An artificial intelligence decision system based on adaptive learning

Abstract

The application provides an artificial intelligence decision system based on adaptive learning, and relates to the technical field of intelligent language processing.The artificial intelligence decision system based on adaptive learning comprises a multi-modal input preprocessing module, an adaptive intention recognition module, a multi-dimensional decision scheduling engine module, an insurance vertical field knowledge base module, a third-party business system docking module, an adaptive learning optimization module and a man-machine interaction bottom-up module.The application provides an artificial intelligence decision system based on adaptive learning, and solves the technical problems of low intention recognition accuracy, lack of full-process decision closed-loop capability, poor vertical field adaptability, insufficient adaptive iteration capability and weak compliance support of the existing insurance intelligent customer service system.

Description

Technical Field

[0001] This invention relates to the field of intelligent language processing technology, specifically to an artificial intelligence decision-making system based on adaptive learning. Background Technology

[0002] With the deepening of digital transformation in the insurance industry, intelligent customer service systems have become one of the core channels for insurance institutions to serve users. However, most intelligent customer service systems in the insurance industry are currently developed based on keyword matching, fixed script templates, and limited rule engines, resulting in many defects. For example, they lack the ability to generalize intent recognition. Most existing systems adopt fixed intent keyword matching modes, which cannot adapt to the diverse and colloquial expressions of users' natural language. The accuracy of recognizing professional terms and synonyms in the insurance vertical field is low, which easily leads to misjudgment of intent and incorrect service response. At the same time, they have the problem of poor adaptability to vertical fields. Existing general intelligent customer service systems cannot deeply adapt to the rules and processes of professional scenarios such as underwriting and claims settlement, clause interpretation, policy maintenance and claims, and health management in the insurance industry, and cannot meet the professional service needs of insurance business. They also lack adaptive learning and iteration capabilities. The intent rules, script templates, and knowledge bases of existing systems need to be updated manually on a regular basis, and cannot be automatically iterated and optimized based on user interaction data. As business scenarios expand, the maintenance costs of the system continue to rise, and the response accuracy cannot be continuously improved.

[0003] Therefore, this invention proposes an artificial intelligence decision-making system based on adaptive learning, which can be used in the insurance vertical field. It has the capabilities of accurate hierarchical intent recognition, fully automated decision scheduling, and adaptive learning iteration, while meeting industry compliance requirements and effectively solving the above-mentioned problems and difficulties. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides an artificial intelligence decision-making system based on adaptive learning. The aim is to overcome the deficiencies of existing technologies and provide an artificial intelligence decision-making system based on adaptive learning to solve the technical problems of low intent recognition accuracy, lack of closed-loop decision-making capabilities throughout the entire process, poor adaptability to vertical domains, insufficient adaptive iteration capabilities, and weak compliance support in existing intelligent insurance customer service systems.

[0005] To achieve the above objectives, the present invention provides the following technical solution: An artificial intelligence decision-making system based on adaptive learning includes a multimodal input preprocessing module, an adaptive intent recognition module, a multidimensional decision scheduling engine module, an insurance vertical domain knowledge base module, a third-party business system integration module, an adaptive learning optimization module, and a human-computer interaction fallback module. The system is characterized in that: the multimodal input preprocessing module receives multimodal input data from users, preprocesses the input data, and outputs standardized user query text. The adaptive intent recognition module is communicatively connected to the multimodal input preprocessing module and is used to perform hierarchical intent recognition on the standardized user query text based on a pre-trained language model in the insurance vertical domain, and output the matching target first-level intent and second-level intent. The multi-dimensional decision scheduling engine module is communicatively connected to the adaptive intent recognition module and has a built-in intent decision link mapping library, which is used to match the corresponding execution rule set according to the target first-level intent and second-level intent. The execution rule set includes the opening connecting words, the subsequent execution logic chain and the closing guiding words. The insurance vertical domain knowledge base module is communicatively connected to the multi-dimensional decision scheduling engine module, and is used to store structured and unstructured knowledge data in the insurance field to provide knowledge support for decision execution. The third-party business system integration module is communicatively connected to the multi-dimensional decision scheduling engine module, and is used to integrate with the third-party service system of insurance business to realize business data query and execution of business operation instructions; The adaptive learning optimization module is communicatively connected to the adaptive intent recognition module, the multi-dimensional decision scheduling engine module, and the insurance vertical domain knowledge base module, respectively, and is used to collect full user interaction data and iteratively optimize the intent recognition model, decision link rules, and knowledge base content based on the interaction data. The human-computer interaction fallback module is connected to the multi-dimensional decision scheduling engine module and the adaptive learning optimization module, respectively, and is used to transfer the user to human customer service and synchronize the full-link interaction data when the system cannot complete the automated processing or the user initiates a manual request.

