AutoGraphex: Zero-shot Biomedical Definition Generation with Automatic Prompting

Anonymous

Abstract: Describing terminologies with definition texts is an important step towards understanding the scientific literature, especially for domains with limited labeled terminologies. Previous works have sought to design supervised neural text generation models to solve the biomedical terminology generation task, but most of them failed to define never-before-seen terminologies in newly emerging research fields. Here, we tackle this challenge by introducing a zero-shot definition generation model based on prompting, a recent approach for eliciting knowledge from pre-trained language models, with automatically generated prompts. Furthermore, we enhanced the biomedical terminology dataset by adding descriptive texts to each biomedical subdiscipline, thus enabling zero-shot learning scenarios. Our model outperformed existing supervised baseline and the baseline pre-trained language model that employs manually crafted prompts by up to 52 and 6 BLEU score, respectively.

Paper Type: short

Consent To Share Data: yes