of further restrictions, an element defined by the phrase "comprising a..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

Claims

1. A method for managing structured knowledge parsing in call transcription, characterized in that, The method includes: Receive upstream call transcription objects and normalize them into multiple round structures; each round structure includes a round number, a general role tag for this round, a start time for this round, an end time for this round, and round text; the general role tag for this round is obtained by mapping the original role name to a pre-configured role mapping table; The multiple round structures are used as dialogues to be segmented, and the segmentation rules are used as segmentation constraints. These are input into a pre-defined large model to segment dialogue units, resulting in multiple business units. Each business unit includes a round range, business type, topic convergence state, overlapping branch state, and overlapping associated units. The segmentation rules include a dividing line rule, a rule to prohibit incorrect segmentation, and an overlap determination rule. The dividing line rule is used to identify the location where the business semantics change during the call. The rule to prohibit incorrect segmentation is used to constrain the large model from breaking up the single-round dialogue content with the same semantics. The overlap determination rule is used to limit the overlap of round ranges of different business units to only be allowed in call scenarios where the main topic is continuous and accompanied by temporary branches in parallel. It prohibits full-scale stacking overlap without business logic and outputs the corresponding overlapping associated unit identifier when overlap exists. For each business unit, the round range of the business unit and the round text corresponding to the round number covered by the round range, together with the preset distillation constraint prompts, are used as distillation inputs and input into the large model to perform business field distillation, thereby obtaining the structured knowledge entries corresponding to each business unit; the structured knowledge entries are used to describe the business communication information contained in the call dialogue within the corresponding business unit. Based on the time attributes and business field information of the business units corresponding to each structured knowledge entry, rule association operations are performed on each structured knowledge entry to obtain the association tag information corresponding to each structured knowledge entry, and the association tag information is written into the corresponding structured knowledge entry to obtain the tagged structured knowledge entry. Each labeled structured knowledge item undergoes quality governance verification to obtain verification results. Based on each verification result, the corresponding quality repair process is performed on the corresponding labeled structured knowledge item to obtain the remediated structured knowledge item.

2. The method according to claim 1, characterized in that, The start time of this round for a business unit is the start time of the round structure corresponding to the smallest round number covered by the round range of the business unit; the end time of this round for a business unit is the end time of the round structure corresponding to the largest round number covered by the round range of the business unit. The process involves performing rule-based association operations on each structured knowledge entry based on the time attributes and business field information of the business unit corresponding to each structured knowledge entry, obtaining association tagging information corresponding to each structured knowledge entry, and writing the association tagging information into the corresponding structured knowledge entry to obtain tagged structured knowledge entries, including: For each structured knowledge entry, the temporal predecessor and successor relationship of the business unit corresponding to the structured knowledge entry is determined based on the start time of the current round, and the temporal predecessor and successor relationship is written into the structured knowledge entry. For each structured knowledge entry, perform interval intersection operation on the round range of the business unit corresponding to the structured knowledge entry to complete the overlapping parallel association verification, obtain the overlapping association unit identifier of the structured knowledge entry, and write the overlapping association unit identifier into the structured knowledge entry. For all structured knowledge entries, perform similar clustering according to the main core topic field within the structured knowledge entry. Structured knowledge entries with the same main core topic are grouped into the same cluster group. A unique cluster group identifier is assigned to each cluster group, and the cluster group identifier is written into the corresponding structured knowledge entry. For all structured knowledge entries, the Jaccard similarity algorithm is used to calculate the text similarity between each pair of structured knowledge entries. When the similarity is greater than or equal to a preset threshold, the corresponding structured knowledge entries are marked as potentially conflicting, and the potentially conflicting marks are written into the corresponding structured knowledge entries. Among them, temporal predecessor-successor relationships, overlapping associated unit identifiers, cluster group identifiers, and suspected conflicts are all association marker information.

3. The method according to claim 2, characterized in that, For each structured knowledge entry, the process of determining the temporal predecessor-successor relationship of the business unit corresponding to that structured knowledge entry based on the start time of the current round, and writing the temporal predecessor-successor relationship into the structured knowledge entry, includes: Sort all current structured knowledge items in ascending order according to the start time of the current round of the corresponding business unit; For each structured knowledge item, if the topic convergence status of the business unit corresponding to the structured knowledge item is not converged, then the business unit whose start time in this round is later than that business unit and whose overlapping branch status is yes is determined as the successor unit of that business unit; if the topic convergence status of the business unit corresponding to the structured knowledge item is converged, then the business unit whose start time in this round is earlier than the end time of this round of that business unit and whose topic convergence status is not converged is determined as the successor unit of that business unit. Based on the correspondence between successor units recorded in all structured knowledge entries, the predecessor units of each structured knowledge entry are derived in reverse. The information of the predecessor and successor units of the structured knowledge entry is written into the structured knowledge entry as a temporal predecessor-successor relationship.

4. The method according to claim 3, characterized in that, The process of deriving the predecessor unit of each structured knowledge item by reverse derivation based on the correspondence between successor units recorded in all structured knowledge items includes: Read the subsequent units of each structured knowledge entry; If the successor unit of a certain structured knowledge entry is a business unit corresponding to another structured knowledge entry, then the business unit corresponding to the structured knowledge entry is determined as the predecessor unit of the other structured knowledge entry. Specifically, for structured knowledge entries that are not pointed to by any successor unit of a structured knowledge entry, their predecessor unit is set to empty.

