Presentation Information
[1Yin-B-05]Improving Text Classification Accuracy with LLM-Generated Structured Explanations
〇Shingo Sugawara1, Yu Okano1 (1. Resonac Corporation.)
Keywords:
LLM,Text Classification,Data Augmentaion
Enterprise purchasing item names vary sufficiently that identical items are recorded under different names, causing entity-resolution errors. Because these names are short and dense with jargon and proper nouns, conventional classifiers perform poorly, and direct classification with large language models (LLM) proved unstable. We therefore prompted GPT 4o to generate structured explanations for each item name. For thousands of items, JSON fields capturing intended use, key specifications, and related domains were concatenated with the original names and used to train an 11 class BERT classifier. This approach improved accuracy, precision, recall, and F1 relative to conventional classifiers and direct LLM classification, providing an effective method for data cleansing and deduplication.