Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Causal-Guided Active Learning for Debiasing Large Language Models

Li Du, Zhouhao Sun, Xiao Ding, Yixuan Ma, Yang Zhao, Kaitao Qiu, Ting Liu, Bing Qin

TL;DR

A casual-guided active learning framework is proposed, which utilizes LLMs itself to automatically and autonomously identify informative biased samples and induce the bias patterns, and a cost-effective and efficient in-context learning based method is employed to prevent LLMs from utilizing dataset biases during generation.

Abstract

Although achieving promising performance, recent analyses show that current generative large language models (LLMs) may still capture dataset biases and utilize them for generation, leading to poor generalizability and harmfulness of LLMs. However, due to the diversity of dataset biases and the over-optimization problem, previous prior-knowledge-based debiasing methods and fine-tuning-based debiasing methods may not be suitable for current LLMs. To address this issue, we explore combining active learning with the causal mechanisms and propose a casual-guided active learning (CAL) framework, which utilizes LLMs itself to automatically and autonomously identify informative biased samples and induce the bias patterns. Then a cost-effective and efficient in-context learning based method is employed to prevent LLMs from utilizing dataset biases during generation. Experimental results show that CAL can effectively recognize typical biased instances and induce various bias patterns for debiasing LLMs.

Causal-Guided Active Learning for Debiasing Large Language Models

TL;DR

A casual-guided active learning framework is proposed, which utilizes LLMs itself to automatically and autonomously identify informative biased samples and induce the bias patterns, and a cost-effective and efficient in-context learning based method is employed to prevent LLMs from utilizing dataset biases during generation.

Abstract

Although achieving promising performance, recent analyses show that current generative large language models (LLMs) may still capture dataset biases and utilize them for generation, leading to poor generalizability and harmfulness of LLMs. However, due to the diversity of dataset biases and the over-optimization problem, previous prior-knowledge-based debiasing methods and fine-tuning-based debiasing methods may not be suitable for current LLMs. To address this issue, we explore combining active learning with the causal mechanisms and propose a casual-guided active learning (CAL) framework, which utilizes LLMs itself to automatically and autonomously identify informative biased samples and induce the bias patterns. Then a cost-effective and efficient in-context learning based method is employed to prevent LLMs from utilizing dataset biases during generation. Experimental results show that CAL can effectively recognize typical biased instances and induce various bias patterns for debiasing LLMs.
Paper Structure (31 sections, 4 equations, 5 figures, 6 tables)

This paper contains 31 sections, 4 equations, 5 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Preliminary
  3. Dataset Bias within Textual Corpus under Causal Perspective
  4. Biases of Generative LLMs
  5. Active learning
  6. Methodology
  7. Causal Invariance Based Biased Instances Identification
  8. Selection of Informative Biased Instances and Bias Pattern Induction
  9. In Context Learning-based Bias Suppression
  10. Experiments
  11. Experimental Details
  12. Evaluation Tasks
  13. Baseline Methods
  14. Main Results
  15. Case Study
  16. ...and 16 more sections

Figures (5)

  • Figure 1: (a) Dataset bias under causal perspective (b) Illustration of the Causal-Guided Active Learning framework.
  • Figure 2: Results of bias pattern induction. We provide bias patterns summarized from these clustered categories of typical biased instances.
  • Figure 3: Case study of the selected counter example pairs for Chatbot, MNLI, and BBQ datasets respectively when experimented with llama2-13B-chat. Example 1 and Example 2 together constitute a counter example pair.
  • Figure 4: Influence of different orders of magnitude for counter example pairs and negative examples. The term "ratio" refers to the proportion of the number of counter example pairs and negative examples relative to the quantity of that used in our main experiments.
  • Figure 5: Prompts for the bias pattern induction procedure for the Chatbot dataset