Intelligent agent telephone disconnection method, device, electronic device and storage medium

Abstract

This application relates to the field of artificial intelligence and provides a method, apparatus, electronic device, and storage medium for intelligent agent telephone hang-up. The method includes: acquiring multi-turn dialogue data of the current call between the intelligent agent and the merchant; based on the multi-turn dialogue data, identifying the intelligent agent's intent, the merchant's intent, and the task completion status; determining the hang-up determination result of the current call based on the intelligent agent's intent, the merchant's intent, and the task completion status; if the hang-up determination result indicates that the current call meets potential termination conditions, initiating a delayed confirmation window, which is configured as a time window mode and / or a round-number window mode according to the application scenario; if it is determined that no new valid speech from the intelligent agent or the merchant is detected during the opening of the delayed confirmation window, determining that the current call has ended and performing a telephone hang-up operation. The intelligent agent telephone hang-up method provided in this application improves the accuracy of telephone hang-up determination through multi-dimensional analysis and a delayed confirmation mechanism.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a method, apparatus, electronic device, and storage medium for intelligent agent telephone disconnection. Background Technology

[0002] Currently, with the development of artificial intelligence technology, intelligent agents are being used more and more widely in voice interaction scenarios such as ordering food and making reservations by phone. When an intelligent agent makes a call on behalf of another person, accurately determining when to hang up is a key step in ensuring task completion and user experience.

[0003] Currently, common methods for determining whether a call will end in a voice conversation include keyword matching, fixed rule-based methods, and simple single-dimensional intent classification. However, these existing technologies have significant limitations in complex real-world dialogue scenarios: keyword matching and simple classification lack a deep understanding of the context of multi-turn conversations; fixed rule-based methods are too rigid. Especially when both parties add crucial information after the call appears to have ended, these existing methods cannot accurately identify the true state of the conversation, resulting in low accuracy in determining whether the call will end, easily leading to false or delayed hang-ups, and negatively impacting the user experience. Summary of the Invention

[0004] This application provides a method, apparatus, electronic device, and storage medium for intelligent agent telephone hanging-up, in order to solve the technical problem that existing intelligent agent telephone hanging-up methods have low accuracy in hanging-up judgment under complex application scenarios.

[0005] In a first aspect, embodiments of this application provide a method for intelligent agent to hang up a telephone call, including: Acquire multi-turn dialogue data of the current call between the agent and the merchant; Based on the multi-turn dialogue data, the agent's intent, the merchant's intent, and the task completion status are identified. Based on the agent's intent, the merchant's intent, and the task completion status, the hang-up determination result of the current call is determined. If the hang-up determination result indicates that the current call meets the potential termination condition, a delayed confirmation window is initiated. The delayed confirmation window is configured as a time window mode and / or a round window mode according to the application scenario. If no new valid message is detected from the agent or the merchant during the period the delayed confirmation window is open, the current call is determined to have ended, and the call is disconnected.

[0006] In some embodiments, identifying agent intent, merchant intent, and task completion status based on the multi-turn dialogue data, and determining the hang-up determination result of the current call based on the agent intent, the merchant intent, and the task completion status, includes: The multi-turn dialogue data is input into the dialogue state recognition model, which identifies the agent's intent, the merchant's intent, and the task completion status based on the multi-turn dialogue data. Based on the agent's intent, the merchant's intent, and the task completion status, the model outputs the hang-up determination result of the current call. The dialogue state recognition model is trained based on multi-round historical dialogue data and historical feedback data between the agent and historical merchants, as well as the corresponding hang-up judgment result labels.

[0007] In some embodiments, the dialogue state recognition model includes an agent intent recognition layer, a merchant intent recognition layer, a task completion state recognition layer, and a hang-up determination layer; the multi-turn dialogue data includes agent dialogue data and merchant dialogue data; The agent intent recognition layer is used to recognize the agent intent based on the agent dialogue data and map the agent intent into a first hang-up signal. The merchant intent recognition layer is used to recognize the merchant intent based on the merchant dialogue data and map the merchant intent into a second hang-up signal; The task completion status recognition layer is used to identify the task completion status based on the multi-turn dialogue data and map the task completion status to a third hang-up signal. The hang-up determination layer is used to obtain the final hang-up signal based on the first hang-up signal, the second hang-up signal, and the third hang-up signal, and to determine the hang-up determination result.

[0008] In some embodiments, after performing the call hanging-up operation, the method further includes: Obtain feedback data, which includes: behavioral data of the intelligent agent and the merchant after the call is disconnected; Based on the feedback data, the parameters of the dialogue state recognition model are optimized.

[0009] In some embodiments, the potential termination conditions include: both the agent's intent and the merchant's intent are to terminate the current call, and the task completion status is that the task of the current call has been completed.

[0010] In some embodiments, the time window mode is: setting a fixed time threshold, allowing the agent and the merchant to continue the conversation within the fixed time threshold, and listening for any new valid statements; the round window mode is: setting a conversation round threshold, allowing the agent and the merchant to continue the conversation during the conversation round threshold, and listening for any new valid statements.

[0011] In some embodiments, after initiating the delayed confirmation window, the method further includes: If a new valid message is detected from the agent or the merchant during the period when the delayed confirmation window is open, it is determined that the current call has not ended, the call termination process is canceled, and normal dialogue interaction continues.

[0012] In some embodiments, determining that no new valid statements from the agent or the merchant are detected during the period the delayed confirmation window is open includes: During the period when the delayed confirmation window is open, new conversation data for the current call is acquired; The new dialogue data is identified. If the new dialogue data is not silent audio, environmental noise, or a statement that only contains a preset ending intention, then it is determined that no new valid speech has been detected.

[0013] In some embodiments, the dialogue state recognition model is trained based on the following steps: Acquire multi-round historical dialogue data and historical feedback data of historical calls between the intelligent agent and historical merchants, and determine the hang-up judgment result label of the historical call; The multi-round historical dialogue data is input into the initial dialogue state recognition model to obtain the hang-up determination prediction result of the historical call output by the initial dialogue state recognition model. Based on the hang-up determination prediction result, the historical feedback data, and the hang-up determination result label, a joint loss function value is calculated. Based on the joint loss function value, the parameters of the initial dialogue state recognition model are iteratively optimized to obtain the dialogue state recognition model.

