Model

arXiv Abstract-to-Title · mT5

Purpose

Give the model a research abstract and it proposes a title for the paper. It learned this task by fine-tuning mT5-small on arXiv abstract–title pairs. The training setup allowed 256 tokens per input. Its published ROUGE scores measure word or sequence overlap with the original titles, so a good suggestion phrased differently can still receive a low score.

Base model or method
mT5-small

Sources

arXiv Abstract-to-Title · mT5