Web Retrieval Agents for Evidence-Based Misinformation Detection
Combining LLMs with web retrieval improves misinformation detection by 20%.
Key Findings
Methodology
The study proposes a framework combining large language models (LLMs) with web retrieval. Claims are decomposed into queries, evidence is fetched via tools like Cohere RAG or DuckDuckGo, and LLMs reason over the evidence iteratively.
Key Results
- Result 1: On the LIAR-New dataset, GPT-4 with retrieval achieved a 20.5% macro F1 improvement, from 47.8% to 68.3%.
- Result 2: DuckDuckGo retrieval with GPT-3.5 summarization performed comparably to Cohere RAG, achieving 70.3% F1 even without PolitiFact data.
- Result 3: More retrieval sources (e.g., k=10) significantly enhanced performance, with low reliance on any single source.
Significance
This research addresses LLM limitations like hallucinations and outdated knowledge by integrating web retrieval. It not only improves detection performance but also provides a scalable framework for future misinformation mitigation systems.
Technical Contribution
The study introduces a novel retrieval-augmented generation (RAG) framework supporting iterative search and reasoning. Compared to methods like HiSS, it demonstrates more stable performance across models and includes in-depth source bias and quality analysis.
Novelty
This is the first approach to combine iterative retrieval and reasoning for misinformation detection, validated across multiple datasets and models.
Limitations
- Limitation 1: Dependence on retrieval tools and internet access may limit applicability in offline scenarios.
- Limitation 2: Limited effectiveness for non-textual evidence like images or videos.
- Limitation 3: Performance may degrade with noisy retrieval results.
Future Work
Future research could explore more efficient retrieval and summarization methods, improve handling of non-textual evidence, and expand to multilingual scenarios.
AI Executive Summary
The spread of misinformation poses a significant societal challenge, exacerbated by advancements in generative AI. Existing large language models (LLMs) struggle with outdated knowledge and hallucinations, limiting their effectiveness in detecting misinformation. This study introduces a novel framework that combines LLMs with web retrieval. By decomposing claims into queries, retrieving evidence using tools like Cohere RAG or DuckDuckGo, and reasoning iteratively, the system achieves significant performance improvements.
For instance, on the LIAR-New dataset, GPT-4 with retrieval improved its macro F1 score by 20.5%, reaching 68.3%. The study also analyzed source bias and quality, finding that the system is not overly reliant on any single source and maintains high credibility. Additionally, DuckDuckGo retrieval proved flexible and effective, even when key sources like PolitiFact were excluded.
Despite limitations such as reliance on internet access and challenges with non-textual evidence, this research offers a robust foundation for misinformation detection. Future directions include optimizing retrieval and summarization, addressing non-textual evidence, and expanding to multilingual contexts, paving the way for scalable, real-world applications.
Deep Analysis
Background
Misinformation poses societal risks, amplified by generative AI. Traditional models like BERT and BART struggle with ambiguous information and lack real-time knowledge. LLMs show promise but suffer from hallucinations and outdated training data.
Core Problem
The core challenge is integrating LLMs' language understanding with real-time external knowledge to build an effective misinformation detection system. Existing methods lack robust retrieval-reasoning integration.
Innovation
Key innovations include: 1) a framework combining iterative retrieval and reasoning; 2) comparison of Cohere RAG and DuckDuckGo retrieval tools; 3) in-depth analysis of source bias and quality.
Methodology
- �� Decompose claims into queries using LLMs.
- �� Retrieve evidence via Cohere RAG or DuckDuckGo.
- �� Feed retrieved evidence into LLMs for iterative reasoning.
- �� Evaluate on datasets like LIAR-New, comparing multiple models.
Experiments
Experiments used datasets like LIAR-New and FEVER to evaluate GPT-4, GPT-3.5, and others with and without retrieval. Ablation studies examined the impact of source quantity and quality on performance.
Results
GPT-4 with retrieval improved macro F1 by 20.5% on LIAR-New. DuckDuckGo retrieval maintained high performance without PolitiFact data. More sources improved robustness significantly.
Applications
Applications include social media content moderation and news fact-checking, particularly for tasks requiring real-time knowledge updates.
Limitations & Outlook
Limitations include reliance on retrieval tools, limited handling of non-textual evidence, and performance sensitivity to noisy retrieval results.
Plain Language Accessible to non-experts
Imagine this system as a 'fact detective.' When someone makes a claim, the detective breaks it into smaller questions, searches the internet for evidence, and carefully evaluates the findings before deciding if the claim is true. For example, if someone says, 'A celebrity donated $1 million,' the detective searches for news articles, finds credible reports, and confirms the claim.
ELI14 Explained like you're 14
Think of this like when your friend says, 'This celebrity donated $1 million!' You're not sure, so you Google it. You find news articles saying it's true, and you tell your friend, 'Yep, it's real!' This research builds an AI to do exactly that—help us quickly check if something is true or not!
Glossary
LLM (Large Language Model)
An AI model trained on vast text data, excelling at language understanding and generation.
Used to decompose claims and reason about their truth.
RAG (Retrieval-Augmented Generation)
A method combining external retrieval with generative models to handle real-time information.
Used to fetch evidence from the web for reasoning.
LIAR-New
A dataset of political statements with factuality labels.
Used to evaluate misinformation detection systems.
Cohere RAG
A tool combining Cohere Chat API with web retrieval.
Used to retrieve evidence for claims.
DuckDuckGo
A privacy-focused search engine with a free API.
Used as a retrieval tool to fetch web evidence.
Open Questions Unanswered questions from this research
- 1 How can misinformation detection work offline without internet access?
- 2 How can non-textual evidence like images or videos be better utilized?
- 3 How to reduce reliance on specific retrieval sources while maintaining accuracy?
Applications
Immediate Applications
Social Media Moderation
Helps platforms detect misinformation quickly, reducing harmful spread.
News Fact-Checking
Provides journalists with real-time claim verification tools, improving reporting accuracy.
Long-term Vision
Multilingual Misinformation Detection
Expands to global contexts, supporting diverse languages and cultures.
Abstract
This paper develops an agent-based automated fact-checking approach for detecting misinformation. We demonstrate that combining a powerful LLM agent, which does not have access to the internet for searches, with an online web search agent yields better results than when each tool is used independently. Our approach is robust across multiple models, outperforming alternatives and increasing the macro F1 of misinformation detection by as much as 20 percent compared to LLMs without search. We also conduct extensive analyses on the sources our system leverages and their biases, decisions in the construction of the system like the search tool and the knowledge base, the type of evidence needed and its impact on the results, and other parts of the overall process. By combining strong performance with in-depth understanding, we hope to provide building blocks for future search-enabled misinformation mitigation systems.