[0014] In some embodiments, calculating the comprehensive loss function value based on the hang-up determination prediction result, the historical feedback data, and the hang-up determination result label includes: Based on the hang-up determination prediction result and the historical feedback data, the first loss function value is calculated; Based on the hang-up determination prediction result and the hang-up determination result label, calculate the second loss function value; The joint loss function value is calculated based on the first loss function value and the second loss function value.

[0015] In some embodiments, the step of inputting the multi-turn historical dialogue data into an initial dialogue state recognition model to obtain the hang-up prediction result of the historical call output by the initial dialogue state recognition model includes: The multi-round historical dialogue data is input into the initial dialogue state recognition model. Based on the multi-round historical dialogue data, the initial dialogue state recognition model identifies the historical agent intent, historical merchant intent, and historical task completion status. Based on the historical agent intent, the historical merchant intent, and the historical task completion status, the model outputs the hang-up prediction result of the historical call.

[0016] Secondly, embodiments of this application provide an intelligent agent telephone hang-up device, comprising: The acquisition unit is used to acquire multi-turn dialogue data of the current call between the agent and the merchant. The identification and determination unit is used to identify the agent's intent, the merchant's intent, and the task completion status based on the multi-turn dialogue data, and to determine the hang-up determination result of the current call based on the agent's intent, the merchant's intent, and the task completion status. The delayed confirmation unit is used to start a delayed confirmation window when it is determined that the hang-up determination result is that the current call meets the potential termination condition. The delayed confirmation window is configured as a time window mode and / or a round window mode according to the application scenario. The hang-up execution unit is used to determine that the current call has ended and to perform a call hang-up operation if no new valid speech is detected from the agent or the merchant during the opening of the delayed confirmation window.

[0017] Thirdly, embodiments of this application provide an electronic device, including a processor and a memory storing a computer program, wherein the processor executes the program to implement the steps of the intelligent agent telephone hanging-up method described in the first or second aspect.

[0018] Fourthly, embodiments of this application provide a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the intelligent agent telephone hanging-up method described in the first or second aspect.

[0019] Fifthly, embodiments of this application provide a computer program product, including a computer program that, when executed by a processor, implements the steps of the intelligent agent telephone hanging-up method described in the first or second aspect.

[0020] The intelligent agent telephone hanging-up method, device, electronic device, and storage medium provided in this application acquire multi-turn dialogue data of the current call between the intelligent agent and the merchant. Based on the intelligent agent's intent, the merchant's intent, and the task completion status, a multi-dimensional comprehensive judgment is made. This allows for accurate understanding of the true state of the dialogue in conjunction with contextual information. When it is determined that the current call meets the potential termination conditions, the call is not immediately interrupted. Instead, a delayed confirmation window, which can be flexibly configured according to the application scenario, is activated. The hanging-up operation is only performed after confirming that there are no new valid statements. The introduced delayed confirmation mechanism reserves a reasonable buffer space for the seemingly ended dialogue, effectively capturing key information supplemented by both parties at any time, improving the accuracy of hanging up. This fundamentally avoids information omissions, task failures, and duplicate dialing problems caused by premature accidental hanging up in complex voice interaction scenarios. It not only effectively saves communication resources but also significantly improves the accuracy of intelligent agent voice interaction, task completion rate, and user experience. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in this application 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 some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is one of the flowcharts illustrating the intelligent agent telephone hanging-up method provided in the embodiments of this application; Figure 2 This is the second flowchart illustrating the intelligent agent telephone hanging-up method provided in the embodiments of this application; Figure 3 This is a flowchart illustrating the training process of the dialogue state recognition model provided in the embodiments of this application; Figure 4 This is a schematic diagram of the structure of the intelligent agent telephone disconnection device provided in the embodiments of this application; Figure 5 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0024] Figure 1This is one of the flowcharts illustrating the intelligent agent telephone hanging-up method provided in an embodiment of this application. (Refer to...) Figure 1 This application provides a method for intelligent agents to hang up telephone calls, which may include: Step 110: Obtain multi-turn dialogue data of the current call between the agent and the merchant.

[0025] In this context, the intelligent agent can be an AI program that makes calls to restaurants based on user requests during a phone ordering interaction. The restaurant can be a restaurant staff member answering the call or an automated customer service system. The current call refers to the ongoing voice interaction process that has not yet ended. Multi-turn dialogue data refers to the voice data or text data converted from speech generated through multiple interactions between the intelligent agent and the restaurant.

[0026] Optionally, there are several ways to obtain multi-turn dialogue data. Real-time voice streams of the agent and the merchant can be converted into text using speech recognition technology. The converted text can be stored as a dialogue history in chronological order. At the same time, role labels can be added to each round of dialogue to distinguish the content of the agent and the merchant's speech.

[0027] Step 120: Based on multi-turn dialogue data, identify the agent's intent, the merchant's intent, and the task completion status. Based on the agent's intent, the merchant's intent, and the task completion status, determine the hanging-up decision for the current call.

[0028] Among these, agent intent refers to the core needs or states expressed by the agent in its speech, such as expressing gratitude or goodbye to end the call, or inquiring or confirming to continue communication. Merchant intent refers to the service status conveyed by the merchant in its speech, such as confirming a successful booking, indicating that service is unavailable, or requesting additional information. Task completion status refers to whether the core purpose information of the current phone interaction has been collected or completed, such as whether key booking information such as time, number of people, and phone numbers has been fully confirmed. Based on the above multi-dimensional recognition results, a comprehensive evaluation is conducted to determine the hanging-up decision for the current call. The hanging-up decision result refers to the preliminary decision made after logical analysis as to whether the call can be terminated.

[0029] Step 130: If the hang-up determination result is that the current call meets the potential termination conditions, start the delayed confirmation window. The delayed confirmation window is configured as a time window mode and / or a round window mode according to the application scenario.

[0030] Among them, the potential termination condition refers to a state in which both parties appear to have completed their communication from the current perspective of dialogue semantics and task progress. Application scenarios include phone ordering, business consultation, and after-sales service.

