A method for removing hallucinations in customer service question answering based on a large model

Abstract

This invention discloses a method for removing illusions in customer service question answering based on a large model, belonging to the field of artificial intelligence and natural language processing technology. The method includes: acquiring the user's original question; inputting the original question into a pre-trained intent correction mini-model to generate a standard question corresponding to the original question; based on the generated standard question, performing vector retrieval in a pre-built knowledge base to obtain a matching standard answer; performing compliance checks on the retrieved standard answer, and outputting the final answer after passing the check. This invention maps user questions to standard questions through a mini-model, transforming open-domain generation into closed-domain retrieval, suppressing illusions at the source. Combined with RAG retrieval and large-model-based detection, it significantly improves the accuracy and security of customer service question answering, making it particularly suitable for heavily regulated industries such as finance and insurance.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and natural language processing technology, specifically a method for removing illusions in customer service question answering based on a large model. Background Technology

[0002] The insurance industry is a heavily regulated financial sector. In the field of AI-generated content, authenticity and accuracy are paramount, especially for customer service Q&A content. It must be both real-time and accurate; otherwise, complaints and even legal disputes may arise. Traditional customer service robots primarily rely on keyword matching. While their accuracy is not high, they avoid simply copying and pasting answers and may refuse to answer altogether when necessary. Customer service Q&A robots based on large models possess stronger semantic understanding and generalization capabilities, but they are also susceptible to the problem of "illusion." If the generated content crosses content boundaries, the consequences can be severe.

[0003] A Chinese patent discloses a large-model question-answering method and system for hallucination suppression (authorization announcement number CN120929579 B). This patented technology is beneficial for improving the quality of question-answering output, and by introducing a large-model hallucination comparison, it helps detect and avoid hallucinations in intelligent question-answering results based on large models within the RAG framework. However, this solution is still a "post-hoc remedial" mechanism; the answer generation process still relies on the large model and fails to fundamentally suppress the generation of hallucinations. In scenarios with strong financial regulation, there are still risks of divergent generated content and insufficient controllability. Therefore, this invention provides a method for removing hallucinations in customer service question-answering based on a large model to solve the problems mentioned in the background. Summary of the Invention

[0004] The purpose of this invention is to provide a method for removing illusions in customer service question-and-answer sessions based on a large model. This invention creates a customer service question-and-answer model applicable to the health insurance industry that reduces illusions and improves accuracy during the question-and-answer process.

[0005] To achieve the above objectives, the present invention provides the following technical solution: This invention provides a method for removing illusions in customer service question-and-answer sessions based on a large model, comprising the following steps: The native problem of obtaining user input; The original question is input into a pre-trained intention correction mini-model to generate a standard question corresponding to the original question; wherein, the intention correction mini-model is obtained by fine-tuning the "original question-standard question" mapping corpus, and the standard question is a natural language question that can be used to retrieve knowledge bases; Based on the generated standard questions, a vector retrieval is performed in a pre-built knowledge base to obtain matching standard answers; The retrieved standard answers undergo compliance checks, and the final answer is output after the check passes.

[0006] This invention also provides a system for removing illusions in customer service question-and-answer sessions based on a large model, comprising: The question retrieval module is configured to retrieve native questions input by the user. The intent correction module is configured as a pre-trained intent correction mini-model, used to map the original question to the corresponding standard question; wherein, the intent correction mini-model is obtained by fine-tuning the "original question-standard question" mapping corpus, and the standard question is a natural language question that can be used to retrieve the knowledge base; The retrieval module is configured to perform vector retrieval in a pre-built knowledge base based on the generated standard question to obtain a matching standard answer. The compliance detection module is configured to perform compliance checks on the retrieved standard answers and output the final answer that passes the check.

[0007] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the above-described method.

[0008] Furthermore, the "native question - standard question" mapping corpus can be generated in several ways: establishing baseline mapping pairs by collecting user questions from real customer service logs; generating user question variants based on a standard question library using a large model; performing quality filtering on the generated mapping pairs through multi-model consistency screening or adversarial negative sample construction; selecting low-confidence samples for manual review through active learning; and adding the reviewed mapping pairs to the training set for iterative optimization of the small model. The fine-tuning method for the intent correction small model includes at least one of SFT full-parameter fine-tuning and LoRA partial parameter freezing fine-tuning.