[0006] Furthermore, the multimodal input preprocessing module includes a speech-to-text unit, a text noise reduction unit, an insurance domain entity extraction unit, and a semantic normalization unit; The speech-to-text unit is used to convert the user's voice input into text data; The text denoising unit is used to remove redundant characters, interjections, and invalid information from text data; The insurance domain entity extraction unit is used to identify and extract insurance domain entities from the text; The entities in the insurance field include policy numbers, insurance types, disease names, hospital names, drug names, and user identity information; The semantic normalization unit is used to normalize synonyms in the text and output standardized user query text.

[0007] Furthermore, the adaptive intent recognition module includes an intent hierarchy management unit, a semantic encoding unit, an intent confidence calculation unit, and an intent error correction unit; The intent hierarchy management unit has a built-in hierarchical intent list in the insurance field. The intent list includes primary intents and several corresponding secondary intents. The primary intents include knowledge Q&A, business inquiry, business processing, products and recommendations, medical support, and health management. The hierarchical intent list includes the following secondary intents: Knowledge Q&A: Secondary intents corresponding to primary intents include insurance industry knowledge Q&A, insurance terms knowledge Q&A, and insurance process knowledge Q&A; Business Inquiry: Secondary intents corresponding to primary intents include policy inquiry, claims record inquiry, and payment record inquiry; Business Processing: Secondary intents corresponding to primary intents include online policy maintenance, online claims, online renewal, online cancellation, general underwriting, disease-type underwriting, general claims processing, disease-related claims processing, hospital-related claims processing, drug-related claims processing, transfer to human agent, and online complaints; Product and Recommendation: Secondary intents corresponding to primary intents include personalized product recommendations and product comparison; Medical Support: Secondary intents corresponding to primary intents include initial disease analysis and guidance; Health Management: Secondary intents corresponding to primary intents include health management service recommendations. The semantic encoding unit is used to perform semantic encoding on standardized user query text based on a pre-trained language model in the insurance vertical domain, and generate semantic vectors. The intent confidence calculation unit is used to input the semantic vector into the fine-tuned intent classification model, calculate the matching confidence of each first-level intent and second-level intent, and output the candidate intent set with the highest confidence. The intent correction unit is used to perform compliance verification and logical error correction on the candidate intent set. When the highest confidence level is higher than the preset threshold, it is determined to be the target first-level intent and second-level intent. When the highest confidence level is lower than the preset threshold, intent clarification script is generated to guide the user to supplement the input information.

[0008] Furthermore, the multi-dimensional decision scheduling engine module includes a link matching unit, a knowledge base scheduling unit, a business system scheduling unit, and a service generation scheduling unit; The link matching unit is used to match the corresponding set of execution rules from the intent decision link mapping library according to the target secondary intent; The knowledge base scheduling unit is used to call the sub-knowledge base corresponding to the insurance vertical domain knowledge base module according to the execution logic chain when the target first-level intent is knowledge question and answer, and obtain the question and answer result. The business system scheduling unit is used to call the corresponding business system interface through the third-party business system docking module to complete the data query or business operation execution when the target first-level intent is business query or business processing, according to the execution logic chain. The service generation and scheduling unit is used to call the corresponding service engine according to the execution logic chain to generate personalized service results when the target primary intent is product and recommendation, medical support or health management.

