Controllable dialogue case generation method based on large model

Abstract

The invention discloses a controllable dialogue case generation method based on a large model, a full-process automatic dialogue case processing system is constructed based on a three-stage assembly line of screening, optimization and COT generation, and the method comprises the following specific steps: a first stage: multi-dimensional dialogue screening, including dialogue quality rating model construction and hierarchical screening strategy; in the second stage, multi-role multi-emphasis parallel rewriting is carried out; and a third stage: COT information generation. According to the method, automatic generation of'excellent dialogue cases + COT information 'is achieved, it is ensured that the generated dialogue cases meet the requirements of logic coherence, role consistency and theme focusing, the COT information matched with the dialogue cases is synchronously generated, and standardized training materials are provided for large model learning; a deep correlation mechanism of COT information and case content is constructed, a complete link of'case text-COT reasoning-knowledge support 'is formed, and the learning efficiency of a large model for complex dialogue logic is improved; and the large-scale generation bottleneck of manual dependence is eliminated.

Description

Technical Field

[0001] The present invention belongs to the technical field of data processing, and specifically relates to a method for generating controllable dialogue cases based on large models. Background Art

[0002] Current techniques for generating dialogue cases based on large models mainly rely on manually setting simple rules or basic parameters for generation. The generated dialogue cases generally have problems such as insufficient logical coherence, inconsistent character behaviors, and deviation from the theme, and it is difficult to directly use them as "excellent cases" for the logical guidance and few-shot learning of large models. Especially when it is necessary to provide the large model with Chain of Thought (COT) information, the existing techniques lack the ability to explicitly generate the reasoning process and decision-making logic in the cases. For example, in customer service dialogue cases, it is impossible to automatically generate the complete COT information of "user question → requirement analysis → knowledge base matching → response strategy selection", and this information is the core reference content for the large model to learn how to handle complex dialogue logic in the few-shot scenario.

[0003] The existing solutions can only manually screen a small number of cases from the generated results and manually supplement the COT information. This process has two major pain points: Low efficiency in producing excellent cases: Manual screening requires reviewing the logical rationality, character compliance, and theme relevance of the dialogue content one by one, and additional COT reasoning steps need to be written, which is time-consuming and laborious (for example, generating 100 high-quality customer service cases with COT requires more than 200 man-hours of manual input); Insufficient standardization of COT information: The COT content supplemented manually depends on individual experience, and the logical granularity and expression methods are not unified, resulting in the large model being difficult to effectively extract general reasoning patterns during learning, affecting the few-shot learning effect.

[0004] The existing technical defects are as follows: Lack of the ability to generate excellent cases and COT information: The existing large model generation process lacks automatic constraints on the core elements of "excellent cases" (logically complete, consistent characters, and focused themes), and no dedicated generation module is designed for COT information, so it is impossible to directly produce high-quality cases containing clear reasoning chains.

[0005] Scalability bottleneck caused by manual intervention: When building a case library for few-shot learning, it is necessary to generate case texts and corresponding COT information simultaneously. The manual processing cost increases linearly with the case scale, and it is difficult to meet the demand for a large amount of high-quality training data for the rapid iteration of large models.

[0006] Weak correlation between COT information and cases: In the existing technology, the generation process of cases and COT information is independent of each other, and the COT reasoning steps are often disconnected from the conversation content (for example, in the case, the customer service directly provides a solution, but COT does not explain how to match the knowledge base), resulting in the inability to establish an effective "behavior-reasoning" mapping when learning large models. Summary of the invention

[0007] In order to solve the defects and shortcomings in the above-mentioned prior art, the present invention provides a method for realizing the automatic generation of "excellent dialogue cases + COT information", ensuring that the generated dialogue cases meet the requirements of logical coherence, role consistency, and topic focus, and synchronously generates COT information matching them, so as to provide standardized training materials for few-shot learning of large models; constructs a deep association mechanism between COT information and case content, so that the generated COT information can accurately reflect the reasoning process of each key decision point in the dialogue case, forming a complete link of "case text → COT reasoning → knowledge support", and improving the learning efficiency of large models for complex dialogue logic; eliminates the bottleneck of large-scale generation due to manual dependence: through automated quality control and COT generation modules, replaces manual screening and information supplementation, realizes rapid expansion of high-quality case library, and meets the few-shot learning needs of large models in different scenarios (such as customer service, education, and medical care). A controllable dialogue case generation method based on a large model is provided.

