Thinking chain data generation method and device, electronic equipment and readable medium

Abstract

The invention provides a thinking chain data generation method and device, electronic equipment and a readable medium, and belongs to the technical field of big data. According to the method, an original problem in a business scene in a target industry can be adopted to obtain corresponding description information; inputting the description information of the original problem into the first large model to obtain a reasoning path corresponding to the original problem output by the first large model; the industry knowledge corresponding to the target industry is fused through the reasoning path to obtain thinking chain data corresponding to the original problem; wherein the first large model is obtained through pre-training based on a historical reasoning scheme related to the target industry. According to the method, a reasoning path is automatically generated based on a pre-trained first large model on the basis of an original problem in a specific business scene in the industry, industry knowledge is further fused, and complete and accurate thinking chain data is efficiently generated, so that development and training of the large model in the professional field are supported by high-quality thinking chain data, and the development and training efficiency is improved. And the large model performance and the reasoning service quality are improved.

Description

Technical Field

[0001] The present disclosure belongs to the field of big data technology, and specifically relates to a method for generating thought chain data, a device for generating thought chain data, an electronic device, and a computer-readable medium. Background Art

[0002] The Chain of Thought (CoT) is the reasoning process that a large model uses to handle complex problems, from initial observation to problem breakdown, logical reasoning, and outputting conclusions. The Chain of Thought includes not only the output answer but also the intermediate steps, as well as the decision-making process and rationale for each step, providing transparent and explainable logical support for the large model.

[0003] On this basis, the further development and training of large models also require the support of thinking chain data. Therefore, accurately and efficiently synthesizing thinking chain data that meets business needs is an urgent problem to be solved in this field. Summary of the Invention

[0004] The purpose of the embodiments of the present disclosure is to provide a thought chain data generation method, a thought chain data generation device, an electronic device and a computer-readable medium, which can automatically, accurately and efficiently synthesize thought chain data that meets business needs to support further research and development and training of large models.

[0005] In order to solve the above technical problems, the present disclosure is implemented as follows:

[0006] In the first aspect, the present disclosure provides a method for generating thought chain data, which may include: collecting original problems in at least one business scenario in the target industry; obtaining descriptive information of the original problem; inputting the descriptive information corresponding to the original problem into a first large model, and obtaining the reasoning path corresponding to the original problem output by the first large model; integrating the industry knowledge corresponding to the target industry with the reasoning path to obtain thought chain data corresponding to the original problem; wherein the first large model is obtained by pre-training based on historical reasoning schemes related to the target industry.

[0007] Optionally, the pre-training steps of the first large model are as follows: collect historical reasoning schemes related to the target industry, which include descriptive information and reasoning paths corresponding to historical problems; use the descriptive information corresponding to historical problems as input and the reasoning paths corresponding to historical problems as output to constitute training data, and pre-train the first large model.

[0008] Optionally, after integrating the industry knowledge corresponding to the target industry with the reasoning path to obtain the thinking chain data corresponding to the original problem, it also includes: performing quality assessment on the thinking chain data based on business indicators; and screening the thinking chain data according to the evaluation results of the quality assessment.

[0009] Optionally, the business indicator includes at least one of reasoning completeness, format correctness, answer correctness, and language uniformity.

[0010] Optionally, screening the thought chain data according to the evaluation results of the quality assessment includes: when the evaluation results of the quality assessment do not meet the sample application conditions, eliminating the thought chain data; or, when the evaluation results of the quality assessment do not meet the sample application conditions, marking the thought chain data as to be modified.

[0011] Optionally, obtaining the descriptive information of the original question includes: inputting the original question into the second largest model, and obtaining the descriptive information corresponding to the original question output by the second largest model.

[0012] Optionally, the industry knowledge corresponding to the target industry is integrated into the reasoning path to obtain the thinking chain data corresponding to the original problem, including: retrieving the corresponding industry knowledge for at least one execution step in the reasoning path; inputting the reasoning path and the industry knowledge into the third largest model, and obtaining the thinking chain data output after the third largest model integrates the execution steps and the corresponding industry knowledge.

[0013] Optionally, after collecting original questions in at least one business scenario in the target industry, the method further includes: performing sample amplification on the original questions, where the sample amplification includes at least one of synonym rewriting, word order adjustment, and sentence structure conversion.

[0014] Optionally, the description information includes at least one of a question type and a key tag.

[0015] Optionally, the reasoning path includes an execution step.

[0016] Optionally, the description information also includes at least one of a problem description and a scenario description.

[0017] Optionally, the reasoning path further includes a step description corresponding to at least one execution step.

[0018] In the second aspect, the present disclosure provides a thinking chain data generation device, which may include: a problem collection module, used to collect original problems in at least one business scenario in the target industry; a problem description module, used to obtain descriptive information of the original problem; a path reasoning module, used to input the descriptive information corresponding to the original problem into the first large model, and obtain the reasoning path corresponding to the original problem output by the first large model; a thinking chain generation module, used to integrate the industry knowledge corresponding to the target industry with the reasoning path to obtain the thinking chain data corresponding to the original problem; wherein, the first large model is obtained by pre-training based on the historical reasoning scheme related to the target industry.

