A dialogue processing method, device, storage medium and electronic equipment

Abstract

The specification discloses a dialogue processing method, device, storage medium and electronic equipment, wherein the method comprises: obtaining service fulfillment reminding scene service fulfillment dialogue data, transcribing the service fulfillment dialogue data to obtain fulfillment reminding interactive text sequence, performing dialogue service quality inspection processing on the fulfillment reminding interactive text sequence based on the fulfillment reminding interactive text sequence and multiple dialogue quality inspection agents to obtain multiple dialogue service quality inspection information, performing quality inspection arbitration processing based on the multiple dialogue service quality inspection information and the fulfillment reminding interactive text sequence and using a quality inspection arbitration agent to obtain a comprehensive dialogue service quality inspection result. The specification improves the accuracy and processing efficiency of service quality verification.

Description

Technical Field

[0001] This specification relates to the field of computer technology, and in particular to a dialogue processing method, apparatus, storage medium, and electronic device. Background Technology

[0002] With the rapid development of digital enterprise services and cloud computing technologies, various integrated service platforms widely provide continuous or phased platform services to external users. When users utilize these platform services, corresponding service fulfillment processes are inevitably involved (such as phased data resource scheduling confirmation, periodic service task node feedback, service renewal, and user service contract execution). To ensure the stable flow of the platform service chain and the effective allocation of service data resources, the server will communicate with users to remind them of service fulfillment requirements. In various service fulfillment reminder scenarios, service platforms generate massive amounts of service fulfillment dialogue data daily (such as full-duplex voice interaction data streams). To ensure the standardization of service interactions and provide users with better platform services, service platforms typically need to perform service quality verification and anomaly risk monitoring on this massive, unstructured dialogue data. Currently, dialogue data processing and quality inspection methods in related technologies mainly rely on manual offline sampling verification and simple keyword and regular expression matching based on a preset static rule base. Summary of the Invention

[0003] This specification provides a dialogue processing method, apparatus, storage medium, and electronic device, the technical solutions of which are as follows: Firstly, this specification provides a dialogue processing method, the method comprising: Obtain service fulfillment dialogue data in the service fulfillment reminder scenario, and transcribe the service fulfillment dialogue data to obtain a sequence of fulfillment reminder interaction text; Based on the aforementioned performance reminder interactive text sequence, multiple dialogue quality inspection agents are used to perform dialogue service quality inspection processing to obtain multiple dialogue service quality inspection information. Based on multiple dialogue service quality inspection information and the performance reminder interactive text sequence, a quality inspection arbitration agent is used to perform quality inspection arbitration processing to obtain a comprehensive dialogue service quality inspection result.

[0004] Secondly, this specification provides a dialogue processing apparatus, the apparatus comprising: The data processing module is used to acquire service performance dialogue data in the service performance reminder scenario, and transcribe the service performance dialogue data to obtain a performance reminder interactive text sequence; The dialogue quality inspection module is used to perform dialogue service quality inspection processing using multiple dialogue quality inspection agents based on the performance reminder interactive text sequence, and obtain multiple dialogue service quality inspection information. The quality inspection arbitration module is used to obtain a comprehensive dialogue service quality inspection result by using a quality inspection arbitration agent to perform quality inspection arbitration processing based on multiple dialogue service quality inspection information and the performance reminder interactive text sequence.

[0005] Thirdly, this specification provides a computer storage medium storing at least one instruction adapted for loading by a processor and executing method steps of one or more embodiments of this specification.

[0006] Fourthly, this specification provides a computer program product storing at least one instruction adapted to be loaded by a processor and to execute the method steps of one or more embodiments of this specification.

[0007] Fifthly, this specification provides an electronic device that may include: a processor and a memory; wherein the memory stores a computer program adapted to be loaded by the processor and to execute the method steps of one or more embodiments of this specification.

[0008] The beneficial effects of the technical solutions provided in some embodiments of this specification include at least the following: In one or more embodiments of this specification, service performance dialogue data in a service performance reminder scenario is acquired and transcribed into a sequence of performance reminder interactive text. Then, based on this sequence, multiple dialogue quality inspection agents are used to perform dialogue service quality inspection to obtain multiple dialogue service quality inspection information. Finally, a quality inspection arbitration agent is used to perform quality inspection arbitration based on the multiple dialogue service quality inspection information and the performance reminder interactive text sequence to obtain a comprehensive dialogue service quality inspection result. This solves the problems of low efficiency, insufficient data coverage, and high false alarm rates and inaccurate identification of deep intent of nodes due to the lack of deep semantic understanding of the dialogue context caused by static rule matching. The fully automated process breaks through the data throughput bottleneck of manual processing, meeting the real-time quality inspection needs of massive concurrent data streams. Simultaneously, through the collaborative computing and arbitration mechanism of multiple agents, accurate parsing of deep semantics of high-dimensional complex interactive data and cross-fusion of multi-dimensional features are achieved, reducing the system's false alarm rate and improving the accuracy of deep intent identification of server and user nodes and the overall efficiency of automated verification of global service quality. Attached Figure Description

[0009] To more clearly illustrate the technical solutions in this specification or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this specification. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 This is a schematic diagram of a dialogue processing system provided in this manual; Figure 2 This is a flowchart illustrating a dialogue service processing method provided in this manual. Figure 3 This is a flowchart illustrating the quality inspection process for a dialogue service provided in this manual. Figure 4 This is a schematic diagram of a quality inspection arbitration process provided in this manual; Figure 5 This is a schematic diagram of the structure of a dialogue processing device provided in this specification; Figure 6 This is a schematic diagram of the structure of an electronic device provided in this specification. Detailed Implementation

[0011] The technical solutions in this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this specification, and not all embodiments. Based on the embodiments in this specification, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this specification.

[0012] In the description of this specification, it should be understood that the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance. In the description of this specification, it should be noted that, unless otherwise expressly specified and limited, "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or devices. Those skilled in the art can understand the specific meaning of the above terms in this specification based on the specific circumstances. Furthermore, in the description of this specification, unless otherwise stated, "multiple" means two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, and B alone. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship.

[0013] In related technologies, manual offline sampling and verification are relied upon, or simple keyword and regular expression matching is performed based on a pre-set static rule base. This approach has significant limitations in practical applications.

[0014] On the one hand, manual sampling has extremely low data processing throughput and insufficient coverage, which cannot meet the processing needs of massive concurrent data streams. On the other hand, data processing methods based on simple rule matching lack the ability to understand the deep semantics of the dialogue context. When faced with high-dimensional and complex interactive data, they are prone to logical false positives, resulting in a high false alarm rate and an inability to accurately identify the deep intentions of users or server nodes.

[0015] The present specification will now be described in detail with reference to specific embodiments.

[0016] Please see Figure 1 This is a schematic diagram of a dialogue processing system provided in this specification. Figure 1 As shown, the dialogue processing system may include at least a client cluster and a service platform 100.

[0017] The client cluster may include at least one client, such as Figure 1 As shown, it specifically includes client 1 corresponding to user 1, client 2 corresponding to user 2, ..., client n corresponding to user n, where n is an integer greater than 0.

[0018] Each client in a client cluster can be an electronic device with communication capabilities, including but not limited to: wearable devices, handheld devices, personal computers, tablets, in-vehicle devices, smartphones, computing devices, or other processing devices connected to a wireless modem. Electronic devices may have different names in different networks, such as: user equipment, access terminal, user unit, user station, mobile station, mobile station, remote station, remote terminal, mobile device, user terminal, electronic device, wireless communication device, user agent or user device, cellular phone, cordless phone, personal digital assistant (PDA), and electronic devices in 5G networks or future evolved networks.

[0019] The service platform 100 can be a standalone server device, such as a rack-mount, blade, tower, or cabinet-type server device, or a workstation, mainframe, or other hardware device with strong computing power; or it can be a server cluster composed of multiple servers. The servers in the service cluster can be composed in a symmetrical manner, wherein each server is functionally and hierarchically equivalent in the transaction chain, and each server can provide services independently. The independent provision of services can be understood as not requiring the assistance of other servers.

[0020] In one or more embodiments of this specification, the service platform 100 can establish a communication connection with at least one client in the client cluster, and complete data interaction during the dialogue processing based on this communication connection. It should be noted that the service platform 100 and at least one client in the client cluster establish a communication connection for interactive communication via a network. This network can be a wireless network or a wired network. Wireless networks include, but are not limited to, cellular networks, wireless LANs, infrared networks, or Bluetooth networks. Wired networks include, but are not limited to, Ethernet, universal serial bus (USB), or controller area network. In one or more embodiments of this specification, technologies and / or formats including Hyper Text Markup Language (HTML), Extensible Markup Language (XML), etc., are used to represent data exchanged over the network (such as target compressed packets). Furthermore, conventional encryption technologies such as Secure Socket Layer (SSL), Transport Layer Security (TLS), Virtual Private Network (VPN), and Internet Protocol Security (IPsec) can be used to encrypt all or some links. In other embodiments, customized and / or dedicated data communication technologies may be used to replace or supplement the aforementioned data communication technologies.

[0021] In a schematic representation, the client's primary function is to act as a data acquisition or presentation end for service fulfillment. When a service fulfillment reminder is initiated (e.g., a resource allocation reminder or transaction status change notification sent to the target response node), the participating client will collect full-duplex voice signals in real time or receive offline audio files and encapsulate them into service fulfillment dialogue data. This service fulfillment dialogue data can be continuously or periodically uploaded to the service platform 100 via a pre-defined application programming interface (API) or streaming media transmission protocol (such as WebSocket, RTMP, etc.). Furthermore, the client cluster may also include a dedicated management client for backend management, used to receive and graphically display the final comprehensive dialogue service quality inspection results issued by the service platform 100, for subsequent closed-loop processing by administrators or manual verification nodes.

[0022] Optionally, the client can be a device used by customer service personnel who maintain the platform services; in some embodiments, the client may also include the device of a user who uses the platform services, which may also be referred to as a user terminal.

[0023] Indicatively, when the service platform 100 receives service fulfillment dialogue data from the client through the communication gateway, the service platform 100 activates the data transcription module. This module invokes the acoustic base model deployed at the platform's underlying layer to perform acoustic feature decoding on the audio stream corresponding to the acquired service fulfillment dialogue data. The service platform 100 utilizes the underlying tensor computing unit to map the unstructured speech analog signal into discrete character information, thereby transcribing the service fulfillment dialogue data and ultimately generating and mounting a fulfillment reminder interactive text sequence in the cache. This text sequence fully preserves the temporal characteristics of the interaction process, providing a basic data structure for subsequent multi-dimensional quality inspection.

[0024] Schematic illustration: The service platform 100 may contain a computing layer deployed based on a multi-agent collaborative architecture. Based on the aforementioned performance reminder interactive text sequence, the service platform 100 utilizes a topological sorting mechanism to concurrently wake up multiple dialogue quality inspection agents in the underlying thread pool.

[0025] Optionally, the multiple dialogue quality inspection agents may perform in-depth feature extraction and computation on different dimensions of service fulfillment interactions (such as node intent state, abnormal interaction signaling, communication topology connectivity, etc.). Alternatively, multiple dialogue quality inspection agents may perform in-depth feature extraction and computation on the same dimension of service fulfillment interactions.

[0026] The service platform 100 allocates independent video memory and computing processes to each dialogue quality inspection agent, driving each agent to perform parallel computing on the performance reminder interactive text sequence in the semantic vector space, thereby performing dialogue service quality inspection processing with high concurrency, and finally obtaining multiple dialogue service quality inspection information at their respective output ports.

[0027] Schematic illustration: Service platform 100 further deploys a quality inspection arbitration agent independent of the aforementioned dialogue quality inspection agent, serving as the decision-making center for global quality verification. After receiving multiple dialogue service quality inspection messages generated by the concurrent computation, based on the multiple dialogue service quality inspection messages and the performance reminder interaction text sequence, service platform 100 invokes the quality inspection arbitration agent to perform quality inspection arbitration processing, and finally calculates the converged comprehensive dialogue service quality inspection result.

