CN-121980351-A - Data labeling method, device, electronic equipment and storage medium

CN121980351ACN 121980351 ACN121980351 ACN 121980351ACN-121980351-A

Abstract

The application provides a data labeling method, a data labeling device, electronic equipment and a storage medium, and relates to the technical field of artificial intelligence. The method comprises the steps of obtaining a data set, wherein the data set comprises input data to be processed and real labeling data corresponding to the input data, carrying out iterative optimization on a labeling strategy of a pre-constructed large language model LLM generator and a judging strategy of the pre-constructed large language model LLM judging device based on the data set, the pre-constructed large language model LLM generator and the pre-constructed large language model LLM judging device to generate a final target labeling strategy and a target judging strategy, and carrying out labeling and judging processing on target data to be labeled by adopting the LLM generator based on the target labeling strategy and the LLM judging device based on the target judging strategy to generate corresponding target labeling data. The method improves the data labeling efficiency. In addition, as the LLM generator and the LLM discriminant only need to iterate the corresponding labeling strategy and the discrimination strategy, the weight of the iterative training model is not needed, and the iteration efficiency is higher.

Inventors

DU LIANG
YANG JIE
ZENG CHUXUAN
LU PEIAN
DENG LING
Lin Lihai

Assignees

中国联合网络通信集团有限公司
联通智能制造科技产业(广东)有限公司

Dates

Publication Date: 20260505
Application Date: 20260123

Claims (10)

1. A method for labeling data, comprising: the method comprises the steps of acquiring a data set, wherein the data set comprises input data to be processed and real annotation data corresponding to the input data; Performing iterative optimization on the marking strategy of the LLM generator and the judging strategy of the LLM judging device based on the data set, the pre-constructed large language model LLM generator and the pre-constructed large language model LLM judging device to generate a final target marking strategy and a final target judging strategy; and marking and distinguishing the target data to be marked by adopting a LLM generator based on the target marking strategy and a LLM distinguishing device based on the target distinguishing strategy to generate corresponding target marking data.
2. The method of claim 1, wherein iteratively optimizing the labeling strategy of the LLM generator and the discrimination strategy of the LLM discriminator based on the dataset, the pre-constructed large language model LLM generator and the pre-constructed large language model LLM discriminator to generate a final target labeling strategy and target discrimination strategy comprises: Inputting the input data into the LLM generator to generate a corresponding labeling result and a labeling interpretation, wherein the labeling interpretation corresponds to the labeling result; Inputting the data set, the labeling result and a pre-constructed initial discrimination strategy into the LLM discriminator to generate a corresponding discrimination result and discrimination interpretation, wherein the discrimination interpretation corresponds to the discrimination result; generating a corresponding optimized labeling strategy and a discrimination strategy according to the discrimination result and the discrimination interpretation optimization corresponding initial labeling strategy and the initial discrimination strategy, wherein the initial labeling strategy is the initial labeling strategy of the LLM generator; Performing iterative optimization on the labeling strategy of the LLM generator and the judging strategy of the LLM judging device until the preset requirement is met; and taking the final iterative labeling strategy as a target labeling strategy of the LLM generator, and taking the final iterative judging strategy as a target judging strategy of the LLM discriminator.
3. The method according to claim 2, wherein said optimizing the corresponding initial labeling strategy and the initial discrimination strategy based on the discrimination result and the discrimination interpretation, generating the corresponding optimized labeling strategy and discrimination strategy, comprises: If the judging result is that the marking result is inconsistent with the real marking data, optimizing the initial marking strategy according to first error data corresponding to the inconsistency and judging interpretation corresponding to the first error data, and generating a corresponding optimized marking strategy; If the judging result is that the labeling result is consistent with the real labeling data, but the labeling result is not consistent with the real labeling data, optimizing the initial judging strategy according to second error data corresponding to the actual inconsistency, labeling explanation corresponding to the second error data and a pre-constructed large language model LLM (logical level model) jetter, and generating a correspondingly optimized judging strategy, wherein the second error data is corresponding to the actual inconsistency in the input data.
4. The method of claim 3, wherein the optimizing the initial discrimination strategy based on the second error data corresponding to the actual inconsistency, the annotation interpretation corresponding to the second error data, and the pre-built large language model LLM disguiser, generating the corresponding optimized discrimination strategy comprises: inputting the second error data and the annotation explanation corresponding to the second error data into the LLM thinker, and generating corresponding thinking-back feedback; inputting the negative feedback, the annotation explanation corresponding to the second error data and the second error data into the LLM discriminator to optimize the initial discrimination strategy and generate a discrimination strategy after the optimization.
5. A method according to claim 3, wherein the step of optimizing the initial labeling strategy based on first error data corresponding to inconsistencies, discrimination interpretations corresponding to the first error data, and generating a corresponding optimized labeling strategy is based on the following algorithm: Wherein, the Representing the labeling strategy after the optimization, Representing a pre-optimization labeling strategy such as an initial labeling strategy, Representing the input data, i.e. the first error data, Representing a discriminant interpretation.
6. The method of claim 1, wherein the acquiring the data set comprises: Collecting the input data; generating real annotation data corresponding to the input data in response to the annotation processing of the input data by a user; and generating the data set according to the input data and the real annotation data.
7. The method of claim 2, wherein the inputting the input data into the LLM generator, prior to generating the corresponding annotation result and annotation interpretation, further comprises: And constructing the LLM generator, an initial labeling strategy corresponding to the LLM generator, the LLM discriminator and an initial discriminating strategy corresponding to the LLM discriminator.
8. A data tagging device, comprising: The device comprises an acquisition module, a processing module and a processing module, wherein the acquisition module is used for acquiring a data set, and the data set comprises input data to be processed and real annotation data corresponding to the input data; the optimization module is used for carrying out iterative optimization on the marking strategy of the LLM generator and the judging strategy of the LLM judging device based on the data set, the pre-constructed large language model LLM generator and the pre-constructed large language model LLM judging device to generate a final target marking strategy and a final target judging strategy; The generating module is used for marking and judging the target data to be marked by adopting a LLM generator based on the target marking strategy and a LLM discriminant based on the target judging strategy, and generating corresponding target marking data.
9. An electronic device is characterized by comprising a memory and a processor; the memory stores computer-executable instructions; The processor executes computer-executable instructions stored in the memory to implement the data tagging method according to any one of claims 1 to 7.
10. A computer-readable storage medium having stored therein computer-executable instructions which, when executed by a processor, are adapted to carry out the data annotation method according to any one of claims 1 to 7.