[0008] To achieve the above purpose, the present invention provides the following technical solutions: a controllable dialogue case generation method based on a large model, based on the three-stage pipeline of "screening → optimization → generation of COT", to build a full-process automated dialogue case processing system, the specific steps are as follows: Phase 1: Multi-dimensional dialogue screening, including: dialogue quality rating model construction and hierarchical screening strategy; Phase II: Parallel rewriting of multiple roles and multiple focuses, as follows: After a typical dialogue enters the rewriting module, the system first determines the scene to which it belongs through the scene classifier, and automatically matches multiple corresponding roles based on the scene-role mapping rules. Each role corresponds to an independent rewriting engine, which performs differentiated rewriting based on its exclusive language style template; The third stage: COT information generation, as follows: After completing the dialogue screening and rewriting, the system enters the thinking chain information (COT) generation stage. This stage focuses on the explicit analysis of the dialogue decision logic, and collaboratively extracts key decision nodes and reasoning paths through multiple reasoning models. It also combines quality control mechanisms to ensure the logical integrity and accuracy of COT information, providing traceable thinking references for large-model few-shot learning.

[0009] Preferably, the construction of the dialogue quality rating model in the first stage includes: input layer: receiving the original dialogue data, preprocessing it through the large model prompt, and extracting key features; rating dimension: achieved through the large model multi-classification task; logical coherence: detecting whether the semantic jumps between dialogue turns are reasonable; topic relevance: judging whether the dialogue revolves around the target scenario; role compliance: verifying whether the role language conforms to the preset personality; information value: evaluating whether the dialogue contains effective decision-making reference content.

[0010] Preferably, the hierarchical screening strategies in the first stage include: excellent dialogues: directly enter the candidate set of the case library; typical dialogues: the user's words belong to the target scenario, but the customer service response has one of the following problems: incomplete information; inconsistent character style; insufficient scene adaptability; useless dialogues: automatically eliminated.

[0011] Preferably, in the second stage, the system determines through a scenario classifier that the scenario to which it belongs is a vehicle model consultation scenario or a car purchase policy scenario. When the scenario to which it belongs is determined to be a vehicle model consultation scenario, the product consultant role engine and the technical engineer engine are matched; when the scenario to which it belongs is determined to be a car purchase policy scenario, the policy interpretation specialist engine and the financial consultant engine are matched; when the scenario to which it belongs is determined to be a customer service scenario, the customer service support role engine is matched.

[0012] Preferably, the product consultant role engine focuses on the professional expression of vehicle parameters, calls the vehicle configuration library to supplement technical details, and adds horizontal comparison data; the policy interpretation specialist engine gives priority to breaking down applicable conditions and interest points when parsing policy terms, links the latest subsidy documents, and quantifies actual user benefits; the customer service support role engine reorganizes the speech according to standardized service processes, adds user care statements, and ensures that the response has both process clarity and emotional affinity.

[0013] Preferably, in the second stage, each role engine runs in parallel, generating 2-3 rewritten versions focusing on different dimensions for the same user's speech, forming a set of candidate responses that include multiple role perspectives, and providing rich decision-making materials for subsequent comprehensive evaluation; finally, the optimal version is automatically screened using preset speech quality indicators to form a standardized case text.

[0014] Preferably, the controllable dialogue case generation method can also use a domain-specific large model trained on automotive industry dialogue data to intelligently screen the original dialogues. The model has a deep understanding of professional terminology, user core needs and high-quality dialogue features in automotive scenarios, and can better screen out dialogue cases that meet the requirements.

[0015] Preferably, the COT information in the third stage includes decision basis, reasoning steps, and knowledge reference path.

