PaperlayerView exact revision on arXiv

Source-linked AI summary

RAEGNet: Relation-Aware Evidence Graph Network for Harm-Aware Multimodal Fake News Detection

Wenbin Shen, Guoxuan Qin, Guangxu Yao, Baodong Wang, Yuanbo Rui, Zhongjie Ba, Zhichao Lian

arXiv:2609.36902v1cs.CLcs.CVcs.MM

TL;DR

Multimodal fake-news detection must contend with event-irrelevant retrieval and limited attention to differences in potential harm. This paper combines event-level evidence retrieval with a relation-aware graph and conditional-harm learning; experiments report overall gains over baselines and strong high-harm detection. The authors also identify dependence on retrieval quality and LLMs, including added inference latency.

  • Problem

    Existing methods often rely on entity-level retrieval that can introduce event-irrelevant noise, while many focus on aggregate performance rather than integrating varying harm judgments into multimodal detection.

    Source: §1 · p. 1

  • Method

    ELERF retrieves evidence using complete news-event semantics, while RAEGNet models news-evidence and evidence-evidence relations and jointly learns authenticity and conditional harm.

    Source: §1 · p. 2

  • Results

    RAEGNet outperforms baselines overall across three datasets and exhibits strong capability in identifying high-harm fake news.

    Source: §5 · p. 8

  • Takeaways & Limitations

    The approach combines authenticity detection with harm-risk assessment and relational evidence reasoning across the evaluated datasets.

    Source: §5 · p. 8

    The system depends on search-engine recall and LLM-assisted stance scoring, and LLM use increases inference latency, posing efficiency challenges for high-concurrency real-time detection.

    Source: Limitations · p. 9

Abstract

from arXiv · show

Existing multimodal fake news detection methods often introduce external information to assist detection. However, most of them rely on entity-level retrieval and are therefore prone to introducing event-irrelevant noise. Meanwhile, existing methods mainly focus on improving overall performance and do not account for differences in the degree of harm posed by different instances of fake news. To address these limitations, we design an Event-Level Evidence Retrieval Framework (ELERF) and propose a Relation-Aware Evidence Graph Network (RAEGNet). ELERF retrieves external evidence based on the complete event semantics of a news item. RAEGNet constructs a directed graph that incorporates news-evidence stance relations and evidence-evidence interaction relations, and introduces a conditional-harm branch to jointly model authenticity and potential harm. Experimental results demonstrate that RAEGNet outperforms multiple baseline methods across all evaluated metrics on Weibo-21, Fakeddit, and our self-constructed SSS dataset.

1 Introduction

The introduction identifies event-irrelevant retrieval noise and insufficient attention to varying fake-news harm as gaps. It presents event-level retrieval, relation-aware evidence reasoning, harm-aware learning, and the SSS dataset as responses.

1 Introduction

  • Multimodal fake-news detection must assess authenticity while prioritizing highly harmful items, whose social costs can be greater than ordinary rumors.The introduction contrasts political deepfakes and public-health rumors with poorly fabricated entertainment gossip.Source: §1 · p. 1
  • Existing methods often retrieve by entity or shallow matching and focus on aggregate classification, leaving event-level relevance and integrated multimodal harm judgments insufficiently addressed.The introduction notes that high-harm misclassification carries greater social cost than misclassifying an ordinary rumor.Source: §1 · p. 1
  • ELERF queries with the complete news text and retrieves external evidence through retrieval, filtering, and stance quantification.The stated aim is to obtain evidence strongly related to the complete news event and reduce semantic noise.Source: §1 · p. 2
  • RAEGNet models news-evidence stance and evidence-evidence interactions in a directed graph with nine edge types.The proposed graph jointly captures relation types and strengths.Source: §1 · p. 2
  • SSS contains 7,997 Chinese and English multimodal news samples with harm labels, external evidence, and evidence stance scores.The dataset is assembled from six public sources and undergoes validity checks, deduplication, semantic filtering, and manual review.Source: §1 · p. 2
  • The harm-aware objective jointly learns authenticity classification, conditional-harm estimation, and ordinal harm consistency.This is presented as a joint optimization objective.Source: §1 · p. 2

2 Related Work

Prior multimodal fake-news work progresses from basic feature fusion toward fine-grained cross-modal interaction, while knowledge-enhanced methods draw on several kinds of external information. The review identifies limits in detecting deeply disguised fake news and in avoiding irrelevant retrieval noise.