Description

Data labeling method, device, electronic equipment and storage medium Technical Field The application belongs to the technical field of artificial intelligence, and particularly relates to a data labeling method, a device, electronic equipment and a storage medium. Background With the continuous development of artificial intelligence technology, artificial intelligence models are increasingly applied in various fields. The model needs to be subjected to model building and model training processes before being put into production and applied. In model training, a large amount of sample data is usually required, and the sample data is labeled, so that model training is performed based on the labeled data. The current labeling mode generally adopts a manual labeling mode, which depends on manual experience and a large amount of manual resources, and meanwhile, the labeling efficiency of the mode is lower, and further optimization is needed. Disclosure of Invention The technical problem to be solved by the application is to provide a data labeling method, a device, electronic equipment and a storage medium aiming at the defects in the prior art, and by using the data labeling method, the dependence on artificial resources can be reduced, and the data labeling efficiency is improved. In a first aspect, an embodiment of the present application provides a data labeling method, including: the method comprises the steps of acquiring a data set, wherein the data set comprises input data to be processed and real labeling data corresponding to the input data; Iterative optimization is carried out on the marking strategy of the LLM generator and the judging strategy of the LLM judging device based on the data set, the pre-constructed large language model LLM generator and the pre-constructed large language model LLM judging device, and a final target marking strategy and a final target judging strategy are generated; And marking and distinguishing the target data to be marked by adopting a LLM generator based on a target marking strategy and a LLM distinguishing device based on a target distinguishing strategy to generate corresponding target marking data. In some embodiments of the first aspect, iteratively optimizing a labeling strategy of the LLM generator and a discrimination strategy of the LLM discriminator based on the dataset, the pre-constructed large language model LLM generator and the pre-constructed large language model LLM discriminator, generating a final target labeling strategy and a target discrimination strategy, comprising: Inputting the input data into an LLM generator to generate a corresponding labeling result and a labeling interpretation, wherein the labeling interpretation corresponds to the labeling result; inputting the data set, the labeling result and the pre-constructed initial discrimination strategy into a LLM discriminator to generate a corresponding discrimination result and discrimination interpretation; optimizing a corresponding initial labeling strategy and an initial judging strategy according to the judging result and the judging explanation, and generating the labeling strategy and the judging strategy after the corresponding optimization; performing iterative optimization on the labeling strategy of the LLM generator and the judging strategy of the LLM judging device until the preset requirement is met; and taking the final iterative labeling strategy as a target labeling strategy of the LLM generator, and taking the final iterative judging strategy as a target judging strategy of the LLM discriminator. In some embodiments of the first aspect, optimizing the corresponding initial labeling strategy and the initial discrimination strategy according to the discrimination result and the discrimination interpretation, generating the corresponding optimized labeling strategy and discrimination strategy includes: if the judging result is that the marking result is inconsistent with the real marking data, optimizing an initial marking strategy according to the first error data corresponding to the inconsistency and the judging interpretation corresponding to the first error data, and generating a corresponding optimized marking strategy; If the judging result is that the labeling result is consistent with the real labeling data, but the actual inconsistency exists between the labeling result and the real labeling data, optimizing an initial judging strategy according to second error data corresponding to the actual inconsistency, labeling explanation corresponding to the second error data and a pre-constructed large language model LLM jetschef, and generating a corresponding optimized judging strategy, wherein the second error data is corresponding to the actual inconsistency in the input data. In some embodiments of the first aspect, optimizing an initial discrimination policy according to second error data corresponding to actual inconsistencies, label interpretations corresponding t