Intelligent dialogue system, method and terminal in insurance scene

Abstract

The invention discloses an intelligent dialogue system, an intelligent dialogue method and an intelligent dialogue terminal in an insurance scene. The core of the intelligent dialogue system is that an emotion recognition model, an intention recognition model, a planning model, a verbal skill generation model and a verbal skill risk assessment model are configured. According to the method, the problem that a current single large model cannot realize efficient and safe dialogue in the insurance industry is solved through cooperation of multiple models. The dialogue system based on the hybrid model can adapt to most insurance businesses, including but not limited to question consultation, business processing, customer activation and the like.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to an intelligent dialogue system, method and terminal in an insurance scenario. Background Art

[0002] Currently, most insurance customer service systems are manual, but the insurance industry usually needs to handle a large number of consultations and transactions, so a large customer service team is required. More advanced customer service systems will use large models to communicate with customers.

[0003] Manual customer service is costly, inefficient, and difficult to provide 24-hour uninterrupted service, so manual methods can no longer meet the needs of current insurance customer service systems. If a large model is used, the insurance business is a complex field involving many professional terms and regulations. Although the large model has a lot of domain knowledge, it cannot fully understand all the details and complexities in the insurance business. The current single large model cannot communicate efficiently and securely in the insurance industry. Summary of the invention

[0004] The purpose of the present invention is to provide an intelligent dialogue system, method and terminal in an insurance scenario, aiming to solve at least one of the technical problems existing in the above-mentioned prior art. To achieve the above-mentioned purpose, the technical solution adopted by the present invention is as follows:

[0005] An intelligent dialogue system in an insurance scenario, comprising:

[0006] The recognition module is used to receive customer messages and identify them based on user profiles and historical messages to obtain customer emotions and intentions;

[0007] The planning module gives specific requirements for speech generation based on the task objectives and customer intent input by the recognition module.

[0008] The speech generation module generates speech accordingly according to the speech generation requirements sent by the planning module;

[0009] The speech risk assessment module evaluates the risk of the speech sent by the speech generation module. If the risk is low, the speech will be sent directly to the customer; if the risk is high, the speech will be sent for manual review or sent to the risk speech database;

[0010] The main control module is used to manage and schedule the identification module, planning module, speech generation module and speech risk assessment module.

[0011] As a further solution of the present invention: the identification module includes:

[0012] Emotion recognition module, which identifies customer emotions based on customer replies, user profiles, and historical messages;

[0013] The intent recognition module obtains customer intent based on customer reply messages and integrated user profiles and historical message recognition.

[0014] The present invention also provides an intelligent dialogue method in an insurance scenario, which is applied to an intelligent dialogue system in an insurance scenario, and includes the following steps:

[0015] S1, obtain the information sent by the customer;

[0016] S2, the recognition model identifies the information sent by the customer based on the user profile and historical messages to obtain customer emotions and customer intentions;

[0017] S3, sending the task goal, customer intention and customer emotion to the planning model. If the planning model determines that the customer refuses to communicate, the process proceeds to step S6 and waits for the customer's reply. Otherwise, the planning model gives specific speech generation requirements and sends them to the speech generation model, and the process proceeds to step S4.

[0018] S4, the speech generation model generates speech according to the input speech generation requirements and sends them to the speech risk assessment model;

[0019] S5, the speech risk assessment model evaluates the risk of the speech sent in; if the risk is low, the speech is sent directly to the customer and the process goes to step S6; if the risk is high, the speech is sent for manual review or sent to the risk speech database;

[0020] S6, after receiving the customer's reply, go to step S2.

[0021] As a further solution of the present invention: the manual review method described in step S5 is:

[0022] If the manual review of the script is passed, the script will be sent directly to the customer and the process goes to step S6;

[0023] If the manual review of the speech fails, the speech is returned to the speech generation model for regeneration, and the process goes to step S4.

[0024] As a further solution of the present invention: in step S5, the high-risk speech is sent to the risk speech database, and after being trained with the risk speech database, it is returned to the speech generation model for regeneration, and then enters step S4.

[0025] As a further solution of the present invention: in step S3, if the planning model determines that the customer refuses to communicate, the customer is added to the return visit list.

[0026] As a further solution of the present invention: after adding the customer to the return visit list, the customer return visit time is determined according to the customer's intention and customer emotion. If the customer expresses current inconvenience, a return visit is made again after a few hours; if the customer expresses strong resistance, a longer return visit time is set.

[0027] As a further solution of the present invention: when executing a customer return visit, a preset activation script is sent to the customer, and after the customer replies to any message, the process proceeds to step S2.

