[Paper Review] Bad Actor, Good Advisor: Exploring the Role of Large Language Models in Fake News Detection

Beizhe Hu, Qiang Sheng+5 more|arXiv (Cornell University)|Sep 21, 2023

Misinformation and Its Impacts44 references15 citations

TL;DR

The paper shows GPT-3.5 underperforms fine-tuned BERT for fake news detection but can serve as an advisor by providing multi-perspective rationales; it proposes ARG to fuse LLM-derived rationales with SLMs and ARG-D as a rationale-free distillation variant.

ABSTRACT

Detecting fake news requires both a delicate sense of diverse clues and a profound understanding of the real-world background, which remains challenging for detectors based on small language models (SLMs) due to their knowledge and capability limitations. Recent advances in large language models (LLMs) have shown remarkable performance in various tasks, but whether and how LLMs could help with fake news detection remains underexplored. In this paper, we investigate the potential of LLMs in fake news detection. First, we conduct an empirical study and find that a sophisticated LLM such as GPT 3.5 could generally expose fake news and provide desirable multi-perspective rationales but still underperforms the basic SLM, fine-tuned BERT. Our subsequent analysis attributes such a gap to the LLM's inability to select and integrate rationales properly to conclude. Based on these findings, we propose that current LLMs may not substitute fine-tuned SLMs in fake news detection but can be a good advisor for SLMs by providing multi-perspective instructive rationales. To instantiate this proposal, we design an adaptive rationale guidance network for fake news detection (ARG), in which SLMs selectively acquire insights on news analysis from the LLMs' rationales. We further derive a rationale-free version of ARG by distillation, namely ARG-D, which services cost-sensitive scenarios without querying LLMs. Experiments on two real-world datasets demonstrate that ARG and ARG-D outperform three types of baseline methods, including SLM-based, LLM-based, and combinations of small and large language models.

Motivation & Objective

• Assess whether large language models (LLMs) can effectively detect fake news compared to small language models (SLMs).
• Investigate how LLM-generated rationales from multiple perspectives can help or hinder detection.
• Develop a practical framework that leverages LLMs as advisors to improve SLM-based fake news detectors.
• Provide a cost-aware variant that removes the need to query LLMs during inference.
• Make a rationale collection from GPT-3.5 for future research publicly available.

Proposed method

• Empirically compare GPT-3.5-turbo prompting approaches (zero-shot, zero-shot CoT, few-shot, few-shot CoT) against fine-tuned BERT on Chinese (Weibo21) and English (GossipCop) datasets.
• Analyze LLM-generated rationales from textual description, commonsense, and factuality perspectives and their impact on detection performance.
• Propose Adaptive Rationale Guidance (ARG): an architecture where a small LM interacts with LLM-derived rationales via cross-attention, reasoning about LLM judgments, and rationale usefulness.
• Introduce ARG-D: a distilled, rationale-free variant that imitates ARG’s decision behavior for cost-sensitive scenarios.
• Provide a learning objective that combines veracity prediction with LLM judgment prediction and rationale usefulness evaluation, along with a distillation objective.
• Evaluate cost-control strategy that uses ARG-D by default and selectively queries ARG to match full ARG performance at reduced cost.

Experimental results

Research questions

• RQ1 Can large language models (LLMs) outperform small, task-tuned models in fake news detection across Chinese and English datasets?
• RQ2 Do LLM-generated rationales from multiple perspectives improve fake news detection when used to guide a small LM?
• RQ3 Is the ARG framework capable of surpassing SLM-only and LLM-only baselines, and how does the distilled ARG-D perform under cost constraints?
• RQ4 What is the contribution of each ARG module (LLM judgment predictor, rationale usefulness evaluator, news-rationale interaction) to overall performance?
• RQ5 Can a rationale-free distillation (ARG-D) retain much of ARG’s benefit while lowering inference cost?

Key findings

• GPT-3.5-turbo generally underperforms fine-tuned BERT across both datasets and prompting schemes.
• Few-shot prompting improves LLM performance but does not consistently surpass the SLM baseline; Chain-of-Thought can help English more than Chinese in some settings.
• LLMs provide informative multi-perspective rationales (textual description, commonsense, factuality) that can aid detection if effectively integrated, but single-perspective analysis can miss key cues.
• ARG consistently outperforms SLM-only, LLM-only, and naive LLM+SLM baselines in macro F1 and accuracy on both datasets.
• ARG-D, the distilled rationale-free variant, also beats most baselines and offers cost-effective competitive performance; with selective querying, ARG can match full ARG performance at reduced cost.
• Ablations show the LLM Judgment Predictor and Rationale Usefulness Evaluator are crucial for ARG’s effectiveness, while the news-rationale interaction structure remains essential.