[0019] Optionally, the device may also include a first large model training module, which is used to collect historical reasoning schemes related to the target industry, wherein the historical reasoning schemes include descriptive information and reasoning paths corresponding to historical problems; the descriptive information corresponding to historical problems is used as input, and the reasoning paths corresponding to historical problems are used as output to constitute training data, and the first large model is pre-trained.

[0020] Optionally, the device may further include a quality assessment module for performing quality assessment on the thought chain data based on business indicators; and screening the thought chain data according to the assessment result of the quality assessment.

[0021] Optionally, the business indicator includes at least one of reasoning completeness, format correctness, answer correctness, and language uniformity.

[0022] Optionally, the quality assessment module is specifically used to eliminate the thought chain data when the evaluation result of the quality assessment does not meet the sample application conditions; or, when the evaluation result of the quality assessment does not meet the sample application conditions, mark the thought chain data as to be modified.

[0023] Optionally, the problem description module is specifically used to input the original problem into the second largest model, and obtain description information corresponding to the original problem output by the second largest model.

[0024] Optionally, the thinking chain generation module is specifically used to retrieve corresponding industry knowledge for at least one execution step in the reasoning path; input the reasoning path and industry knowledge into the third largest model, and obtain the thinking chain data output after the third largest model integrates the execution steps and the corresponding industry knowledge.

[0025] Optionally, the question collection module is further used to perform sample amplification on the original question, where the sample amplification includes at least one of synonym rewriting, word order adjustment, and sentence structure conversion.

[0026] Optionally, the description information includes at least one of a question type and a key tag.

[0027] Optionally, the reasoning path includes an execution step.

[0028] Optionally, the description information also includes at least one of a problem description and a scenario description.

[0029] Optionally, the reasoning path further includes a step description corresponding to at least one execution step.

[0030] In a third aspect, the present disclosure provides an electronic device comprising a processor, a memory, and a program or instruction stored in the memory and executable on the processor, wherein the program or instruction, when executed by the processor, implements the steps of the thought chain data generation method of the first aspect.

[0031] In a fourth aspect, the present disclosure provides a computer-readable medium storing a program or instruction, which, when executed by a processor, implements the steps of the thought chain data generation method of the first aspect.

[0032] In a fifth aspect, the present disclosure provides a chip comprising a processor and a communication interface, wherein the communication interface and the processor are coupled, and the processor is used to run programs or instructions to implement the steps of the thinking chain data generation method of the first aspect.

[0033] In a sixth aspect, the present disclosure provides a computer program product comprising instructions, which, when executed on a computer, enables the computer to execute the steps of the method for generating thought chain data according to the first aspect.

[0034] The thought chain data generation method provided by the present disclosure can adopt the original problem in the business scenario of the target industry and obtain the descriptive information corresponding to the original problem; on this basis, the descriptive information of the original problem is input into the first large model, and the reasoning path corresponding to the original problem output by the first large model is obtained; and then the reasoning path is integrated with the industry knowledge corresponding to the target industry to obtain the thought chain data corresponding to the original problem; wherein, the first large model is pre-trained based on the historical reasoning scheme related to the target industry. In the embodiment of the present disclosure, based on the original problem in the specific business scenario within the industry, the reasoning path can be automatically generated based on the pre-trained first large model, and the industry knowledge can be further integrated to efficiently generate complete and accurate thought chain data, thereby supporting the development and training of large models in professional fields with high-quality thought chain data, reducing the dependence and cost on manual labeling, and improving the performance of large models and the quality of reasoning business. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 This is a flowchart of one of the steps of the method for generating thought chain data provided in an embodiment of the present disclosure;

[0036] Figure 2 A flowchart of the pre-training steps of the first large model provided in an embodiment of the present disclosure;

[0037] Figure 3 This is a second flowchart of the method for generating thought chain data according to an embodiment of the present disclosure;

[0038] Figure 4 A structural block diagram of a thought chain data generating device provided in an embodiment of the present disclosure;

[0039] Figure 5 A schematic structural diagram of an electronic device provided in an embodiment of the present disclosure;

[0040] Figure 6 A hardware schematic diagram of an electronic device provided in an embodiment of the present disclosure. DETAILED DESCRIPTION

[0041] The following will clearly and completely describe the technical solutions in the embodiments of the present disclosure in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present disclosure, not all of them. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present disclosure without making any creative efforts shall fall within the scope of protection of the present disclosure.

[0042] The terms "first", "second", etc. in the specification and claims of the present disclosure are used to distinguish similar objects, and are not used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable under appropriate circumstances, so that the embodiments of the present disclosure can be implemented in an order other than those illustrated or described herein, and the objects distinguished by "first", "second", etc. are generally of the same type, and the number of objects is not limited. For example, the first object can be one or more. In addition, "and / or" in the specification and claims represents at least one of the connected objects, and the character " / " generally indicates that the objects related to each other are in an "or" relationship.

[0043] Figure 1 This is one of the flow charts of the method for generating thought chain data provided by the embodiment of the present disclosure. Figure 1 As shown, the method may include the following steps 101 to 104.