[0028] Optionally, after completing the above calculations, the service platform 100 can persistently write the comprehensive dialogue service quality inspection results into the underlying distributed database, or push the result data packet to the designated management client in the client cluster in real time through the communication link, so as to trigger subsequent manual intervention processes or automated service correction actions, thereby forming a complete, efficient, and low-latency intelligent verification mechanism for dialogue processing and service performance quality.

[0029] The dialogue processing system embodiments provided in this specification and the dialogue processing methods described in one or more embodiments belong to the same concept. The execution entity corresponding to the dialogue processing method involved in one or more embodiments in this specification can be the aforementioned service platform 100; the execution entity corresponding to the dialogue processing method involved in one or more embodiments in this specification can also be the electronic device corresponding to the client, specifically determined based on the actual application environment. The implementation process of the dialogue processing system embodiments can be detailed in the following method embodiments, and will not be repeated here.

[0030] based on Figure 1 The following is a detailed description of the dialogue processing method provided by one or more embodiments of this specification, as illustrated in the scene diagram.

[0031] Please see Figure 2 This specification provides a flowchart illustrating a dialogue processing method according to one or more embodiments. This method can be implemented using a computer program and can run on a dialogue processing device based on the von Neumann architecture. The computer program can be integrated into an application or run as a standalone utility application. The dialogue processing device can be a service platform.

[0032] Specifically, the dialogue processing method includes: S102: Obtain service performance dialogue data in the service performance reminder scenario, and transcribe the service performance dialogue data to obtain a performance reminder interaction text sequence; Service fulfillment reminder scenario refers to the dialogue and data interaction environment (such as voice dialogue environment, video dialogue environment, chat dialogue environment) established between the customer service terminal (i.e., service dispatch terminal node) of the service platform and the communication terminal (i.e. target response terminal node) held by the user to be fulfilled in the computer network communication architecture, in order to promote the closed loop of specific transaction services (such as reminding users to complete resource service renewal, contract service fulfillment, etc.).

[0033] Service fulfillment dialogue data refers to the communication data captured in the above-mentioned call scenarios between customer service and users whose services are pending fulfillment, which includes complete interactions between the two parties.

[0034] Transcription is the process of mapping unstructured service fulfillment dialogue data into a sequence of fulfillment reminder interactive text.

[0035] The performance reminder interactive text sequence refers to the structured data set generated after the above transcription and decoding. In some embodiments, it includes a timestamp aligned with the original dialogue, node role identifiers that distinguish customer service and users to be fulfilled, and corresponding natural language streaming media (such as text) data.

[0036] In one feasible implementation, a deep neural network model for transcribing dialogue data can be used to achieve high-precision transcription. First, the original two-channel audio file is physically decomposed, separating the audio channels corresponding to the customer service representative and the user awaiting fulfillment into two independent mono audio streams. Then, an acoustic channel data decoding engine is used to perform acoustic decoding and text mapping operations on the waveform data in these two independent audio streams. After decoding, the text content generated independently by the two channels is sequentially merged and logically recombined according to a unified global timeline, ultimately generating a fulfillment reminder interactive text sequence containing complete interaction timestamps and identifying the roles of the customer service representative and the user awaiting fulfillment.

[0037] Optionally, the specific execution of step S102, which involves transcribing the service fulfillment dialogue data to obtain the fulfillment reminder interactive text sequence, can be performed in the following manner: Step A2: Drive the acoustic base model through the data transcription agent to extract speaker separation features from the audio stream corresponding to the service fulfillment dialogue data to obtain the node role identifier that distinguishes the service scheduling end node and the target response end node, and decouple the audio stream into independent audio channels corresponding to the service scheduling end node and the target response end node. A data transcription agent is an independent artificial intelligence entity deployed on distributed computing nodes. Unlike traditional static executable programs, this agent uses a deep neural network or a large language model (LLM) as its acoustic foundation model. In some embodiments, the data transcription agent's architecture integrates an environment-aware interface, a task planning engine, tool invocation operators, and a context state memory cache. In this embodiment, the data transcription agent can autonomously monitor the input state of service fulfillment dialogue data through the perception interface, dynamically generate transcription strategies based on the task planning engine, and actively drive (invoke) the underlying acoustic foundation model as an external tool to execute specific waveform decoding and feature separation tasks as a control node. Simultaneously, the agent maintains the full lifecycle state of the interaction sequence in memory, achieving efficient mapping and autonomous assembly of unstructured audio streams into structured character matrices.

[0038] Speaker separation feature extraction refers to the acoustic signal processing process of extracting voiceprint feature vectors from a mixed audio stream and determining who spoke and when through clustering.

[0039] An independent audio channel refers to a data stream or memory block that contains only the voice signal of a single specific character after the original audio stream has been masked or physically separated.

[0040] Step A4: Decode the acoustic channel data of the independent audio channel to generate a sequence of interactive text messages for fulfilling obligations, including timestamps and the node role identifier.

[0041] Acoustic channel data decoding refers to the computational process of mapping continuous acoustic feature frames within a specific channel into discrete natural language text characters.

[0042] Indicatively, after acquiring service fulfillment dialogue data (such as historical recording files), the control data transcription agent initializes and deploys the acoustic base model accordingly, and then processes it: During step A2, the complete audio stream file is first loaded into memory. The acoustic base model then uses a global self-attention mechanism to perform global feature analysis on the entire temporal audio sequence. By executing a speaker separation feature extraction algorithm based on spectral clustering, the acoustic base model can fully utilize global contextual information to accurately segment the speech boundaries of different speakers and assign node role identifiers to distinguish between the service scheduling node and the target response node. Using the boundary information corresponding to these node role identifiers, a Boolean mask matrix is ​​generated at the underlying level for the original tensor data, decoupling the original audio stream at the mathematical tensor level into independent audio channels corresponding to the service scheduling node and the target response node. When executing step A4, the control data transcription agent drives the acoustic base model to decode the acoustic channel data of each independent audio channel. After the data decoding maps the text characters, it accurately calculates and generates the timestamp of each speech segment based on the global timeline of the audio file. Finally, it aligns and structures these timestamps, node role identifiers and decoded text to generate a complete sequence of performance reminder interactive text.

[0043] This specification utilizes a data transcription agent to drive the underlying acoustic base model to perform speaker separation, precisely decoupling the full-duplex mixed audio into independent channels. This effectively eliminates acoustic feature cross-interference and improves the accuracy of the underlying text transcription. Simultaneously, high-precision timestamps and node role identifiers are bound during the decoding process, transforming unstructured audio signals into structured interactive text sequences with temporal attributes. This provides high-quality data for downstream multi-agent high-concurrency feature extraction and accurate semantic analysis.

[0044] Optionally, the following describes the model training process of an acoustic pedestal model driven by a data transcription agent: Model creation phase: The neural network topology of the acoustic base model is initialized and constructed based on the machine learning model. For example, considering the high temporal dependence of audio data, the acoustic base model in this embodiment can adopt a deep multi-task encoder-decoder architecture. Specifically, the low-level feature extraction front end of the model contains several layers of convolutional neural networks to extract frequency domain features of the acoustic dimension from the input raw audio waveform; after the feature extraction front end, multiple sequence modeling modules based on self-attention mechanisms (such as Conformer encoding layers) are cascaded to capture long-range contextual dependencies between temporal feature tensors; at the output end of the model, the network structure is designed as a two-branch topology: the first branch is a fully connected layer for voiceprint feature mapping, used to output voiceprint clustering vectors representing different roles; the second branch is a character probability distribution decoding layer, used to output the natural language character prediction results corresponding to the acoustic frames.

[0045] Sample training material acquisition stage: A large number of interactive voice communication records are extracted from the historical interaction database as raw training materials; In the sample data annotation phase: an expert-based system constructs training sample pairs consisting of input and output audio tensors and supervised label data from the original training materials. The specific supervised label data includes three aligned logical dimensions: the first dimension is a timestamp boundary label, used to precisely define the absolute start and end points of each consecutive pronunciation segment on the timeline; the second dimension is a node role label, assigning a clear category attribute to the corresponding pronunciation segment (e.g., labeling it as a service scheduling node or a target response node); the third dimension is a text sequence label, which is the accurate transcribed character sequence corresponding to the pronunciation segment in human natural language semantics. The system serializes and encapsulates the above three dimensions of supervised label data, strictly aligning and binding it with the audio tensor.

[0046] Model training phase: According to the preset batch size, the constructed training samples are input into the input layer of the acoustic pedestal model. In each forward propagation, the acoustic pedestal model calculates and outputs the voiceprint clustering prediction matrix and character transcription probability distribution matrix under the current network weights based on the input audio tensor. Then, the underlying tensor calculation unit is called to calculate the joint loss value of the prediction matrix and the supervision label data generated in the sample data annotation phase using a preset loss function (such as contrastive loss function, cross-entropy loss function, Euclidean distance loss function).

[0047] After calculating the global joint loss scalar, the backpropagation algorithm is executed to derive the gradient tensor matrix of the global joint loss scalar with respect to each neuron node in the network. Then, using a preset deep learning optimizer (e.g., the Adam optimizer), the weight parameter tensors of the underlying acoustic pedestal model are updated along the gradient descent direction according to a dynamically decaying learning rate. The above forward propagation and backpropagation parameter update process will be continuously iterated on the full training dataset until the joint loss value calculated by the model (on the validation set) converges and stabilizes within a preset tolerance threshold, or until a preset maximum number of iterations is reached. This completes the training of the acoustic pedestal model.

[0048] Optionally, the preset loss function may include a joint objective function of a connection-temporal classification loss term and a cross-entropy loss term. During calculation, the connection-temporal classification loss value, used to address the misalignment between the input audio frame length and the output text character length, and the cross-entropy loss value, used to constrain the accuracy of voiceprint role classification, are calculated separately. The connection-temporal classification loss value is multiplied by the weight hyperparameters to obtain the first loss component. The difference between a constant and the weight hyperparameters is calculated, and the cross-entropy loss value is multiplied by this difference to obtain the second loss component. The first and second loss components are summed, and the sum represents the global joint loss scalar of a single forward computation of the acoustic pedestal model. This global joint loss scalar, through the preset weight hyperparameters, comprehensively balances the error proportions of different acoustic tasks in backpropagation.

[0049] S104: Based on the performance reminder interactive text sequence, multiple dialogue quality inspection agents are used to perform dialogue service quality inspection processing to obtain multiple dialogue service quality inspection information. The dialogue quality inspection agent uses a deep neural network or a large language model (LLM) as its inference control center. In related technologies, the agent's architecture integrates a temporal data perception interface, a feature extraction planning engine, a model invocation operator, and a dialogue context state memory cache. In this embodiment, multiple dialogue quality inspection agents can autonomously listen to and read the performance reminder interaction text sequence generated in the underlying memory object pool through the perception interface. Based on their respective built-in task planning engines, they dynamically generate feature calculation strategies for specific quality inspection dimensions and actively drive (invoke) the corresponding underlying large language model as an inference tool, independently performing deep text parsing and feature mapping operations in the implicit high-dimensional semantic space. Simultaneously, the agent maintains the context temporal state of multi-turn dialogues in real time in the underlying memory, autonomously reducing the dimensionality of the input structured character matrix or transforming it into a standardized dialogue service quality inspection information data packet including specific quality inspection features and their corresponding confidence levels.

[0050] Dialogue service quality inspection information refers to a structured data set output by various dialogue quality inspection agents after processing. In the underlying data structure, it is represented by quality inspection features of specific dimensions (such as intent labels, risk hit status, topology scores) and their corresponding confidence levels.