2.1 Multimodal Fake News Detection

  • Early multimodal methods used CNNs and RNNs for visual and textual features, fusing them with concatenation or attention.The cited examples are Jin et al. (2017) and Yang et al. (2018).Source: §2.1 · p. 2
  • Later methods capture cross-modal ambiguity or inconsistency with fuzzy reasoning, consistency graphs, and contrastive learning, but struggle with deeply disguised fake news.Examples include CAFE, InfoSurgeon, and CCGN.Source: §2.1 · p. 2

2.2 Knowledge-Enhanced Detection Methods

  • Knowledge-enhanced detection uses static knowledge graphs, open-domain retrieval, and LLM or multi-agent systems.The review gives CompareNet and AKA-Fake, MUSER, and FKA-Owl and RAMA as examples, respectively.Source: §2.2 · p. 2
  • Many knowledge-enhancement strategies still use entity-level or shallow semantic matching, making them prone to irrelevant noise that can degrade overall performance.The review identifies this as a limitation across the discussed strategies.Source: §2.2 · p. 2

3 Methodology

RAEGNet combines event-level evidence retrieval with multimodal feature fusion and a relation-aware graph for authenticity reasoning. A conditional-harm branch is jointly optimized with authenticity classification to model harm magnitude and ordering.

3.1 Overview

  • RAEGNet encodes news text, image, and evidence, then uses a relation-aware evidence graph for authenticity reasoning and a harm-risk branch to estimate conditional harm.Source: §3.1 · p. 2

3.2 Event-Level Evidence Retrieval

  • ELERF retrieves snippets using the complete news text, removes explicit and implicit label leakage, and assigns each remaining evidence item a discrete stance score.Source: §3.2 · p. 3

3.3 Feature Encoding and Fusion

  • The model encodes news text, image, and evidence in a unified space, then computes a fused news feature using learnable text and image feature weights.Source: §3.3 · p. 3

3.4 Relation-Aware Evidence Graph Network

  • The graph represents news–evidence stance relations and evidence–evidence interactions, encoding edge strength and relation type for attention-based message passing and pooling.Source: §3.4 · p. 4

3.5 Joint Optimization with Harm Risk

  • A joint objective trains authenticity classification and conditional-harm estimation, using pointwise harm fitting and ordinal ranking to capture harm magnitude and order.Source: §3.5 · p. 5

4 Experiments

Experiments compare RAEGNet with task-specific and LLM-based baselines across three datasets, then examine overall performance, component contributions, harm-level behavior, retrieval granularity, sensitivity, and case studies.

4.1.1 Datasets

  • Weibo-21 and Fakeddit are Chinese and English multimodal datasets, while SSS is a Chinese–English dataset curated from six sources through deduplication, quality screening, and manual review.Human annotators systematically assigned harm labels for all three datasets; Table 1 reports news and automatically retained evidence counts.Source: §4.1.1 · p. 5, Table 1 · p. 5

4.1 Experimental Setup

  • Experiments use five-fold cross-validation, conventional accuracy, precision, recall, and F1, plus HHF metrics for held-out samples with hi ≥0.7.The high-harm subset treats fake news as positive and high-harm real news as negative; RAEGNet-B is the main early-detection setting and RAEGNet-A the post-hoc setting.Source: §4.1.2 · p. 6, §4.1.3 · p. 6, §4.1.4 · p. 6

4.2 Overall Performance

  • RAEGNet-B consistently outperforms baselines in early detection, RAEGNet-A performs strongest post hoc, and both event-level variants exceed RAEGNet-E in accuracy and F1 across all three datasets.Table 2 presents the overall comparison; RAEGNet-B uses pre-news event-level evidence, RAEGNet-A uses pre- and post-news evidence, and RAEGNet-E uses entity-level evidence.Source: Table 2 · p. 6, §4.2 · p. 7

4.3 Ablation Study

  • Removing the evidence graph, edge types, or edge weights generally reduces accuracy and HHF F1, while removing harm-aware components reduces HHF F1 and HHF Recall.The nine ablation variants also show substantial contributions from news text, image, and external evidence.Source: §4.3 · p. 7, Table 3 · p. 7

4.4 Harm-Level Analysis

  • RAEGNet achieves higher detection accuracy and better overall recognition performance for fake news with greater potential harm.The harm-risk branch and ordinal ranking loss model conditional-harm magnitude and ordinal structure, providing complementary supervision for high-harm fake-news recognition.Source: §4.4 · p. 7, Figure 3 · p. 7