[0044] Step 101: Collect original questions in at least one business scenario in the target industry.

[0045] In the disclosed embodiment, the generation of thinking chain data starts with the determination of the problem. Based on the needs of business scenarios, model development, etc., original problems with industry expertise and business specificity can be collected in one or more business scenarios of the target industry. Among them, the target industry can be determined based on the needs of large-scale model development in multiple industries based on historical experience, business characteristics or custom divisions. For example, finance can be determined as the target industry in multiple industries such as finance, medical care, education, media, industrial manufacturing, and infrastructure. Business scenarios can be scenarios related to different institutions, enterprises, customer groups, and business needs in the industry. For example, taking the financial industry as an example, it can include marketing scenarios, risk control scenarios, etc. Taking the infrastructure industry as an example, it can include new construction scenarios, expansion and reconstruction scenarios, relocation scenarios, and restoration scenarios, etc.

[0046] Step 102: Obtain description information of the original question.

[0047] In the disclosed embodiments, the descriptive information may be information that expands, supplements, or refines the original question. The type and content of the descriptive information may be selected based on actual business needs and data processing conditions. The descriptive information may be obtained through extended retrieval, tag matching, model annotation, manual annotation, and other methods.

[0048] Step 103: Input the description information corresponding to the original question into the first large model, and obtain the reasoning path corresponding to the original question output by the first large model.

[0049] In the disclosed embodiment, the description information corresponding to the original problem can be input into the first large model, and the first large model performs reasoning and planning based on the description information to output a reasoning path corresponding to the original problem. The reasoning path may include a step-by-step solution to the original problem.

[0050] Step 104: Integrate the industry knowledge corresponding to the target industry with the reasoning path to obtain the thinking chain data corresponding to the original problem.

[0051] In the disclosed embodiment, the reasoning path provides a basic framework for solving the original problem, in which there may be situations where the description is vague, brief or wrong. On this basis, the description in the reasoning path can be further refined, supplemented or corrected by integrating the industry knowledge corresponding to the target industry to obtain the thinking chain data corresponding to the original problem. Exemplarily, the entity described in the reasoning path can be supplemented with information about the entity based on the integration of industry knowledge; the indicator referenced in the reasoning path can be supplemented with the representation meaning and calculation method of the indicator based on the integration of industry knowledge. On this basis, the thinking chain data uses the reasoning path as the basic reasoning framework, and further includes complete and rich analysis content and professional judgment, thereby improving data quality, and can better support the research and development and training of large models in line with business needs for the target industry.

[0052] Among them, the first model is pre-trained based on historical reasoning solutions related to the target industry.

[0053] In the disclosed embodiments, the first large model can be pre-trained based on historical reasoning solutions related to the target industry. These historical reasoning solutions can include historical problems and their corresponding practical solutions, large model reasoning solutions, and the like. Those skilled in the art can collect corresponding historical reasoning solutions based on business needs to pre-train the first large model, and this disclosed embodiment does not impose any specific limitations on this.

[0054] In an optional embodiment of the present disclosure, the pre-training step of the first large model in the aforementioned step 103 can be as follows: Figure 2 As shown, the process includes the following steps 201 to 202.

[0055] Step 201: Collect historical reasoning solutions related to the target industry, where the historical reasoning solutions include description information and reasoning paths corresponding to historical problems.

[0056] In the disclosed embodiments, historical reasoning solutions can be obtained by collecting and standardizing actual solutions related to the target industry, expert historical experience, business documents, industry cases, and other large-scale model reasoning solutions. Standardization can be achieved by using natural language processing (NLP) and annotation to determine the descriptive information corresponding to the historical problem and the reasoning path of the historical problem as the historical reasoning solution.

[0057] Step 202: Use the descriptive information corresponding to the historical question as input and the reasoning path corresponding to the historical question as output to constitute training data, and pre-train the first large model.

[0058] In the disclosed embodiments, during the pre-training process of the first large model, the descriptive information corresponding to the historical question can be used as input, and the reasoning path corresponding to the historical question can be used as output to form training data, allowing the first large model to be structured and fine-tuned based on the training data. Based on this, the pre-trained first large model can dynamically reason and plan based on the question descriptive information, generating a reasoning path for the question.

[0059] Figure 3 This is the second step flow chart of the method for generating thought chain data provided by the embodiment of the present disclosure. Figure 3 As shown, the method may include the following steps 301 to 306.

[0060] Step 301: Collect original questions in at least one business scenario in the target industry.

[0061] In the embodiment of the present disclosure, step 301 may correspond to the relevant description of the aforementioned step 101, and will not be described again here to avoid repetition.

[0062] In the embodiment of the present disclosure, the original question can be collected from sources such as business systems, historical reasoning data, and relevant industry reports, and the embodiment of the present disclosure does not impose specific restrictions on this.

[0063] In an optional embodiment of the present disclosure, the following step A may be further included after step 301 .

[0064] Step A: amplify the sample of the original question, where the sample amplification includes at least one of synonym rewriting, word order adjustment, and sentence structure conversion.

