Auto-regressive Text Generation with Pre-Trained Language Models: An Empirical Study on Question-type Short Text Generation

Anonymous

Abstract: We present a multi-way parallel math word problem dataset, which covers English, Tamil and Sinhala. We employ this dataset in an empirical analysis of GPT-2, BART, and T5, as well as mT5 and mBART in auto-regressive text generation. Our findings show that BART and T5 perform noticeably better than GPT-2 for the considered task, and text generation with mBART50 and mT5 provides very promising results even for languages under-represented in these pre-trained models.

Paper Type: short