Unsupervised Distractor Generation via Large Language Model Distilling and Counterfactual Contrastive Decoding

TL;DR

This paper introduces an unsupervised distractor generation framework for reading comprehension that uses large language models for knowledge distillation and counterfactual contrastive decoding, outperforming larger models with fewer resources.

Contribution

It presents a novel unsupervised approach combining LLM-based knowledge distillation and counterfactual decoding for distractor generation, reducing reliance on annotated data and large models.

Findings

Outperforms GPT-3.5-turbo with fewer parameters

Uses dual task training with pseudo distractors and answer info

Effective in practical reading comprehension applications

Abstract

Within the context of reading comprehension, the task of Distractor Generation (DG) aims to generate several incorrect options to confuse readers. Traditional supervised methods for DG rely heavily on expensive human-annotated distractor labels. In this paper, we propose an unsupervised DG framework, leveraging Large Language Models (LLMs) as cost-effective annotators to enhance the DG capability of smaller student models. Specially, to perform knowledge distilling, we propose a dual task training strategy that integrates pseudo distractors from LLMs and the original answer in-formation as the objective targets with a two-stage training process. Moreover, we devise a counterfactual contrastive decoding mechanism for increasing the distracting capability of the DG model. Experiments show that our unsupervised generation method with Bart-base greatly surpasses GPT-3.5-turbo performance with only 200 times fewer model parameters. Our proposed unsupervised DG method offers a cost-effective framework for practical reading comprehension applications, without the need of laborious distractor annotation and costly large-size models