Source-linked AI summary
Improving Fake News Detection by Using an Entity-enhanced Framework to Fuse Diverse Multimodal Clues
Peng Qi, Juan Cao, Xirong Li, Huan Liu, Qiang Sheng, Xiaoyue Mi, Qin He, Yongbiao Lv, Chenyang Guo, Yingchao Yu
TL;DR
Multimodal fake news detection is limited by insufficient modeling of image semantics and text-image relationships. EM-FEND extracts visual entities and embedded text to model entity inconsistency, mutual enhancement, and text complementation, and experiments report superiority over state-of-the-art methods.
Problem
Existing multimodal detectors often model only basic image semantics as a supplement to text, while overlooking diverse cross-modal correlations and high-level visual semantics.
Method
EM-FEND extracts textual and visual entities, OCR text, and visual features, then jointly models text complementation, mutual enhancement, and entity inconsistency.
Results
Extensive experiments, including offline and online evaluations, demonstrate EM-FEND’s superiority over state-of-the-art methods.
Takeaways & Limitations
Visual entities help represent news-related high-level image semantics and bridge semantic correlations between news text and images.
Takeaways & Limitations
The model assumes a single image per news item for simplification, although it is applicable to news containing multiple images.
Abstract
from arXiv · showhide
Recently, fake news with text and images have achieved more effective diffusion than text-only fake news, raising a severe issue of multimodal fake news detection. Current studies on this issue have made significant contributions to developing multimodal models, but they are defective in modeling the multimodal content sufficiently. Most of them only preliminarily model the basic semantics of the images as a supplement to the text, which limits their performance on detection. In this paper, we find three valuable text-image correlations in multimodal fake news: entity inconsistency, mutual enhancement, and text complementation. To effectively capture these multimodal clues, we innovatively extract visual entities (such as celebrities and landmarks) to understand the news-related high-level semantics of images, and then model the multimodal entity inconsistency and mutual enhancement with the help of visual entities. Moreover, we extract the embedded text in images as the complementation of the original text. All things considered, we propose a novel entity-enhanced multimodal fusion framework, which simultaneously models three cross-modal correlations to detect diverse multimodal fake news. Extensive experiments demonstrate the superiority of our model compared to the state of the art.
1 INTRODUCTION
Multimodal fake news detection must address shortcomings in basic text-image fusion by modeling diverse cross-modal clues and high-level visual semantics. EM-FEND extracts entities and embedded image text to jointly model these relationships.
- Existing multimodal detectors often concatenate textual features with basic VGG19-derived image features, insufficiently modeling multimodal content.
- Three useful correlations are entity inconsistency, mutual enhancement, and text complementation between news text and images.Inconsistency can arise from wrongly reposted images; images can highlight key textual entities; embedded image text can complete the story.
- More than 20% of multimodal fake news in preliminary Weibo statistics spreads in image form, with embedded text often carrying the complete story.
- Basic visual objects create a semantic gap with named entities in news text, making cross-modal clues difficult to reason about.Recognizing a celebrity only as “person” can hide inconsistency between the text and image.
- EM-FEND extracts textual entities, visual entities, embedded image text, and visual features, then fuses the three cross-modal correlations for detection.Visual entities capture high-level image semantics and bridge text-image semantic correlations.
- Offline and online evaluations demonstrate EM-FEND’s superiority compared with the state of the art.
2 RELATED WORK
Prior multimodal fake news studies commonly combine general visual features with text, while other work addresses event generalization or cross-modal inconsistency and enhancement.
- A common framework extracts general VGG19 image features and concatenates them with textual features.
- Some approaches add event classification or meta neural processes to improve detection across events and emergent events.
- Other studies measure multimodal inconsistency using irrelevant-image detection or image-caption similarity.
- A cited example illustrates irrelevant-image reuse through a comment about a favorite actor paired with an image context.
Image Input
EM-FEND’s image-input and fusion design combines visual representations with textual and entity-level signals, while prior methods largely use basic visual features or caption-based consistency.
- EM-FEND extracts OCR text and visual CNN features alongside textual and visual entities before fusion.
- The framework models text complementation, mutual enhancement, and entity inconsistency, then concatenates the resulting multimodal features for classification.
- Prior inconsistency methods compare original text with generated image captions, but captioning performance can be limited by corpus mismatch with real-world news.
- Earlier mutual-enhancement methods focus on unidirectional or object-level fusion and ignore cross-modal enhancement on high-level semantics.
- The paper addresses these gaps by explicitly extracting visual entities and embedded image text to fuse diverse multimodal clues.
3 ENTITY-ENHANCED MULTIMODAL FAKE NEWS DETECTION
EM-FEND detects multimodal fake news by jointly modeling textual and visual information through entity extraction, embedded-text complementation, co-attention, and entity inconsistency measurement.
- 3.1 Model Overview: EM-FEND extracts textual entities, visual entities, embedded image text, and regional CNN features before multimodal fusion.Visual entities include celebrities, landmarks, organizations, salient concepts, objects, and scenes; embedded text is extracted with OCR.
- 3.1 Model Overview: The framework models entity inconsistency, mutual enhancement, and text complementation as three cross-modal correlations for binary fake-news classification.Its multimodal representation combines text, visual entities, image features, and entity-consistency features before classification.
- 3.3.1 Text Complementation: Original and embedded image text are jointly encoded with BERT, allowing image text to complement the original news text.The two text sources are separated by [SEP], producing contextual features for the composed text.