4.5 Effectiveness of Event-Level Retrieval

  • Event-level retrieval improves accuracy and fake/real F1 across all three datasets compared with entity-level evidence under the same RAEGNet architecture.The comparison uses leakage-filtered evidence retrieved by ELERF for the event-level setting.Source: §4.5 · p. 8, Table 4 · p. 8

4.6 Sensitivity Analysis

  • λharm = 0.25 and λrank = 0.25 yield relatively balanced results across all three datasets.Weights that are too small weaken conditional-harm supervision, whereas excessively large weights may impair basic authenticity classification.Source: §4.6 · p. 8, Figure 4 · p. 8

4.7 Case Study

  • In both case studies, RAEGNet correctly classifies the news using retrieved evidence, while MSACA predicts incorrectly.The examples concern Aloha Airlines Flight 243 and a false claim about mail-in ballots in the 2020 U.S. election.Source: §4.7 · p. 8, Figure 5 · p. 8

5 Conclusion

RAEGNet combines relation-aware evidence reasoning with conditional-harm learning for fake-news detection. Across three datasets, it outperforms baselines overall and shows strong capability in identifying high-harm fake news.

5 Conclusion

  • RAEGNet unifies a relation-aware evidence graph with conditional-harm learning for fake-news detection.The model uses multiple relation types and edge weights to support news-evidence interaction while jointly modeling harm-risk assessment and authenticity classification.Source: §5 · p. 8
  • Across three datasets, RAEGNet outperforms baselines overall and shows strong capability in identifying high-harm fake news.The conclusion describes the overall comparison without reporting metric values.Source: §5 · p. 8
  • The model jointly addresses harm-risk assessment and authenticity classification.The conclusion presents these as jointly modeled tasks.Source: §5 · p. 8

Limitations

RAEGNet’s evidence acquisition and stance scoring depend on external resources and LLMs, and LLM use adds latency that challenges real-time, high-concurrency detection.

Limitations

  • Evidence acquisition depends on the initial recall of search engines.The limitation concerns the initial-recall stage of event-level evidence acquisition.Source: Limitations · p. 9
  • Stance-score quantification requires LLM assistance.The model therefore relies on LLMs for this evidence-processing step.Source: Limitations · p. 9
  • LLM use increases inference latency, challenging real-time streaming detection under high concurrency.The authors identify computational efficiency as the concern in this deployment setting.Source: Limitations · p. 9

Ethical Considerations

The paper describes safeguards for research data use and a staged process for constructing SSS, including semantic screening, human review, and event-aware fold assignment.

Ethical Considerations

  • The study uses public datasets and web-accessible evidence for academic misinformation research, with data use following original access, licensing, and privacy requirements.The framework is intended for defensive analysis, not surveillance, censorship, or automated decisions without appropriate human oversight.Source: Ethical Considerations · p. 9

A Construction of the SSS Dataset

  • SSS was assembled to address low-quality samples that may not faithfully reflect model capability.The authors screened and assembled the subset from several existing public datasets.Source: A Construction of the SSS Dataset · p. 12

A.1 Basic Data Cleaning

  • Initial SSS cleaning removed text shorter than 15 words or Chinese characters, damaged or unavailable images, and residual markup or garbling; multimodal near-duplicates above 0.85 similarity were manually checked.A pair was removed as duplicate only when it expressed the same factual claim and contained the same or nearly identical visual context.Source: §A.1 · p. 12

A.2 LLM-Based Semantic Filtering

  • LLM-based semantic filtering jointly evaluates each news text and image against structured screening criteria.The criteria and examples are presented in Table 5.Source: §A.2 · pp. 12–13, Table 5 · p. 12

A.3 Manual Review

  • Five reviewers independently assessed samples after LLM filtering, retaining them with explicit approval from at least four reviewers and discussing uncertain cases.Reviewers selected 7,893 of 8,597 samples judged usable by the LLM and 104 of 312 uncertain samples.Source: §A.3 · p. 13

A.4 Event Grouping and Fold Construction

  • Event-related candidates above 0.85 multimodal similarity were judged by reviewers, and each event group was kept within one fold to reduce direct event overlap.Reviewers assessed event identity using entities, action, time, location, and factual content; folds also preserved distributions of authenticity labels, languages, sources, and harm levels.Source: §A.4 · p. 13, Table 7 · p. 13

A.5 Analysis of Dataset Composition

The paper reports strong agreement for human harm labels and substantial adjudication was uncommon. SSS is a bilingual dataset, and harm is defined and annotated conditionally on the news being false.