[0016] Compared with the prior art, the present invention has the following beneficial effects: 1. Automatically generate "excellent dialogue cases + COT information": Ensure that the generated dialogue cases meet the requirements of logical coherence, consistent roles, and focused themes through technical solutions, and simultaneously generate matching COT information (such as decision-making basis, reasoning steps, knowledge citation paths) to provide standardized training materials for few-shot learning of large models. 2. Build a deep association mechanism between COT information and case content: Enable the generated COT information to accurately reflect the reasoning process of each key decision point in the dialogue case, forming a complete link of "case text → COT reasoning → knowledge support" to improve the learning efficiency of large models for complex dialogue logic. 3. Eliminate the bottleneck of large-scale generation relying on manual labor: Replace manual screening and information supplementation through an automated quality control and COT generation module to rapidly expand the high-quality case library and meet the few-shot learning needs of large models in different scenarios (such as customer service, education, and healthcare). BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is a flowchart of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0018] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0019] Embodiment 1 As Figure 1 shown, a method for generating controllable dialogue cases based on a large model constructs a full-process automated dialogue case processing system based on a three-stage pipeline of "screening → optimization → generating COT", and the core technical route is as follows: The first stage: Multi-dimensional dialogue screening 1. Construction of a dialogue quality rating model Input layer: Receive original dialogue data (such as customer service chat records, user forum Q&A), preprocess it through a large model prompt, and extract key features (such as dialogue turns, role labels, topic keywords).

[0020] Rating dimensions (implemented through multi-classification tasks of large models): Logical coherence: Detect whether the semantic jumps between dialogue turns are reasonable; Thematic relevance: Determine whether the dialogue focuses on the target scenario (such as topics like "vehicle model parameters", "purchase policies", "after-sales services" in the automotive industry); Role compliance: Verify whether the role language conforms to the preset personality Information value: Evaluate whether the conversation contains effective decision-making reference content 2. Hierarchical screening strategy Excellent dialogue: directly enter the candidate set of the case library, such as customer service dialogue that includes a complete problem-solving link (user question → demand analysis → accurate response); Typical conversation: The user's words belong to the target scenario, but the customer service response has one of the following problems: Incomplete information (e.g. only answering “yes” without specifying subsidy conditions); Inconsistent character styles (e.g. customer service staff use overly technical language, deviating from the “friendly” persona); Insufficient scenario adaptability (e.g., the regional subsidies where the user is located are not mentioned in the car purchase policy scenario); Useless conversations: Automatically remove conversations such as repetitive small talk, confusing topics, and conversations containing offensive language.

[0021] Phase 2: Parallel rewriting with multiple roles and multiple focuses After a typical conversation enters the rewriting module, the system first determines the scenario to which it belongs (such as car model consultation, car purchase policy, etc.) through the scenario classifier, and automatically matches multiple corresponding roles based on the scenario-role mapping rules (such as car model consultation scenario matches "product consultant" and "technical engineer", car purchase policy scenario matches "policy interpretation specialist" and "financial consultant", etc.). Each role corresponds to an independent rewriting engine, which performs differentiated rewriting based on its exclusive language style template: Product consultant role engine: Focus on the professional expression of vehicle parameters, call the vehicle configuration library to supplement technical details (such as "NEDC range" and "motor peak power"), and add horizontal comparison data (such as "range is improved by 15% compared with the same level of models"); Policy interpretation specialist engine: When analyzing policy terms, it prioritizes the applicable conditions and benefits, links the latest subsidy documents (such as "after the national subsidy is reduced, the local government will provide an additional 10,000 yuan subsidy for new energy vehicle purchases"), and quantifies the actual benefits of users; Customer Service Support Role Engine: Reorganize the script according to the standardized service process and add user care statements (such as "Your warranty application has been accepted and an inspection is expected to be arranged within 24 hours") to ensure that the response has both process clarity and emotional affinity.

[0022] Each role engine runs in parallel, generating 2-3 rewritten versions focusing on different dimensions (such as parameter detailed version, policy analysis version, and service guidance version) for the same user's speech, forming a set of candidate responses that include multiple role perspectives, providing rich decision-making materials for subsequent comprehensive evaluation.

[0023] Finally, the optimal version is automatically screened using preset discourse quality indicators (such as the coverage rate of professional terms and the user satisfaction prediction model) to form a standardized case text.

[0024] Phase 3: COT Information Generation After the dialogue screening and rewriting are completed, the system enters the stage of generating Chain of Thought (COT) information. This stage focuses on the explicit analysis of the dialogue decision logic, collaboratively extracts key decision nodes and reasoning paths through multiple reasoning models, and combines with a quality control mechanism to ensure the logical integrity and accuracy of the COT information, providing a traceable thinking reference for few-shot learning of large models.