- 3.3.2 Mutual Enhancement: A two-stream co-attention transformer exchanges queries across modalities to produce image-enhanced text and text-enhanced visual features.Co-attention is applied between textual features and visual entities, then between enhanced text and visual CNN features.
- 3.3.3 Entity Inconsistency Measurement: Entity inconsistency compares textual and visual person, location, and context entities in a shared textual feature space.The model treats news as entity-inconsistent only when aligned multimodal entities are absent, accounting for unaligned textual entities.
4 EXPERIMENTS
The experiments evaluate EM-FEND’s classification performance, feature and correlation effectiveness, and online robustness across Chinese and English multimodal news datasets.
- Evaluation Questions: The experiments address classification performance, the contribution of visual features and cross-modal correlations, and online fake-news detection.These are defined as EQ1, EQ2, and EQ3, respectively.
- Datasets: Experiments use two real-world datasets in different languages to evaluate generalization.The Chinese dataset comes from Sina Weibo, while the English dataset contains multimodal articles from credibility-assessed and authoritative news websites.
- Datasets: The English dataset contains 2,844 fake and 2,825 real news articles, each paired with an image.Text-only, non-English, and unavailable-image items were removed.
- Dataset Splitting and Metrics: Event-cluster splitting prevents overlap among training, validation, and testing sets, which are approximately distributed at a 3:1:1 ratio.Accuracy, precision, recall, and fake-news F1 are used as evaluation metrics.
4.2 Implementation Details
Implementation uses pretrained language and vision models, compares EM-FEND with representative baselines, and includes a reduced Word2Vec/Bi-LSTM variant for fair comparison.
- Implementation: BERT models provide textual representations, with a maximum sequence length of 256 for both datasets.Chinese and English use bert-base-chinese and bert-base-uncased, respectively.
- Implementation: The co-attention transformer uses 8 heads, while EM-FEND and EM-FEND-base use hidden sizes of 256 and 128, respectively.EM-FEND-base also uses a 128-dimensional LSTM.
- Comparison Methods: EM-FEND is compared with single-modality and multimodal representative methods for multimodal fake-news detection.The comparison includes Bi-LSTM, BERT, VGG19, attRNN, MVAE, SAFE, and SpotFake.
- Comparison Methods: EM-FEND-base replaces BERT with Bi-LSTM and pretrained Word2Vec embeddings to make comparisons fairer.The reduced variant also uses Word2Vec for textual and visual entity embeddings.
4.4 Performance Comparison (EQ1)
EM-FEND substantially outperforms representative methods on both datasets, with or without BERT, demonstrating the value of modeling multimodal clues. Multimodal methods generally outperform single-modality methods, while BERT further improves EM-FEND.
- EM-FEND and EM-FEND-base outperform corresponding state-of-the-art methods by at least 3.8 and 3.4 percentage points in accuracy, respectively.
- EM-FEND performs much better than other methods on both datasets regardless of whether BERT is used as the textual feature extractor.
- Multimodal methods generally outperform single-modality methods, indicating complementarity between textual and visual features.
- BERT improves EM-FEND performance, attributed mainly to transformers’ contextual modeling and the knowledge injected by pre-training.
4.5 Ablation Study (EQ2)
The ablation study finds that all three visual feature types and the modeled cross-modal correlations contribute to EM-FEND. Their relative importance differs across datasets, and replacing co-attention with averaging reduces accuracy.
- Visual Features: All three visual feature types are important, but fine-tuned VGG features matter most in the Chinese dataset while visual entities matter most in the English dataset.
- Visual Features: The differing feature importance reflects dataset sources: Chinese social-media images are lower quality, whereas English news images are higher-quality and more informative.
- Cross-modal Correlations: Removing any cross-modal correlation reduces performance, showing that the correlations collectively contribute to EM-FEND’s best results.
- Cross-modal Correlations: Replacing the co-attention transformer with averaging lowers accuracy by at least 1.4 percentage points, supporting its role in multimodal alignment.
- Cross-modal Correlations: Entity consistency has less influence than other cross-modal correlations, possibly because visual entities are sparse and entity detectors introduce noise.
4.6 Robustness to Imbalanced Online Data (EQ3)
EM-FEND is evaluated on a severely imbalanced online dataset containing far fewer fake than real news posts. Its ROC curves show better performance than CARMN in this setting.
- The online dataset contains 217 multimodal fake news posts and 3353 real news posts, approximately a 1:15 ratio.
- EM-FEND outperforms CARMN on the imbalanced online dataset according to their ROC curves.
- The comparison evaluates EM-FEND’s robustness to imbalanced online data against CARMN, identified as its best competitor.
4.7 Case Study
The case study examines multimodal fake news with low person consistency. These examples illustrate how EM-FEND’s entity inconsistency module measures disagreement between textual and visual person entities.
- The case study presents multimodal fake news measured as having low person consistency by EM-FEND’s entity inconsistency module.
- The module provides interpretable explanations by measuring multimodal entity inconsistency in these cases.
- In the examples, the person entity mentioned in the text is inconsistent with the person recognized in the image.
5 CONCLUSION
The paper identifies three cross-modal correlations in multimodal fake news and proposes EM-FEND to model them using visual entities.
- EM-FEND simultaneously models entity inconsistency, mutual enhancement, and text complementation.
- Visual entities help capture news-related visual semantics and multimodal clues.
- Extensive experiments demonstrate EM-FEND’s effectiveness.