[0009] Furthermore, the insurance vertical domain knowledge base module includes multiple sub-knowledge bases, namely the insurance industry knowledge sub-base, the insurance terms knowledge sub-base, the insurance process knowledge sub-base, the underwriting and claims rules sub-base, the insurance product knowledge sub-base, and the disease and health management knowledge sub-base. Each sub-base corresponds to the secondary intent under the knowledge question and answer category, as well as the knowledge support needs of the intents of business processing, product and recommendation, medical treatment support, and health management.

[0010] Furthermore, the third-party business system integration module has a built-in interface adaptation unit for integrating with the core insurance business system, policy management system, claims business system, policy maintenance business system, user information management system, health management service system, and insurance product recommendation engine system, enabling bidirectional data transmission and closed-loop execution of business instructions.

[0011] Furthermore, the adaptive learning optimization module includes an interactive data acquisition unit, a data annotation unit, a model fine-tuning unit, a rule optimization unit, a knowledge base update unit, and a full-link log retention unit; The interactive data acquisition unit is used to collect all user interaction data, including user input data, intent recognition results, decision execution process, user feedback data, data transferred to human assistant, and business processing success rate data. The data annotation unit is used to automatically annotate and manually verify the collected interactive data, and generate model fine-tuning datasets and rule optimization datasets. The model fine-tuning unit is used to periodically fine-tune the pre-trained language model and intent classification model in the insurance vertical domain based on the labeled dataset, thereby improving the accuracy of intent recognition. The rule optimization unit is used to optimize the set of execution rules in the intent decision link mapping library based on the labeled dataset, thereby improving the matching degree and success rate of decision execution. The knowledge base update unit is used to update the content of the insurance vertical domain knowledge base based on the new knowledge requirements in the interactive data. The full-link log retention unit is used to retain full-link traceable logs of all algorithm decisions in the system, meeting the compliance requirements for algorithm filing in the insurance industry.

[0012] Furthermore, the triggering conditions for the human-computer interaction fallback module include: the user explicitly initiates a request to transfer to a human agent, the highest confidence level of intent recognition is lower than a preset minimum threshold, the decision-making process fails, and the user is dissatisfied with the system's response. When the human-computer interaction fallback module transfers the user to a human customer service representative, it simultaneously synchronizes the current user's full-link interaction data, user profile data, and executed decision-making processes to the human agent's terminal. After the human service is completed, it synchronizes the human service interaction data to the adaptive learning optimization module.

[0013] This invention provides an artificial intelligence decision-making system based on adaptive learning. It has the following beneficial effects: 1. This invention provides an artificial intelligence decision-making system based on adaptive learning, which has accurate intent recognition capabilities in the insurance vertical domain. Based on a pre-trained language model in the insurance vertical domain, a hierarchical insurance business intent system is constructed. Through semantic encoding and confidence calculation, accurate intent recognition of user natural language input is achieved, which greatly improves the generalization recognition capability of colloquial and diverse expressions and reduces the intent misjudgment rate.

[0014] 2. This invention provides an AI decision-making system based on adaptive learning, which has the capability of decision-making closed loop across all scenarios and processes. Through a multi-dimensional decision scheduling engine, it constructs a mapping system between intent and decision-making links, realizing automated decision scheduling across all insurance scenarios, from knowledge Q&A, business inquiry, business processing to product recommendation, medical support, and health management. It completes the entire closed loop from user intent recognition to business execution, significantly improving the service automation rate and reducing manual operation costs.