[0031] Optionally, a buffer timer or round counter can be enabled to initiate the delayed confirmation window. The delayed confirmation window can be configured based on user input or historical experience data, or its configuration can be adaptively adjusted based on multi-round dialogue data from the current call. Configuration information includes at least the window type and threshold. The time window mode sets a fixed time period as the buffer period. The round window mode sets a certain number of dialogue rounds as the buffer period.

[0032] Step 140: If no new valid message is detected from the agent or merchant during the period when the delayed confirmation window is open, determine that the current call has ended and perform a call termination operation.

[0033] It should be noted that during the delayed confirmation window, the system continuously monitors and detects any new valid statements from agents or merchants. New valid statements refer to supplementary content with actual business significance proposed by both parties in the call, excluding meaningless background noise, simple interjections, and standard closing remarks.

[0034] Real-time speech detection and semantic analysis are performed on newly input audio streams. If no new valid speech is detected within the set time window or round window, the current call is considered to have been completely ended, and the communication interface is then controlled to hang up the call and release communication resources.

[0035] In this embodiment, by identifying the multi-turn dialogue state from multiple dimensions such as agent intent, merchant intent, and task completion status, the true state of the dialogue can be fully and accurately understood, improving the accuracy of hang-up judgment. Furthermore, the introduction of a delayed confirmation mechanism, which reserves a buffer window after the potential termination conditions are met, can effectively capture key information temporarily supplemented by both parties after the call seems to have ended. This fundamentally avoids the problems of information omission, task failure, and repeated dialing caused by premature hang-up in complex voice interaction scenarios, significantly improving the completion rate of agent telephone tasks and user experience.

[0036] In some embodiments, based on multi-turn dialogue data, the agent's intent, the merchant's intent, and the task completion status are identified. Based on the agent's intent, the merchant's intent, and the task completion status, the hang-up decision for the current call is determined, including: Multi-turn dialogue data is input into the dialogue state recognition model. Based on the multi-turn dialogue data, the dialogue state recognition model identifies the agent's intent, the merchant's intent, and the task completion status. Based on the agent's intent, the merchant's intent, and the task completion status, it outputs the hanging-up determination result of the current call. The dialogue state recognition model is trained based on multi-round historical dialogue data and historical feedback data between the agent and historical merchants, as well as the corresponding hang-up judgment result labels.

[0037] Specifically, the dialogue state recognition model can be a machine learning network model or a deep learning network model. The dialogue text, including role annotations, obtained from speech recognition, can be converted into computer-processable feature vectors or word embeddings and then input into the model. This dialogue state recognition model can use algorithms such as self-attention mechanisms or recurrent neural networks to perform deep semantic feature extraction on multi-turn dialogue data, thereby autonomously analyzing the agent's intent, the merchant's intent, and the task completion status, and ultimately outputting the result of whether the current call is terminated.

[0038] Among them, multi-turn historical dialogue data refers to the complete text records of past interactions between agents and various call center service personnel, collected during the model training phase. Historical feedback data refers to a series of subsequent behavioral facts that occurred after a historical call was hung up, such as whether the agent dialed back shortly after the call was hung up, whether the merchant proactively called back or left a message, etc. These behavioral facts can truly reflect whether the historical hang-up operation was a mistaken hang-up.

[0039] The hang-up decision label refers to a standard reference value, either labeled by experts or automatically generated based on the system's business logic, indicating whether a historical dialogue node can truly be hung up. Using a large amount of multi-round historical dialogue data as input samples for the model, and combining historical feedback data and the hang-up decision label as supervision signals, the network weight parameters within the dialogue state recognition model are continuously adjusted through optimization algorithms. This ensures that the model's output prediction results can best fit the actual business patterns and the correct decision boundaries.

[0040] In this embodiment, by introducing a dialogue state recognition model jointly trained based on rich historical multi-turn dialogue records and real historical feedback data to perform state recognition and hang-up decision, the powerful understanding of deep semantics of the context by the large model can be fully utilized, making the system's judgment on complex situations such as implicit termination or supplementary information more intelligent and accurate, and significantly improving the accuracy of hang-up determination results.

[0041] In some embodiments, the dialogue state recognition model includes an agent intent recognition layer, a merchant intent recognition layer, a task completion state recognition layer, and a hang-up determination layer; the multi-turn dialogue data includes agent dialogue data and merchant dialogue data. The agent intent recognition layer is used to recognize agent intents based on agent dialogue data and map agent intents to a first hang-up signal. The merchant intent recognition layer is used to identify merchant intent based on merchant dialogue data and map the merchant intent into a second hang-up signal; The task completion status recognition layer is used to identify the task completion status based on multi-turn dialogue data and map the task completion status to a third hang-up signal; The hang-up determination layer is used to obtain the final hang-up signal based on the first hang-up signal, the second hang-up signal, and the third hang-up signal, and to determine the hang-up determination result.

[0042] Specifically, multi-turn dialogue data can be broken down and separated after role labeling. Agent dialogue data refers to the text statements belonging to the agent in the multi-turn dialogue data. Merchant dialogue data refers to the text statements belonging to the merchant who answered the phone in the multi-turn dialogue data.

[0043] The agent intent recognition layer is a network branch within the dialogue state recognition model used for semantic classification of agent statements. It can extract features from agent dialogue data to determine whether it contains ending intents such as farewell or stop. A numerical encoding mechanism can be used to represent the agent intent. Mapped to the first hang-up signal For example, when the identified agent's intention is to end or refuse, the intention is mapped to a first hang-up signal with a value of 1; when the identified agent's intention is to inquire or confirm, the intention is mapped to a first hang-up signal with a value of 0.

[0044] when , ; when , .

[0045] The merchant intent recognition layer is a dedicated network branch in the model for analyzing the semantics of merchant responses. It can determine whether the merchant has confirmed a successful booking, indicates an inability to provide service, or requests additional information based on the merchant's dialogue data. A numerical encoding mechanism can be used to represent the merchant's intent. Mapped to the second hang-up signal For example, when the merchant's intent is identified as ending, the intent is mapped to a second hang-up signal with a value of 1; when the merchant's intent is identified as inquiring, confirming, or asking for supplementary information, it is mapped to a second hang-up signal with a value of 0.