[0065] In the disclosed embodiments, sample augmentation can be performed on the original questions collected to increase data coverage and diversity. Sample augmentation can be achieved through templated and regularized text processing, or through text generation using a large language model. For example, sample augmentation can include synonym rewriting, word order adjustment, and sentence structure conversion.

[0066] Step 302: Obtain description information of the original question.

[0067] In the embodiment of the present disclosure, step 302 may correspond to the relevant description of the aforementioned step 102, and will not be described again here to avoid repetition.

[0068] In an optional embodiment of the present disclosure, the description information includes at least one of a question type and a key tag.

[0069] In the embodiment of the present disclosure, the description information may be the question type of the original question. In the business scenario of a specific target industry, different question types may be divided according to business needs and purposes, thereby categorizing the original question.

[0070] For example, the original question of the infrastructure industry, “How to choose an address when building public facilities in the urban public domain?” can be of the question type “Public facility planning and site selection.”

[0071] In the embodiment of the present disclosure, the description information may be a key tag. The key tag may be one or more keywords extracted from the original question, or may include other keywords that are expanded based on the keywords extracted from the original question and are related to technology or field.

[0072] For example, the key tags of the original question of the infrastructure industry, “How to choose the address when building public facilities in the urban public domain?” can be “public facility planning, site selection decision, GIS (Geographic Information System)”.

[0073] In an optional embodiment of the present disclosure, the description information further includes at least one of a problem description and a scenario description.

[0074] In the disclosed embodiment, the description information may also include a description of the question, such as the question content, purpose, requirements for answers, possible technologies and fields involved, and industry-related evaluations.

[0075] For example, the original problem of the infrastructure industry, "How to choose the address when building public facilities in the urban public domain?" is described as follows:

[0076] This derivation scheme aims to provide a systematic reasoning path for the site selection of urban public facilities, from demand research, data collection, spatial screening, comprehensive evaluation, to final decision-making. Through quantitative data analysis, GIS spatial decision-making, multi-metric comprehensive scoring, and on-site verification, combined with expert and public participation, a closed-loop feedback mechanism is established to ensure that site selection decisions are scientifically sound and highly adaptable to the industry.

[0077] In the embodiment of the present disclosure, the description information may further include a scenario description, such as an extended description of the applicable scenario of the problem.

[0078] For example, the original problem of the infrastructure industry, “How to choose the address when building public facilities in the urban public domain?” is described as follows:

[0079] It is suitable for use by city governments, planning departments and related consulting agencies when formulating site selection plans for new public facilities (such as community centers, public libraries, medical centers, etc.).

[0080] In an optional embodiment of the present disclosure, step 302 may include the following step B.

[0081] Step B: input the original question into the second largest model, and obtain the descriptive information corresponding to the original question output by the second largest model.

[0082] In the disclosed embodiment, the original question can be input into the second largest model. After the second largest model deeply analyzes the original question and abstractly extracts key information, the second largest model outputs a description of the original question. The format and content of the description information can be referred to in the previous description and will not be repeated here to avoid repetition.

[0083] Step 303: Input the description information corresponding to the original question into the first large model, and obtain the reasoning path corresponding to the original question output by the first large model.

[0084] In the embodiment of the present disclosure, step 303 may correspond to the relevant description of the aforementioned step 103, and will not be described again here to avoid repetition.

[0085] In an optional embodiment of the present disclosure, the reasoning path includes an execution step.

[0086] In the disclosed embodiments, a reasoning path may include one or more execution steps, and executing the aforementioned execution steps in sequence along the reasoning path may serve as a solution to the aforementioned original problem. When the number of execution steps in the reasoning path is greater than or equal to 2, the different execution steps may have reasoning logic between them, thereby expressing an execution order, such as a sequential order of execution, a parallel order, etc.

[0087] For example, the steps for solving the original problem of the infrastructure industry, "How to choose a site when building public facilities in the urban public domain?", can be as follows:

[0088] Step 1: Demand research and planning goal confirmation.

[0089] Step 2: Data collection, integration and spatial preprocessing.

[0090] Step 3: Preliminary screening and spatial modeling.

[0091] Step 4: Multi-factor comprehensive evaluation and decision-making model construction.

[0092] Step 5: Confirmation of site selection decision and formulation of implementation plan.

[0093] In an optional embodiment of the present disclosure, the reasoning path further includes a step description corresponding to at least one execution step.

[0094] In the embodiment of the present disclosure, the execution steps of the reasoning path may also include step descriptions corresponding to one or more execution steps. The step descriptions may be an extension, refinement, and supplement to the execution steps, specifically describing the purpose, object, and action of the execution steps, as well as other required information.

[0095] For example, the original question of the capital construction industry, "How to choose an address when building public facilities in the urban public domain?" can be solved by the following steps and step descriptions:

[0096] Step 1: Demand research and planning goal confirmation.

[0097] Steps: Gather city development plans, regional economic data, population distribution, existing public service layout, and resident needs research reports to clarify the target functions, service scope, and key indicators of new public facilities. At the same time, through questionnaires and expert discussions, combined with planning documents, a detailed needs analysis report and planning objectives will be developed.

[0098] Step 2: Data collection, integration and spatial preprocessing.

