Multiclass text classification remains a difficult task, primarily due to semantic ambiguity and differences in input length. This study evaluates RoBERTa and GPT-based models for multiclass text classification, focusing on how prompting strategies and document length affect accuracy and robustness. Experiments were conducted using the OSDG Community Dataset, which contains approximately 15,000 labeled samples. The dataset was partitioned into four subsets based on input length: short, medium, long, and all combined. Three GPT variants (zero-shot, few-shot, and fine-tuned) were compared against a RoBERTa baseline. Fine-tuning was implemented via OpenAI’s supervised API with prompt-response formatting. Performance was assessed through F1-score, precision, recall, and balanced accuracy. Fine-tuned GPT achieved the strongest results in all settings, with a macro F1-score of 0.9204 on the all-combined dataset, representing a 4.61% improvement over RoBERTa. Consistent gains were also observed across short (8.63%), medium (3.83%), and long (20.31%) texts. The largest improvement occurred on long documents, while medium-length inputs provided the most stable performance across models. These findings highlight the effectiveness of task-specific fine-tuning in enhancing GPT’s capability to classify SDG-related texts across diverse input lengths.
Hasanah et al. (Wed,) studied this question.