Training data generation method and device, equipment and medium
By comparing the generated test questions and answer data, the agent response data with high consistency was selected as training data, which solved the problem of noise and bias in the fine-tuning of the AI agent model and improved the accuracy of the reference response.
Patent Information
- Application Number
- CN202511696207.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-18
- Publication Date
- 2026-02-06
AI Technical Summary
In existing technologies, during the model fine-tuning process of AI agents, the misalignment between agent responses and internal inference links may introduce data noise and cognitive biases, leading to a decrease in the accuracy of reference responses.
The first AI agent generates exam question data and standard answer data, while the second AI agent generates candidate answer data. Based on the comparison between the standard answers and candidate answers, the agent response data that is highly consistent with the relevant data is selected as training data for fine-tuning the AI agent model.
It effectively avoids the introduction of data noise and cognitive bias, improves the accuracy of reference responses, and ensures that the responses generated by the AI agent are consistent with relevant data.
Smart Images
Figure CN121479313A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of artificial intelligence, in particular to a training data generation method and device, equipment and medium. BACKGROUND
[0002] With the deep integration of Large Language Model (LLM) and Retrieval Augmented Generation (RAG) technology in the insurance field, artificial intelligence agents (AI agents) have been gradually applied to intelligent customer service, agent assistance and other core business scenarios. Referring to Figure 1 , the figure is a schematic diagram of applying an AI agent. First, the customer submits a customer question related to the insurance business; then, the customer question is input into the AI agent to enable the AI agent to search for relevant information in the preset knowledge base; then, the AI agent reads and reasons the relevant information based on deep semantic understanding technology, generates and outputs a reference reply as an assistant suggestion for the agent to respond to the customer.
[0003] Currently, in order to continuously optimize the quality of the reference reply output by the AI agent, the agent reply actually adopted by the agent is usually taken as feedback data and returned to the training data set to fine-tune the AI agent model.
[0004] However, when the agent reply and the internal reasoning link of the AI agent cannot be completely aligned, this model fine-tuning process may introduce data noise and cognitive bias, which may cause the reference reply output by the AI agent to have hallucinations, thereby reducing the accuracy of the reference reply. SUMMARY
[0005] Based on the above problems, the present application provides a training data generation method, device, equipment and medium, which can improve the accuracy of the reference reply.
[0006] The present application embodiment discloses the following technical scheme:
[0007] In a first aspect, the present application discloses a training data generation method, which comprises:
[0008] Obtaining agent reply data and relevant information data corresponding to the agent reply data;
[0009] Generating, by a first artificial intelligence agent, examination question data and standard answer data corresponding to the agent reply data;
[0010] Using a second artificial intelligence agent, candidate answer data corresponding to the test question data is generated based on the relevant data.
[0011] Training data is generated based on the comparison results between the standard answer data and the candidate answer data; the training data includes the agent response data.
[0012] Optionally, generating the test question data and standard answer data corresponding to the agent response data includes:
[0013] Knowledge points are extracted from the agent response data; the knowledge points include entities and relationships.
[0014] Generate test question data and standard answer data corresponding to the knowledge points; the test question data includes at least one of the following types: single-choice questions, multiple-choice questions, fill-in-the-blank questions, and true / false questions.
[0015] Optionally, the candidate answer data includes confidence levels; generating training data based on the comparison results between the standard answer data and the candidate answer data includes:
[0016] If the confidence level is higher than the preset confidence threshold, training data is generated based on the comparison between the standard answer data and the candidate's answer data.
[0017] Optionally, the standard answer data includes the standard answer and the basis for the standard answer, and the candidate answer data includes the candidate answer and the basis for the candidate answer;
[0018] The step of generating training data based on the comparison results between the standard answer data and the candidate answer data includes:
[0019] If the candidate's answer is completely consistent with the standard answer, and the basis of the candidate's answer is completely consistent with the basis of the standard answer, then training data is generated.
[0020] Optionally, generating training data based on the comparison results between the standard answer data and the candidate answer data includes:
[0021] If the candidate's answer is completely consistent with the standard answer, and the similarity between the candidate's answer and the standard answer is higher than or equal to a preset similarity threshold, then training data is generated.
[0022] Secondly, this application discloses a training data generation device, the device comprising: a data acquisition module, a first generation module, a second generation module, and a result comparison module;
[0023] The data acquisition module is used to acquire agent response data and related information data corresponding to the agent response data;
[0024] The first generation module is used to generate test question data and standard answer data corresponding to the agent response data through a first artificial intelligence agent;
[0025] The second generation module is used to generate candidate answer data corresponding to the test question data based on the relevant data through a second artificial intelligence agent;
[0026] The result comparison module is used to generate training data based on the comparison results between the standard answer data and the candidate's answer data; the training data includes the agent response data.