[0028] As a further solution of the present invention: the recognition model includes an emotion recognition model and an intention recognition model. The training method of the emotion recognition model and the intention recognition model is: first, unsupervised fine-tuning is used to preliminarily strengthen the model, and then supervised fine-tuning is performed using high-quality industry data. The trained intention recognition model and emotion recognition model can combine the context, accurately understand customer intentions, and judge customer emotions.

[0029] And, an intelligent dialogue terminal in an insurance scenario, comprising:

[0030] one or more processors;

[0031] a memory having one or more programs stored therein;

[0032] When the one or more programs are executed by the one or more processors, the one or more processors implement the intelligent dialogue method in any insurance scenario.

[0033] Beneficial effects of the present invention:

[0034] This invention solves the problem that the current single large model cannot communicate efficiently and safely in the insurance industry by using multiple models. The dialogue system based on the hybrid model can adapt to most insurance businesses, including but not limited to problem consultation, business processing, customer activation, etc.

[0035] The present invention benefits from the massive data of the insurance industry. It first uses unsupervised fine-tuning to preliminarily strengthen the model, and then uses high-quality data in the industry for supervised fine-tuning to train an intention recognition model and an emotion recognition model that professionally recognize customer intentions. It can combine the context to accurately understand customer intentions and judge customer emotions, thus avoiding the problem of subsequent process errors caused by misunderstandings.

[0036] The present invention independently configures a speech risk assessment model. By marking risky speech, the speech risk assessment model can evaluate all generated speech. Risky speech will be manually reviewed and released, and will be added to the risk speech library to continuously optimize the model detection capability.

[0037] Based on the intention recognition model and the emotion recognition model, the present invention determines the timing of return visits according to the customer's final intention and emotion, and customizes different activation plans for them, thereby discovering potential customers in a personalized way;

[0038] The present invention proposes a separation of model planning and execution in the final model execution process. The execution process planning is performed by the classification model, and the specific interaction is completed by the intention, speech generation and other models. The interaction process is isolated from the planning model, which greatly reduces the interference between models and improves the model capabilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] The present invention will be further described below in conjunction with the accompanying drawings.

[0040] Figure 1 This is a flow chart of an intelligent dialogue method in an insurance scenario of the present invention;

[0041] Figure 2 It is a schematic diagram of the structure of an intelligent dialogue system in an insurance scenario of the present invention. DETAILED DESCRIPTION

[0042] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0043] See also Figure 1 As shown, the first embodiment of the present invention provides an intelligent dialogue method in an insurance scenario, taking the example of a customer actively sending a message, including the following steps:

[0044] S1, obtain any information sent by the client.

[0045] S2, the recognition model identifies the information sent by the customer based on historical conversations and user profiles and historical messages in the customer information database to obtain customer emotions and customer intentions.

[0046] In this embodiment, the recognition model is divided into an emotion recognition model and an intention recognition model. When training the emotion recognition model and the intention recognition model, benefiting from the massive data of the insurance industry, the model is firstly strengthened using unsupervised fine-tuning, and then supervised fine-tuning is performed using high-quality data in the industry to train an intention recognition model that professionally identifies customer intentions and an emotion recognition model that identifies customer emotions. The recognition model can combine the context, accurately understand customer intentions, judge customer emotions, and avoid situations where misunderstandings lead to errors in subsequent processes.

[0047] Use specifically trained intent recognition models and emotion recognition models to solve the problem in human-computer interaction where a large model has difficulty understanding the customer's true intentions and emotions.

[0048] S3, the task goal, customer intention and customer emotion are sent to the planning model together, but not directly to the original customer dialogue. If the planning model determines that the customer refuses to communicate, the process proceeds to step S6, waiting for the customer's reply; if the planning model determines that the customer is willing to communicate, the planning model gives specific speech generation requirements, and sends the specific speech generation requirements to the speech generation model, and proceeds to step S4.

[0049] In this embodiment, if the planning model determines that the customer refuses to communicate, the customer is added to the return visit list, and the customer return visit time is determined according to the customer's intention and emotion. If the customer expresses current inconvenience, a return visit is made again after a few hours (2-10 hours); if the customer expresses strong resistance, a longer return visit time (24-48 hours) is set. When executing the customer return visit, the preset activation script is sent to the customer, and after waiting for the customer to reply to any message, step S2 is entered.

[0050] Provide different activation plans for different customers and discover potential customers in a personalized way.