[0009] Furthermore, to handle complex user questions, before inputting the original question into the intent correction mini-model, complexity prediction and question parsing operations are also included: for multi-intent questions, they are divided into several independent sub-questions for separate processing; for context-dependent questions, reference resolution and omission completion are performed by combining dialogue history; for conditional questions, conditional phrases and core questions are extracted, and the conditional phrases are used for subsequent answer filtering.

[0010] Furthermore, when there are multiple questions to be processed, the standard questions corresponding to each sub-question are retrieved in parallel to obtain a list of candidate answers, and the answers are fused according to logical relationships: parallel questions are concatenated in order; logically related questions are reorganized based on the large model to restructure the retrieval results, and the large model is constrained to only restructure and not create.

[0011] Furthermore, the compliance detection adopts a two-level architecture: first, the rule engine quickly filters sensitive words and formats; for complex cases that the rules cannot judge, the large model is called to perform semantic compliance detection, and the judgment basis is recorded for audit traceability.

[0012] Furthermore, it also includes a feedback optimization loop: collecting user feedback on the final answer, identifying cases of missed or incorrect detection, and using this information to update compliance detection rules or fine-tune small models.

[0013] Furthermore, the knowledge base construction and maintenance methods include: collecting and cleaning historical Q&A data of enterprises to form standard QA pairs, and storing the question fields in a vector database; adopting a double-buffer update strategy to achieve atomic switching; and binding a knowledge base version number to each session to ensure consistency within the session.

[0014] Furthermore, when the confidence level of the small model intended to correct the standard question is lower than a preset threshold, a transfer to human customer service is triggered to enhance the robustness of the system.

[0015] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention primarily addresses how to reduce illusions in customer service question-and-answer systems based on large models in the insurance field. Besides leveraging flexible semantic understanding capabilities, it further improves the overall accuracy of the model's question-and-answer process and makes the generated content more secure. The overall solution is based on large model, small model fine-tuning, and RAG technology, representing a systematic approach to illusion removal.

[0016] 2. By mapping user questions to standard questions, the open-domain generation task is transformed into a closed-domain retrieval task. The small model only performs "multiple-choice questions," fundamentally avoiding the illusion caused by content divergence. The retrieval path based on standard questions highly matches the knowledge base content. Combined with compliance checks, the question-answering accuracy is significantly improved (experiments show a BLEU score of over 96%). Attached Figure Description

[0017] Figure 1 This is a system framework diagram for a method to remove illusions in customer service question answering based on a large model. Detailed Implementation

[0018] 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.

[0019] Please see Figure 1In this embodiment of the invention, a method for removing hallucinations based on a large model customer service Q&A is presented. This invention creates an integrated hallucination removal solution by combining multiple technologies based on large model, small model fine-tuning, and RAG technology.

[0020] Example 1: A basic method for removing illusions in customer service question-and-answer systems based on a large model This embodiment provides a method for removing illusions in customer service question-and-answer systems based on a large model, applied to online customer service systems in heavily regulated industries such as finance and insurance. The overall process includes four core stages: intent recognition, knowledge base maintenance, answer acquisition, and answer polishing and detection.

[0021] Step 1: Intent Recognition Users submit native questions (such as "Does this insurance cover colds?") on the customer service interface. These questions are first input into a pre-trained intent correction mini-model. This mini-model is fine-tuned using a "native question - standard question" mapping corpus. Its training objective is to map users' casual, colloquial questions to pre-maintained standard questions (e.g., outputting the standard question: "Does XX critical illness insurance cover outpatient expenses for the common cold?"). This mapping method transforms an open-domain generation task into a closed-domain selection task, suppressing the generation of illusions at the source.

[0022] Step 2: Knowledge Base Maintenance Past customer service Q&A data is collected, cleaned, and organized to form standard QA question-and-answer pairs. The bge-embedding model is used to convert standard questions into vector representations, which are then stored in the Milvus vector database. The knowledge base is maintained independently of the model, supporting real-time addition, deletion, modification, and querying, and can respond to business changes without retraining the model.

[0023] Step 3: Obtaining the Answer Based on the standard questions generated in step 1, a similarity search is performed in the Milvus vector library to obtain matching QA pairs, from which the standard answer is extracted. This step ensures that all answers come from our own knowledge base, eliminating the risk of divergence caused by model-generated content.