[0027] Optionally, the first generation module is specifically used to: extract knowledge points from the agent response data; the knowledge points include entities and relationships; generate test question data and standard answer data corresponding to the knowledge points; the test question data includes at least one of single-choice questions, multiple-choice questions, fill-in-the-blank questions, and true / false questions.
[0028] Optionally, the result comparison module is specifically used to: if the confidence level value is higher than a preset confidence threshold, generate training data based on the comparison result between the standard answer data and the candidate's answer data.
[0029] Optionally, the standard answer data includes the standard answer and the basis for the standard answer, and the candidate answer data includes the candidate answer and the basis for the candidate answer; the result comparison module is specifically used to: generate training data if the candidate answer is completely consistent with the standard answer, and the basis for the candidate answer is completely consistent with the basis for the standard answer.
[0030] Optionally, the result comparison module is specifically used to: generate training data if the candidate's answer is completely consistent with the standard answer, and the similarity between the candidate's answer and the standard answer is higher than or equal to a preset similarity threshold.
[0031] Thirdly, this application discloses a training data generation device, the device comprising: a memory and a processor;
[0032] The memory is used to store programs;
[0033] The processor is configured to execute the program to implement the various steps of the training data generation method as described in the first aspect.
[0034] Fourthly, this application discloses a computer-readable medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the training data generation method as described in the first aspect.
[0035] Compared with the prior art, this application has the following beneficial effects:
[0036] This application provides a method, apparatus, device, and medium for generating training data. The method includes: acquiring agent response data and related information data corresponding to the agent response data; generating test question data and standard answer data corresponding to the agent response data through a first artificial intelligence agent; generating candidate answer data corresponding to the test question data through a second artificial intelligence agent based on the related information data; and generating training data based on the comparison results of the standard answer data and the candidate answer data. The training data includes the agent response data. Therefore, this method can select agent response data that is highly consistent with the related information data and internal inference chain as training data for fine-tuning the original AI agent model. This avoids the introduction of data noise and cognitive bias, reduces the risk of the original AI agent's output reference responses being misleading, and thus improves the accuracy of the reference responses. Attached Figure Description
[0037] To more clearly illustrate the technical solutions in the embodiments of 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 only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0038] Figure 1 This is a schematic diagram of an AI agent application;
[0039] Figure 2 A flowchart illustrating a method for generating training data provided in an embodiment of this application;
[0040] Figure 3 A schematic diagram illustrating a method for generating training data according to an embodiment of this application;
[0041] Figure 4 A schematic diagram of a training data generation device provided in an embodiment of this application;
[0042] Figure 5 This is a schematic diagram of a computer-readable medium provided in an embodiment of this application. Detailed Implementation
[0043] As described earlier, currently, to continuously optimize the quality of reference responses output by AI agents, actual agent responses adopted by agents are typically used as feedback data and fed back into the training dataset for model fine-tuning of the AI agent. However, when agent responses are not perfectly aligned with the internal inference chain of the AI agent, this model fine-tuning process may introduce data noise and cognitive biases, potentially leading to illusions in the reference responses output by the AI agent, thereby reducing the accuracy of the reference responses.
[0044] Through research, the inventors proposed a method, apparatus, device, and medium for generating training data. This method can select agent response data that is highly consistent with relevant data and internal inference links as training data for fine-tuning the original AI agent model. This avoids the introduction of data noise and cognitive bias, reduces the risk of the original AI agent's output reference response being illusory, and thus improves the accuracy of the reference response.
[0045] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present application.
[0046] See Figure 2 This figure is a flowchart of a training data generation method provided in an embodiment of this application. See also... Figure 3 The figure is a schematic diagram of a training data generation method provided in an embodiment of this application. The method includes:
[0047] S201: Obtain agent response data and related information data corresponding to the agent response data.
[0048] Agent response data refers to the responses that are ultimately adopted and sent to customers during interactions between human agents and customers. In one specific implementation, agent response data can be obtained from the log module of the intelligent agent system. For example, agent response data could be: This insurance product covers 20 critical illnesses, including cancer and heart disease.
[0049] Relevant data refers to prior knowledge retrieved related to agent responses (from the original AI agent), such as structured or unstructured data obtained by retrieving information from a pre-defined knowledge base using RAG technology. For example, relevant data could be: This insurance product covers 20 critical illnesses including cancer, heart disease, and stroke, with a waiting period of 90 days.