[0046] when , ; when , .

[0047] The task completion status recognition layer is the logical module in the model responsible for global business progress analysis. It extracts key information such as reservation time, number of diners, and reservation status based on complete multi-turn dialogue data to assess whether current business information has been fully collected. A numerical encoding mechanism can be used to represent the task completion status. Mapped to the third hang-up signal For example, when the extracted key information is complete and the task completion status is identified as "task completed", it is mapped to a third hang-up signal with a value of 1. When the task is not completed or is in a blocked state, it is mapped to a third hang-up signal with a value of 0.

[0048] when , ; when , .

[0049] The hang-up determination layer is the module responsible for coordinating and outputting the conclusion. It can perform logical operations, such as multiplication, on the first hang-up signal, the second hang-up signal, and the third hang-up signal to obtain the final hang-up signal.

[0050] Optionally, the final hang-up signal is calculated using the following formula: ; in, This represents the final hang-up signal. When all three dimensions of the hang-up signal are 1, indicating that the agent's intention is to end, the merchant's intention is to end, and the task has been completed, the calculated final hang-up signal is 1. In this case, the hang-up determination result is that the current call meets the potential end condition. Conversely, if any dimension of the hang-up signal is 0, the final hang-up signal is 0. In this case, the hang-up determination result is that the current call does not meet the potential end condition.

[0051] In this embodiment, the dialogue state recognition model is decoupled and subdivided into independent recognition layers targeting agent intent, merchant intent, and overall business task state. The multi-dimensional recognition results are then mapped to standardized numerical hang-up signals for unified joint judgment. This structure makes the complex deep learning semantic understanding process structured and possesses a clear logical link. This multi-branch, independent feature extraction and comprehensive judgment structure not only ensures that information processing in each dimension does not interfere with each other, fully considering both parties' communication intentions and objective business progress, but also significantly improves the accuracy, robustness, and interpretability of the model's analysis of potential hang-up conditions.

[0052] In some embodiments, after performing the call hang-up operation, the method further includes: Obtain feedback data, which includes: behavioral data of the agent and the merchant after the call is disconnected; Based on the feedback data, optimize the parameters of the dialogue state recognition model.

[0053] Behavioral data refers to the records of actions taken by the agent and the merchant after a phone call ends, which have practical business implications. For example, behavioral data may include whether the agent redials the merchant's number within a short period, whether the agent supplements the merchant's information via SMS or other online channels, and whether the merchant proactively calls back or leaves a message to supplement information. This real-time behavioral data objectively reflects the accuracy and rationality of the previous hang-up decision. For instance, if a redial or merchant call-back occurs within a short period, it usually means that the previous call was ended too early, i.e., the model made an incorrect hang-up decision.

[0054] The behavioral data reflecting the actual hang-up effect are used as evaluation feedback signals. These feedback signals are then correlated with the original dialogue records to construct high-quality feedback samples. Subsequently, these feedback samples are re-inputted into the dialogue state recognition model. Using machine learning algorithms such as backpropagation, the weight parameters in the network nodes within the dialogue state recognition model are fine-tuned and updated, enabling the model to focus on relearning and adapting to the logical features of complex dialogues that lead to misjudgments.

[0055] In this embodiment, by automatically collecting real behavioral data of the agent and the merchant after the call is hung up, the parameters of the dialogue state recognition model are continuously optimized. This gives the model a strong self-iteration and evolution capability, enabling the system to form a complete closed-loop learning mechanism in actual business operations. This effectively strengthens the model's ability to identify and process difficult samples and edge dialogue cases that are prone to misjudgment. As a result, it can continuously correct its judgment deviation based on actual usage and steadily improve the accuracy of call hang-up judgment.

[0056] In some embodiments, potential termination conditions include: both the agent's intent and the merchant's intent are to end the current call, and the task completion status is that the task of the current call has been completed.

[0057] Specifically, "agent's intent to end the current call" means that semantic analysis identifies the agent expressing a desire to say goodbye or stop the communication during the interaction, such as the agent saying "goodbye" or "thank you," which are statements indicating the end of the call. "Merchant's intent to end the current call" means that the merchant has completed the current service response and conveyed the intention to end the communication, such as the merchant confirming the booking was successful or explicitly stating that they cannot provide service. "Task completion status" means that the core business process information carried by the current call has been extracted and verified, confirming that all necessary key node information has been obtained and processed completely. For example, in a telephone food ordering business, core elements such as the reservation time, number of diners, contact number, and final order status are all clear and complete.

[0058] In this embodiment of the application, by strictly combining the two-way subjective intentions of both parties in the call with the objective completion status of business tasks as a prerequisite for determining potential termination conditions, it is possible to filter out polite small talk or one-sided pauses in the conversation to the greatest extent, which greatly enhances the rigor and scientific nature of the hang-up determination, effectively prevents premature entry into the hang-up preparation stage, and improves the overall accuracy of the hang-up determination.

[0059] In some embodiments, the time window mode is: setting a fixed time threshold, allowing the agent and the merchant to continue the conversation within the fixed time threshold, and listening for any new valid statements; the round window mode is: setting a conversation round threshold, allowing the agent and the merchant to continue the conversation during the conversation round threshold, and listening for any new valid statements.

[0060] The time window mode refers to a buffering mechanism based on the objective passage of time. A fixed time threshold is a specific duration limit pre-configured by the system or set according to scenario characteristics, such as three or five seconds. Within this fixed time threshold, the system maintains the current communication link in an connected state, allowing the agent and the merchant to continue voice communication, while continuously collecting and monitoring the voice stream in the communication link to determine if any new valid statements have occurred.

[0061] The round-window mode refers to a buffering mechanism based on the number of back-and-forth interactions in a dialogue. The dialogue round threshold is a set limit on the number of dialogue rounds, such as one or two rounds. A single round consists of one complete exchange between the agent's voice and the merchant's response. During this threshold period, the system keeps the connection active, allowing both parties to continue the conversation and continuously monitoring whether any new, valid statements are made by either party within the defined dialogue rounds.

[0062] It should be noted that in practical applications, the time window mode or the round window mode can be used separately according to the specific business dialogue rhythm, or the two modes can be configured and used in combination.

