Fine-Tuning GPT-3 for Synthetic Danish News Generation

TY - GEN

T1 - Fine-Tuning GPT-3 for Synthetic Danish News Generation

AU - Almasi, Mina

AU - Schiønning, Anton Drasbæk

N1 - Conference code: 16

PY - 2023/9

Y1 - 2023/9

N2 - While GPT-3 has garnered significant attention for its capabilities in natural language generation, research on its use outside of English is still relatively limited. We focus on how GPT-3 can be fine-tuned for generating synthetic news articles in a low-resource language, namely Danish. The model’s performance is evaluated on the dimensions of human and machine detection in two separate experiments. When presented with either a real or GPT-3 generated news article, human participants achieve a 58.1% classification accuracy. Contrarily, a fine-tuned BERT classifier obtains a 92.7% accuracy on the same task. This discrepancy likely pertains to the fine-tuned GPT-3 model oversampling high-likelihood tokens in its text generation. Although this is undetectable to the human eye, it leaves a statistical discrepancy for machine classifiers to detect. We address how decisions in the experimental design favoured the machine classifiers over the human evaluators, and whether the produced synthetic articles are applicable in a real-world context.

AB - While GPT-3 has garnered significant attention for its capabilities in natural language generation, research on its use outside of English is still relatively limited. We focus on how GPT-3 can be fine-tuned for generating synthetic news articles in a low-resource language, namely Danish. The model’s performance is evaluated on the dimensions of human and machine detection in two separate experiments. When presented with either a real or GPT-3 generated news article, human participants achieve a 58.1% classification accuracy. Contrarily, a fine-tuned BERT classifier obtains a 92.7% accuracy on the same task. This discrepancy likely pertains to the fine-tuned GPT-3 model oversampling high-likelihood tokens in its text generation. Although this is undetectable to the human eye, it leaves a statistical discrepancy for machine classifiers to detect. We address how decisions in the experimental design favoured the machine classifiers over the human evaluators, and whether the produced synthetic articles are applicable in a real-world context.

KW - natural language processing

KW - large language models

KW - fine-tuning

KW - machine learning

U2 - 10.18653/v1/2023.inlg-main.4

DO - 10.18653/v1/2023.inlg-main.4

M3 - Article in proceedings

SN - 979-8-89176-001-1

SP - 54

EP - 68

BT - Proceedings of the 16th International Natural Language Generation Conference

A2 - Keet, C. Maria

A2 - Lee, Hung-Yi

A2 - Zarrieß, Sina

PB - Association for Computational Linguistics

T2 - 16th International Natural Language Generation Conference

Y2 - 11 September 2023 through 15 September 2023

ER -