[0015] 3. This invention provides an artificial intelligence decision-making system based on adaptive learning. This system is deeply adapted to the vertical insurance industry. It has built a dedicated vertical domain knowledge base and business system interface for professional scenarios in the insurance industry. It is deeply adapted to professional business processes such as underwriting and claims settlement, policy interpretation, policy maintenance and claims settlement, and health management, so as to meet the professional service needs of the insurance industry. Attached Figure Description

[0016] Figure 1 This is a system flowchart of the AI ​​decision-making system based on adaptive learning according to the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Example 1: like Figure 1 As shown in the figure, this embodiment of the invention provides an artificial intelligence decision-making system based on adaptive learning. The specific steps of the system are as follows: Step 1: Input preprocessing; Receive multimodal input data from users, preprocess the input data, and output standardized user query text; Step 2: Hierarchical Intent Recognition; Input standardized user query text into a pre-trained language model in the insurance vertical domain, perform semantic encoding, and then input it into an intent classification model to calculate the matching confidence of each level of intent and determine the candidate intent set; Step 3: Validation of intent; Verify the candidate intent set. If the highest confidence level is higher than the preset threshold, determine the target primary intent and secondary intent, and proceed to Step 4. If the highest confidence level is lower than the preset threshold, generate and output intent clarification script to guide the user to supplement information, and return to Step 1. If it is detected that the user has clearly requested to transfer to a human agent, proceed directly to Step 7. Step 4: Decision link matching; Based on the determined primary and secondary intentions of the target, match the corresponding set of execution rules from the intention decision link mapping library. The set of execution rules includes the opening connecting words, the subsequent execution logic chain, and the closing guiding words. Step 5: Decision Execution; According to the matching subsequent execution logic chain, the corresponding knowledge base, business system or service engine is scheduled to complete the corresponding knowledge query, business operation or service generation, and the execution result content is obtained; Step 6: Interactive Output and Data Retention; Integrate the opening transition script, execution result content, and closing guiding script into a response content output to the user, while retaining the full-link data of this interaction, and enter the adaptive learning process; Step 7: Human-machine backup handling; transfer the user to a human customer service representative, synchronize the current user's end-to-end interaction data to the human agent's terminal, and after completing the human service, synchronize the human service data to the adaptive learning dataset. Step 8: Adaptive Iterative Optimization; Based on the full interactive dataset, the intent recognition model, decision execution rule set, and insurance vertical domain knowledge base are periodically iterated and optimized to achieve adaptive learning of the system.

[0019] Example 2: like Figure 1 As shown, this embodiment of the invention provides an artificial intelligence decision-making system based on adaptive learning. In the input preprocessing process of step 1 of the system, the preprocessing specifically includes speech-to-text processing of the voice input to obtain the original text data, noise reduction processing of the original text data to remove redundant characters, interjections and invalid information, entity extraction of the text data in the insurance field to identify and extract policy number, insurance type name, disease name, hospital name, drug name and user identity information entity, semantic normalization processing of synonyms in the text, and output of standardized user query text. In step 5, during decision execution, the corresponding scheduling logic is executed based on the different objectives and intentions, specifically including: If the primary objective is knowledge-based question answering, the corresponding sub-knowledge base of the insurance vertical domain knowledge base is called according to the execution logic chain, and the question answering results are obtained through retrieval and matching. If the primary objective is business inquiry or business processing, the corresponding insurance business system interface is called according to the execution logic chain to complete the closed-loop execution of business data inquiry or business operation; If the primary objective is product and recommendation, medical support, or health management, the corresponding service engine is invoked according to the execution logic chain to generate personalized service results. Step 8, adaptive iterative optimization, specifically includes: Collect all user interaction data to build model fine-tuning datasets and rule optimization datasets; Based on the fine-tuning dataset, the pre-trained language model and intent classification model for the insurance vertical domain are fine-tuned regularly to improve the accuracy of intent recognition. Optimize the execution rule set in the intent decision link mapping library based on the rule optimization dataset to improve the success rate of decision execution; Update the content of the insurance vertical domain knowledge base based on the new knowledge requirements in the interactive data. Retain a complete, traceable log of all algorithm decisions to meet the compliance requirements for algorithm filing in the insurance industry.

[0020] The following points should be noted in this article: 1. The accompanying drawings of the embodiments disclosed herein only relate to the structures involved in the embodiments disclosed herein; other structures can be referred to in general design.

[0021] 2. Where there is no conflict, the embodiments of this disclosure and the features in the embodiments can be combined with each other to obtain new embodiments.