[0063] In this embodiment, by providing two different dimensions of delayed confirmation mechanism, namely time window mode and round window mode, it can flexibly adapt to the speaking habits and voice interaction rhythm under different business scenarios. It not only leaves a reasonable buffer for both parties to supplement information when thinking or pausing briefly, but also avoids unlimited blind waiting caused by no one speaking for a long time. Thus, while ensuring the integrity of the call content to the greatest extent, it also takes into account the overall execution efficiency of the voice service and the utilization rate of communication resources.

[0064] In some embodiments, after initiating the delayed confirmation window, the method further includes: If a new valid message is detected from an agent or merchant during the period when the delayed confirmation window is open, the current call is determined not to have ended, the call termination process is canceled, and normal conversation continues.

[0065] Specifically, during the system's set time buffer or round buffer period, the audio stream input of both communicating parties is continuously monitored. Through real-time speech recognition and intent analysis, if the agent or merchant raises new statements with practical business discussion value, it is determined that a new valid statement from the agent or merchant has been detected during the delayed confirmation window, indicating that the current call has not truly ended. The underlying operation command for releasing the communication link, which is currently being prepared or queued, is immediately terminated, seamlessly restoring the dialogue state to normal business communication mode. The agent can then perform normal semantic understanding and response to the merchant's newly supplemented information.

[0066] In this embodiment of the application, by monitoring the new messages of both parties in real time during the delay buffer period and setting up corresponding circuit breakers and recovery mechanisms, it is ensured that when the dialogue has not been completely finished, the temporary supplementary information input can be responded to in a timely manner. This not only gives the agent extremely high interactive flexibility, but also completely avoids the omission of key information or the failure of the overall task due to forced premature disconnection.

[0067] In some embodiments, determining that no new valid statements from the agent or merchant are detected during the period when the delayed confirmation window is open includes: During the period when the delayed confirmation window is open, acquire new conversation data for the current call; The new dialogue data is identified. If the new dialogue data is not silent audio, environmental noise, or a statement that only contains a preset ending intention, then it is determined that no new valid speech has been detected.

[0068] Specifically, during the delay buffer countdown or round counting, the latest audio stream signal and its converted text, captured in real time by the microphone or communication interface, are used to obtain new dialogue data. Acoustic and language models are then used to determine the attributes of the new dialogue data.

[0069] Silent audio refers to audio segments lacking effective speech energy generated by human vocal cord vibration, such as a state of complete silence. Ambient noise refers to audio segments containing only background noise unrelated to the business conversation, such as the clattering of dishes in a restaurant or the sound of traffic on the street. Pre-defined closing phrases refer to pre-defined polite closing words that do not contain substantial incremental business information, such as goodbye or bye-bye.

[0070] In this embodiment, by performing fine-grained attribute identification and interference elimination on newly generated dialogue data within the delayed confirmation window, meaningless silences, background noise, and purely polite closing remarks without any incremental information can be intelligently filtered out. This ensures that the hanging-up decision mechanism will not be mistakenly interrupted or indefinitely delayed due to accidental noise or polite repetition of farewells. Thus, while ensuring the integrity of the dialogue, communication resources can be released in a timely and decisive manner, significantly improving the overall service operation efficiency and resource utilization of the intelligent agent.

[0071] Figure 2 This is a second flowchart illustrating the intelligent agent telephone hanging-up method provided in an embodiment of this application. (Refer to...) Figure 2 This application provides a method for intelligent agents to hang up telephone calls, which may include: First, dialogue data is collected. Then, the agent's intent, the merchant's intent, and the task completion status are identified in parallel or sequentially. Based on the identification results across these three dimensions, the system outputs a comprehensive judgment result. Subsequently, the system uses this judgment result to determine whether the current call meets potential termination conditions.

[0072] If the determination is negative, return to the step of continuing to collect dialogue data. If the determination is positive, start the delayed confirmation window and monitor new messages within the delayed window in real time. Then determine whether a valid message has been detected within the delayed confirmation window.

[0073] If a valid message is detected within the delayed confirmation window, the call continues, the hang-up process is canceled, and the process returns to the step of collecting dialogue data to continue normal interaction. If no valid message is detected, the hang-up decision is executed.

[0074] After the decision to hang up is executed, a background self-optimization process begins. This process includes collecting feedback data, constructing a feedback sample loss function based on the feedback data, constructing a joint loss function by combining the main task loss, optimizing the model using the joint loss function, and deploying the optimized model. Deploying the optimized model forms a complete closed-loop learning mechanism, enabling continuous improvement in the accuracy and robustness of subsequent call hang-up decisions based on actual business feedback.

[0075] Figure 3 This is a flowchart illustrating the training process of the dialogue state recognition model provided in an embodiment of this application. Figure 3 As shown, in some embodiments, the dialogue state recognition model is trained based on the following steps: Step 310: Obtain multi-round historical dialogue data and historical feedback data of historical calls between the intelligent agent and historical merchants, and determine the hang-up judgment result label of the historical call; Step 320: Input the multi-round historical dialogue data into the initial dialogue state recognition model to obtain the historical call hang-up prediction results output by the initial dialogue state recognition model; Step 330: Based on the hang-up determination prediction result, historical feedback data, and hang-up determination result label, calculate the joint loss function value. Based on the joint loss function value, iteratively optimize the parameters of the initial dialogue state recognition model to obtain the dialogue state recognition model.

[0076] Specifically, the initial dialogue state recognition model refers to a basic neural network model that has not yet been fine-tuned with specific feedback data or whose network parameters have not yet reached their optimal state. The hang-up decision label refers to the true classification judgment used as the standard answer in supervised learning. The parameters of the initial dialogue state recognition model can be iteratively optimized using backpropagation optimization algorithms such as gradient descent. The weights of the network nodes within the model are continuously updated based on the joint loss function value until the loss function value converges to an acceptable minimum range. Finally, the parameters are solidified to obtain a mature dialogue state recognition model.