[0051] Schematic illustration: Multiple dialogue quality inspection agents can perform the same "violation risk scanning" task, using the same sequence of performance reminder interaction texts as a shared tensor input, distributed to the aforementioned heterogeneous dialogue quality inspection agents. The first dialogue quality inspection agent invokes a basic large language model with hundreds of billions of parameters, using a multi-head self-attention mechanism to extract generalized features from the text sequence based on the full semantic space, calculating the violation probability tensor of the first dimension. Simultaneously, the second dialogue quality inspection agent invokes a model with tens of billions of parameters fine-tuned with vertical domain data, injecting a specific thought chain reasoning template, forcing the model to perform layer-by-layer Boolean logic derivation according to a preset compliance logic tree, calculating the violation probability tensor of the second dimension. Concurrently, the third dialogue quality inspection agent, without relying on the large model's generation capabilities, acts as a knowledge graph query agent, mapping the performance reminder interaction text sequence into entity relation triples, and calculating the shortest distance of the graph path in a pre-set compliance knowledge graph, deriving the violation matching tensor of the third dimension.

[0052] The aforementioned multiple dialogue quality inspection agents, each with completely different underlying architectures, performed entirely independent latent space vector mappings and tensor operations on the same sequence of performance reminder interaction texts. Due to the heterogeneity of their underlying architectures and inference paths, the inherent systematic biases and blind spots of a single structure are fundamentally eliminated. After computation, each heterogeneous agent generates an independent result structure (containing its own determined feature labels and feature confidence levels) for the same quality inspection dimension. Subsequently, these multiple heterogeneous result structures derived from different inference paths are packaged together to generate "multiple dialogue service quality inspection information" corresponding to that single quality inspection dimension.

[0053] S106: Based on the multiple dialogue service quality inspection information and the performance reminder interactive text sequence, a quality inspection arbitration agent is used to perform quality inspection arbitration processing to obtain a comprehensive dialogue service quality inspection result.

[0054] Quality inspection arbitration agent: refers to an autonomous artificial intelligence entity with a global perspective and the ability to resolve conflicts. Optionally, after collecting multi-source heterogeneous local quality inspection data and global text data through quality inspection arbitration processing, the quality inspection arbitration agent performs feature alignment, confidence comparison, weight allocation, and high-dimensional matrix fusion at the underlying level to eliminate the complex computational process of local model logical illusions or feature conflicts.

[0055] The comprehensive dialogue service quality inspection result refers to the structured evaluation data package with global uniqueness, convergence and high confidence output after the above arbitration operation, which may include global interaction quality score and risk rating parameters.

[0056] In a schematic manner, the control arbitration agent parses discrete feature labels and their corresponding floating-point confidence scores from multiple dialogue service quality inspection information into a standardized feature vector array. It then extracts global contextual parameters (e.g., calculating the dialogue rounds throughout the entire lifecycle and extracting specific macro-level transaction service type identifiers) using the complete performance reminder interaction text sequence. These environmental parameters are input into a preset dynamic weight evaluation function to calculate dynamic evaluation weights for different feature dimensions in real time (e.g., dynamically increasing the weight component of the sensitive word monitoring agent's output feature when a "hard rejection" type is identified in the text sequence). When logical conflicts are detected in the feature information output by different dialogue quality inspection agents (e.g., intent recognition is deemed benign, but sensitive word monitoring is deemed high-risk), the arbitration agent's arithmetic logic unit, based on the aforementioned dynamic evaluation weights and the underlying confidence scores of each piece of information, performs weighted fusion operations and conflict resolution logic to converge the multi-dimensional discrete features into a global quantitative indicator, generating and outputting a comprehensive dialogue service quality inspection result that can serve as the basis for final business decisions.

[0057] This specification describes a method for obtaining service fulfillment dialogue data in service fulfillment reminder scenarios and transcribing it into a sequence of interactive text. Multiple dialogue quality inspection agents are then employed to process this text sequence, generating multiple dialogue service quality inspection information. Finally, a quality inspection arbitration agent is used to arbitrate the multiple dialogue service quality inspection information and the sequence of interactive text to obtain a comprehensive dialogue service quality inspection result. This method addresses the technical problems of low efficiency, insufficient data coverage, and high false alarm rates caused by static rule matching lacking deep semantic understanding of the dialogue context, thus failing to accurately identify the deep intent of nodes. The fully automated process overcomes the data throughput bottleneck of manual processing, meeting the real-time quality inspection needs of massive concurrent data streams. Furthermore, the collaborative computation and arbitration mechanism of multiple agents enables accurate parsing of the deep semantics of high-dimensional complex interactive data and cross-fusion of multi-dimensional features, reducing the system's false alarm rate and improving the accuracy of identifying the deep intent of server and user nodes, as well as the overall efficiency of automated verification of global service quality.

[0058] Optional, please see Figure 3 , Figure 3 This is a flowchart illustrating a dialogue service quality inspection process proposed in one or more embodiments of this specification. Specifically, the process of using multiple dialogue quality inspection agents to perform dialogue service quality inspection based on the fulfillment reminder interactive text sequence, and obtaining multiple dialogue service quality inspection information, can be performed in the following manner: S202: Determine the dialogue quality inspection agent corresponding to multiple reference quality inspection dimensions of the performance reminder interactive text sequence; Reference quality inspection dimensions refer to specific evaluation aspects or business objectives used to measure the quality of service fulfillment dialogue interactions, such as user fulfillment intent dimension, customer service script violation dimension, dialogue topology connectivity dimension, etc.

[0059] S204: Based on the performance reminder interactive text sequence, each of the dialogue quality inspection agents drives the large language model to perform dialogue service quality inspection processing from the reference quality inspection dimension, thereby obtaining multiple dialogue service quality inspection information.

[0060] Schematic, during step S202, the metadata header of the performance reminder interactive text sequence is first parsed to extract the business type identifier and real-time environment variables of the current interaction. Based on the above identifier and variables, multiple reference quality inspection dimensions required for the current scenario are dynamically determined by querying the high-frequency configuration registry in memory, and corresponding multiple dialogue quality inspection agents are instantiated in the underlying memory pool accordingly. In step S204, a thread pool with dynamic scaling capability is allocated to the main control process. Multiple dialogue quality inspection agents are submitted to the thread pool as independent computation tasks to achieve concurrent execution optimization. Each awakened dialogue quality inspection agent, within its own independent thread, encapsulates the performance reminder interactive text sequence into a context tensor of a specific dimension and attaches inference instructions set for its own reference quality inspection dimensions. Subsequently, each dialogue quality inspection agent concurrently drives the underlying deployed large language model (which can be the same model instance sharing GPU memory or multiple distributed heterogeneous model nodes). After receiving input, the large language model performs forward attention calculation on the underlying high-dimensional feature vector set and outputs a feature tensor containing specific quality inspection labels and probability confidence. Each agent uses its built-in parsing operator to deserialize the feature tensor returned by the large language model, thereby obtaining the dialogue service quality inspection information for that dimension.

[0061] This specification demonstrates how multiple reference quality inspection dimensions and corresponding dialogue quality inspection agents are dynamically determined for interactive text sequences, thereby achieving decoupling of the underlying logic and precise resource allocation for complex quality inspection tasks. Furthermore, by leveraging multiple agents to concurrently drive a large language model, each independently performing deep semantic feature extraction and parsing within its own dedicated dimension, data concurrency throughput is improved, ensuring high accuracy in multi-dimensional feature parsing.

[0062] Optionally, the dialogue quality inspection intelligent agent includes an intent recognition intelligent agent, a sensitive word monitoring intelligent agent, and an effective communication recognition intelligent agent. Specifically, the execution of S204 can be performed in the following manner: Step C2: Drive the first language model through the intent recognition agent, call the service scenario knowledge graph based on the performance reminder interactive text sequence to perform intent recognition, and obtain the role intent label and the first dialogue service quality inspection information with the first confidence level; An intent recognition agent refers to an agent specifically designed to analyze the underlying semantic purpose and behavioral motivation in conversations between customer service representatives (service scheduling endpoints) and users (target response endpoints). This agent is not merely a static functional module, but rather an independent computing power scheduling node integrating environmental awareness, task planning, tool invocation, and context state memory.

[0063] The first major language model refers to a pre-trained deep neural network invoked by an intent-recognition agent as an external reasoning tool, which possesses powerful natural language understanding and latent space semantic mapping capabilities.

[0064] A service scenario knowledge graph refers to a structured relationship network pre-built in a graph database, containing standardized script nodes, transaction state transition topology logic, and preset intent categories under specific customer service fulfillment reminder services.

[0065] Role intent tags refer to the discretized transaction semantic classification identifiers of the current session fragment.

[0066] The first confidence level refers to the numerical value that represents the probability of accurately predicting the character's intention label.

[0067] The first dialogue service quality inspection information refers to the structured feature data package, including role intent labels and first confidence level, which is finally instantiated and output by the intent recognition intelligent agent.

[0068] Schematic illustration: After acquiring the current round's fulfillment reminder interaction text sequence, the intent recognition agent captures this data in real time through its underlying temporal data perception interface and locks the current dialogue context in its memory state cache. The intent recognition agent activates its internal feature extraction planning engine, actively assembling a prompt word engineering matrix for the intent recognition dimension, and issues tool call commands through a dedicated application programming interface to drive the first language model to perform forward propagation computation. During this process, the first language model, as an external tool of the agent, uses a self-attention mechanism to extract semantic vector representations of the text sequence and, based on the graph retrieval parameters issued by the intent recognition agent, calls the service scenario knowledge graph deployed externally in the graph database in real time. The first language model calculates the cosine similarity between the text semantic vector and each node and state transition path in the service scenario knowledge graph in the graph vector space. Based on the cosine similarity and similarity threshold, it filters similar graph paths and similar nodes, and performs intent recognition based on the similar graph paths and similar nodes to obtain the original tensor result of the predicted probability distribution for the intent. Then, the first language model returns the original tensor result containing the predicted probability distribution to the intent recognition agent. The intent recognition agent performs deserialization and normalization logic to reverse map the original tensor result to obtain the specific role intent label, and extracts the corresponding maximum conditional probability floating-point value as the first confidence level, and finally encapsulates and generates the first dialogue service quality inspection information.

[0069] Optionally, the following illustrates the model training process of a first-class language model: Model creation phase: Initialize and construct the neural network topology of the first large language model based on authorized general-purpose large language models (such as Qianwen Large Model, DeepSeek Large Model, GPT Large Model). For example, considering the strong long-range contextual dependencies in natural language conversation, the first large language model can also adopt an autoregressive generative deep neural network architecture based on a self-attention mechanism. Specifically, the model's input layer comprises a word embedding layer and a positional encoding layer, used to map discrete symbols of the input text into continuous, high-dimensional, semantically dense vectors X, and inject temporal positional features. Following the word embedding layer, multiple Transformer decoder layers are cascaded, each containing a multi-head causal self-attention network and a feedforward neural network, used to perform deep semantic representation extraction of the text in the inter-layer latent space. At the model's output, the network structure is designed as a dual-branch computation head: the first branch is a label prediction output layer, used to output the conditional probability distribution of the next text label to maintain the model's language generation and understanding generalization foundation; the second branch is an intent feature pooling layer and a fully connected layer, used to compress and map the contextual semantics of the entire text into a fixed-dimensional global semantic feature vector, facilitating subsequent cosine similarity alignment calculations with graph nodes in the service scenario knowledge graph based on the cosine similarity function.