[0022] Although embodiments of the present invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

Claims

1. An artificial intelligence decision-making system based on adaptive learning, comprising a multimodal input preprocessing module, an adaptive intent recognition module, a multi-dimensional decision scheduling engine module, an insurance vertical domain knowledge base module, a third-party business system integration module, an adaptive learning optimization module, and a human-computer interaction fallback module, characterized in that: The multimodal input preprocessing module is used to receive multimodal input data from users, preprocess the input data, and output standardized user query text. The adaptive intent recognition module is communicatively connected to the multimodal input preprocessing module. It is used to perform hierarchical intent recognition on standardized user query text based on a pre-trained language model in the insurance vertical domain, and output the matching target first-level intent and second-level intent. The multi-dimensional decision scheduling engine module is communicatively connected to the adaptive intent recognition module and has a built-in intent decision link mapping library, which is used to match the corresponding execution rule set according to the target first-level intent and second-level intent. The execution rule set includes the opening connecting words, the subsequent execution logic chain and the closing guiding words. The insurance vertical domain knowledge base module is communicatively connected to the multi-dimensional decision scheduling engine module, and is used to store structured and unstructured knowledge data in the insurance field to provide knowledge support for decision execution. The third-party business system integration module is communicatively connected to the multi-dimensional decision scheduling engine module, and is used to integrate with the third-party service system of insurance business to realize business data query and execution of business operation instructions; The adaptive learning optimization module is connected to the adaptive intent recognition module, the multi-dimensional decision scheduling engine module, and the insurance vertical domain knowledge base module, respectively. It is used to collect full user interaction data and iteratively optimize the intent recognition model, decision link rules and knowledge base content based on the interaction data. The human-computer interaction fallback module is connected to the multi-dimensional decision scheduling engine module and the adaptive learning optimization module, respectively, and is used to transfer the user to human customer service and synchronize the full-link interaction data when the system cannot complete the automated processing or the user initiates a manual request.

2. The artificial intelligence decision-making system based on adaptive learning according to claim 1, characterized in that: The multimodal input preprocessing module includes a speech-to-text unit, a text noise reduction unit, an insurance domain entity extraction unit, and a semantic normalization unit; The speech-to-text unit is used to convert the user's voice input into text data; The text denoising unit is used to remove redundant characters, interjections, and invalid information from text data; The insurance domain entity extraction unit is used to identify and extract insurance domain entities from the text; The entities in the insurance field include policy numbers, insurance types, disease names, hospital names, drug names, and user identity information; The semantic normalization unit is used to normalize synonyms in the text and output standardized user query text.

3. The artificial intelligence decision-making system based on adaptive learning according to claim 1, characterized in that: The adaptive intent recognition module includes an intent hierarchy management unit, a semantic encoding unit, an intent confidence calculation unit, and an intent error correction unit. The intent hierarchy management unit has a built-in hierarchical intent list in the insurance field. The intent list includes first-level intents and several corresponding second-level intents. First-level intents include knowledge Q&A, business inquiry, business processing, products and recommendations, medical treatment support, and health management. The hierarchical intent list includes the following secondary intents: Knowledge Q&A: Secondary intents corresponding to primary intents include insurance industry knowledge Q&A, insurance terms knowledge Q&A, and insurance process knowledge Q&A; Business Inquiry: Secondary intents corresponding to primary intents include policy inquiry, claims record inquiry, and payment record inquiry; Business Processing: Secondary intents corresponding to primary intents include online policy maintenance, online claims, online renewal, online cancellation, general underwriting, disease-type underwriting, general claims processing, disease-related claims processing, hospital-related claims processing, drug-related claims processing, transfer to human agent, and online complaints; Product and Recommendation: Secondary intents corresponding to primary intents include personalized product recommendations and product comparison; Medical Support: Secondary intents corresponding to primary intents include initial disease analysis and guidance; Health Management: Secondary intents corresponding to primary intents include health management service recommendations. The semantic encoding unit is used to perform semantic encoding on standardized user query text based on a pre-trained language model in the insurance vertical domain, and generate semantic vectors. The intent confidence calculation unit is used to input the semantic vector into the fine-tuned intent classification model, calculate the matching confidence of each first-level intent and second-level intent, and output the candidate intent set with the highest confidence. The intent correction unit is used to perform compliance verification and logical error correction on the candidate intent set. When the highest confidence level is higher than the preset threshold, it is determined to be the target first-level intent and second-level intent. When the highest confidence level is lower than the preset threshold, an intent clarification message is generated to guide the user to supplement the input information.