[0024] Step 4: Answer polishing and testing The retrieved standard answers are not directly returned to the user, but are instead sent to an internally deployed large model for compliance checks. The large model performs a baseline judgment on the answers based on pre-defined semantic safety rules (such as compliance, accuracy, and sensitivity), and only answers that pass the checks are output. The large model acts as a "reviewer" rather than a "creator," further isolating the generation risks.

[0025] Corpus generation and model fine-tuning The training corpus for the aforementioned intention correction mini-model was generated using the following method: Utilizing Python scripts and the reasoning capabilities of a general large model, diverse variations of user questions were generated based on a standard question library; combined with seed data from real customer service logs, a high-quality "user question - standard question" mapping pair was constructed through multi-model consistency screening, adversarial negative sample construction, and active learning with manual review. Fine-tuning supports SFT full-parameter fine-tuning and LoRA partial parameter freezing fine-tuning, which can be flexibly switched according to the corpus size and accuracy requirements. After training, the model achieved an accuracy of over 96% on the BLEU metric.

[0026] Example 2: Enhanced Implementation for Complex Scenarios This embodiment, based on embodiment 1, enhances aspects such as complex user issues, dynamic maintenance of the knowledge base, and system self-evolution to further improve the robustness and applicability of the system.

[0027] Complex problem parsing mechanism Before inputting the original problem into the small model intended for correction, a complexity prediction and parsing module is added: Multi-intent question: The user question is divided into several independent sub-questions by intent segmentation model, which are then mapped to standard questions and retrieved answers and merged in sequence.

[0028] Context-dependent issues: Combine the dialogue state tracker to resolve the references and complete the omissions of the current question, and then send the completed question to the small model for processing.

[0029] Conditional questions: The conditional extraction model identifies conditional phrases and core questions in the questions, and uses the conditional phrases as additional information for filtering and ranking in subsequent answer retrieval.

[0030] Two-level compliance testing The compliance testing employs a two-tier architecture: first, a rule engine quickly filters standard answers (e.g., sensitive word detection, format checks); for complex cases that the rules cannot handle, a larger model is invoked for semantic compliance testing, and the basis for the judgment is recorded for audit traceability. This architecture improves the accuracy of judgments in complex semantic scenarios while ensuring efficiency.

[0031] Knowledge base dynamic maintenance A dual-buffer update strategy is adopted: two vector databases are maintained, one primary and one backup. During updates, the index is rebuilt in the backup database and then switched atomically, achieving seamless updates for users. Simultaneously, a knowledge base version number is bound to each user session to ensure consistency within the session. For minor data changes, incremental vector insertion is supported to improve update efficiency.

[0032] Low confidence fallback and feedback optimization When the confidence level of the small model in generating standard questions falls below a preset threshold, it triggers a transfer to human customer service or refuses to answer, thus avoiding incorrect mapping. The system also collects user feedback on the final answer and regularly analyzes negative feedback cases to update compliance detection rules or fine-tune the small model, forming a closed loop of continuous evolution.

[0033] Currently, this method supports customer service Q&A scenarios and intent recognition scenarios, and will switch script parameters according to the amount of data and accuracy requirements.

[0034] Accuracy evaluation. The trained model is mainly evaluated using BLUE and ROUGE. Currently, the accuracy rate under BLUE reaches 96%+.

[0035] This invention utilizes a large-scale model inference, a small-scale model fine-tuning (a method for generating user-to-standard question corpus), and a supporting RAG technology system. Its differences from current technologies are as follows: Small model corpus preparation. Currently, mainstream customer service question-answering models generally generate training corpora based on questions and specific answers. The content generated by this approach is prone to divergence. This invention adopts a method of mapping user questions to standard questions to fix the generated content and reduce the generation of illusions. Question-and-answer technology process construction. This invention creates an efficient customer service question-and-answer technology process and system. It uses a user-question-standard-question-knowledge-base answer retrieval method to eliminate model illusion, which can both understand user intent and improve question-and-answer accuracy.

