Developing a Model for Extracting Actual Product Names From Order Item Descriptions Using Generative Language Models

Abstract

This study proposes a cost-effective method for developing a model that extracts actual product names from order item descriptions in food delivery requests processed by a delivery agency platform. First, we utilize GPT-4, a state-of-the-art generative language model, to generate actual product names corresponding to the original order item descriptions, creating an annotated dataset. Next, we process this dataset into a Named Entity Recognition (NER) format to generate a labeled dataset. Based on this, we fine-tune Korean and multilingual pre-trained language models to develop models for extracting actual product names. Experimental results show that the Korean model achieves the highest performance, with an F1 score of 0.8750 or higher on a 10% holdout dataset, securing a product name extraction model that meets the desired level of accuracy. The developed model has been successfully integrated into the food delivery platform’s data analysis process, where it is being utilized for various data management tasks, such as building a master product name table and enhancing the retrieval quality of similar products. This study confirms the effectiveness of an approach that leverages generative language models for annotated dataset creation and fine-tuning of pre-trained language models for developing models tailored to domain-specific problems.

Keywords

• Annotated dataset
• delivery agency platform
• extracting actual product names
• generative language model
• labeled dataset