4. The artificial intelligence decision-making system based on adaptive learning according to claim 1, characterized in that: The multi-dimensional decision scheduling engine module includes a link matching unit, a knowledge base scheduling unit, a business system scheduling unit, and a service generation scheduling unit. The link matching unit is used to match the corresponding set of execution rules from the intent decision link mapping library according to the target secondary intent; The knowledge base scheduling unit is used to call the sub-knowledge base corresponding to the insurance vertical domain knowledge base module according to the execution logic chain when the target first-level intent is knowledge question and answer, and obtain the question and answer result. The business system scheduling unit is used to call the corresponding business system interface through the third-party business system docking module to complete the data query or business operation execution when the target first-level intent is business query or business processing, according to the execution logic chain. The service generation and scheduling unit is used to call the corresponding service engine according to the execution logic chain to generate personalized service results when the target primary intent is product and recommendation, medical support or health management.

5. The artificial intelligence decision-making system based on adaptive learning according to claim 1, characterized in that: The insurance vertical domain knowledge base module includes multiple sub-knowledge bases, including an insurance industry knowledge sub-base, an insurance terms knowledge sub-base, an insurance process knowledge sub-base, an underwriting and claims rules sub-base, an insurance product knowledge sub-base, and a disease and health management knowledge sub-base. Each sub-base corresponds to a secondary intent under the knowledge question and answer category, as well as the knowledge support needs for intents related to business processing, products and recommendations, medical treatment support, and health management.

6. The artificial intelligence decision-making system based on adaptive learning according to claim 1, characterized in that: The third-party business system integration module has a built-in interface adaptation unit for integrating with core insurance business systems, policy management systems, claims business systems, policy maintenance business systems, user information management systems, health management service systems, and insurance product recommendation engine systems, enabling bidirectional data transmission and closed-loop execution of business instructions.

7. The artificial intelligence decision-making system based on adaptive learning according to claim 1, characterized in that: The adaptive learning optimization module includes an interactive data acquisition unit, a data annotation unit, a model fine-tuning unit, a rule optimization unit, a knowledge base update unit, and a full-link log retention unit. The interactive data acquisition unit is used to collect all user interaction data, including user input data, intent recognition results, decision execution process, user feedback data, data transferred to human assistant, and business processing success rate data. The data annotation unit is used to automatically annotate and manually verify the collected interactive data, and generate model fine-tuning datasets and rule optimization datasets. The model fine-tuning unit is used to periodically fine-tune the pre-trained language model and intent classification model in the insurance vertical domain based on the labeled dataset, thereby improving the accuracy of intent recognition. The rule optimization unit is used to optimize the set of execution rules in the intent decision link mapping library based on the labeled dataset, thereby improving the matching degree and success rate of decision execution. The knowledge base update unit is used to update the content of the insurance vertical domain knowledge base based on the new knowledge requirements in the interactive data. The full-link log retention unit is used to retain full-link traceable logs of all algorithm decisions in the system, meeting the compliance requirements for algorithm filing in the insurance industry.

8. The artificial intelligence decision-making system based on adaptive learning according to claim 1, characterized in that: The triggering conditions for the human-computer interaction fallback module include: the user explicitly initiates a manual request, the highest confidence level of intent recognition is lower than the preset minimum threshold, the decision-making process fails, and the user is dissatisfied with the system's response. When the human-computer interaction fallback module transfers the user to a human customer service representative, it simultaneously synchronizes the current user's full-link interaction data, user profile data, and executed decision-making processes to the human agent's terminal. After the human service is completed, it synchronizes the human service interaction data to the adaptive learning optimization module.

Images