[0036] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for removing hallucinations in large model-based customer service question answering, characterized in that, Includes the following steps: Step 1: Obtain the user's original question input; Step 2: Input the original question into the pre-trained intention correction mini-model to generate a standard question corresponding to the original question; wherein, the intention correction mini-model is obtained by fine-tuning the "original question-standard question" mapping corpus, and the standard question is a natural language question that can be used to retrieve knowledge bases; Step 3: Based on the generated standard question, perform vector retrieval in the pre-built knowledge base to obtain the matching standard answer; Step 4: Perform a compliance check on the retrieved standard answer, and output the final answer after the check passes.

2. A system for removing hallucinations in large model based customer service question answering, characterized in that, include: 1) Question retrieval module, configured to retrieve native questions input by the user; 2) Intent correction module, configured as a pre-trained intent correction mini-model, used to map the original question to the corresponding standard question; wherein, the intent correction mini-model is obtained by fine-tuning the "original question-standard question" mapping corpus, and the standard question is a natural language question that can be used to retrieve the knowledge base; 3) Retrieval module, configured to perform vector retrieval in a pre-built knowledge base based on the generated standard question to obtain the matching standard answer; 4) Compliance detection module, configured to perform compliance checks on the retrieved standard answers and output the final answer that passes the check.

3. The method of claim 1, wherein, The "native question - standard question" mapping corpus is generated through at least one of the following methods: Collect user issues from real customer service logs and establish benchmark mapping pairs; Generate user question variations based on a standard question library using a large model; Quality filtering of generated mapping pairs is performed through multi-model consistency screening or adversarial negative sample construction. Low-confidence samples are selected for manual review through active learning; The verified mapping pairs are added to the training set, and the small model is iteratively optimized. The fine-tuning methods for the intended small model correction include at least one of SFT full-parameter fine-tuning and LoRA partial parameter freeze fine-tuning.

4. The method of claim 1, wherein, Before inputting the original problem into the intention correction mini-model, the process also includes a complexity prediction of the original problem and a problem parsing operation based on the prediction result, including at least one of the following: When the original problem contains multiple parallel or progressive sub-problems, the original problem is divided into several independent sub-problems through the intent segmentation model, and each sub-problem is input into the intent correction mini-model as a problem to be processed. When the original question depends on the dialogue history, the dialogue history information is maintained by the dialogue state tracker, the original question is resolved by reference and omitted and completed, and the completed question is input into the intention correction small model; When the original question contains a conditional adverbial clause, the conditional phrases and core question in the original question are identified by the conditional extraction model. The core question is input into the intention correction mini-model, and the conditional phrases are used as additional information for filtering in subsequent answer retrieval.

5. The method of claim 1, wherein, The standard question generated is used to perform vector retrieval in a pre-built knowledge base to obtain a matching standard answer, including: When there are multiple questions to be processed, vector retrieval is performed on the standard questions corresponding to each question in parallel to obtain a list of candidate answers for each question. Based on the logical relationships between the questions to be processed, multiple candidate answers are merged to generate a merged answer; The fusion includes: for multiple pending questions with parallel relationships, concatenating the candidate answers according to the order in which the pending questions appear in the original questions; for multiple pending questions with logical relationships, reorganizing the retrieved candidate answers based on the large model to generate a coherent response, and constraining the large model to only reorganize based on the retrieved candidate answers.

6. The method of claim 1, wherein, The compliance check of the retrieved standard answers includes: A rule engine is used to quickly filter standard answers; For complex cases that the rules engine cannot determine, a large model is invoked to perform semantic compliance detection; Record the basis for detection and judgment for compliance audit traceability.

7. The method of claim 1, wherein, include: Collect user feedback data on the final answer; Based on the feedback data, cases of missed or false positives are identified and detected; The identified cases can be used to update the compliance detection rules or fine-tune the intent correction mini-model.

8. The method of claim 1, wherein, The knowledge base is constructed and maintained in the following ways: Collect, clean, and organize past Q&A data from enterprises to form standard QA question-and-answer pairs; Convert the question fields in the standard QA question-answer pair into vector representations and store them in a vector database; The vector database is updated using a double-buffer update strategy; Bind a knowledge base version number to each user session to ensure consistency of questions and answers within the same session.

9. The method of claim 1, wherein, When the confidence level of the intention correction mini-model for the generated standard question is lower than a preset threshold, a transfer to human customer service is triggered.

10. A computer-readable storage medium having a computer program stored thereon, the computer program, when executed by a processor, implementing the method as claimed in any one of claims 1 to 9.

Images