Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Mitigating Hallucination in Fictional Character Role-Play

Nafis Sadeq, Zhouhang Xie, Byungkyu Kang, Prarit Lamba, Xiang Gao, Julian McAuley

TL;DR

RoleFact is proposed, a role-playing method that mitigates hallucination by modulating the influence of parametric knowledge using a pre-calibrated confidence threshold and improves the factual precision of generated responses for adversarial questions.

Abstract

Role-playing has wide-ranging applications in customer support, embodied agents, and computational social science. The influence of parametric world knowledge of large language models (LLMs) often causes role-playing characters to act out of character and to hallucinate about things outside the scope of their knowledge. In this work, we focus on the evaluation and mitigation of hallucination in fictional character role-play. We introduce a dataset with over 2,000 characters and 72,000 interviews, including 18,000 adversarial questions. We propose RoleFact, a role-playing method that mitigates hallucination by modulating the influence of parametric knowledge using a pre-calibrated confidence threshold. Experiments show that the proposed method improves the factual precision of generated responses by 18% for adversarial questions with a 44% reduction in temporal hallucination for time-sensitive interviews. The code and the dataset are available at https://github.com/NafisSadeq/rolefact.git.

Mitigating Hallucination in Fictional Character Role-Play

TL;DR

RoleFact is proposed, a role-playing method that mitigates hallucination by modulating the influence of parametric knowledge using a pre-calibrated confidence threshold and improves the factual precision of generated responses for adversarial questions.

Abstract

Role-playing has wide-ranging applications in customer support, embodied agents, and computational social science. The influence of parametric world knowledge of large language models (LLMs) often causes role-playing characters to act out of character and to hallucinate about things outside the scope of their knowledge. In this work, we focus on the evaluation and mitigation of hallucination in fictional character role-play. We introduce a dataset with over 2,000 characters and 72,000 interviews, including 18,000 adversarial questions. We propose RoleFact, a role-playing method that mitigates hallucination by modulating the influence of parametric knowledge using a pre-calibrated confidence threshold. Experiments show that the proposed method improves the factual precision of generated responses by 18% for adversarial questions with a 44% reduction in temporal hallucination for time-sensitive interviews. The code and the dataset are available at https://github.com/NafisSadeq/rolefact.git.

Paper Structure

This paper contains 18 sections, 7 figures, 7 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Methodology
  3. The SGR Dataset
  4. Experiments
  5. Baselines
  6. Evaluation Metrics
  7. Results
  8. Hyper-parameter Tuning
  9. Ablation study
  10. Human Evaluation
  11. Case Study
  12. Related Work
  13. Language Models for Role-play
  14. Factuality in Natural Language Generation
  15. Retrieval Augmented Methods
  16. ...and 3 more sections

Figures (7)

  • Figure 1: Example of cross-universe hallucination (Hiccup should not answer questions about Hogwarts) and temporal hallucination (Harry should not talk about the Patronus charm in his first year) in character role-play.
  • Figure 2: Factual precision degrades when we minimize parametric knowledge by anonymizing the prompts.
  • Figure 3: Factual precision degrades with decreasing character popularity (left to right), shown for characters associated with the 'How to Train Your Dragon' series.
  • Figure 4: An overview of RoleFact, which performs parametric and non-parametric verification of atomic facts.
  • Figure 5: Performance by role popularity (decreasing popularity left to right, adversarial task, GPT-3.5).
  • ...and 2 more figures