[0077] In this embodiment, by synchronously introducing historical feedback data reflecting real business scenarios during the model training phase to calculate joint loss and optimize parameters, the limitation of traditional models relying solely on static labels for training is broken. This allows the model to specifically learn from real cases of accidental hang-ups, greatly enhancing the generalization ability and robustness of the final generated dialogue state recognition model for difficult samples and edge dialogue scenarios.

[0078] In some embodiments, a comprehensive loss function value is calculated based on the hang-up determination prediction result, historical feedback data, and hang-up determination result label, including: Based on the hang-up determination prediction results and historical feedback data, calculate the value of the first loss function; Based on the hang-up determination prediction result and the hang-up determination result label, calculate the value of the second loss function; The joint loss function value is calculated based on the first loss function value and the second loss function value.

[0079] Specifically, calculating the first loss function value refers to calculating the loss of the feedback samples. During this calculation, a weighted cross-entropy algorithm can be used. For difficult samples in the historical feedback data that are indicated as incorrect hang-ups, a higher weight parameter is assigned than for correctly hung-up samples when calculating the error. This forces the model to pay more attention to these error-prone key samples during training. Calculating the second loss function value refers to calculating the main task loss, which is simply evaluating the general classification bias between the model's prediction and the standard hang-up judgment label. Calculating the joint loss function value involves fusing the first and second loss function values. A hyperparameter can be introduced during fusion to control the relative importance ratio of the two loss components, and this hyperparameter can be dynamically adjusted based on the specific number of feedback samples.

[0080] Optionally, the formula for calculating the value of the first loss function is as follows: ; in, This represents the value of the first loss function. It is the amount of historical feedback data. It is the first The label for the hang-up determination result of a historical call. It is the first Predicted results for hanging up in a historical call For the first The weight of each historical call is assigned, with a higher weight for historical calls that were mistakenly hung up and a weight of 1 for historical calls that were correctly hung up.

[0081] Optionally, the formula for calculating the joint loss function value is as follows: ; in, This represents the value of the joint loss function. This represents the value of the second loss function. is a hyperparameter used to control the weights of the first and second loss function values.

[0082] Understandably, calculating the second loss function ensures the model possesses basic classification capabilities, while calculating the first loss function guides the model to better handle misjudgments in real-world applications. By dynamically adjusting the weight ratio of both and calculating the joint loss function, the model's ability to handle edge cases can be gradually improved while maintaining overall performance, resulting in more accurate hang-up judgments. Fine-tuning the model using the joint loss function and deploying the tuned model, along with continuous monitoring of its performance in real-world scenarios, allows for timely identification and adjustment of potential problems, continuously improving the accuracy of subsequent call hang-up judgments.

[0083] In some embodiments, multi-turn historical dialogue data is input into an initial dialogue state recognition model to obtain the hang-up prediction result of the historical call output by the initial dialogue state recognition model, including: The initial dialogue state recognition model is input into the multi-turn historical dialogue data. Based on the multi-turn historical dialogue data, the initial dialogue state recognition model identifies the historical agent intent, historical merchant intent, and historical task completion status. Based on the historical agent intent, historical merchant intent, and historical task completion status, it outputs the historical call hang-up prediction result.

[0084] Specifically, during the training phase, the initial dialogue state recognition model needs to execute multi-dimensional feature extraction logic. It can utilize internal branch networks to identify the historical agent's intent, the historical merchant's intent, and the objective historical task completion status contained in historical call records. Based on the comprehensive analysis of these three historical dimensions, it ultimately outputs a prediction result regarding the termination of the historical call.

[0085] The intelligent agent telephone hanging-up device provided in the embodiments of this application is described below. The intelligent agent telephone hanging-up device described below can be referred to in correspondence with the intelligent agent telephone hanging-up method described above.

[0086] Figure 4 This is a schematic diagram of the structure of the intelligent agent telephone disconnection device provided in an embodiment of this application. Figure 4 As shown, this application embodiment provides an intelligent agent telephone disconnection device 400, including: Acquisition unit 410 is used to acquire multi-turn dialogue data of the current call between the agent and the merchant; The identification and judgment unit 420 is used to identify the agent's intent, the merchant's intent, and the task completion status based on multi-turn dialogue data, and to determine the hang-up judgment result of the current call based on the agent's intent, the merchant's intent, and the task completion status. The delayed confirmation unit 430 is used to start a delayed confirmation window when the hang-up determination result is that the current call meets the potential termination condition. The delayed confirmation window is configured as a time window mode and / or a round window mode according to the application scenario. The hang-up execution unit 440 is used to determine that the current call has ended and to perform a call hang-up operation if no new valid speech from the agent or merchant is detected during the opening of the delayed confirmation window.

[0087] Optionally, based on multi-turn dialogue data, the agent's intent, the merchant's intent, and the task completion status are identified. Based on the agent's intent, the merchant's intent, and the task completion status, the decision to hang up the current call is determined, including: Multi-turn dialogue data is input into the dialogue state recognition model. Based on the multi-turn dialogue data, the dialogue state recognition model identifies the agent's intent, the merchant's intent, and the task completion status. Based on the agent's intent, the merchant's intent, and the task completion status, it outputs the hanging-up determination result of the current call. The dialogue state recognition model is trained based on multi-round historical dialogue data and historical feedback data between the agent and historical merchants, as well as the corresponding hang-up judgment result labels.

[0088] Optionally, the dialogue state recognition model includes an agent intent recognition layer, a merchant intent recognition layer, a task completion state recognition layer, and a hang-up determination layer; the multi-turn dialogue data includes agent dialogue data and merchant dialogue data; The agent intent recognition layer is used to recognize agent intents based on agent dialogue data and map agent intents to a first hang-up signal. The merchant intent recognition layer is used to identify merchant intent based on merchant dialogue data and map the merchant intent into a second hang-up signal; The task completion status recognition layer is used to identify the task completion status based on multi-turn dialogue data and map the task completion status to a third hang-up signal; The hang-up determination layer is used to obtain the final hang-up signal based on the first hang-up signal, the second hang-up signal, and the third hang-up signal, and to determine the hang-up determination result.

[0089] Optionally, the intelligent agent call termination device also includes a feedback unit, which is used to acquire feedback data after the call termination operation is performed. The feedback data includes: behavioral data of the intelligent agent and the merchant after the call is terminated; and to optimize the parameters of the dialogue state recognition model based on the feedback data.