[0051] S4, the speech generation model generates corresponding speech according to the input speech generation requirements, and sends the generated speech to the speech risk assessment model.

[0052] In this embodiment, if the current customer is willing to communicate but is in a bad mood, the planning model gives instructions to the dialogue generation model, explicitly requiring the customer to be comforted first and then the specific business to be handled; if the customer's intention is not very clear, the dialogue generation model can be instructed to further inquire to clarify the customer's intention.

[0053] S5, the speech risk assessment model assesses the risk level of the speech sent to the speech generation model; if the risk is low, the speech is sent directly to the customer and enters step S6; if the risk is high, the speech is sent for manual review or sent to the risk speech database.

[0054] In this embodiment, when manual review is adopted, if the manual review of the script is passed, the script is directly sent to the customer and the process goes to step S6;

[0055] If the manual review of the speech fails, the speech is returned to the speech generation model for regeneration, and the process goes to step S4.

[0056] When the riskier speech is sent to the risk speech database, the risk speech is trained in the risk speech database and then returned to the speech generation model for regeneration, and the process goes to step S4.

[0057] Using a specifically trained speech risk assessment model can significantly reduce the risk issues in large-model speech generation and prevent large models from outputting inappropriate information.

[0058] S6, after receiving the customer's reply, go to step S2.

[0059] The first embodiment provides an intelligent dialogue method in an insurance scenario: benefiting from the massive data of the insurance industry, the model is first initially strengthened using unsupervised fine-tuning, and then supervised fine-tuning is performed using high-quality data in the industry to train an intent recognition model and an emotion recognition model that professionally identify customer intentions. The model can accurately understand customer intentions and judge customer emotions based on the context, thus avoiding the problem of subsequent process errors caused by misunderstandings.

[0060] This method independently configures a speech risk assessment model. By marking risky speech, the speech risk assessment model can evaluate all generated speech. Risky speech will be manually reviewed and released, and will be added to the risky speech library to continuously optimize the model detection capability.

[0061] At the same time, in order to discover potential customers, based on the intention recognition model and emotion recognition model, the timing of return visits is determined according to the customer's final intention and emotion, and different activation plans are customized for them, so as to discover potential customers in a personalized way.

[0062] This method proposes a separation of model planning and execution in the final model execution process. The execution process planning is performed by the classification model, and the specific interaction is completed by the intention, speech generation and other models. The interaction process is isolated from the planning model, which greatly reduces the interference between models and improves the model capabilities.

[0063] See also Figure 2 As shown, the second embodiment of the present invention provides an intelligent dialogue system in an insurance scenario, which corresponds to the first embodiment providing an intelligent dialogue method in an insurance scenario. The system may at least include a recognition module, a planning module, a speech generation module, a speech risk assessment module and a main control module.

[0064] In this embodiment, the above-mentioned recognition module includes an emotion recognition module and an intention recognition module. The emotion recognition module is responsible for building an emotion recognition model, and identifying customer emotions based on customer messages, user profiles, and historical messages. The intention recognition module is responsible for building an intention recognition model, and identifying and obtaining customer intentions based on customer messages, user profiles, and historical messages.

[0065] The above-mentioned planning module is responsible for building a planning model, and judging whether the customer is willing to communicate based on the task objectives and the customer intentions and emotions sent by the recognition module. If the planning module determines that the customer refuses to communicate, the customer will be added to the return visit list and will be returned at an appropriate time. If the planning module determines that the customer is willing to communicate, the planning module will give specific speech generation requirements and send the specific speech generation requirements to the speech generation module.

[0066] The above-mentioned speech generation module is responsible for constructing a speech generation model and generating corresponding speech according to the specific speech generation requirements input by the planning module.

[0067] If the current customer is willing to communicate but is in a bad mood, the planning module will give instructions to the dialogue generation module, explicitly requiring the customer to be comforted first and then to handle specific business; if the customer's intention is not very clear, the dialogue generation module can be instructed to further inquire to clarify the customer's intention.

[0068] The above-mentioned speech risk assessment module is responsible for constructing a speech risk assessment model to evaluate the risk level of the speech sent into the speech generation module; if the risk is low, the speech will be sent directly to the customer; if the risk is high, the speech will be sent for manual review or sent to the risk speech database.

[0069] The main control module is used to manage and schedule the identification module, planning module, speech generation module and speech risk assessment module.

[0070] The second embodiment provides an intelligent dialogue system in an insurance scenario, which is a dialogue system that combines multiple models and uses multiple models to solve the problem that the current single large model cannot communicate efficiently and safely in the insurance industry. The dialogue system based on the hybrid model can adapt to most insurance businesses, including but not limited to problem consultation, business processing, customer activation, etc.