[0099] Steps: After confirming requirements, conduct comprehensive data collection using multiple sources (such as transportation networks, land use, and population), cleanse, normalize, and spatialize the data. Combined with a GIS platform, perform spatial overlay and visualization analysis of these data types, building a unified, standardized dataset to provide accurate data support for subsequent preliminary screening and quantitative evaluation of candidate areas.

[0100] Step 3: Preliminary screening and spatial modeling.

[0101] Step Description: Based on the spatial data obtained in Step 2, set basic site selection criteria (such as accessibility, environmental safety, land policies, and development potential). Utilize an automated spatial decision-making model to conduct a preliminary screening of regions and identify several potential candidate sites. This stage combines quantitative thresholds with expert scoring to automatically eliminate areas that do not meet the criteria, laying the foundation for a comprehensive, multi-criteria evaluation.

[0102] Step 4: Multi-factor comprehensive evaluation and decision-making model construction.

[0103] Steps: For the initially selected candidate areas, a multi-criteria decision analysis method (such as AHP, TOPSIS, or GIS multi-criteria decision evaluation) is used to quantitatively score and weight each area's accessibility, resource utilization, construction costs, environmental impact, and future development potential. Combined with the qualitative opinions of the expert reviewers, a comprehensive scoring model is established to generate a ranking of the candidate areas, clarifying each area's competitive advantages and disadvantages, and forming a decision-making model based on comprehensive data and expert judgment.

[0104] Step 5: Confirmation of site selection decision and formulation of implementation plan.

[0105] Steps: Based on the aforementioned comprehensive assessment results, the on-site inspection report, and expert and public feedback, the optimal location for public facility construction will be selected and a detailed implementation plan will be developed. This plan will include a site selection report, feasibility analysis, budget and construction planning, risk warning measures, and subsequent operational management recommendations.

[0106] Step 304: Integrate the industry knowledge corresponding to the target industry with the reasoning path to obtain the thinking chain data corresponding to the original problem.

[0107] In the embodiment of the present disclosure, step 304 may correspond to the relevant description of the aforementioned step 104, and will not be described again here to avoid repetition.

[0108] In an optional embodiment of the present disclosure, step 304 may correspond to the following steps C1 to C2.

[0109] Step C1: Retrieve corresponding industry knowledge for at least one execution step in the reasoning path.

[0110] In the embodiment of the present disclosure, industry knowledge bases corresponding to different industries can be pre-built to integrate the corresponding relevant industry knowledge. On this basis, after obtaining the reasoning path, the corresponding industry knowledge can be retrieved in the corresponding industry knowledge base for the execution step. Among them, the corresponding industry knowledge can be retrieved separately for each execution step, or the corresponding industry knowledge can be selectively retrieved for some execution steps according to actual needs, such as selecting the execution step whose text length is greater than or equal to the length threshold, the number of entity nouns contained is greater than or equal to the number threshold, and the execution step is determined to be a key step based on the reasoning logic.

[0111] Step C2: input the reasoning path and industry knowledge into the third model, and obtain the thinking chain data output after the third model integrates the execution steps and the corresponding industry knowledge.

[0112] In the disclosed embodiment, the reasoning path and industry knowledge can be input into the third model, which will analyze and integrate the industry knowledge and the corresponding execution steps in the reasoning path, fill in text details, supplement the extended description content, and perform verification and correction processing, thereby outputting complete and accurate thinking chain data.

[0113] Among them, the first model is pre-trained based on historical reasoning solutions related to the target industry.

[0114] In the embodiment of the present disclosure, the first large model can refer to the aforementioned Figure 1 To avoid repetition, the description of the first model will not be repeated here.

[0115] Step 305: Perform quality assessment on the thinking chain data based on business indicators.

[0116] In the embodiments of the present disclosure, a quality assessment can be performed based on the obtained thought chain data to determine whether the thought chain data meets the requirements of subsequent development and training of large models. One or more business indicators can be set to evaluate the quality of thought chain data. Those skilled in the art can select the content and number of business indicators based on actual needs, and the embodiments of the present disclosure do not impose specific limitations on this.

[0117] In an optional embodiment of the present disclosure, the business indicator includes at least one of reasoning completeness, format correctness, answer correctness, and language uniformity.

[0118] In the embodiment of the present disclosure, the business indicator may be reasoning completeness, which can be used to evaluate whether the execution steps in the reasoning path are complete, whether there are omissions of key steps in the reasoning logic, and whether there are incomplete text structures.

[0119] In the embodiment of the present disclosure, the business indicator may be format correctness, and in this case, it is possible to detect whether there is a structured separation in the generated thought chain data. For example, whether the reasoning path is correctly embedded in the standard label (such as <think>< / think> ) in the package area, whether the answers are displayed separately and whether the data format meets the predefined standardized format requirements.

[0120] In the disclosed embodiment, the business indicator may be the correctness of the answer. In this case, attention may be paid to the output answer, such as evaluating the correctness of the output answer by comparing it with the correct reference answer, or evaluating the correctness of the output answer by using a model cross-validation method to obtain the accuracy of the output answer.

