Enhancing Multimodal Misinformation Detection by Replaying the Whole Story from Image Modality Perspective

AI summary

Overview

Research area: Multimodal misinformation detection (MMD) — automated detection of fake versus real social media posts that pair text with an image, combining computer vision, NLP, and information theory.

Technical level: Advanced. The paper assumes familiarity with stable diffusion text-to-image generation, graph neural networks, cross-attention fusion, and mutual-information objectives.

Scope in one sentence: The paper argues that the image modality contributes far less than text to MMD systems, and proposes a method, RetSimd, that splits a post's text into segments, generates an image for each segment so the whole story is "replayed" visually, and fuses all images (plus the original) through a graph neural network before classification.

What This Paper Is About

Existing MMD models generally treat text and image as equally informative, but the authors observe that a post's text usually narrates the entire event or story while its image typically shows only one partial scene. Because of this asymmetry, the image modality adds little to detection. RetSimd addresses the problem by expanding the image side: each text is chopped into segments, each segment is fed to a text-to-image generator, and the resulting sequence of generated images — together with the original image — is fused to predict whether the post is real or fake.

Key Contributions

1. An empirical validation, across five state-of-the-art MMD methods and three benchmark datasets, that the text modality contributes substantially more to MMD accuracy than the image modality, supported by an information-theoretic "contribution degree" metric.
2. The RetSimd framework, which splits a post's text into a sequence of segments, feeds each segment into a pre-trained text-to-image generator, and augments the post with images intended to replay the whole story rather than a single partial scene.
3. Two information-theoretic regularizations tuning the generator — text–image mutual information and image–label mutual information — combined with post-training on an auxiliary text-to-image dataset to preserve image quality.
4. A graph-based multimodal fusion network built on three heuristic image relationships (central, temporal, and dependency), encoded with a graph neural network and combined with text through cross-attention.

Main Findings

• Text dominates in preliminary ablations: Across ResNet + BERT, MAE + DeBERTaV3, CLIP, CAFE, and BMR on GossipCop, Weibo, and Twitter, the accuracy gap between the full variant and the text-only / image-replaced variants was consistently much smaller than the gap for the image-only / text-replaced variants, and the image-only and text-replaced variants were in some cases "even ineffective to some extent."
• Quantified contribution gap: Under the paper's contribution-degree metric based on information gain, information gains from image-only/text-replaced variants exceeded those from text-only/image-replaced variants, which the authors read as text carrying the most substantial information for veracity prediction and images the least, "or even negative gain."
• RetSimd improves baselines on most metrics: On GossipCop, adding RetSimd raised ResNet + BERT accuracy from 87.17 ± 0.4 to 88.13 ± 0.2 (Δ 1.21), R&B + SAFE from 87.14 ± 0.6 to 88.30 ± 1.1 (Δ 1.29), R&B + BMR from 87.32 ± 0.3 to 88.42 ± 0.4 (Δ 1.02), and R&B + GAMED from 87.03 ± 0.5 to 88.30 ± 0.7 (Δ 1.81).
• Weibo results: ResNet + BERT improved from 90.38 ± 0.7 to 91.48 ± 0.7 (Δ 1.01); R&B + GAMED from 90.24 ± 0.5 to 92.08 ± 0.2 (Δ 1.84); R&B + MCAN from 90.58 ± 0.5 to 91.76 ± 0.3 (Δ 1.17).
• Twitter results and headline gain: The paper states that on Twitter, RetSimd outperforms its baseline model BMR by approximately 2.86 on the average of all metrics. In Table 1, R&B + BMR accuracy moves from 66.71 ± 1.4 to 69.57 ± 2.7 (Δ 2.87), and R&B + GAMED from 66.53 ± 2.5 to 69.77 ± 2.5 (Δ 3.56).
• Information-gain metric improves on most datasets: The reported values of G(y, x^v) fall relative to baselines in most cases, for example GossipCop ResNet + BERT from 0.0349 to 0.0301, Weibo ResNet + BERT from 0.0760 to 0.0632, and Twitter R&B + GAMED from 0.1122 to 0.0838. The authors interpret this as RetSimd raising the image modality's contribution and enabling more confident predictions.
• One noted exception: On Twitter with the BMR model, the G(y, x^v) result "did not achieve the optimal result." The authors explain that this model, when text is removed, tends to predict all samples as the real class with high confidence, producing low prediction entropy, and that RetSimd mitigates this tendency.
• Ablation confirms every module matters: Removing graph fusion, mutual information, or augmented images always hurt accuracy. On Twitter the decreases were 1.63 (w/o graph fusion), 2.15 (w/o mutual information), and 2.72 (w/o augmented images); on Weibo 0.68, 0.91, and 1.15; on GossipCop 0.63, 0.79, and 0.98. The consistent ranking was w/o graph fusion > w/o mutual information > w/o augmented image.