[0070] Sample data acquisition phase: A large number of customer service interaction work order records, standard fulfillment reminder text pairs, and user feedback behavior data awaiting fulfillment are extracted in batches from a distributed historical database as raw training text materials. Expert-side services are then used to label and bind multi-dimensional supervision signals to these raw training text materials, thereby constructing standard model training sample pairs. Optionally, the specific supervision label data includes two core dimensions: the first dimension is the causal language model label Y. lm (i.e., the next real character prediction tag), used to maintain the model's natural language-based understanding probability distribution during pre-training or full parameter fine-tuning; the second dimension is the global role intention classification label Y. intent The label strictly follows the standard knowledge nodes and transaction state transition paths in the service scenario knowledge graph for logical alignment and classification (for example, the user's reply is accurately labeled as a standard intent category code such as a promise to fulfill a clear future node, a temporary difficulty to apply for an extension, or an unwilling malicious refusal).

[0071] Model training phase: Following a preset batch computation capacity, the constructed training sample pairs are iteratively pushed into the input layer of the first large language model. In each forward propagation, the model calculates and outputs the autoregressive text flow prediction probability matrix P based on the input symbol index sequence and the current model parameter weights. lm And the global intent classification prediction probability distribution tensor P intentThe prediction results and the supervised label data generated during the sample data annotation stage are used to calculate the first joint loss value using a preset loss function (such as the contrastive loss function, the cross-entropy loss function, or the Euclidean distance loss function).

[0072] Optionally, to simultaneously optimize the general language modeling capabilities of the large language model and the intent classification accuracy in specific service fulfillment reminder scenarios, a model incorporating the causal language model cross-entropy loss term L is constructed. lm Cross-entropy loss term L with intention classification supervision intent The joint objective function (which could be the cross-entropy loss term L of a causal language model) lm Cross-entropy loss term L with intention classification supervision intent The weighted sum (in the form of a weighted sum) is used as the preset loss function. When specifically calculating the joint loss value, the causal text autoregressive cross-entropy loss value, which constrains the model's basic natural language prediction ability, and the classification cross-entropy loss value, which constrains the model's accuracy in recognizing role intent, are calculated separately.

[0073] Causal text autoregressive cross-entropy loss value L lm The following is a calculation based on the model's predicted output (the underlying label prediction output layer calculates the conditional probability distribution of the next text label on the full vocabulary based on the currently input sequence of preceding symbols) and the supervision label data (the next real text label in the symbol index sequence corresponding to the segmented original conversation text corpus), using the cross-entropy loss function, as follows:

[0074] Among them, L lm The autoregressive cross-entropy loss value of causal text is represented; N represents the total length of the symbol index sequence corresponding to the segmented conversation text corpus; y t The next real text tag (i.e., supervised label data) corresponding to the t-th time step in the representation sequence; y <t P(y) represents the sequence of preceding symbols that have been input up to the t-th time step; t |y <t The characterization model's underlying label prediction output layer predicts the conditional probability of the true text label based on the preceding symbol sequence during forward computation, which is the autoregressive text stream prediction probability matrix P. lm .

[0075] The intent classification supervised cross-entropy loss term, Lintent, is obtained by performing a log-likelihood check between the predicted probability values ​​for each preset intent category obtained from the model's forward inference and the real intent classification supervised label data, as follows:

[0076] Among them, Lintent The cross-entropy loss value represents the intention classification supervision value, where K represents the total number of global role intention categories preset by the system; k represents the index value of the k-th index in the intention category sequence; Y intent，k P represents the true classification label value corresponding to the k-th intent category in the supervised label data. intent，k The representation model outputs the normalized predicted conditional probability value of the current text belonging to the k-th intent category in the forward inference computation.

[0077] Optionally, the first joint loss value L1 can be calculated using the following joint objective function: L1=a1*L lm +(1-a1)*L intent Where a1 represents the preset equilibrium hyperparameter; After calculating the first joint loss value L1, the automatic differentiation engine (i.e., a general gradient calculation component for backpropagation, such as PyTorch / TensorFlow Autograd) is invoked to execute the backpropagation algorithm to calculate the gradient tensor matrix of the first joint loss value L1 with respect to each synaptic weight node in the large model network. Then, a preset deep learning optimizer (e.g., the AdamW optimizer) is used to update the model weight parameters until model training ends. For example, the preset deep learning optimizer updates the weight parameter tensors of the bottom layer of the first large language model along the gradient descent direction using a dynamically decaying learning rate. The above forward propagation and backpropagation parameter update process will be continuously iterated on the full training dataset until the global joint loss scalar calculated on the validation set converges and stabilizes within a preset tolerance threshold, or until a preset maximum number of iterations is reached. At this point, the weight parameters in the network topology are frozen, completing the training of the first large language model.

[0078] Step C4: Drive the second language model through the sensitive word monitoring agent, and perform violation risk scanning based on the performance reminder interactive text sequence to obtain second dialogue service quality inspection information including sensitive risk hit information and second confidence level; Sensitive word monitoring agents refer to autonomous artificial intelligence entities (Agents) deployed at the computer quality inspection execution layer, specifically designed to identify and capture violations and unauthorized behaviors in text interactions. These agents integrate underlying control architectures such as environmental perception, task planning, model invocation, and context state memory, enabling them to autonomously transform the input text sequence stream into inference instructions targeting compliance dimensions.

[0079] The second major language model refers to the pre-trained deep neural network model called by the sensitive word monitoring agent as an external computing tool, which is used to perform high-dimensional feature extraction and risk probability calculation of text in the latent semantic space.

[0080] Violation risk scanning refers to the computational process of performing semantic mapping, compliance boundary verification, and long-range dependency analysis on text sequences during service interactions to uncover hidden non-compliant features such as illegal promises, malicious threats, or excessive language.

[0081] Sensitive risk hit information refers to discrete label data that represents what preset compliance red lines or risk categories a text sequence specifically triggered.

[0082] The second confidence level refers to the floating-point conditional probability value used to quantify the accuracy of the prediction of the sensitive risk hit information.

[0083] In a schematic manner, after acquiring the current round of compliance reminder interaction text sequence, the sensitive word monitoring agent intercepts this text data through its underlying real-time data perception interface. The sensitive word monitoring agent calls a preset compliance feature extraction planning engine in its main control kernel, autonomously assembles a system prompt word engineering matrix targeting the violation scanning dimension, and drives the second language model to perform forward propagation computation. During this process, the second language model, as an external tool, receives the sequence input and uses a multi-head self-attention mechanism to calculate the implicit associations between characters in the text along the time step dimension to identify hidden variant violation expressions that can be covered by non-regular plaintext matching. After the calculation is completed, the second language model feeds back the original feature tensor containing the probability of compliance red line triggering to the sensitive word monitoring agent. The sensitive word monitoring agent performs threshold filtering and nonlinear mapping in its state memory buffer, classifying the violation type with the highest probability as sensitive risk hit information, and using the normalized floating-point value of this highest probability as the second confidence level.

[0084] Optionally, the following illustrates the training process of a second-largest language model: Model creation phase: Initialize and construct the neural network topology of the second large language model based on the general large language model.

[0085] Optionally, the second language model in this embodiment can adopt an autoregressive generative deep neural network architecture based on a self-attention mechanism. Specifically, the input layer of this model includes a text embedding operator and a positional encoding matrix, used to map the input dialogue symbol sequence into a continuous high-dimensional latent space dense tensor. After the input layer, a feedforward network composed of multiple dual-scalar attention blocks is cascaded to strip and extract context compliance dependency features of the conversational text at different computational depths. At the output end of the model, the network topology is designed as a bidirectional parallel task computation head: the first computation head is a vocabulary probability mapping layer, used to output the conditional probability distribution of the next text tag to maintain the basic semantic understanding network of the model; the second computation head is a multi-label risk classification layer, which, by introducing a pooling operator, compresses the feature matrix of the entire text into a fixed-dimensional risk representation vector to predict the probability of hitting multiple violation red lines in parallel.

[0086] Sample data acquisition phase: A large number of historical customer service violation recordings transcribed into text, compliance dispute scripts, and highly sensitive adversarial dialogue samples are extracted in batches from the distributed risk control database as raw training corpus. Expert-side services are invoked to bind corresponding multi-dimensional supervision signals to the raw training corpus, thereby constructing standard risk training sample pairs. Optional supervision label data includes two core dimensions: the first dimension is a conventional autoregressive text prediction matrix, used as a mask supervision signal during large model training to maintain basic natural language logic; the second dimension is a multi-label violation risk matrix Y. risk The matrix performs independent multi-hot encoding labeling for M different pre-set violation risk categories (such as verbal threats, excessive promises, privacy leaks, and inducement to consume), that is, it marks the hit on the corresponding violation risk category dimension.

[0087] Model training phase: According to the preset single-batch computation capacity, training sample pairs are input into the input layer of the second large language model. In each forward propagation, the model calculates and outputs the autoregressive character prediction probability matrix under the current parameter weights and the normalized conditional probability vector P for each violation risk category based on the input symbol index sequence. risk The prediction results and the supervised label data generated during the sample data annotation stage are used to calculate the second global joint loss value using a preset loss function.

[0088] Optionally, in order to simultaneously optimize language modeling capabilities and sensitive risk recall in an imbalanced environment (i.e., when the proportion of violating samples in the global data is extremely low), a joint objective function including a causal language model cross-entropy loss term and a multi-label focusing loss term was constructed.

[0089] The cross-entropy loss term L in the causal language model lm2The cross-entropy loss function is calculated based on the model's predicted output (the conditional probability distribution of the next text tag predicted by the vocabulary probability mapping layer based on the conversation context) and the supervised label data (the predicted tag of the next real character in the input dialogue symbol sequence).

[0090] Among them, L lm2 The cross-entropy loss value represents the causal language model; n represents the total length of the index sequence of the symbol identifiers; x t The next real text tag (i.e., supervised label data) corresponding to the t-th time step in the representation sequence; x <t P(y) represents the sequence of preceding symbols that have been input up to the t-th time step; t |y <t The underlying vocabulary probability mapping layer of the representation model predicts the conditional probability of the real text tag based on the preceding symbol sequence in the forward computation, that is, the predicted probability value P. risk .

[0091] Multi-label focused loss term L focal The probability values ​​predicted by the model (the probability values ​​predicted independently by the multi-label risk classification layer for multiple different violation risk categories (such as verbal threats, over-promises, etc.) after pooling and forward matrix multiplication) and the supervision label data (the real violation risk matrix labeled with multi-label multi-hot encoding for the conversation text) are calculated by inputting them into the cross-entropy loss function.

[0092] Among them, L focal The multi-label focusing loss value is represented by M, which represents the total number of pre-defined violation risk categories, and r represents the pre-defined focusing exponent constant. Its function is to dynamically decay the loss weight of easily distinguishable samples (i.e., high-confidence samples), and guide the large language model to focus the computational power of gradient update on difficult-to-distinguish high-risk variant violation samples during backpropagation. P t,m This represents the equivalent predicted probability constructed for the m-th violation risk category. Its specific value is dynamically determined by the label values ​​of the real multi-label violation risk matrix, and its logical mapping relationship is as follows: When the label value of the m-th category in the real multi-label violation risk matrix is ​​a constant (i.e., the risk actually exists), the equivalent prediction probability P t,m The value of is equal to the normalized probability value (normalized conditional probability vector P) of the multi-label risk classification layer's prediction output for that category. risk The vector value for that category); when the label value is a constant zero (i.e., the risk does not actually exist), the equivalent prediction probability P t,mThe value of is equal to the difference between the constant 1 and the normalized probability value.

[0093] The second global joint loss value is based on the cross-entropy loss term L of the causal language model. lm2 and multi-label focused loss term L focal The weighted average is obtained.

[0094] After calculating the second global joint loss, the underlying automatic differentiation engine (i.e., a general gradient calculation component that automatically performs backpropagation based on the computational graph chain rule in related technologies) is invoked to execute the backpropagation algorithm to derive the gradient tensor matrix of the global joint loss scalar with respect to each synaptic weight node in the large model network. Then, a preset deep learning optimizer is used to update the weight parameter tensors of the second large language model's underlying structure along the gradient descent direction. This parameter update process is continuously iterated on the full training dataset until the second global joint loss value calculated on the validation set converges to a preset tolerance range. At this point, the weight parameters in the network topology are frozen, completing the training of the second large language model.