[0121] In the embodiment of the present disclosure, the business indicator can be language uniformity. At this time, it can be detected whether the text of the thinking chain data maintains the output of a single language type to avoid mixing of multiple languages, thereby ensuring the consistency and professionalism of the thinking chain data in the language expression style.

[0122] Step 306: Filter the thought chain data according to the quality assessment results.

[0123] In the disclosed embodiment, the evaluation results of the quality assessment may include whether the business indicators of the thinking chain data meet the requirements of subsequent development and training of the large model. When screening based on the evaluation results, if it meets the requirements, it can be retained to support the subsequent related development of the large model.

[0124] In an optional embodiment of the present disclosure, step 306 includes the following step D.

[0125] Step D: When the evaluation result of the quality assessment does not meet the sample application conditions, the thinking chain data will be eliminated.

[0126] In the embodiment of the present disclosure, the sample application condition may be a critical condition set based on whether the business indicator meets the requirements of subsequent research and development and training of the large model. When the business indicator meets the sample application condition, it can be considered to meet the requirements, otherwise it does not meet the requirements. For example, when the business indicator is reasoning integrity, the sample application condition may be that the reasoning integrity is greater than or equal to the integrity threshold. Then, when the reasoning integrity of the thinking chain data is less than the integrity threshold, it can be considered that the evaluation result of the quality evaluation of the thinking chain data does not meet the sample application condition. The same can be said for other business indicators.

[0127] In the disclosed embodiment, when the evaluation result of the quality assessment does not meet the sample application conditions, it can be considered that the thinking chain data does not meet the requirements of subsequent research and development and training of the large model, so the thinking chain data can be directly eliminated to avoid interference with the subsequent research and development and training of the large model.

[0128] In an optional implementation of the present disclosure, step 306 includes the following step E.

[0129] Step E: When the evaluation result of the quality evaluation does not meet the sample application conditions, the thinking chain data is marked as to be modified.

[0130] In the embodiment of the present disclosure, the sample application conditions may correspond to the relevant description of the aforementioned step D, and will not be described again here to avoid repetition.

[0131] In the embodiment of the present disclosure, when the evaluation result of the quality assessment does not meet the sample application conditions, it can be considered that the thinking chain data does not meet the requirements of subsequent research and development and training of the large model, so that the thinking chain data can be marked as to be modified. The thinking chain data can be modified based on the mark, such as modification based on the analysis and evaluation of the large model, or modification relying on manual review, etc.

[0132] The thought chain data generation method provided by the present disclosure can adopt the original problem in the business scenario of the target industry and obtain the descriptive information corresponding to the original problem; on this basis, the descriptive information of the original problem is input into the first large model, and the reasoning path corresponding to the original problem output by the first large model is obtained; and then the reasoning path is integrated with the industry knowledge corresponding to the target industry to obtain the thought chain data corresponding to the original problem; wherein, the first large model is pre-trained based on the historical reasoning scheme related to the target industry. In the embodiment of the present disclosure, based on the original problem in the specific business scenario within the industry, the reasoning path can be automatically generated based on the pre-trained first large model, and the industry knowledge can be further integrated to efficiently generate complete and accurate thought chain data, thereby supporting the development and training of large models in professional fields with high-quality thought chain data, reducing the dependence and cost on manual labeling, and improving the performance of large models and the quality of reasoning business.

[0133] Figure 4 This is a structural block diagram of a thinking chain data generation device 400 provided in an embodiment of the present disclosure. The device can be applied to, and the device may include: a problem collection module 401, used to collect original problems in at least one business scenario in the target industry; a problem description module 402, used to obtain descriptive information of the original problem; a path reasoning module 403, used to input the descriptive information corresponding to the original problem into the first large model, and obtain the reasoning path corresponding to the original problem output by the first large model; a thinking chain generation module 404, used to integrate the industry knowledge corresponding to the target industry with the reasoning path to obtain the thinking chain data corresponding to the original problem; wherein, the first large model is obtained by pre-training based on the historical reasoning scheme related to the target industry.

[0134] In an optional embodiment of the present disclosure, the device may also include a first large model training module for collecting historical reasoning schemes related to the target industry, wherein the historical reasoning schemes include descriptive information and reasoning paths corresponding to historical problems; the descriptive information corresponding to historical problems is used as input, and the reasoning paths corresponding to historical problems are used as output to constitute training data, and the first large model is pre-trained.

[0135] In an optional embodiment of the present disclosure, the device may further include a quality assessment module for performing quality assessment on the thought chain data based on business indicators; and screening the thought chain data according to the assessment result of the quality assessment.

[0136] In an optional embodiment of the present disclosure, the business indicator includes at least one of reasoning completeness, format correctness, answer correctness, and language uniformity.

[0137] In an optional embodiment of the present disclosure, the quality assessment module is specifically used to eliminate the thinking chain data when the evaluation result of the quality assessment does not meet the sample application conditions; or, when the evaluation result of the quality assessment does not meet the sample application conditions, mark the thinking chain data as to be modified.

[0138] In an optional embodiment of the present disclosure, the problem description module 402 is specifically configured to input the original problem into the second largest model, and obtain description information corresponding to the original problem output by the second largest model.