[0090] Optionally, potential termination conditions include: both the agent's intent and the merchant's intent are to end the current call, and the task completion status is that the task of the current call has been completed.

[0091] Optionally, the time window mode sets a fixed time threshold, allowing the agent and merchant to continue the conversation within the fixed time threshold, and listens for any new valid statements; the round window mode sets a conversation round threshold, allowing the agent and merchant to continue the conversation during the conversation round threshold period, and listens for any new valid statements.

[0092] Optionally, the intelligent agent telephone disconnection device further includes a disconnection cancellation unit, which is used for: If a new valid message is detected from an agent or merchant during the period when the delayed confirmation window is open, the current call is determined not to have ended, the call termination process is canceled, and normal conversation continues.

[0093] Optionally, determine that no new valid statements from the agent or merchant are detected during the period the delayed confirmation window is open, including: During the period when the delayed confirmation window is open, acquire new conversation data for the current call; The new dialogue data is identified. If the new dialogue data is not silent audio, environmental noise, or a statement that only contains a preset ending intention, then it is determined that no new valid speech has been detected.

[0094] Optionally, the dialogue state recognition model is trained based on the following steps: Acquire multi-turn historical dialogue data and historical feedback data of historical calls between the intelligent agent and historical merchants, and determine the hang-up judgment result label of the historical call; Input multi-round historical dialogue data into the initial dialogue state recognition model to obtain the historical call hang-up prediction results output by the initial dialogue state recognition model; Based on the hang-up determination prediction results, historical feedback data, and hang-up determination result labels, the joint loss function value is calculated. Based on the joint loss function value, the parameters of the initial dialogue state recognition model are iteratively optimized to obtain the dialogue state recognition model.

[0095] Optionally, based on the hang-up determination prediction result, historical feedback data, and hang-up determination result label, a comprehensive loss function value is calculated, including: Based on the hang-up determination prediction results and historical feedback data, calculate the value of the first loss function; Based on the hang-up determination prediction result and the hang-up determination result label, calculate the value of the second loss function; The joint loss function value is calculated based on the first loss function value and the second loss function value.

[0096] Optionally, multi-turn historical dialogue data is input into the initial dialogue state recognition model to obtain the hang-up prediction results of the historical calls output by the initial dialogue state recognition model, including: The initial dialogue state recognition model is input into the multi-turn historical dialogue data. Based on the multi-turn historical dialogue data, the initial dialogue state recognition model identifies the historical agent intent, historical merchant intent, and historical task completion status. Based on the historical agent intent, historical merchant intent, and historical task completion status, it outputs the historical call hang-up prediction result.

[0097] Figure 5 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 5 As shown, the electronic device may include a processor 510, a communication interface 520, a memory 530, and a communication bus 540, wherein the processor 510, the communication interface 520, and the memory 530 communicate with each other via the communication bus 540. The processor 510 may call a computer program in the memory 530 to execute the steps of the agent's telephone hang-up method, such as: acquiring multi-turn dialogue data of the current call between the agent and the merchant; based on the multi-turn dialogue data, identifying the agent's intent, the merchant's intent, and the task completion status; based on the agent's intent, the merchant's intent, and the task completion status, determining the hang-up determination result of the current call; if the hang-up determination result indicates that the current call meets the potential termination conditions, starting a delayed confirmation window, which is configured as a time window mode and / or a round-number window mode according to the application scenario; if it is determined that no new valid speech from the agent or the merchant is detected during the opening of the delayed confirmation window, determining that the current call has ended and executing the telephone hang-up operation.

[0098] Furthermore, the logical instructions in the aforementioned memory 530 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0099] On the other hand, this application also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can perform the steps of the intelligent agent telephone hang-up method provided in the above embodiments, such as: acquiring multi-turn dialogue data of the current call between the intelligent agent and the merchant; identifying the intelligent agent's intent, the merchant's intent, and the task completion status based on the multi-turn dialogue data; determining the hang-up determination result of the current call based on the intelligent agent's intent, the merchant's intent, and the task completion status; starting a delayed confirmation window when the hang-up determination result is that the current call meets the potential termination conditions, wherein the delayed confirmation window is configured as a time window mode and / or a round number window mode according to the application scenario; and determining that the current call has ended and performing a telephone hang-up operation when it is determined that no new valid speech from the intelligent agent or the merchant is detected during the opening of the delayed confirmation window.

[0100] On the other hand, embodiments of this application also provide a processor-readable storage medium storing a computer program. The computer program is used to cause a processor to execute the steps of the intelligent agent telephone hang-up method provided in the above embodiments. For example, it includes: acquiring multi-turn dialogue data of the current call between the intelligent agent and the merchant; identifying the intelligent agent's intent, the merchant's intent, and the task completion status based on the multi-turn dialogue data; determining the hang-up determination result of the current call based on the intelligent agent's intent, the merchant's intent, and the task completion status; if the hang-up determination result indicates that the current call meets the potential termination conditions, starting a delayed confirmation window, the delayed confirmation window being configured as a time window mode and / or a round-number window mode according to the application scenario; if it is determined that no new valid speech from the intelligent agent or the merchant is detected during the opening of the delayed confirmation window, determining that the current call has ended and performing a telephone hang-up operation.

[0101] The processor-readable storage medium can be any available medium or data storage device that the processor can access, including but not limited to magnetic memory (e.g., floppy disk, hard disk, magnetic tape, magneto-optical disk (MO)), optical memory (e.g., CD, DVD, BD, HVD), and semiconductor memory (e.g., ROM, EPROM, EEPROM, non-volatile memory (NAND FLASH), solid-state drive (SSD)).

[0102] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0103] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0104] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for intelligent agents to hang up telephone calls, characterized in that, include: Acquire multi-turn dialogue data of the current call between the agent and the merchant; Based on the multi-turn dialogue data, the agent's intent, the merchant's intent, and the task completion status are identified. Based on the agent's intent, the merchant's intent, and the task completion status, the hang-up determination result of the current call is determined. If the hang-up determination result indicates that the current call meets the potential termination condition, a delayed confirmation window is initiated. The delayed confirmation window is configured as a time window mode and / or a round window mode according to the application scenario. If no new valid message is detected from the agent or the merchant during the period the delayed confirmation window is open, the current call is determined to have ended, and the call is disconnected.