[0050] S202: Through the first artificial intelligence agent, generate test question data and standard answer data corresponding to the agent's response data.
[0051] First, a primary AI agent (typically a large language model with powerful understanding and generation capabilities) extracts knowledge points from agent response data. These knowledge points include entities and relationships. Entities refer to key objects or concepts mentioned in the agent response data, while relationships describe the associations between entities. In the example above, the entities include disease names (cancer, heart disease) and quantities (20 types), and the relationship is an inclusion relationship.
[0052] In one specific implementation, the entity recognition model (e.g., an entity recognition model based on BERT or BiLSTM-CRF) included in the first artificial intelligence agent can be used to identify entities in the knowledge points; the relation extraction model of the first artificial intelligence agent can be used to extract relations in the knowledge points.
[0053] Subsequently, test question data and standard answer data corresponding to the knowledge points are generated. The test question data can be of at least one type: multiple choice, fill-in-the-blank, and true / false. The standard answer data includes the standard answer and its rationale; the standard answer data can be directly determined based on the knowledge points or derived from them.
[0054] In one specific implementation, knowledge points can be populated into a preset template to obtain exam question data. For example, a multiple-choice question template could be:
[0055] {
[0056] Question: What are the attributes of an insurance product?
[0057] Options: A. {Error value 1} B. {Correct value} C. {Error value 2}
[0058] }
[0059] For example, exam question data could be:
[0060] {
[0061] Question: Does this insurance product cover cancer?
[0062] Options: A. Yes B. No
[0063] }
[0064] For example, standard answer data could be:
[0065] {
[0066] Standard Answer: A
[0067] The standard answer is based on the data from the agent's response, which clearly states that cancer is among the 20 diseases covered.
[0068] }
[0069] S203: Through a second artificial intelligence agent, generate candidate answer data corresponding to the test question data based on relevant data.
[0070] Relevant data (such as insurance terms and product manuals) and exam questions are input into a second AI agent. This allows the agent to perform deep reading and understanding of the data, and to reason about the exam questions to generate corresponding student answer data. It's understood that the second AI agent is a dialogue model based on a retrieval-enhanced generative architecture. Its core technical requirement is that the generation process must be strictly limited to the information boundaries constituted by the provided data, prohibiting the use of any prior knowledge that may exist in its internal parameters, or any speculation or conjecture beyond the content of the data.
[0071] The candidate's answer data includes the candidate's answer, the basis for the answer (the specific original text or logic cited from relevant data to trace the source of the answer and ensure interpretability), and the confidence level (a quantitative value of the second AI agent's confidence in the candidate's answer, typically a value between 0 and 1). For example, the candidate's answer data could be:
[0072] {
[0073] Student's answer: A
[0074] The candidate's answer was based on the fact that relevant data clearly states that cancer is among 20 diseases.
[0075] Confidence level: 0.99
[0076] }
[0077] It should be noted that if the second AI agent generates contradictory student answer data for the same test question based on different related data (for example, one related data indicates selection A, while another related data indicates selection B), an alarm will be triggered. The alarm indicates that there is an inconsistency within the preset knowledge base, and this set of data is temporarily unsuitable for training. Human intervention is required first to verify and correct the source problem of the preset knowledge base.
[0078] It should also be noted that if the second AI agent cannot generate candidate answer data corresponding to the test question data based on the relevant data (i.e., "unable to answer" or generates an empty answer), or if the confidence value of the generated candidate answer data is lower than the preset confidence threshold (e.g., 0.7), an alarm will be triggered. Conversely, if the second AI agent can generate candidate answer data and the confidence value is higher than the preset confidence threshold, then step S204 can be executed.
[0079] S204: Generate training data based on the comparison results between the standard answer data and the candidate's answer data; the training data includes agent response data.
[0080] Firstly, if the comparison between the standard answer data and the candidate's answer data indicates that they are completely consistent (i.e., the candidate's answer is completely consistent with the standard answer, and the basis for the candidate's answer is completely consistent with the basis for the standard answer), then a strong positive training data (HIGH_CONFIDENCE) of "seat operator response data - related information data" is directly generated. Strong positive training data indicates that the seat operator response data is not only correct in its conclusions, but its underlying reasoning path and knowledge basis are also completely aligned with the related information data. This represents the highest quality and most reliable training sample.
[0081] Secondly, if the comparison between the standard answer data and the candidate's answer data indicates that they are inconsistent, then it should be determined first whether the candidate's answer is consistent with the standard answer.