[0139] In an optional embodiment of the present disclosure, the thought chain generation module 404 is specifically used to retrieve corresponding industry knowledge for at least one execution step in the reasoning path; input the reasoning path and industry knowledge into the third largest model, and obtain the thought chain data output after the third largest model integrates the execution steps and the corresponding industry knowledge.

[0140] In an optional embodiment of the present disclosure, the question collection module 401 is further configured to perform sample amplification on the original question, where the sample amplification includes at least one of synonym rewriting, word order adjustment, and sentence pattern conversion.

[0141] In an optional embodiment of the present disclosure, the description information includes at least one of a question type and a key tag.

[0142] In an optional embodiment of the present disclosure, the reasoning path includes an execution step.

[0143] In an optional embodiment of the present disclosure, the description information further includes at least one of a problem description and a scenario description.

[0144] In an optional embodiment of the present disclosure, the reasoning path further includes a step description corresponding to at least one execution step.

[0145] The thought chain data generation device provided by the present disclosure can adopt the original problem in the business scenario of the target industry and obtain the descriptive information corresponding to the original problem; on this basis, the descriptive information of the original problem is input into the first large model, and the reasoning path corresponding to the original problem output by the first large model is obtained; and then the industry knowledge corresponding to the target industry is integrated with the reasoning path to obtain the thought chain data corresponding to the original problem; wherein, the first large model is pre-trained based on the historical reasoning scheme related to the target industry. In the embodiment of the present disclosure, based on the original problem in the specific business scenario within the industry, the reasoning path can be automatically generated based on the pre-trained first large model, and the industry knowledge can be further integrated to efficiently generate complete and accurate thought chain data, thereby supporting the development and training of large models in professional fields with high-quality thought chain data, reducing the dependence and cost on manual labeling, and improving the performance of large models and the quality of reasoning business.

[0146] Figure 5 This is a structural diagram of an electronic device 500 provided in an embodiment of the present disclosure, such as Figure 5 As shown, the electronic device 500 may include a processor 501, a memory 502, and a program or instruction stored in the memory 502 and executable on the processor 501. When the program or instruction is executed by the processor 501, the various processes of the above-mentioned signaling intercommunication embodiment are implemented and the same technical effect can be achieved. To avoid repetition, they will not be described here.

[0147] It should be noted that Figure 5 The electronic device 500 shown is only an example and should not limit the functions and scope of use of the embodiments of the present disclosure.

[0148] Figure 6 A hardware diagram of an electronic device 600 provided in an embodiment of the present disclosure is shown in FIG. Figure 6 As shown, electronic device 600 includes a central processing unit (CPU) 601, which can perform various appropriate actions and processes according to the program stored in ROM (Read Only Memory) 602 or the program loaded from storage part 608 into RAM (Random Access Memory) 603. Various programs and data required for system operation are also stored in RAM 603. CPU 601, ROM 602 and RAM 603 are connected to each other via bus 604. I / O (Input / Output) interface 605 is also connected to bus 604.

[0149] The following components are connected to the I / O interface 605: an input section 606 including a keyboard, a mouse, and the like; an output section 607 including devices such as a CRT (Cathode Ray Tube), an LCD (Liquid Crystal Display), and a speaker; a storage section 608 including devices such as a hard disk; and a communication section 609 including a network interface card such as a LAN (Local Area Network) card or a modem. The communication section 609 performs communication processing via a network such as the Internet. A drive 610 is also connected to the I / O interface 605 as needed. Removable media 611, such as a magnetic disk, an optical disk, a magneto-optical disk, or a semiconductor memory, is installed in the drive 610 as needed, so that computer programs read from the removable media can be installed in the storage section 608 as needed.

[0150] In particular, according to an embodiment of the present disclosure, the process described below with reference to the flowchart can be implemented as a computer software program. For example, an embodiment of the present disclosure includes a computer program product, which includes a computer program carried on a computer-readable medium, and the computer program includes a program code for executing the method shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from a network via the communication section 609, and / or installed from a removable medium 611. When the computer program is executed by the central processing unit (CPU 601), various functions defined in the system of the present application are executed.

[0151] The embodiment of the present disclosure also sends a computer-readable medium, on which a program or instruction is stored. When the program or instruction is executed by the processor, the various processes of the above-mentioned signaling intercommunication embodiment are implemented and the same technical effect can be achieved. To avoid repetition, it will not be repeated here.

[0152] The processor is the processor in the electronic device in the above embodiment. The computer-readable medium includes computer-readable media such as ROM, RAM, magnetic disk or optical disk.

[0153] The embodiment of the present disclosure also provides a chip, which includes a processor and a communication interface. The communication interface and the processor are coupled. The processor is used to run programs or instructions to implement the various processes of the above-mentioned signaling intercommunication embodiment and can achieve the same technical effect. To avoid repetition, it will not be repeated here.

[0154] It should be understood that the chip mentioned in the embodiments of the present disclosure can also be called a system-level chip, a system chip, a chip system, or a system-on-chip chip, etc.

[0155] The embodiment of the present disclosure provides a computer program product containing instructions, which, when executed on a computer, enables the computer to execute the steps of signaling intercommunication as described above and achieve the same technical effects. To avoid repetition, they will not be described here.