[0071] In addition, a third embodiment of the present invention provides an intelligent dialogue terminal in an insurance scenario, including:

[0072] one or more processors;

[0073] a memory having one or more programs stored therein;

[0074] When the one or more programs are executed by the one or more processors, the one or more processors implement an intelligent dialogue method in an insurance scenario provided in the second embodiment.

[0075] The above is a detailed description of the preferred embodiments of the present invention, which cannot be considered to limit the scope of the present invention. All equivalent changes and improvements made within the scope of the present invention should still fall within the scope of the present invention.

Claims

1. An intelligent dialogue system in an insurance scenario, characterized in that: include: The recognition module is used to receive customer messages and identify them based on user profiles and historical messages to obtain customer emotions and intentions; The planning module gives specific requirements for speech generation based on the task objectives and customer intent input by the recognition module. The speech generation module generates speech accordingly according to the speech generation requirements sent by the planning module; The speech risk assessment module evaluates the risk of the speech sent by the speech generation module; If the risk is low, the script will be sent directly to the customer; if the risk is high, the script will be sent for manual review or entered into the risk script database; The main control module is used to manage and schedule the identification module, planning module, speech generation module and speech risk assessment module.

2. The intelligent dialogue system in an insurance scenario according to claim 1, characterized in that: The identification module comprises: Emotion recognition module, which identifies customer emotions based on customer replies, user profiles, and historical messages; The intent recognition module obtains customer intent based on customer reply messages and integrated user profiles and historical message recognition.

3. An intelligent dialogue method in an insurance scenario, applied to an intelligent dialogue system in an insurance scenario as claimed in claim 1, characterized in that: The following steps are involved: S1, obtain the information sent by the customer; S2, the recognition model identifies the information sent by the customer based on the user profile and historical messages to obtain customer emotions and customer intentions; S3, sending the task goal, customer intention and customer emotion to the planning model. If the planning model determines that the customer refuses to communicate, the process proceeds to step S6 and waits for the customer's reply. Otherwise, the planning model gives specific speech generation requirements and sends them to the speech generation model, and the process proceeds to step S4. S4, the speech generation model generates speech according to the input speech generation requirements and sends them to the speech risk assessment model; S5, the speech risk assessment model assesses the risk level of the input speech; If the risk is low, the script is sent directly to the customer and the process goes to step S6; if the risk is high, the script is sent for manual review or is entered into a risk script database; S6, after receiving the customer's reply, go to step S2.

4. The intelligent dialogue method in an insurance scenario according to claim 3, characterized in that: The manual review method described in step S5 is: If the manual review of the script is passed, the script will be sent directly to the customer and the process goes to step S6; If the manual review of the speech fails, the speech is returned to the speech generation model for regeneration, and the process goes to step S4.

5. The intelligent dialogue method in an insurance scenario according to claim 3 is characterized by: In step S5, the high-risk speech is sent to the risk speech database, and after training with the risk speech database, it is returned to the speech generation model for regeneration, and then enters step S4.

6. The intelligent dialogue method in an insurance scenario according to claim 3, characterized in that: In step S3, if the planning model determines that the customer refuses to communicate, the customer is added to the return visit list.

7. The intelligent dialogue method in an insurance scenario according to claim 6, characterized in that: After adding the customer to the follow-up list, determine the time for the follow-up visit based on the customer's intention and emotion. If the customer expresses current inconvenience, then make another follow-up visit after a few hours; if the customer expresses strong resistance, then set a longer follow-up time.

8. The intelligent dialogue method in an insurance scenario according to claim 7, characterized in that: When performing a customer return visit, a preset activation script is sent to the customer. After the customer replies to any message, the process proceeds to step S2.

9. The intelligent dialogue method in an insurance scenario according to claim 3, characterized in that: The recognition model includes an emotion recognition model and an intent recognition model. The training method of the emotion recognition model and the intent recognition model is: first use unsupervised fine-tuning to preliminarily strengthen the model, and then use high-quality industry data for supervised fine-tuning. The trained intent recognition model and emotion recognition model can combine the context, accurately understand customer intentions, and judge customer emotions.

10. An intelligent dialogue terminal in an insurance scenario, characterized in that: include: one or more processors; a memory having one or more programs stored therein; When the one or more programs are executed by the one or more processors, the one or more processors implement an intelligent dialogue method in an insurance scenario as described in any one of claims 3-9.

Images