2. The intelligent agent telephone hanging-up method according to claim 1, characterized in that, The step of identifying the agent's intent, the merchant's intent, and the task completion status based on the multi-turn dialogue data, and determining the hang-up decision result of the current call based on the agent's intent, the merchant's intent, and the task completion status, includes: The multi-turn dialogue data is input into the dialogue state recognition model, which identifies the agent's intent, the merchant's intent, and the task completion status based on the multi-turn dialogue data. Based on the agent's intent, the merchant's intent, and the task completion status, the model outputs the hang-up determination result of the current call. The dialogue state recognition model is trained based on multi-round historical dialogue data and historical feedback data between the agent and historical merchants, as well as the corresponding hang-up judgment result labels.

3. The intelligent agent telephone hanging-up method according to claim 2, characterized in that, The dialogue state recognition model includes an agent intent recognition layer, a merchant intent recognition layer, a task completion state recognition layer, and a hang-up determination layer; the multi-turn dialogue data includes agent dialogue data and merchant dialogue data. The agent intent recognition layer is used to recognize the agent intent based on the agent dialogue data and map the agent intent into a first hang-up signal. The merchant intent recognition layer is used to recognize the merchant intent based on the merchant dialogue data and map the merchant intent into a second hang-up signal; The task completion status recognition layer is used to identify the task completion status based on the multi-turn dialogue data and map the task completion status to a third hang-up signal. The hang-up determination layer is used to obtain the final hang-up signal based on the first hang-up signal, the second hang-up signal, and the third hang-up signal, and to determine the hang-up determination result.

4. The intelligent agent telephone hanging-up method according to claim 2, characterized in that, After the call is disconnected, the process also includes: Obtain feedback data, which includes: behavioral data of the intelligent agent and the merchant after the call is disconnected; Based on the feedback data, the parameters of the dialogue state recognition model are optimized.

5. The intelligent agent telephone hanging-up method according to claim 1, characterized in that, The potential termination conditions include: both the agent's intent and the merchant's intent are to end the current call, and the task completion status is that the task of the current call has been completed.

6. The intelligent agent telephone hanging-up method according to claim 1, characterized in that, The time window mode is as follows: a fixed time threshold is set, within which the agent and the merchant are allowed to continue the dialogue, and the system listens for any new valid statements. The round window mode is as follows: a dialogue round threshold is set, during which the agent and the merchant are allowed to continue the dialogue, and the system listens for any new valid statements.

7. The intelligent agent telephone hanging-up method according to claim 1, characterized in that, Following the startup delay confirmation window, the system also includes: If a new valid message is detected from the agent or the merchant during the period when the delayed confirmation window is open, it is determined that the current call has not ended, the call termination process is canceled, and normal dialogue interaction continues.

8. The intelligent agent telephone hanging-up method according to claim 1, characterized in that, The determination that no new valid statements from the agent or the merchant were detected during the period when the delayed confirmation window was open includes: During the period when the delayed confirmation window is open, new conversation data for the current call is acquired; The new dialogue data is identified. If the new dialogue data is not silent audio, environmental noise, or a statement that only contains a preset ending intention, then it is determined that no new valid speech has been detected.

9. The intelligent agent telephone hanging-up method according to claim 2, characterized in that, The dialogue state recognition model is trained based on the following steps: Acquire multi-round historical dialogue data and historical feedback data of historical calls between the intelligent agent and historical merchants, and determine the hang-up judgment result label of the historical call; The multi-round historical dialogue data is input into the initial dialogue state recognition model to obtain the hang-up determination prediction result of the historical call output by the initial dialogue state recognition model. Based on the hang-up determination prediction result, the historical feedback data, and the hang-up determination result label, a joint loss function value is calculated. Based on the joint loss function value, the parameters of the initial dialogue state recognition model are iteratively optimized to obtain the dialogue state recognition model.

10. The intelligent agent telephone hanging-up method according to claim 9, characterized in that, The step of calculating the comprehensive loss function value based on the hang-up determination prediction result, the historical feedback data, and the hang-up determination result label includes: Based on the hang-up determination prediction result and the historical feedback data, the first loss function value is calculated; Based on the hang-up determination prediction result and the hang-up determination result label, calculate the second loss function value; The joint loss function value is calculated based on the first loss function value and the second loss function value.

11. The intelligent agent telephone hanging-up method according to claim 9, characterized in that, The step of inputting the multi-turn historical dialogue data into the initial dialogue state recognition model to obtain the hang-up prediction result of the historical call output by the initial dialogue state recognition model includes: The multi-round historical dialogue data is input into the initial dialogue state recognition model. Based on the multi-round historical dialogue data, the initial dialogue state recognition model identifies the historical agent intent, historical merchant intent, and historical task completion status. Based on the historical agent intent, the historical merchant intent, and the historical task completion status, the model outputs the hang-up prediction result of the historical call.

12. A smart agent telephone disconnection device, characterized in that, include: The acquisition unit is used to acquire multi-turn dialogue data of the current call between the agent and the merchant. The identification and determination unit is used to identify the agent's intent, the merchant's intent, and the task completion status based on the multi-turn dialogue data, and to determine the hang-up determination result of the current call based on the agent's intent, the merchant's intent, and the task completion status. The delayed confirmation unit is used to start a delayed confirmation window when it is determined that the hang-up determination result is that the current call meets the potential termination condition. The delayed confirmation window is configured as a time window mode and / or a round window mode according to the application scenario. The hang-up execution unit is used to determine that the current call has ended and to perform a call hang-up operation if no new valid speech is detected from the agent or the merchant during the opening of the delayed confirmation window.

13. An electronic device comprising a processor and a memory storing a computer program, characterized in that, When the processor executes the computer program, it implements the intelligent agent telephone hanging-up method according to any one of claims 1 to 11.

14. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the intelligent agent telephone hanging-up method as described in any one of claims 1 to 11.

15. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the intelligent agent telephone hanging-up method according to any one of claims 1 to 11.

Images