[0025] Example 2 The method for generating controllable dialogue cases based on a large model can use a domain-specific large model trained on automotive industry dialogue data to intelligently screen the original dialogue. This model deeply understands the professional terms, core user needs, and high-quality dialogue features in automotive scenarios, and is more capable of screening out dialogue cases that meet the requirements.

[0026] Finally, it should be noted that the above are only the preferred embodiments of the present invention and are not used to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some of the technical features. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A method for generating controllable dialogue cases based on large models, characterized in that: Based on the three-stage pipeline of "screening → optimization → generation of COT", a fully automated dialogue case processing system is built. The specific steps are as follows: Phase 1: Multi-dimensional dialogue screening, including: dialogue quality rating model construction and hierarchical screening strategy; Phase II: Parallel rewriting of multiple roles and multiple focuses, as follows: After a typical dialogue enters the rewriting module, the system first determines the scene to which it belongs through the scene classifier, and automatically matches multiple corresponding roles based on the scene-role mapping rules. Each role corresponds to an independent rewriting engine, which performs differentiated rewriting based on its exclusive language style template; The third stage: COT information generation, as follows: After completing the dialogue screening and rewriting, the system enters the thinking chain information (COT) generation stage. This stage focuses on the explicit analysis of the dialogue decision logic, and collaboratively extracts key decision nodes and reasoning paths through multiple reasoning models. It also combines quality control mechanisms to ensure the logical integrity and accuracy of COT information, providing traceable thinking references for large-model few-shot learning.

2. The method for generating controllable dialogue cases based on large models according to claim 1, wherein: The construction of the dialogue quality rating model in the first stage includes: input layer: receiving the original dialogue data, preprocessing it through the large model prompt, and extracting key features; rating dimensions: implemented through large model multi-classification tasks; logical coherence: detecting whether the semantic jumps between dialogue rounds are reasonable; topic relevance: judging whether the dialogue revolves around the target scenario; role compliance: verifying whether the role language conforms to the preset personality; information value: evaluating whether the dialogue contains effective decision-making reference content.

3. The method for generating controllable dialogue cases based on a large model according to claim 1, wherein: The hierarchical screening strategies in the first stage include: Excellent dialogues: directly enter the candidate set of the case library; Typical dialogues: the user's words belong to the target scenario, but the customer service response has one of the following problems; incomplete information; inconsistent character style; insufficient scene adaptability; Useless dialogues: automatically eliminated.

4. The method for generating controllable dialogue cases based on large models according to claim 1, wherein: In the second stage, the system uses the scenario classifier to determine whether the scenario it belongs to is a vehicle model consultation scenario or a car purchase policy scenario. When it is determined that the scenario it belongs to is a vehicle model consultation scenario, it matches the product consultant role engine and the technical engineer engine; when it is determined that the scenario it belongs to is a car purchase policy scenario, it matches the policy interpretation specialist engine and the financial consultant engine; when it is determined that the scenario it belongs to is a customer service scenario, it matches the customer service support role engine.

5. The method for generating controllable dialogue cases based on a large model according to claim 4, wherein: The product consultant role engine focuses on the professional expression of vehicle parameters, calls the vehicle configuration library to supplement technical details, and adds horizontal comparison data; the policy interpretation specialist engine prioritizes the breakdown of applicable conditions and interest points when parsing policy terms, links the latest subsidy documents, and quantifies the actual benefits of users; the customer service support role engine reorganizes the script according to standardized service processes, adds user care statements, and ensures that the response has both process clarity and emotional affinity.

6. The method for generating controllable dialogue cases based on a large model according to claim 5, wherein: In the second stage, each role engine runs in parallel, generating 2-3 rewritten versions focusing on different dimensions for the same user's words, forming a set of candidate responses that include multiple role perspectives, providing rich decision-making materials for subsequent comprehensive evaluation; finally, the preset speech quality indicators are used to automatically select the optimal version to form a standardized case text.

7. The method for generating controllable dialogue cases based on a large model according to claim 1, characterized in that: The controllable dialogue case generation method can also use a domain-specific large model trained on automotive industry dialogue data to intelligently screen the original dialogue. This model deeply understands the professional terms, core user needs, and high-quality dialogue features in the automotive scenario, and is more capable of screening out dialogue cases that meet the requirements.

8. The method for generating controllable dialogue cases based on a large model according to claim 1, wherein: In the third stage, the COT information includes the basis for decision-making, reasoning steps, and knowledge citation path.

Images