5. The method according to claim 1, characterized in that, The quality governance verification includes format verification, segmentation correctness verification, business type correctness verification, and conflict verification; Among them, format verification failure is judged as mechanical problem, segmentation correctness verification failure is judged as large model quality problem, business type correctness verification failure is judged as system problem, and conflict verification failure is classified into mechanical problem, model quality problem or system problem according to the problem category. The quality repair process for the aforementioned mechanical problems is: automatic repair via program execution; The quality repair process for the aforementioned model quality issues is to re-call the large model to perform splitting or distillation processing. The quality repair process for the system-related issues is as follows: modify the configuration and perform reclassification.

6. The method according to claim 1, characterized in that, The method further includes: The structured knowledge entries, after quality repair processing, are subjected to quality governance verification again. The structured knowledge entries that pass the second verification will be marked as active for external retrieval. The structured knowledge items that fail the re-verification will be stored in the preset rerun database, and a rerun command will be generated.

7. A structured knowledge parsing and management device for call transcription, characterized in that, The device includes: The round normalization unit is used to receive upstream call transcription objects and normalize the upstream call transcription objects into multiple round structures; each round structure includes a round number, a general role tag for this round, a start time for this round, an end time for this round, and round text; the general role tag for this round is obtained by mapping the original role name to a pre-configured role mapping table; The dialogue segmentation unit is used to segment the multiple round structures as dialogues to be segmented, and the segmentation rules as segmentation constraints, inputting them into a preset large model to segment the dialogue units, resulting in multiple business units. Each business unit includes a round range, business type, topic convergence state, overlapping branch state, and overlapping associated units. The segmentation rules include a dividing line rule, a rule to prohibit incorrect segmentation, and an overlap determination rule. The dividing line rule is used to identify the location where the business semantics change during the call. The rule to prohibit incorrect segmentation is used to constrain the large model from breaking up the single-round dialogue content with the same semantics. The overlap determination rule is used to limit the overlap of round ranges of different business units to only be allowed in call scenarios where the main topic is continuous and accompanied by temporary branches in parallel, prohibiting full-scale stacking overlap without business logic, and outputting the corresponding overlapping associated unit identifier when overlap exists. The field distillation unit is used to take the round range of the business unit and the round text corresponding to the round number covered by the round range, together with the preset distillation constraint prompts, as distillation inputs and input them into the large model to perform business field distillation, so as to obtain the structured knowledge entries corresponding to each business unit; the structured knowledge entries are used to describe the business communication information contained in the call dialogue within the corresponding business unit. The rule association operation unit is used to perform rule association operations on each structured knowledge entry based on the time attribute and business field information of the business unit corresponding to each structured knowledge entry, to obtain the association tag information corresponding to each structured knowledge entry, and write the association tag information into the corresponding structured knowledge entry to obtain the tagged structured knowledge entry. The quality management processing unit is used to perform quality management verification on each marked structured knowledge item to obtain the verification result, and to perform corresponding quality repair processing on the corresponding marked structured knowledge item based on each verification result to obtain the managed structured knowledge item.

8. The apparatus according to claim 7, characterized in that, The start time of this round for a business unit is the start time of the round structure corresponding to the smallest round number covered by the round range of the business unit; the end time of this round for a business unit is the end time of the round structure corresponding to the largest round number covered by the round range of the business unit. The rule association operation unit specifically includes: The temporal relationship determination unit is used to determine the temporal predecessor and successor relationship of each structured knowledge entry based on the start time of the current round of the business unit corresponding to the structured knowledge entry, and write the temporal predecessor and successor relationship into the structured knowledge entry. The overlapping association verification unit is used to perform interval intersection operation on the round range of the business unit corresponding to each structured knowledge entry, complete the overlapping parallel association verification, obtain the overlapping association unit identifier of the structured knowledge entry, and write the overlapping association unit identifier into the structured knowledge entry. The topic clustering unit is used to perform similar clustering for all structured knowledge entries according to the main core topic field within the structured knowledge entry. Structured knowledge entries with the same main core topic are grouped into the same cluster group, a unique cluster group identifier is assigned to each cluster group, and the cluster group identifier is written into the corresponding structured knowledge entry. The similarity conflict marking unit is used to calculate the text similarity between any two structured knowledge entries using the Jaccard similarity algorithm. When the similarity is greater than or equal to a preset threshold, the corresponding structured knowledge entry is marked as potentially conflicting, and the potentially conflicting mark is written into the corresponding structured knowledge entry. Among them, temporal predecessor-successor relationships, overlapping associated unit identifiers, cluster group identifiers, and suspected conflicts are all association marker information.

9. A call transcription structured knowledge parsing and management device, characterized in that, include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, it implements the call transcription structured knowledge parsing and governance method as described in any one of claims 1-6.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores instructions that, when executed on a terminal device, cause the terminal device to perform the call transcription structured knowledge parsing and governance method as described in any one of claims 1-6.

Images