[0082] If the candidate's answer is inconsistent with the standard answer, it indicates that the agent's response data is incorrect in its conclusion and cannot be aligned. It will be directly marked as INVALID and cannot be used to generate training data.
[0083] If the candidate's answer is consistent with the standard answer, the cosine similarity between the candidate's answer and the standard answer will be further determined by a pre-trained semantic similarity model (such as BERT-base or Sentence-BERT).
[0084] If the similarity between the candidate's answer and the standard answer is higher than or equal to a preset similarity threshold (e.g., 0.85), then weak positive training data (LOW_CONFIDENCE) is generated, consisting of "seat response data - related data". Weak positive training data indicates that the seat response data is correct in its conclusion, and the reasoning path and knowledge basis behind it are very similar to those in the related data, perhaps only differing in expression, and thus also have training value.
[0085] If the similarity between the candidate's answer and the standard answer is lower than the preset similarity threshold, it indicates that although the agent's response data happens to be correct in its conclusion, the underlying reasoning path and knowledge basis deviate significantly from the relevant data, posing a potential risk of misleading. It is directly marked as INVALID and cannot be used to generate training data.
[0086] Understandably, strong positive training data is used to enhance the original AI agent's ability to recognize explicit evidence, ensuring the accuracy and reliability of its generated reference responses and fundamentally suppressing illusions. Weak positive training data is used to improve the original AI agent's ability to generalize synonymous expressions. By exposing the original AI agent to training data of "the same knowledge in different ways," it can be trained to understand that as long as the core conclusion is correct, the specific expression can be diverse.
[0087] It is understood that training data (whether strongly positive or weakly positive) can be a structured data pair, such as: {"Related Materials": [Related Materials Data], "Ideal Response": [Agent Response Data]}. It can also be a dataset including related materials data, agent response data, and reference response data. This application does not impose any limitations on this.
[0088] In summary, the embodiments of this application provide a method for generating training data. This method can select agent response data that is highly consistent with relevant data and internal inference links as training data for fine-tuning the original AI agent model. This avoids the introduction of data noise and cognitive bias, reduces the risk that the reference response output by the original AI agent may be illusory, and thus improves the accuracy of the reference response.
[0089] See Figure 4 The figure is a schematic diagram of a training data generation device provided in an embodiment of this application. The training data generation device 400 includes: a data acquisition module 401, a first generation module 402, a second generation module 403, and a result comparison module 404.
[0090] Data acquisition module 401 is used to acquire agent response data and related data corresponding to the agent response data;
[0091] The first generation module 402 is used to generate test question data and standard answer data corresponding to the agent's response data through the first artificial intelligence agent;
[0092] The second generation module 403 is used to generate candidate answer data corresponding to the test question data based on relevant data through the second artificial intelligence agent;
[0093] The result comparison module 404 is used to generate training data based on the comparison results between the standard answer data and the candidate's answer data; the training data includes agent response data.
[0094] In one specific implementation, the first generation module 402 is specifically used to: extract knowledge points from agent response data; the knowledge points include entities and relationships; generate test question data and standard answer data corresponding to the knowledge points; the test question data includes at least one of single-choice questions, multiple-choice questions, fill-in-the-blank questions, and true / false questions.
[0095] In one specific implementation, the result comparison module 404 is specifically used to generate training data based on the comparison results between the standard answer data and the candidate's answer data if the confidence value is higher than the preset confidence threshold.
[0096] In one specific implementation, the standard answer data includes the standard answer and the basis for the standard answer, and the candidate answer data includes the candidate answer and the basis for the candidate answer; the result comparison module 404 is specifically used to generate training data if the candidate answer is completely consistent with the standard answer and the basis for the candidate answer is completely consistent with the basis for the standard answer.
[0097] In one specific implementation, the result comparison module 404 is specifically used to generate training data if the candidate's answer is completely consistent with the standard answer and the similarity between the candidate's answer and the standard answer is higher than or equal to a preset similarity threshold.
[0098] In summary, the embodiments of this application provide a training data generation device that can select agent response data that is highly consistent with relevant data and internal inference links as training data for fine-tuning of the original AI agent model. This avoids the introduction of data noise and cognitive bias, reduces the risk of the original AI agent's output reference response being illusory, and thus improves the accuracy of the reference response.
[0099] This application also provides a corresponding training data generation device and a computer-readable medium for implementing the training data generation method provided in this application.
[0100] The training data generation device includes a memory and a processor. The memory is used to store instructions or code, and the processor is used to execute the instructions or code so that the device performs a training data generation method according to any embodiment of this application.