[0156] It should be noted that, in this article, the terms "comprise", "include" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the sentence "comprises a..." does not exclude the presence of other identical elements in the process, method, article or device comprising the element. In addition, it should be pointed out that the scope of the methods and devices in the embodiments of the present disclosure is not limited to performing functions in the order shown or discussed, and may also include performing functions in a substantially simultaneous manner or in the opposite order according to the functions involved. For example, the described method may be performed in an order different from that described, and various steps may also be added, omitted, or combined. In addition, the features described with reference to certain examples may be combined in other examples.

[0157] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus the necessary general hardware platform, and of course can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present disclosure, or the part that contributes to the prior art, can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk), and includes a number of instructions for enabling a terminal (which can be a mobile phone, computer, electronic device, air conditioner, or network device, etc.) to execute the methods of the various embodiments of the present disclosure.

[0158] The embodiments of the present disclosure are described above in conjunction with the accompanying drawings, but the present disclosure is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the guidance of the present disclosure, ordinary technicians in this field can also make many forms without departing from the scope of protection of the purpose of the present disclosure and the claims, all of which are protected by the present disclosure.

Claims

1. A method for generating thought chain data, characterized in that: The method comprises: Collect original questions from at least one business scenario in the target industry; Obtain description information of the original problem; Inputting the description information corresponding to the original problem into the first large model, and obtaining the reasoning path corresponding to the original problem output by the first large model; Integrating the industry knowledge corresponding to the target industry with the reasoning path to obtain the thinking chain data corresponding to the original problem; Among them, the first large model is obtained by pre-training based on historical reasoning solutions related to the target industry.

2. The method according to claim 1, characterized in that The pre-training steps of the first model are as follows: Collecting the historical reasoning solutions related to the target industry, wherein the historical reasoning solutions include description information and reasoning paths corresponding to historical problems; The first large model is pre-trained using the descriptive information corresponding to the historical question as input and the reasoning path corresponding to the historical question as output to constitute training data.

3. The method according to claim 1, characterized in that After fusing the industry knowledge corresponding to the target industry with the reasoning path to obtain the thought chain data corresponding to the original question, the method further includes: Performing a quality assessment on the thought chain data based on business indicators; The thought chain data is screened according to the evaluation result of the quality evaluation.

4. The method according to claim 3, characterized in that The business indicator includes at least one of reasoning completeness, format correctness, answer correctness, and language uniformity.

5. The method according to claim 3, characterized in that The screening of the thought chain data according to the evaluation result of the quality evaluation includes: When the evaluation result of the quality evaluation does not meet the sample application conditions, the thought chain data is eliminated; or, When the evaluation result of the quality evaluation does not meet the sample application condition, the thought chain data is marked as to be modified.

6. The method according to any one of claims 1 to 5, characterized in that: The obtaining of the description information of the original question includes: The original question is input into the second largest model, and the description information corresponding to the original question output by the second largest model is obtained.

7. The method according to any one of claims 1 to 5, characterized in that: The step of fusing the industry knowledge corresponding to the target industry with the reasoning path to obtain the thought chain data corresponding to the original question includes: Retrieving the corresponding industry knowledge for at least one execution step in the reasoning path; The reasoning path and the industry knowledge are input into the third model, and the thinking chain data output by the third model after fusing the execution steps with the corresponding industry knowledge is obtained.

8. The method according to any one of claims 1 to 5, characterized in that: After collecting the original questions in at least one business scenario in the target industry, the following is also included: The original question is subjected to sample amplification, where the sample amplification includes at least one of synonym rewriting, word order adjustment, and sentence structure conversion.

9. The method according to any one of claims 1 and 2, characterized in that: The description information includes at least one of the question type and key tags; and / or, The reasoning path includes an execution step.

10. The method according to claim 9, characterized in that The description information also includes at least one of a problem description and a scenario description; and / or, The reasoning path further includes a step description corresponding to at least one of the execution steps.

11. A thought chain data generating device, characterized in that: The device comprises: The problem collection module is used to collect original problems in at least one business scenario in the target industry; A problem description module is used to obtain description information of the original problem; a path reasoning module, configured to input the description information corresponding to the original problem into the first large model, and obtain the reasoning path corresponding to the original problem output by the first large model; A thought chain generation module is used to integrate the industry knowledge corresponding to the target industry with the reasoning path to obtain thought chain data corresponding to the original problem; Among them, the first large model is obtained by pre-training based on historical reasoning solutions related to the target industry.

12. An electronic device, characterized in that: It includes a processor, a memory, and a program or instruction stored in the memory and executable on the processor, wherein the program or instruction, when executed by the processor, implements the thought chain data generating method as described in any one of claims 1 to 10.

13. A computer-readable medium, characterized in that The computer-readable medium stores a program or instruction, and when the program or instruction is executed by a processor, the method for generating thought chain data according to any one of claims 1 to 10 is implemented.

14. A computer program product, characterized in that When the computer program product is run on a computer, the computer is caused to execute the thought chain data generating method according to any one of claims 1 to 10 .

Images