Demonstration of Service for Determining Compliance/Non- Compliance with Employment Rules

Overview

Labor inspectors faced difficulties in determining the compliance or non-compliance of employment rules, as the process involved significant time spent searching internal documents and utilizing existing knowledge, and sometimes led to confusion due to the provision of incorrect information. Therefore, the Ministry of Employment and Labor sought to develop a hyper-scale AI model to demonstrate a service that would enable labor inspectors to enhance the speed and accuracy of their employment rule judgment tasks.

Project description

The training of the service model utilized 120 employment rule classification sample data points. Data augmentation was performed based on the sample data to increase the training volume, followed by model training.

The implementation procedure involved constructing training data through data augmentation based on the employment rule classification sample data. The trained model was then fine-tuned to enhance accuracy while adding API functionality. Subsequently, the model was applied to a separately developed demo page, implementing the service after model training.

This service was implemented to draft employment rules that workplaces are legally required to prepare, ensuring compliance with current laws and regulations. It then assesses the drafted rules against legal standards to determine their suitability (whether they are compliant or non- compliant).

Key technological features

• Training Data Construction: Built training data through data augmentation using 120 employment rule classification sample datasets
• Fine-tuning Process: Developed an employment rule judgment model and added API functionality through model fine-tuning

Expected outcomes

In this proof-of-concept, to develop a super-large AI model, this project conducted data augmentation of approximately 10,000 samples based on 120 employment rule classification samples. After training the model, the model was fine-tuned model fine-tuning, enabling the development of an effective employment rule judgment model. Furthermore, during the demonstration period, to develop an effective natural language processing model, the performance of representative NLP models such as 'Polyglot' and 'LLaMA-2' was compared and analyzed. This demonstration confirmed that the Polyglot model outperformed LLaMA-2. It is anticipated that applying the demonstrated model to the actual work of labor inspectors will enable swift response to public inquiries and accurate information provision.

Launch year: 2023

Attribution

Case study courtesy of the National Information Society Agency, Republic of Korea, Apolitical's Content Partner. Sourced from Use Cases of Public AI Service, Vol.1,  https://eng.nia.or.kr/site/nia_eng/ex/bbs/View.do?cbIdx=31975&bcIdx=28997&parentSeq=29026