Events

Name: Swati Kanwal
Date: 25th April 2024
Time: 2 - 3 PM
Supervisor: Laks Lakshmanan
Location: Zoom
https://ubc.zoom.us/j/69333584806?pwd=c2dlR0I0TUQxTFVYcXRZaFhwL3IyZz09

Meeting ID: 693 3358 4806
Passcode: 996226

Title: Exploring the potential of LLMs for Biomedical Relation Extraction

Abstract:
Though the current state-of-the-art models for biomedical relation extraction rely on encoder-only LLMs that are pre-trained on domain-specific data and then fine-tuned using large amounts of annotated data, there is a compelling argument to explore decoder-only LLMs and alternative task formulation paradigms to investigate if similar or superior results can be achieved for this task without the need for large amounts of annotated data and computational resources.

Surprisingly, there has been limited exploration of decoder-only LLMs for biomedical relation extraction. This study aims to address this gap, presenting BioREPS, a novel method that enhances decoder-only LLMs through semantic similarity and chain-of-thought prompting, yielding promising outcomes. This study highlights the effectiveness of instruction training and self-generated chain-of-thought prompts in enhancing the reasoning abilities of decoder-only LLMs for biomedical relation extraction. Additionally, this research investigates various task formulation paradigms and the empirical advantages of domain-specific training for biomedical relation extraction through a series of experiments. The results confirm that ”general” decoder-only Language Models (LLMs) hold immense potential for the task of biomedical relation extraction and come close to state-of-the-art performance with a fractional amount of data and no additional training.