Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

A Comprehensive Study of Implicit and Explicit Biases in Large Language Models

Fatima Kazi, Alex Young, Yash Inani, Setareh Rafatirad

TL;DR

The paper tackles explicit and implicit biases in large language models (LLMs) arising from internet-scale training data. It introduces a Bias-Characterization Framework and a Bias-Identification Framework (BIF) that combine preprocessing, targeted fine-tuning, MCSB prompting, and data augmentation to diagnose and mitigate biases on StereoSet and CrowSPairs. It demonstrates that augmented fine-tuning and prompting can reduce biases and improve cross-dataset robustness, with notable gains (up to ~20-30% in certain setups) and insights from Bag-of-Words analyses into bias-driving vocabulary. The work highlights practical implications for responsibly deploying generative AI, proposes a framework for ongoing bias detection, and discusses limitations and ethical considerations inherent to benchmarking biases in LLMs.

Abstract

Large Language Models (LLMs) inherit explicit and implicit biases from their training datasets. Identifying and mitigating biases in LLMs is crucial to ensure fair outputs, as they can perpetuate harmful stereotypes and misinformation. This study highlights the need to address biases in LLMs amid growing generative AI. We studied bias-specific benchmarks such as StereoSet and CrowSPairs to evaluate the existence of various biases in multiple generative models such as BERT and GPT 3.5. We proposed an automated Bias-Identification Framework to recognize various social biases in LLMs such as gender, race, profession, and religion. We adopted a two-pronged approach to detect explicit and implicit biases in text data. Results indicated fine-tuned models struggle with gender biases but excelled at identifying and avoiding racial biases. Our findings illustrated that despite having some success, LLMs often over-relied on keywords. To illuminate the capability of the analyzed LLMs in detecting implicit biases, we employed Bag-of-Words analysis and unveiled indications of implicit stereotyping within the vocabulary. To bolster the model performance, we applied an enhancement strategy involving fine-tuning models using prompting techniques and data augmentation of the bias benchmarks. The fine-tuned models exhibited promising adaptability during cross-dataset testing and significantly enhanced performance on implicit bias benchmarks, with performance gains of up to 20%.

A Comprehensive Study of Implicit and Explicit Biases in Large Language Models

TL;DR

The paper tackles explicit and implicit biases in large language models (LLMs) arising from internet-scale training data. It introduces a Bias-Characterization Framework and a Bias-Identification Framework (BIF) that combine preprocessing, targeted fine-tuning, MCSB prompting, and data augmentation to diagnose and mitigate biases on StereoSet and CrowSPairs. It demonstrates that augmented fine-tuning and prompting can reduce biases and improve cross-dataset robustness, with notable gains (up to ~20-30% in certain setups) and insights from Bag-of-Words analyses into bias-driving vocabulary. The work highlights practical implications for responsibly deploying generative AI, proposes a framework for ongoing bias detection, and discusses limitations and ethical considerations inherent to benchmarking biases in LLMs.

Abstract

Large Language Models (LLMs) inherit explicit and implicit biases from their training datasets. Identifying and mitigating biases in LLMs is crucial to ensure fair outputs, as they can perpetuate harmful stereotypes and misinformation. This study highlights the need to address biases in LLMs amid growing generative AI. We studied bias-specific benchmarks such as StereoSet and CrowSPairs to evaluate the existence of various biases in multiple generative models such as BERT and GPT 3.5. We proposed an automated Bias-Identification Framework to recognize various social biases in LLMs such as gender, race, profession, and religion. We adopted a two-pronged approach to detect explicit and implicit biases in text data. Results indicated fine-tuned models struggle with gender biases but excelled at identifying and avoiding racial biases. Our findings illustrated that despite having some success, LLMs often over-relied on keywords. To illuminate the capability of the analyzed LLMs in detecting implicit biases, we employed Bag-of-Words analysis and unveiled indications of implicit stereotyping within the vocabulary. To bolster the model performance, we applied an enhancement strategy involving fine-tuning models using prompting techniques and data augmentation of the bias benchmarks. The fine-tuned models exhibited promising adaptability during cross-dataset testing and significantly enhanced performance on implicit bias benchmarks, with performance gains of up to 20%.

Paper Structure

This paper contains 19 sections, 1 figure, 7 tables.

Table of Contents

  1. Introduction
  2. Related Works
  3. Preliminary Results
  4. Methodology
  5. Models and Fine-tuning
  6. Datasets
  7. Prompting Techniques
  8. Data Augmentation
  9. Results
  10. General Results - Implicit Bias Prompting
  11. General Results - Explicit Bias Prompting
  12. Augmented Fine-Tuning Results - Implicit Bias Prompting
  13. Augmented Fine-Tuning Results - Explicit Bias Prompting
  14. Bag of Words
  15. BoW Fintetuning and System Role
  16. ...and 4 more sections

Figures (1)

  • Figure 1: The proposed framework comprises of 4 key components: Preprocessing - involving filtering StereoSet and CrowSPairs to create MCSBQ and splitting them into training and testing data. Finetuning - utilizing both original and augmented training data to fine-tune the LLMs. Evaluation - entailing testing of fine-tuned LLMs using testing data and baseline model using MCSBQ dataset. Analysis - analyzing results with quantitative (graphs), comparative (tables), and qualitative (BoW) techniques to uncover potential biases.