[0095] Step C6: Drive the third language model through the effective communication identification agent, perform dialogue topology analysis based on the performance reminder interactive text sequence, and obtain third dialogue service quality inspection information including effective communication score and third confidence level.

[0096] Effective communication identification agents refer to agents deployed to quantitatively evaluate the efficiency of information transmission, the connectivity of conversation logic, and the quality of transaction service loops between customer service and users.

[0097] Optionally, the agent integrates a time-series data perception interface, a task planning engine, a tool invocation operator, and a dialogue state memory cache, enabling it to autonomously transform the input text time-series stream into inference instructions targeting the communication topology dimension.

[0098] The third major language model refers to a pre-trained deep neural network that effectively communicates and identifies intelligent agents as external computing tools. It has powerful capabilities for resolving long-range contextual dependencies and extracting high-dimensional features.

[0099] Dialogue topology analysis refers to the computational process of deconstructing and mapping a linear text time sequence into a dynamic interactive graph flow (such as a finite state machine or decision topology tree) containing multiple transaction service state nodes, and quantitatively evaluating session connectivity by calculating the state transition efficiency between nodes, standard protocol path coverage, and response latency.

[0100] Effective communication score is a quantifiable scalar value that characterizes the effectiveness of communication, the completeness of information, and the quality of conversation closure between customer service and the customer to be served. Third confidence level is a conditional probability value used to quantify the predictive robustness and accuracy of the effective communication score.

[0101] In a schematic manner, after acquiring the current round of fulfillment reminder interaction text sequence, the effective communication recognition agent intercepts this data through its underlying real-time data perception interface. Within its main control kernel, the effective communication recognition agent invokes a pre-defined graph topology analysis and planning engine to autonomously assemble a system prompt word engineering matrix targeting the communication effectiveness verification dimension, and drives the underlying third-largest language model to perform forward propagation computation. During this process, the third-largest language model, acting as an external tool, receives the sequence input and utilizes a multi-head self-attention mechanism to calculate the bidirectional question-and-answer pairing relationship between the human customer service agent terminal and the user terminal awaiting fulfillment along the time step dimension. The model deconstructs the dialogue sequence into a dynamic interaction decision tree, calculates the temporal response intervals between nodes, and performs dialogue topology analysis. After the calculation is complete, the third-largest language model feeds back the original feature tensor, containing session connectivity features and node jump probabilities, to the effective communication recognition agent. Effective communication identification agents combine the historical evolution states of the current session in their state memory cache, perform feature dimensionality reduction and numerical normalization or perform compilation quantization to generate evaluation data, classify it as an effective communication score, and use the maximum probability value that characterizes the accuracy of the score as the third confidence level.

[0102] Optionally, the training process for a third-largest language model is given below: Model creation phase: The neural network topology of the third language model is initialized and constructed based on the general large language model. For example, considering that dialogue topology analysis requires in-depth analysis of the temporal alignment features of the text and the multi-turn dialogue transition logic, the third language model adopts an autoregressive generative deep neural network architecture based on a self-attention mechanism. Specifically, the model's input layer includes a text multimodal embedding operator and a high-precision timestamp encoding matrix, used to map the input time-labeled dialogue sequence into a high-dimensional latent space dense tensor containing temporal span features. Following the input layer, a semantic feature extraction layer composed of multiple bidirectional attention blocks and a temporal warping network is cascaded to deconstruct and capture the contextual temporal topological dependency features of the conversational text at different computational depths. At the model's output, the network topology is designed as a bidirectional parallel task computation head: the first computation head is a basic vocabulary probability mapping layer, used to output the conditional probability distribution of the next text tag to maintain the model's basic language logic and semantic generalization network; the second computation head is a topology regression and classification layer, which, by introducing a graph feature pooling layer (i.e., performing weighted global average pooling or max pooling operations on the hidden layer feature vectors at different time steps), compresses and maps the temporal feature matrix of the entire text into a fixed-dimensional topological representation vector. After mapping through a fully connected layer, this vector is used to predict the overall effective communication prediction score scalar S and the corresponding third confidence prediction probability C of the conversation in parallel.

[0103] Sample data acquisition phase: A large number of customer service historical call records, fulfillment reminder text sequences with strictly time-aligned timestamps, and final business closure status data are extracted in batches from the distributed interactive database as raw training corpus. Expert-side services are invoked to bind the raw training corpus with its corresponding multi-dimensional supervision signals, thereby constructing standard topology training sample pairs. Optionally, the specific supervision label data includes two core dimensions: the first dimension is a conventional autoregressive text prediction matrix, used as a mask supervision signal in the forward computation of the large model to maintain the basic natural language generation and understanding logic; the second dimension is a globally effective communication topology label, which is manually quantified based on the historical true closure effect of the corresponding session, the coverage of standard service flow nodes (such as whether identity verification, transaction service announcement, solution achievement, and other standard topology nodes were fully experienced), and the average response latency, accurately labeling the true effective communication score scalar value S corresponding to the session. real And based on the consistency probability of historical quality inspection reviews, it is assigned a corresponding confidence level label C. real .

[0104] Model training phase: According to the preset single-batch computation capacity, the training samples are input into the input layer of the third language model. In each forward propagation, the model calculates and outputs the autoregressive character prediction probability matrix under the current parameter weights based on the input symbol index sequence and temporal feature tensor, as well as the effective communication score prediction value S and the third confidence prediction value C for the conversation topology. Subsequently, the prediction results and the supervision label data generated in the sample data annotation phase are used to calculate the third global joint loss value using a preset loss calculation function.

[0105] Optionally, in order to simultaneously optimize the basic language modeling capabilities of the large language model, the absolute accuracy of the topological regression score, and the robustness of the confidence prediction, a joint objective function is constructed, which includes the causal language model cross-entropy loss term, the score mean square error loss term, and the confidence cross-entropy loss term.

[0106] The third global joint loss value is the causal language model cross-entropy loss term L. lm3 The result is calculated using the cross-entropy function of the character conditional probability distribution based on frame-by-frame prediction and the input of the next real character tag (refer to the relevant parts in the first and second large-scale language models mentioned above, the process is similar), as follows:

[0107] Among them, L lm3 The cross-entropy loss value represents the causal language model; n represents the total length of the index sequence of the symbol identifiers; Z represents the total length of the index sequence of the causal language model. tThe next real text tag (i.e., supervised label data) corresponding to the t-th time step in the representation sequence; Z <t P(Z) represents the sequence of preceding symbols that have been input up to the t-th time step; t |Z <t The vocabulary probability mapping layer at the bottom of the characterization model predicts the character conditional probability of the real text tag in the forward computation based on the preceding symbol sequence.

[0108] Score Mean Square Error Loss Term L msc The continuous floating-point predicted score scalar S, representing communication effectiveness, based on the topological regression and classification layer forward output of the large model, is compared with the true effective communication score scalar S in the supervised label data. real The mean square error function in the input correlation technique is used to calculate: L msc =(S real -S) 2 Confidence classification loss term L conf The third-confidence predicted probability C, based on the topological regression and classification layer outputs, is used to evaluate the accuracy and robustness of this scoring, and is compared with the true confidence label value C in the supervised label data. real The result is calculated using the binary cross-entropy loss function in the relevant input technology: L conf =-(C real logC + (1-C) real log(1-C)) 2 The third global joint loss value is the cross-entropy loss term L of the causal language model. lm3 Score mean square error loss term L msc And confidence classification loss term L conf The result is obtained after weighting.

[0109] Furthermore, after calculating the third global joint loss value, the underlying automatic differentiation engine (i.e., a general gradient calculation component that automatically performs backpropagation based on the chain rule of computational graphs in related technologies) is invoked to execute the backpropagation algorithm to derive the gradient tensor matrix of the third global joint loss value with respect to each synaptic weight node in the network (i.e., a multidimensional array representing the rate of change of the loss function in each parameter dimension). Then, a preset deep learning optimizer is used to update the underlying weight parameter tensor of the third language model along the direction of gradient descent. This parameter update process will continue to iterate on the full training dataset until the global joint loss scalar calculated by the model on the validation set converges to a preset convergence threshold. At this point, the network parameters are frozen, and the training of the third language model is complete.

[0110] In this specification, steps C2-C6 are used to achieve underlying decoupling and parallel scheduling of the quality inspection dimension by combining the intent recognition agent with the knowledge graph to determine intent, the sensitive word monitoring agent to scan for risks, and the effective communication recognition agent to analyze the dialogue topology. This transforms unstructured text into a multi-dimensional structured feature tensor that integrates role intent, compliance risks, and communication efficiency, and simultaneously outputs confidence indicators, eliminating the logical blind spots of single judgments and providing high-quality data for downstream quality inspection arbitration.

[0111] In one feasible implementation, the specific execution step C4, which involves driving the second language model through the sensitive word monitoring agent to perform a violation risk scan based on the performance reminder interactive text sequence, and obtaining second dialogue service quality inspection information including sensitive risk hit information and a second confidence level, can be performed in the following manner: S1002: Drive the second language model through the sensitive word monitoring agent, call the sensitive pattern matching engine to perform regular rule scanning on the performance reminder interactive text sequence, obtain the language sensitive feature scanning result, extract the interactive semantic features of the performance reminder interactive text sequence, and perform variant expression feature recognition and unauthorized resource allocation feature recognition based on the interactive semantic features to obtain the service violation feature recognition result; Sensitive pattern matching engine text rule retrieval component. This engine is built on efficient multi-pattern matching algorithms (such as the Aho-Corasick automaton algorithm and the fast RE2 regular expression engine), and can perform streaming comparison of the input text character stream with a pre-built static sensitive word library.

[0112] The language-sensitive feature scanning result refers to the structured rule hit data packet output by the sensitive pattern matching engine after performing explicit rule retrieval. It is usually represented as a feature structure, which includes the hit rule ID, the plaintext string of the hit keyword, the physical offset of the character in the original text sequence, and the preset risk-based weight components.

[0113] Variant expression feature recognition refers to the judgment of implicit violation risks that are not explicitly triggered in the text. For example, its core lies in identifying evasive violations packaged through homophonic substitution, word splitting and assembly, pinyin separation, synonym variants, malicious omission, and sarcastic tone.

[0114] Unauthorized resource allocation feature identification refers to the judgment of whether a session node arbitrarily promises core controlled resources of the system outside of its authorized boundaries. For example, its core is to identify unauthorized resource allocation features such as customer service representatives making unauthorized promises that exceed their current work order permissions.

[0115] S1004: Based on the language sensitive feature scanning results and the service violation feature identification results, generate second dialogue service quality inspection information including sensitive risk hit information and second confidence level.

[0116] As an illustration, after receiving the fulfillment reminder interactive text sequence with timestamps and node role identifiers, the sensitive word monitoring agent initiates a dual-path parallel computing pipeline, as follows: In the first pipeline: the sensitive word monitoring agent takes the sequence of interactive text for the performance reminder as the raw character stream and directly inputs it into a pre-set sensitive pattern matching engine (such as a regular expression matching component based on the Aho-Corasick automaton algorithm) through a second large language model that has not been pre-trained. This engine performs an explicit hard scan of sensitive words and outputs a language sensitive feature scan result containing the hit rule ID and the risk-based weight bias. This process accurately detects whether the text contains high-risk words (such as abusive or insulting words, etc., explicitly marked in red) that are pre-set in the violation hash dictionary through rule matching. After the scan is completed, the engine outputs the language sensitive feature scan result.

