LEMMA: Towards LVLM-Enhanced Multimodal Misinformation Detection with External Knowledge Augmentation

TL;DR

This paper proposes LEMMA: LVLM-Enhanced Multimodal Misinformation Detection with External Knowledge Augmentation, which leverages LVLM intuition and reasoning capabilities while augmenting them with external knowledge to enhance the accuracy of misinformation detection.

Abstract

The rise of multimodal misinformation on social platforms poses significant challenges for individuals and societies. Its increased credibility and broader impact compared to textual misinformation make detection complex, requiring robust reasoning across diverse media types and profound knowledge for accurate verification. The emergence of Large Vision Language Model (LVLM) offers a potential solution to this problem. Leveraging their proficiency in processing visual and textual information, LVLM demonstrates promising capabilities in recognizing complex information and exhibiting strong reasoning skills. In this paper, we first investigate the potential of LVLM on multimodal misinformation detection. We find that even though LVLM has a superior performance compared to LLMs, its profound reasoning may present limited power with a lack of evidence. Based on these observations, we propose LEMMA: LVLM-Enhanced Multimodal Misinformation Detection with External Knowledge Augmentation. LEMMA leverages LVLM intuition and reasoning capabilities while augmenting them with external knowledge to enhance the accuracy of misinformation detection. Our method improves the accuracy over the top baseline LVLM by 7% and 13% on Twitter and Fakeddit datasets respectively.

Figures (14)

• Figure 1: Comparison of performance metrics across various LVLMs/LLMs (GPT-3.5, GPT-4, GPT-4V, LLaVA, and InstructBLIP) and prompting methods (DIRECT and CoT) on two different datasets (Twitter and Fakeddit).
• Figure 2: An example of a real Fakeddit post where GPT-4V makes a correct prediction based on successfully extracting cross-modal alignment, while GPT-4 fails.
• Figure 3: An example of a fabricated Twitter tweet that shares subtle discrepancies in two modalities, misleading GPT-4V to answer "presence of misinformation"
• Figure 4: The pipeline of the proposed method (LEMMA). The process hinges on two key inputs: multimodal data and selectively filtered evidence gathered from external sources. Components marked with the OpenAI LOGO are developed using the LVLM (GPT-4V).
• Figure 5: A fake news example that image retrieval exposes as a reused promotional image