The First Framework To Utilize LLM To Detect Fake News Is Now Available!
3 main points
✔️ Proposes STEEL, a new framework for automatic fake news detection leveraging LLM
✔️ Provides open source designed for out-of-the-box use without complex data processing or model training
✔️ Large-scale experiments on three real-world data sets demonstrate STEEL's Demonstrated effectiveness
Re-Search for The Truth: Multi-round Retrieval-augmented Large Language Models are Strong Fake News Detectors
written by Guanghua Li, Wensheng Lu, Wei Zhang, Defu Lian, Kezhong Lu, Rui Mao, Kai Shu, Hao Liao
(Submitted on 14 Mar 2024)
Comments: Published on arxiv.
Subjects: Computation and Language (cs.CL); Artificial Intelligence(cs.AI)
The images used in this article are from the paper, the introductory slides, or were created based on them.
Introduction
In recent years, the proliferation of fake news has had far-reaching negative effects on politics, the economy, and society as a whole, and various fake news detection methods have long been developed to solve this problem.
On the other hand, these existing methods have the disadvantage that they often obtain information from static repositories, such as Wikipedia, and thus cannot address new news and claims in particular.
This trend has recently led to the use of the superior inference and generation capabilities of Large Language Models (LLMs ) to detect such fake news.
However, these LLM-based solutions, like traditional methods, have the disadvantage of outdated information and suffer from challenges such as low-quality information retrieval capabilities and context length limitations.
Against this background, this paper describes a paper that proposes STEEL, the first framework that leverages LLMs to detect fake news by using automatic information retrieval that takes advantage of LLMs' reasoning capabilities.
STEELOverview
LLM has demonstrated remarkable capabilities in a variety of areas, including the detection of fake news through the use of RAG (Retrieval-Augmented Generation), a method that retrieves relevant documents from a vast external knowledge base.
However, there are a number of challenges, such as the limitations of relying on limited data sources and the difficulty of real-time updates in an ever-changing news environment.
STEEL (STrategic rEtrieval Enhanced with Large Language Model) proposed in this paperby a multi-round LLM-based RAG framework that retrieves evidence directly from the Internet via a search module and search engine, It solves these challenges.
The overall picture of STEEL is shown in the figure below.
As shown in the figure, STEEL consists of two main modules, the Retrieval module and theReasoning module, which are integrated into a comprehensive framework of re-search mechanisms.
Each of these will be explained.
Retrieval module
The Retrieval module searches for sources of evidence that can be determined to be fake news via a search engine, and then sorts the retrieved documents and the input information, the Claim, based on similarity.
The source implements a basic filtering mechanism, using a list of 1044 known fake news websites as filters based on existing research.
Reasoning module
Sources that can be determined to be fake news retrieved from the web are aggregated into prompts and provided to LLM for inference.
The LLM then performs an evaluation based on the given source, including a determination of whether a re-search is necessary, and outputs a result of either true (true), false (false), or NEI (Not Enough Information = not enough information ) .
Re-Search Mechanism
If the aforementioned Reasoning module outputs a NEI, as shown in the figure below, it is judged that there is insufficient information to determine that it is fake news, and a re-investigation is conducted.
The resurvey first merges the sources collected in the initial search and adds them to a pool named "established evidence " for reference.
Next, "updated queries " are set to retrieve additional relevant information, to which the new information is added.
By repeating this approach, the model gradually accumulates evidence that can be used to determine fake news, thereby improving the model's ability to discern the truth or falsity of news.
Experiments
To evaluate the performance of STEEL, we conducted extensive experiments on three real-world datasets, consisting of the LIAR andPolitiFact English language datasets and the CHEF Chinese language dataset. (These datasets are organized into real and fake news categories.)
In addition, this experiment was conducted using a total of 11 models, consisting of seven evidence-based and four LLM-based methods, as shown below.
- Evidence base (G1):7 DeClarE, HAN, EHIAN, MAC, GET, MUSER, ReRead
- LLM-based (G2): GPT-3.5-Turbo, Vicuna-7B, WEBGLM, and ProgramFC
Fake news detection was posed to these models as a binary classification problem, with F1, Precision, Recall, F1 Macro, and F1 Micro as the evaluation criteria.
The experimental results are shown in the table below.
The tableconfirms that STEEL scores the highest of all methods, improving both F1 Macro and F1 Micro scores by more than 5% across the three real-world data sets.
From this experiment, it is clear that STEEL is very effective in detecting fake news and has significant advantages in both inference and accuracy.
Summary
How was it? In this article, we described a paper that proposed STEEL, the first framework that leverages LLMs to detect fake news by using automatic information retrieval based on LLMs' inference capabilities.
While the large-scale experiments conducted in this paper show that STEEL outperforms existing fake news detection methods, there is a concern that this paper deals only with text data.
Given the complexity of fake news, there is a need to expand the framework's capabilities to integrate information contained in text, images, video, and audio in the future.
On the other hand, working in these areas will not only improve the accuracy of fake news detection, but will also improve the reliability of news.
The details of the framework and experimental results presented here can be found in this paper for those interested.
Categories related to this article