[0117] In the second parallel pipeline: the sensitive word monitoring agent drives the pre-trained second language model to perform implicit semantic soft decision-making. The second language model transforms the performance reminder interactive text sequence into a symbolic index sequence, and extracts dense latent vectors containing deep long-range dependencies of the global context through inter-layer hidden layer matrix multiplication, as interactive semantic features. The second language model utilizes its pre-trained converged multi-label risk classification layer to process the aforementioned interactive semantic features. Specifically, this multi-label risk classification layer compresses and maps the interactive semantic features into fixed-dimensional risk representation vectors through its internal feature pooling operator, and predicts in parallel during forward propagation the normalized violation conditional probability vectors for multiple different preset risk categories (e.g., latent variant expression violation risk, unauthorized resource allocation violation risk, etc.). This multi-dimensional conditional probability vector directly output by the multi-label risk classification layer constitutes the service violation feature identification result.

[0118] Furthermore, during step S1004, the sensitive word monitoring agent executes multi-source heterogeneous feature alignment logic, splicing and fusing the language sensitive feature scanning results (hard decision based on explicit rules) and the service violation feature identification results (soft decision based on implicit semantics of the second large language model) in the feature channel dimension. The agent's built-in confidence joint calibration function is used for identification: the specific function calculation process is to first retrieve the risk base weight component in the language sensitive feature scanning results as the base probability bias, then perform a weighted summation of the normalized conditional probability values ​​of each category in the service violation feature identification results, and then superimpose the weighted summation result onto the base probability bias to finally obtain the calibrated comprehensive probability distribution. Based on the comprehensive probability distribution, the risk violation enumeration value with the highest probability is defined as the sensitive risk hit information, and the maximum normalized floating-point value is used as the second confidence, completing the instantiation and encapsulation of the second dialogue service quality inspection information.

[0119] In this specification, through the steps S1002 to S1004 above, the sensitive word monitoring agent successfully utilizes a heterogeneous dual-drive architecture that combines explicit rule screening with implicit variant / unauthorized reasoning of a large model to achieve accurate perception and high-confidence labeling of complex adversarial violation statements, providing support for the subsequent global resolution by the quality inspection arbitration agent.

[0120] In one feasible implementation, the specific execution step C6, which involves driving the third language model through the effective communication identification agent and performing dialogue topology analysis based on the fulfillment reminder interactive text sequence to obtain third dialogue service quality inspection information including effective communication score and third confidence level, can be performed in the following manner: S2002: Through the effective communication identification agent, the third language model is driven to transform the performance reminder interactive text sequence into an interactive decision-making process topology diagram. The interactive decision-making process topology graph refers to a non-linear directed graph data structure controlled and dominated by an effective communication and identification agent, and constructed in the underlying memory space. In this topology graph, the vertices represent specific business state nodes or dialogue action types achieved in the session process (e.g., "identity authentication node, transaction user commitment node"); the directed edges represent the temporal jump relationships between different transaction service states; and the weight components or additional attributes of the edges are used to carry the corresponding communication metadata (e.g., timestamp difference, word length).

[0121] S2004: The third language model extracts the node information exchange frequency, standard interaction protocol node coverage, and response end feedback latency from the topology graph of the interaction decision process as topology feature vectors. Based on the topology feature vectors, the effectiveness of information transmission is quantitatively evaluated, and third dialogue service quality inspection information including effective communication score and third confidence level is generated.

[0122] Node information exchange frequency refers to the number of data interactions between the service scheduling node and the target response node on the corresponding state nodes of the topology graph within a unit time step or a preset dialogue round period. At the underlying level, it is represented as the algebraic sum of the in-degree and out-degree of a specific vertex, used to quantify the interaction density of the session flow in a core transaction service segment.

[0123] Standard interaction protocol node coverage refers to the proportion of intersection between the set of transaction service state vertices in the currently constructed interaction decision process topology graph and the set of required vertices in the preset standard compliance transaction service flowchart (standard protocol state machine). This metric is used to assess whether customer service agents strictly follow established compliance procedures to advance the conversation from the perspective of graph topology integrity.

[0124] Response latency refers to the directed algebraic difference between the timestamps of two vertices in the interactive decision-making process topology graph, from the time the service scheduling node issues a guiding state vertex (e.g., asks a question) to the time the target response node generates a corresponding response state vertex (e.g., provides an answer). This latency data is used to evaluate the user's communication cooperation and response smoothness from a temporal response dimension.

[0125] Topological feature vectors refer to low-dimensional compact numerical feature matrices generated by normalizing the nonlinear multidimensional topological indicators such as node information exchange frequency, standard interaction protocol node coverage, and response end feedback delay extracted above, and then sequentially concatenating and splicing them in the channel dimension.

[0126] The quantitative evaluation of information transmission effectiveness refers to the calculation process of scalar measurement and deterministic derivation of the overall communication effectiveness and closed-loop quality of a conversation.

[0127] Indicatively, the effective communication recognition agent continuously captures the streaming sequence of performance reminder interactive text through its real-time data perception interface. The agent's main control kernel invokes its internal graph construction planning operator to assemble a system prompt word engineering matrix extracted for the dialogue topology, actively issuing external tool invocation commands to drive the underlying third language model to perform forward propagation computation. The third language model, acting as an external computing tool, receives the sequence input, uses an attention mechanism to identify the verbal action attributes and node transition boundaries of each round of dialogue in the latent semantic space, and returns the identified state node IDs, transition directions, and corresponding absolute timestamps to the effective communication recognition agent in the form of a structured list. Furthermore, it dynamically allocates an adjacency matrix in its memory state cache, using the aforementioned state nodes as vertices and transition directions as directed edges for pointer-level assembly, thereby transforming the non-linear conversational text stream into a standard interactive decision-making process topology graph.

[0128] Further, during step S2004, the effective communication identification agent performs a traversal calculation on the data structure of the currently generated interactive decision-making process topology graph: first, it counts the number of connected edges of each vertex to calculate the frequency of node information exchange; second, it retrieves the standard interaction protocol template corresponding to the current business ID, compares the intersection cardinality of the vertex set of the current topology graph with the required vertex set of the standard template, and calculates the standard interaction protocol node coverage rate; simultaneously, it takes the timestamp difference between adjacent directed vertices and calculates the average value as the response end feedback delay. After extracting the above three-dimensional features, it performs normalization transformation and assembles them to generate a high-density topology feature vector v. The effective communication identification agent inputs the topology feature vector v into the pre-trained converged topology regression and classification layer (i.e., the information transmission effectiveness quantification evaluation operator) in the third language model to obtain the third dialogue service quality inspection information, including the effective communication score S and the third confidence level C. Optionally, the calculation process for the effectiveness quantification operator of the third language model can be referenced as follows: S=tanh(W) eval *v+b eval ) Where S represents the final effective communication score scalar value; W eval and b eval The nonlinear mapping weight matrix and bias vector, respectively, represent the converged nonlinear mappings of the pre-trained third language model; tanh is the hyperbolic tangent activation function, used to normalize and map the final score to a preset percentage range. After calculating the effective communication score, the Softmax classification head at the end of the third language model is used to perform conditional probability prediction on the numerical range of the score, and the maximum probability value is extracted as the third confidence score. The generated effective communication score and third confidence score are structured and packaged, and the corresponding third dialogue service quality inspection information is instantiated and output.

[0129] In this specification, through the detailed steps S2002 to S2004 described above, the intelligent agent successfully utilizes an autonomous control architecture that combines text sequence to graph topology mapping, graph feature vectorization extraction, and nonlinear multi-task evaluation to achieve accurate and in-depth quantification of logical integrity and transmission efficiency in unstructured conversation streams.

[0130] Optional, such as Figure 4 As shown, Figure 4 This is a schematic diagram of a quality inspection arbitration process. Specifically, based on multiple dialogue service quality inspection information and the performance reminder interactive text sequence, a quality inspection arbitration agent is used to perform quality inspection arbitration to obtain a comprehensive dialogue service quality inspection result. The following method can be used as a reference: S302: Parse the multiple dialogue service quality inspection information, and extract the sub-dimensional quality inspection information and information confidence level corresponding to each dialogue service quality inspection information; Sub-dimensional quality inspection information can be, for example, structured feature data independently calculated and output by various front-end dialogue quality inspection agents (such as intent recognition agent, sensitive word monitoring agent, and effective communication recognition agent) for a specific reference quality inspection dimension.

[0131] S304: Based on the sub-dimensional quality inspection information, the information confidence level, and the performance reminder interactive text sequence, the fourth language model is driven by the quality inspection arbitration agent to perform quality inspection arbitration scoring processing from a preset number of sub-quality inspection arbitration dimensions, determine the quality inspection arbitration scoring information corresponding to each sub-quality inspection arbitration dimension, and obtain the dimension evaluation weight corresponding to each sub-quality inspection arbitration dimension. The global interaction quality information is obtained by the quality inspection arbitration agent based on each quality inspection arbitration scoring information and the dimension evaluation weight. The quality inspection arbitration agent refers to an agent with global context awareness and conflict resolution computing capabilities. After receiving local sub-dimensional quality inspection information and information confidence from multiple preceding dialogue quality inspection agents, this agent combines the performance reminder interactive text sequence with the fourth language model to perform quality inspection arbitration scoring processing from a preset number of sub-quality inspection arbitration dimensions. It determines the quality inspection arbitration scoring information corresponding to each sub-quality inspection arbitration dimension, and performs verification fusion and nonlinear resolution of the above heterogeneous feature data using a weighted matrix. This effectively eliminates logical false positives caused by the local blind spots of the factor agents, and finally converges to output a comprehensive dialogue service quality inspection result with global determinism and high robustness.

[0132] The dimensional assessment weight refers to the relative importance coefficients that are preset or dynamically generated for different sub-quality inspection arbitration dimensions.

[0133] Sub-quality inspection arbitration dimensions refer to multiple evaluation aspects set during the global quality inspection arbitration stage to comprehensively assess the overall quality of the performance reminder session. In this embodiment, the preset number of sub-quality inspection arbitration dimensions typically includes four macro dimensions: compliance dimension, communication effectiveness dimension, user intent response dimension, and service performance achievement dimension.

[0134] Global interaction quality information refers to the comprehensive service quality information that represents the entire lifecycle of the entire session, obtained after verification and fusion calculation. This may include: comprehensive quality inspection score, problem classification, comprehensive confidence level, and whether manual review is required.

[0135] Indicatively, the quality inspection arbitration agent reads the performance reminder interactive text sequence generated in the pre-processing steps in real time through its underlying memory bus interface. It simultaneously acquires sub-dimensional quality inspection information (such as specific intent tags, sensitive word hit enumeration values, etc.) and the information confidence scores corresponding to each piece of information from multiple heterogeneous agents. It serializes all the aforementioned heterogeneous data and concatenates them according to a preset arbitration prompt word template. The concatenation process maps the long-range text (performance reminder interactive text sequence) into a global context tensor, maps the discrete sub-dimensional quality inspection information and information confidence scores into numerical metadata vectors, and the composite data packet formed by concatenating the global context tensor and the numerical metadata vectors is used as a unified input tensor and input into the fourth language model.

[0136] The fourth language model is driven to perform forward propagation computation. After receiving the input tensor, the fourth language model uses its internal multi-layer Transformer decoder and cross-attention convergence network to deeply analyze the logical dependencies and potential conflicts between the previous multi-party local quality inspection conclusions and the global original interactive text in the latent semantic space to extract deep features. After deep feature extraction, multiple regression and classification layers at the end of the fourth language model force the deep features to be mapped to a preset number of sub-quality inspection arbitration dimensions. For each sub-quality inspection arbitration dimension, the fourth language model forward computes and outputs a continuous normalized floating-point value. The numerical sequence composed of these normalized floating-point values ​​is the quality inspection arbitration score information corresponding to each sub-quality inspection arbitration dimension. At the same time, the fourth language model also outputs the global confidence (coefficient) corresponding to all quality inspection arbitration score information.