[0101] See Figure 5 This figure is a schematic diagram of a computer-readable medium provided in an embodiment of this application. The computer-readable medium 500 stores a computer program 511, which, when executed by a processor, implements the above-described... Figure 1The steps involved in generating training data.
[0102] It should be noted that, in the context of this application, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. Machine-readable media can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.
[0103] It should be noted that the machine-readable medium described above in this application can be a computer-readable signal medium or a computer-readable storage medium, or any combination of the two. A computer-readable storage medium can be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this application, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In this application, a computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium can be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to: wires, optical fibers, RF (radio frequency), etc., or any suitable combination thereof.
[0104] The aforementioned computer-readable medium may be included in the aforementioned electronic device; or it may exist independently and not assembled into the electronic device.
[0105] Although the subject matter has been described using language specific to structural features and / or methodological logic, it should be understood that the subject matter defined in the appended claims is not necessarily limited to the specific features or actions described above. Rather, the specific features and actions described above are merely illustrative examples of implementing the claims.
[0106] While several specific implementation details are included in the foregoing discussion, these should not be construed as limiting the scope of this application. Certain features described in the context of individual embodiments may also be implemented in combination in a single embodiment. Conversely, various features described in the context of a single embodiment may also be implemented individually or in any suitable sub-combination in multiple embodiments.
[0107] The above description is merely a preferred embodiment of this application and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of disclosure in this application is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-described concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this application.
Claims
1. A method for generating training data, characterized in that, The method includes: Obtain agent response data and related information data corresponding to the agent response data; The first artificial intelligence agent generates test question data and standard answer data corresponding to the agent's response data; Using a second artificial intelligence agent, candidate answer data corresponding to the test question data is generated based on the relevant data. Training data is generated based on the comparison results between the standard answer data and the candidate answer data; the training data includes the agent response data.
2. The method according to claim 1, characterized in that, The generation of test question data and standard answer data corresponding to the agent response data includes: Knowledge points are extracted from the agent response data; the knowledge points include entities and relationships. Generate test question data and standard answer data corresponding to the knowledge points; the test question data includes at least one of the following types: multiple choice, multiple selection, fill-in-the-blank, and true / false questions.
3. The method according to claim 1, characterized in that, The candidate answer data includes confidence levels; the step of generating training data based on the comparison results between the standard answer data and the candidate answer data includes: If the confidence level is higher than the preset confidence threshold, training data is generated based on the comparison between the standard answer data and the candidate's answer data.
4. The method according to claim 1, characterized in that, The standard answer data includes the standard answer and the basis for the standard answer; the candidate answer data includes the candidate answer and the basis for the candidate answer. The step of generating training data based on the comparison results between the standard answer data and the candidate answer data includes: If the candidate's answer is completely consistent with the standard answer, and the basis of the candidate's answer is completely consistent with the basis of the standard answer, then training data is generated.
5. The method according to claim 4, characterized in that, The step of generating training data based on the comparison results between the standard answer data and the candidate answer data includes: If the candidate's answer is completely consistent with the standard answer, and the similarity between the candidate's answer and the standard answer is higher than or equal to a preset similarity threshold, then training data is generated.
6. A training data generation apparatus, characterized in that, The device includes: a data acquisition module, a first generation module, a second generation module, and a result comparison module; The data acquisition module is used to acquire agent response data and related information data corresponding to the agent response data; The first generation module is used to generate test question data and standard answer data corresponding to the agent response data through a first artificial intelligence agent; The second generation module is used to generate candidate answer data corresponding to the test question data based on the relevant data through a second artificial intelligence agent; The result comparison module is used to generate training data based on the comparison results between the standard answer data and the candidate's answer data; the training data includes the agent response data.
7. The apparatus according to claim 6, characterized in that, The first generation module is specifically used for: extracting knowledge points from the agent response data; the knowledge points include entities and relationships; generating test question data and standard answer data corresponding to the knowledge points; the test question data includes at least one of single-choice questions, multiple-choice questions, fill-in-the-blank questions, and true / false questions.
8. The apparatus according to claim 6, characterized in that, The result comparison module is specifically used to: if the confidence level is higher than a preset confidence threshold, generate training data based on the comparison results between the standard answer data and the candidate's answer data.
9. A training data generation device, characterized in that, The device includes: a memory and a processor; The memory is used to store programs; The processor is configured to execute the program to implement the steps of the training data generation method as described in any one of claims 1 to 5.
10. A computer-readable medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the training data generation method as described in any one of claims 1 to 5.