Classification of Descriptive Text Data using Large Language Models
Name
kim_tylerkim_meng_eecs_2025_thesis.pdf
Size
2.04 MB
Format
Adobe PDF
Checksum (MD5)
396bfd00bed12d6cd842d4a5e21a8ac3
Author(s)
Kim, Tyler D.
Advisor(s)
Fernández, John E.
Date Issued
May 2025
Publisher
Massachusetts Institute of Technology
Abstract
This paper explores the application of large language models (LLMs) for text classification tasks. We focus on the classification of company descriptions into business model archetypes as defined by the literature on business model taxonomies. We use instruction fine-tuning to train a RoBERTa model (Liu et al., 2019) for classification. The methodology addresses challenges with limited training data in two key steps: (1) reducing the output dimension by mapping 55 business models (Gassmann et al., 2017) to 8 sustainable archetypes (Bocken et al., 2014), and (2) utilizing synthetic data generation to create a balanced training dataset. The results demonstrate significant improvements in performance, with the selection model achieving 73.08% test accuracy (16.8x improvement) after balancing the dataset. We additionally present evidence of diminishing returns of synthetic generation on test accuracy. This work contributes to practical applications in applying machine learning on tabular data and the modeling of company data.
MIT Department
Massachusetts Institute of Technology. Department of Electrical Engineering and Computer Science
Terms of Use
Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0)
Copyright retained by author(s)
Persistent DSpace Link