[0137] After determining the quality inspection arbitration scoring information, the quality inspection arbitration agent obtains the dimension evaluation weights used to characterize the relative importance of each dimension. In a first optional lightweight implementation, the dimension evaluation weights are preset static scalar arrays stored in the system configuration registry (e.g., compliance, communication effectiveness, user intent response, and service fulfillment each account for 25%). In a second optional dynamic adaptive implementation, the quality inspection arbitration agent can further drive the fourth major language model (or call an independent dynamic weight adjustment network), taking the current runtime macro-environmental variables (such as real-time regulatory severity level and overall customer complaint trends) as input conditions, and directly predicting and outputting a set of dynamic weight vectors with a constant sum of one through the Softmax weight regression head at the end of the model. This dynamic weight vector serves as the dimension evaluation weight corresponding to each of the sub-quality inspection arbitration dimensions.

[0138] The quality inspection arbitration agent performs verification and fusion based on the quality inspection arbitration scores and the dimension evaluation weights to comprehensively obtain global interaction quality information. The quality inspection arbitration agent uses the extracted quality inspection arbitration scores as basic scalars and performs tensor weighted dot product operations with the corresponding dimension evaluation weights to obtain the comprehensive quality inspection score.

[0139] The quality inspection arbitration agent triggers parallel feature deconstruction and label mapping of quality inspection arbitration scoring information. The quality inspection arbitration scores of each sub-dimension are compared with preset safety thresholds. When the quality inspection arbitration score of a certain sub-dimension is determined to be lower than the safety threshold, the violation enumeration label corresponding to the quality inspection arbitration dimension in which the quality inspection arbitration score of that sub-dimension is located is written into the problem classification set. At the same time, the global confidence coefficient is used to determine the interval between the built-in confidence level interval threshold, and the corresponding confidence level is established according to the target interval hit. Further retrieve the contents of the comprehensive quality inspection score, confidence level, and problem classification set. When at least one of the following conditions is met—the comprehensive quality inspection score is below the passing baseline, the confidence level is low, and the problem classification set is not empty—the review control signal is set to true to confirm that manual review is required; otherwise, it is set to false. Finally, the generated comprehensive quality inspection score, problem classification set, confidence level, and manual review control signal are cascaded and assembled into global interactive quality information.

[0140] Optionally, the training process for a fourth major language model is given below: Model creation phase: The neural network topology of the fourth language model is initialized and constructed based on the general large language model. For example, considering that the model needs to process unstructured long-range text sequences and structured heterogeneous quality inspection metadata (i.e., sub-dimensional quality inspection information and information confidence corresponding to the quality inspection information of each dialogue service), the fourth language model adopts an autoregressive generative deep neural network based on a self-attention mechanism as the basic skeleton, and superimposes and fuses a cross-attention convergence network on the basic skeleton.

[0141] For example, the input of this fourth language model is configured with a heterogeneous feature alignment embedding layer, which maps the text tag index sequence and the discretized sub-dimensional quality inspection feature vectors to a unified high-dimensional dense semantic space. After a multi-layer causal self-attention layer, a cascaded cross-attention convergence network is deployed to use the global context tensor of the session as the query matrix Query, and the quality inspection feature tensors of each sub-dimensional as the key matrix Key and value matrix Value to perform non-linear mapping interaction. At the output of the model, the network topology is designed as a bidirectional parallel driving computation head: the first computation head is a conventional basic language model mask tag prediction head, which is used to output the conditional probability distribution of the next text tag to maintain the general semantic generalization and contextual reasoning ability of the underlying layer; the second computation head is a multi-dimensional scoring regression classification layer, which is used to directly output continuous floating-point quality inspection arbitration score information and global confidence (coefficient) for a preset number of sub-quality inspection arbitration dimensions (i.e., compliance, communication effectiveness, user intent response, and service fulfillment).

[0142] Sample data acquisition phase: Massive amounts of historical reminder interaction text sequences, historical sub-dimensional quality inspection information and corresponding information confidence levels calculated by each pre-processor agent for these text sequences, and historical decision-making behavior trajectories of the human expert group during final adjudication conflict resolution are extracted in batches from the historical risk control archive database and the human expert arbitration log database as raw training sample materials. The expert-side service is then used to label the raw training sample materials and bind them to dual supervision signals strictly aligned with their time sequence, thereby constructing standard arbitration fine-tuning training sample pairs. The supervision label data includes two core dimensions: The first dimension is the causal language model label: that is, the identifier index TokenID of the next real character in the text sequence, which is used as a mask supervision signal in the forward propagation of the large model to constrain and maintain the basic natural language logic and generalization reasoning foundation of the model.

[0143] The second dimension is the optimal expert arbitration weight distribution label: this label is a set of normalized target probability vectors W. e Based on the historical human expert group's comprehensive review of multiple reference quality inspection dimensions such as compliance, communication effectiveness, user intent response, and service performance results during the final review, the actual expert score scalar array is given.

[0144] Model training process: According to the preset batch computation capacity, the training sample pairs are iteratively pushed into the input layer of the fourth language model. In each forward propagation computation, the model performs depthwise cross-attention mapping within the latent space of the internal neural network based on the input symbol identifier index sequence and metadata feature tensor. Simultaneously, it outputs the autoregressive character prediction probability distribution under the current network parameter weights and the predicted quality inspection arbitration score vector S for a preset quantity of quality inspection arbitration dimensions. The aforementioned prediction results and the supervision label data generated during the sample data labeling stage are used to calculate the fourth global joint loss value using a preset loss function (such as Euclidean distance function, contrastive loss function, etc.). Based on the fourth global joint loss value, the model parameters of the fourth language model are adjusted until the model training ends, resulting in a well-trained fourth language model.

[0145] Optionally, the joint objective function for the fourth global joint loss value includes the cross-entropy loss term L from the causal language model. lm4 The loss term L is relative to the weight distribution's entropy (i.e., KL divergence). KL : Cross-entropy loss term L in causal language model lm4 The result is calculated using the character conditional probability distribution based on frame-by-frame prediction and the next real character tag (refer to the cross-entropy loss term of the relevant causal language model in the first, second, and third large-scale language models mentioned above; the process is similar), as follows:

[0146] Among them, L lm4 The cross-entropy loss value represents the causal language model; n represents the total length of the index sequence of the symbol identifiers; D t The next real text tag (i.e., supervised label data) corresponding to the t-th time step in the representation sequence; D <t P(D) represents the sequence of preceding symbols that have been input up to the t-th time step. t |D <t The vocabulary probability mapping layer at the bottom of the characterization model predicts the character conditional probability of the real text tag in the forward computation based on the preceding symbol sequence.

[0147] Score Mean Square Error Regression Loss Term L kl The numerical deviation between the model's predicted score and the expert's actual score is measured as follows:

[0148] Where K is the total number of preset sub-quality inspection arbitration dimensions, and k is the label of the k-th dimension; S e, k S represents the true expert score scalar in the k-th dimension of the optimal expert arbitration scoring benchmark label.k The second computational head (multidimensional scoring regression classification layer) of the fourth language model predicts the quality inspection arbitration score information of the k-th dimension of the output in forward inference.

[0149] The fourth global joint loss value is obtained by applying the cross-entropy loss term L of the causal language model. lm4 Regression loss term L with mean square error of score kl We obtain the weighted sum.

[0150] After calculating the fourth global joint loss value, the underlying automatic differentiation engine (i.e., a computational component that automatically performs backpropagation based on the chain rule of the computational graph) is invoked to execute the backpropagation algorithm to derive the gradient tensor matrix of the fourth global joint loss value with respect to each synaptic weight node in the network (i.e., a multidimensional array of derivatives representing the local rate of change of the loss function in each parameter dimension). Then, using a preset deep learning optimizer and following a set learning rate decay strategy, the underlying parameter tensors of the fourth language model are updated along the direction of gradient descent. This parameter fine-tuning process is continuously iterated and looped on the full training dataset until the global joint loss scalar calculated on the independent validation set converges and stabilizes within a preset convergence threshold range. At this point, the system freezes the network parameters, completes the training of the fourth language model, and persistently stores it for use by the quality inspection and arbitration agent.

[0151] In one feasible implementation, the dimensional evaluation weights can be dynamic, and can be referenced in the following manner: 1) Information on the quality inspection environment requirements for monitoring service performance reminder scenarios; The quality inspection environment requirements include, but are not limited to, data that characterizes the external constraints of the current service, service trend information, and macro-control objectives.

[0152] This example illustrates how, through underlying hook functions or an event-driven bus, the system monitors in real-time the quality inspection environment requirements injected into currently running service fulfillment reminder scenarios. If a new service control command is issued by an external management terminal, or if the current service configuration center refreshes its global policy, this quality inspection environment requirement information is immediately captured and deserialized into a high-dimensional environment feature input vector.

[0153] 2) Based on the quality inspection environment requirements information, update the dimension evaluation weights corresponding to the quality inspection information of each sub-dimension.

[0154] In a schematic manner, the quality inspection arbitration agent invokes and activates a preset weight adjustment network. The aforementioned generated environmental feature input vector is input into the preset weight adjustment network (which can employ the aforementioned fourth language model, by constructing weight adjustment prompts for weight adjustment, and inputting the weight adjustment prompts and environmental feature input vectors into the fourth language model), triggering its hidden layer matrix to perform forward propagation multiplication and addition operations. The weight adjustment network dynamically derives the relative importance features of aspects such as compliance, communication effectiveness, user intent response, and performance reminder results in the current real-time spatiotemporal context by calculating the nonlinear spatial distance between the current environmental feature vector and the preset safety boundary vector. The output layer of the weight adjustment network combines the relative importance features to normalize the weight values ​​of each sub-dimension, obtaining new dimension evaluation weights corresponding to each sub-dimension of quality inspection information. S306: Generate a comprehensive dialogue service quality inspection result based on the global interaction quality information.

[0155] The comprehensive dialogue service quality inspection result refers to the final structured decision data package, which is used to directly drive downstream systems to execute control instructions such as automatic archiving, high-risk early warning transfer, and customer service point deduction.

[0156] Indicatively, the global interaction quality information is deconstructed into a runtime object consisting of a comprehensive score, a problem classification list, a confidence level, and a manual review control signal. This triggers conditional branching routing logic, which performs a conditional state determination on the Boolean signal indicating whether manual review is required within the global interaction quality information. When the Boolean signal indicating whether manual review is required is true, the high-risk work order triage control program is activated. First, the routing header generation component dynamically encapsulates the current work order data into a communication header pointing to the asynchronous review message queue, and forcibly sets the work order's flow priority status bit to high priority. Then, all object data, including the comprehensive score, problem classification, and confidence level, is converted into standard binary stream data packets, thereby generating the comprehensive dialogue service quality inspection result. Finally, through the network transport layer interface, this comprehensive dialogue service quality inspection result is pushed into the directed queue of the external message middleware in an asynchronous publish-subscribe format to immediately trigger a pop-up reminder on the manual review terminal.

[0157] When the Boolean signal indicating whether manual review is required is false, the automatic archiving control program is activated. The workflow status bit of the work order is assigned a fully qualified archiving tag, and the order is packaged using a serialization component to generate a comprehensive dialogue service quality inspection result with an automatic release signature. This result is then routed to the automated archiving service interface, completing the closed-loop control of the streaming online quality inspection.

[0158] This specification utilizes a quality inspection arbitration agent to align heterogeneous features, driving the large model to dynamically adjust weights based on environmental information and perform verification fusion, thus eliminating conflicts and false positives caused by local blind spots of sub-agents. Simultaneously, a decision-making chain including scoring, classification, confidence levels, and Boolean control for manual review is constructed, enabling intelligent interception and precise routing of abnormal work orders, reducing decision-making error rates, and improving robustness in complex environments.

[0159] The following will combine Figure 5 This manual provides a detailed description of the dialogue processing device provided. It should be noted that... Figure 5 The dialogue processing device shown is used to execute this specification. Figures 1-4 The methods of the embodiments shown are illustrated only in connection with this specification for ease of explanation. For specific technical details not disclosed, please refer to this specification. Figures 1-4 The example shown.