Methodology in Plain English

A post is represented as a text–image pair with a real/fake label. RetSimd has four modules. First, feature encoders use BERT for text and a ResNet34 pre-trained on ImageNet for images, with linear layers projecting both into a shared space. Second, a text-to-image generator — a stable diffusion model — takes the text, which is divided into K equal-length segments using a fixed-number sliding window (the paper reports that other segmentation strategies were evaluated in the Appendix and sliding windows performed best), and produces one generated image per segment. To keep those generated images faithful and label-relevant, the generator is tuned with two mutual-information objectives (text–image and image–label) in addition to the standard text-to-image training objective on an auxiliary generation dataset such as LAION-2B. Third, a multimodal fusion network builds a graph whose nodes are the K generated images plus the original image, with edges encoding three heuristic relationships: a central relationship connecting every generated image to the original image (assumed to show the most central scene), a temporal relationship connecting images whose text segments are adjacent, and a dependency relationship derived from a spaCy dependency tree whose token nodes are merged into segment-level nodes. A graph neural network with a fixed two layers encodes this graph into a fused image feature, which is combined with the text feature through a cross-attention network. Fourth, a veracity classifier (two linear layers plus an activation) outputs the prediction. Training alternates between optimizing the generator and optimizing the detector, with the detector loss being cross-entropy plus a weighted regularization term. Evaluation uses nine metrics: eight prediction metrics (accuracy, Macro F1, and precision/recall/F1 for the real and fake classes) plus the information-gain quantity G(y, x^v) that measures how much the image modality contributes.

Why This Matters

Impact on research. The paper challenges the common assumption that text and image deserve equal weight in MMD, and provides both an experimental and an information-theoretic argument for the asymmetry. It suggests a new direction: rather than only detecting inconsistencies between existing modalities, restructure the under-contributing modality by generating content that matches the information content of the stronger one.

Real-world applications (suggested by the paper's framing, not evaluated in it):

• Content moderation and trust-and-safety pipelines at social platforms such as Twitter and Weibo, which the paper names as major misinformation channels.
• Fact-checking workflows, where a richer visual representation of a claim's full narrative could support quicker verification.
• Automated early-warning systems for posts about hot events and breaking stories, which the paper identifies as especially prone to misinformation.
• Downstream forensic tooling that needs to reason about what an image does and does not show relative to its caption.

Industry relevance. Any organization running large-scale multimodal classifiers can treat RetSimd as a plug-in augmentation layer: the paper frames its contribution as consistently improving baseline models rather than replacing them, and its code is released at https://github.com/wangbing1416/RETSIMD. The main practical caveats are cost — every post requires running a diffusion model over K text segments — and the reliance on generated imagery, quality of which the authors explicitly protect with post-training and mutual-information regularizers.

Future Directions

• Better text segmentation and generator control. The paper fixes the segment count K and uses a sliding window, noting only that alternatives were tested in the Appendix; learned or content-aware segmentation is an open route.
• Extending the modality gap idea beyond tweet-like posts. The same story-versus-partial-scene asymmetry plausibly applies to video, audio, and longer documents, none of which are tested here.
• Scaling and multilingual validation. The three benchmarks used are GossipCop, Weibo, and Twitter; whether RetSimd's gains hold on larger, more diverse, or non-English corpora is not reported.
• Understanding the information-gain behavior. The paper documents an exception where the image-contribution metric did not improve optimally (BMR on Twitter) and attributes it to low-entropy "all real" predictions; a deeper analysis of when image augmentation helps or hurts is left open.

Target Audience

Researchers and graduate students in multimodal machine learning, misinformation detection, and social media analysis; practitioners building trust-and-safety or content-moderation classifiers; and anyone interested in generative augmentation and graph-based multimodal fusion. The paper's information-theoretic argument and ablation design make it especially useful to readers who want evidence about how much each modality actually contributes to a multimodal detector, rather than an assumed equal weighting.

Authors’ abstract

Multimodal Misinformation Detection (MMD) refers to the task of detecting social media posts involving misinformation, where the post often contains text and image modalities. However, by observing the MMD posts, we hold that the text modality may be much more informative than the image modality because the text generally describes the whole event/story of the current post but the image often presents partial scenes only. Our preliminary empirical results indicate that the image modality exactly contributes less to MMD. Upon this idea, we propose a new MMD method named RETSIMD. Specifically, we suppose that each text can be divided into several segments, and each text segment describes a partial scene that can be presented by an image. Accordingly, we split the text into a sequence of segments, and feed these segments into a pre-trained text-to-image generator to augment a sequence of images. We further incorporate two auxiliary objectives concerning text-image and image-label mutual information, and further post-train the generator over an auxiliary text-to-image generation benchmark dataset. Additionally, we propose a graph structure by defining three heuristic relationships between images, and use a graph neural network to generate the fused features. Extensive empirical results validate the effectiveness of RETSIMD.