[0160] Please see Figure 5 This diagram illustrates the structure of the dialogue processing device described in this specification. The dialogue processing device 1 can be implemented as all or part of a device through software, hardware, or a combination of both. According to some embodiments, the dialogue processing device 1 includes a data processing module 11, a dialogue quality inspection module 12, and a quality inspection arbitration module 13, specifically used for: Data processing module 11 is used to acquire service performance dialogue data in the service performance reminder scenario, and transcribe the service performance dialogue data to obtain a performance reminder interactive text sequence; Dialogue quality inspection module 12 is used to perform dialogue service quality inspection processing using multiple dialogue quality inspection intelligent agents based on the performance reminder interactive text sequence, and obtain multiple dialogue service quality inspection information. The quality inspection arbitration module 13 is used to obtain a comprehensive dialogue service quality inspection result by using a quality inspection arbitration agent to perform quality inspection arbitration processing based on multiple dialogue service quality inspection information and the performance reminder interactive text sequence.

[0161] It should be noted that the dialogue processing device provided in the above embodiments is only illustrated by the division of the above functional modules when executing the dialogue processing method. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. In addition, the dialogue processing device and the dialogue processing method embodiments provided in the above embodiments belong to the same concept, and the implementation process is detailed in the method embodiments, which will not be repeated here.

[0162] The serial numbers in this specification are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0163] This specification also provides a computer storage medium capable of storing multiple instructions adapted to be loaded and executed by a processor as described above. Figures 1-4 The dialogue processing method described in the illustrated embodiment can be found in the following document for a detailed execution process. Figures 1-4 The specific details of the illustrated embodiments will not be elaborated here.

[0164] This specification also provides a computer program product that stores at least one instruction, said at least one instruction being loaded and executed by the processor as described above. Figures 1-4 The dialogue processing method described in the illustrated embodiment can be found in the following document for a detailed execution process. Figures 1-4 The specific details of the illustrated embodiments will not be elaborated here.

[0165] Please refer to Figure 6 This is a structural block diagram of an electronic device provided in an embodiment of this specification. The electronic device in this specification may include one or more of the following components: a processor 1010, a memory 1020, an input device 1030, an output device 1040, and a bus 1050. The processor 1010, memory 1020, input device 1030, and output device 1040 may be connected to each other via the bus 1050.

[0166] Processor 1010 may include one or more processing cores. Processor 1010 connects to various parts of the electronic device using various interfaces and lines, and performs various functions and processes data by running or executing instructions, programs, code sets, or instruction sets stored in memory 1020, and by calling data stored in memory 1020. Optionally, processor 1010 may be implemented using at least one hardware form of Digital Signal Processing (DSP), Field-Programmable Gate Array (FPGA), or Programmable Logic Array (PLA). Processor 1010 may integrate one or a combination of several of the following: Central Processing Unit (CPU), Graphics Processing Unit (GPU), and modem. The CPU primarily handles the operating system, user interface, and applications; the GPU is responsible for rendering and drawing the displayed content; and the modem handles wireless communication. It is understood that the modem may also not be integrated into processor 1010 and may be implemented separately through a communication chip.

[0167] The memory 1020 may include random access memory (RAM) or read-only memory (ROM). Optionally, the memory 1020 may include a non-transitory computer-readable medium. The memory 1020 may be used to store instructions, programs, code, code sets, or instruction sets.

[0168] The input device 1030 is used to receive input instructions or data, and includes, but is not limited to, a keyboard, mouse, camera, microphone, or touch device. The output device 1040 is used to output instructions or data, and includes, but is not limited to, a display device and a speaker. In this embodiment, the input device 1030 can be a temperature sensor for acquiring the operating temperature of the electronic device. The output device 1040 can be a speaker for outputting audio signals.

[0169] In addition, those skilled in the art will understand that the structure of the electronic device shown in the above figures does not constitute a limitation on the electronic device. The electronic device may include more or fewer components than shown, or combine certain components, or have different component arrangements. For example, the electronic device may also include radio frequency circuits, input units, sensors, audio circuits, Wireless Fidelity (WIFI) modules, power supplies, Bluetooth modules, etc., which will not be described in detail here.

[0170] In the embodiments of this specification, the executing entity for each step can be the electronic device described above. Optionally, the executing entity for each step can be the operating system of the electronic device. The operating system can be Android, iOS, or other operating systems; this specification does not limit this.

[0171] exist Figure 6 In the electronic device, the processor 1010 can be used to call a program stored in the memory 1020 and execute it to implement the dialogue processing method as described in the various method embodiments of this specification.

[0172] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory, or random access memory, etc.

[0173] It should be noted that the information (including but not limited to user device information, user personal information, etc.), data (including but not limited to data used for analysis, stored data, displayed data, etc.), and signals involved in the embodiments of this specification are all authorized by the user or fully authorized by all parties, and the collection, use, and processing of related data must comply with the relevant laws, regulations, and standards of the relevant countries and regions. For example, the features, data, and information involved in this specification were all obtained under full authorization.

[0174] The above-disclosed embodiments are merely preferred embodiments of this specification and should not be construed as limiting the scope of this specification. Therefore, any equivalent variations made in accordance with the claims of this specification shall still fall within the scope of this specification.

Claims

1. A dialogue processing method, characterized in that, The method includes: Obtain service fulfillment dialogue data in the service fulfillment reminder scenario, and transcribe the service fulfillment dialogue data to obtain a sequence of fulfillment reminder interaction text; Based on the aforementioned performance reminder interactive text sequence, multiple dialogue quality inspection agents are used to perform dialogue service quality inspection processing to obtain multiple dialogue service quality inspection information. Based on multiple dialogue service quality inspection information and the performance reminder interactive text sequence, a quality inspection arbitration agent is used to perform quality inspection arbitration processing to obtain a comprehensive dialogue service quality inspection result.

2. The method according to claim 1, characterized in that, The process of performing dialogue service quality inspection based on the performance reminder interactive text sequence uses multiple dialogue quality inspection agents to obtain multiple dialogue service quality inspection information, including: Determine the dialogue quality inspection agent corresponding to multiple reference quality inspection dimensions of the performance reminder interactive text sequence; Based on the performance reminder interactive text sequence, the large language model is driven by each dialogue quality inspection agent to perform dialogue service quality inspection processing from the reference quality inspection dimension, thereby obtaining multiple dialogue service quality inspection information.

3. The method according to claim 2, characterized in that, The dialogue quality inspection intelligent agent includes an intent recognition intelligent agent, a sensitive word monitoring intelligent agent, and an effective communication recognition intelligent agent. Based on the performance reminder interactive text sequence, each dialogue quality inspection agent drives a large language model to perform dialogue service quality inspection processing from the reference quality inspection dimension, obtaining multiple dialogue service quality inspection information, including: The intent recognition agent drives the first language model, and based on the performance reminder interactive text sequence, it calls the service scenario knowledge graph to perform intent recognition, and obtains the role intent label and the first dialogue service quality inspection information with the first confidence level. The sensitive word monitoring agent drives the second language model to perform a violation risk scan based on the performance reminder interactive text sequence, and obtains second dialogue service quality inspection information including sensitive risk hit information and second confidence level. The effective communication identification agent drives the third language model, and the dialogue topology analysis is performed based on the performance reminder interactive text sequence to obtain third dialogue service quality inspection information, including effective communication score and third confidence level.

4. The method according to claim 1, characterized in that, The comprehensive dialogue service quality inspection result is obtained by using a quality inspection arbitration agent to perform quality inspection arbitration processing based on multiple dialogue service quality inspection information and the performance reminder interaction text sequence, including: Analyze the multiple dialogue service quality inspection information, and extract the sub-dimensional quality inspection information and information confidence level corresponding to each dialogue service quality inspection information; Based on the sub-dimensional quality inspection information, the information confidence level, and the performance reminder interactive text sequence, the fourth language model is driven by the quality inspection arbitration agent to perform quality inspection arbitration scoring processing from a preset number of sub-quality inspection arbitration dimensions, determine the quality inspection arbitration scoring information corresponding to each sub-quality inspection arbitration dimension, and obtain the dimension evaluation weight corresponding to each sub-quality inspection arbitration dimension. The quality inspection arbitration agent obtains global interaction quality information by comprehensively considering the quality inspection arbitration scoring information and the dimension evaluation weights, and generates a comprehensive dialogue service quality inspection result based on the global interaction quality information.

5. The method according to claim 4, characterized in that, The method further includes: Information on quality inspection environment requirements for monitoring service performance reminder scenarios; Based on the quality inspection environment requirements variable parameters, update the dimension evaluation weights corresponding to the quality inspection information of each sub-dimension.

6. The method according to claim 3, characterized in that, The process involves using the sensitive word monitoring agent to drive a second language model, which scans for violations based on the performance reminder interaction text sequence, to obtain second dialogue service quality inspection information, including sensitive risk hit information and a second confidence level. The sensitive word monitoring agent drives the second language model, calls the sensitive pattern matching engine to perform regular rule scanning on the performance reminder interactive text sequence, obtains the language sensitive feature scanning result, extracts the interactive semantic features of the performance reminder interactive text sequence, and performs variant expression feature recognition and unauthorized resource allocation feature recognition based on the interactive semantic features to obtain the service violation feature recognition result. Based on the language-sensitive feature scanning results and the service violation feature identification results, second dialogue service quality inspection information including sensitive risk hit information and a second confidence level is generated.

7. The method according to claim 3, characterized in that, The method involves using the effective communication identification agent to drive a third language model, performing dialogue topology analysis based on the fulfillment reminder interaction text sequence, and obtaining third dialogue service quality inspection information including an effective communication score and a third confidence level, including: The effective communication identification agent drives the third language model to transform the performance reminder interactive text sequence into an interactive decision-making process topology graph. The third language model extracts node information exchange frequency, standard interaction protocol node coverage, and response end feedback latency from the topology graph of the interaction decision process as topology feature vectors. Based on the topology feature vectors, the effectiveness of information transmission is quantitatively evaluated, and third dialogue service quality inspection information including effective communication score and third confidence level is generated.

8. The method according to claim 1, characterized in that, The process of transcribing the service fulfillment dialogue data to obtain a fulfillment reminder interactive text sequence includes: By driving the acoustic base model through the data transcription agent, the speaker separation feature extraction is performed on the audio stream corresponding to the service fulfillment dialogue data to obtain the node role identifier that distinguishes the service scheduling end node and the target response end node, and the audio stream is decoupled into independent audio channels corresponding to the service scheduling end node and the target response end node. Furthermore, the independent audio channel is acoustically decoded to generate a sequence of interactive text messages for fulfilling obligations, including timestamps and node role identifiers.

9. A dialogue processing device, characterized in that, The device includes: The data processing module is used to acquire service performance dialogue data in the service performance reminder scenario, and transcribe the service performance dialogue data to obtain a performance reminder interactive text sequence; The dialogue quality inspection module is used to perform dialogue service quality inspection processing using multiple dialogue quality inspection agents based on the performance reminder interactive text sequence, and obtain multiple dialogue service quality inspection information. The quality inspection arbitration module is used to obtain a comprehensive dialogue service quality inspection result by using a quality inspection arbitration agent to perform quality inspection arbitration processing based on multiple dialogue service quality inspection information and the performance reminder interactive text sequence.

10. A computer storage medium storing a plurality of instructions adapted for loading by a processor and executing the steps of the method as claimed in any one of claims 1 to 8.

11. A computer program product storing at least one instruction, said at least one instruction being loaded by a processor and executing the steps of the method as claimed in any one of claims 1 to 8.

12. An electronic device, comprising: A processor and a memory; wherein the memory stores a computer program adapted to be loaded by the processor and to execute the steps of the method as described in any one of claims 1 to 8.

Images