A.5 Analysis of Dataset Composition

  • SSS contains 7,997 samples—3,836 real and 4,161 fake—with 2,614 Chinese and 5,383 English samples and an average harm coefficient of 0.489.The authors report that the relatively balanced authenticity distribution mitigates severe class imbalance and the bilingual composition supports evaluation across linguistic contexts.Source: §A.5 · p. 13

B.1 Definition of Harm Labels

  • Conditional harm is the potential social impact of news if it were false and disseminated through social networks, independent of its current authenticity.Source: §B.1 · p. 14

B.2 Manual Annotation

  • Five annotators assigned each item one of five ordered harm levels, mapped to 0.1, 0.3, 0.5, 0.7, or 0.9.They averaged assignments and mapped to the nearest level, discussing cases whose ratings spanned at least two levels.Source: §B.2 · p. 14

B.3 Automated LLM-Based Annotation

  • The auxiliary LLM procedure outputs a continuous harm score under a false-and-widely-disseminated assumption, then maps it to one of five levels; experiments use human labels.The authors describe automated annotation as a way to validate scaling feasibility and obtain initial labels for larger datasets.Source: §B.3 · p. 14

B.4 Annotation Agreement and Validation of the Automated Procedure

  • α > 0.90 inter-annotator agreement held on all three datasets, adjudication was below 1.5% on each, and mean LLM-human agreement was close to or above 90%.The authors interpret these results as indicating operational harm-level definitions, rare substantial disagreement, and the ability of automated annotation to approximate human risk judgments.Source: §B.4 · p. 15

B.5 Distribution of Harm Coefficients

This section describes harm patterns in SSS and the evidence pipeline used to evaluate them. It also distinguishes early detection from post-hoc verification by what evidence is available.

B.5 Distribution of Harm Coefficients

  • Real-news samples cluster mainly at harm coefficients 0.3–0.7, while fake-news counts rise markedly at 0.7 and 0.9.Figure 6 shows the distribution of real- and fake-news samples across discrete harm coefficients.Source: §B.5 · p. 15, Figure 6 · p. 15
  • Under harm-matched evaluation, the full model remains stronger than the variant without the Harm-aware Branch, beyond marginal harm-label associations.Real and fake samples are balanced within each harm level in the held-out set.Source: §B.5 · p. 15

C.1 Retrieval

  • ELERF uses the complete news text to retrieve the top 10 snippets from a predefined source pool, then partitions candidates by publication time.The pool is fixed before evaluation and covers encyclopedic resources, media, institutions, academic sources, specialized domains, and public communication platforms.Source: §C.1 · p. 15

C.2 Filtering

  • Leakage control excludes verdict-revealing material through automatic filtering and LLM screening, while allowing independent post-news facts that still require claim comparison.The criteria include direct verdicts, dataset annotations, and fact-checking conclusions; publication time alone does not make evidence leakage.Source: §C.2 · pp. 16–17

C.3 Stance Quantification

  • For leakage-free candidates, ELERF assigns continuous stance scores and maps them into five discrete values from 0.1 to 0.9.Lower scores indicate refutation, higher scores support, and middle scores indicate unclear stance, insufficient relevance, or inadequate information.Source: §C.3 · p. 17

C.4 Post-hoc Validation of Evidence Filtering

  • Post-hoc validation checks relevance, leakage, and stance consistency without exposing annotators to authenticity labels or fold assignments.Reviewers treated independent factual developments as admissible but direct benchmark-label or final-verdict disclosure as leakage.Source: §C.4 · p. 17
  • In SSS, real news has more retrieved evidence than fake news and more evidence in the high-support interval.Figure 7 presents the stance-score distribution; the authors interpret the pattern as greater external corroboration for real-world events.Source: §C.5 · p. 18, Figure 7 · p. 18

C.6 Evidence Settings and Intended Scenarios

  • RAEGNet-B uses leakage-controlled pre-news evidence for early detection, whereas RAEGNet-A adds admissible post-news evidence for post-hoc verification.Both settings filter direct benchmark-label and final-verdict disclosure; their distinction is information availability, not filtering quality.Source: §C.6 · pp. 18–19

C.7 Retrieval Case Analysis

The retrieval cases contrast event-level evidence with entity-focused background retrieval. The appendix also specifies the LLMs and generation settings used across core modules and baselines.

C.7 Retrieval Case Analysis

  • Event-level queries more often retrieve documents about the news item’s specific action, claim, or factual circumstance than entity-level queries do.The cases are qualitative examples, not a separate large-scale relevance audit.Source: §C.7 · p. 20

D Statement on the Use of LLMs

  • The study describes its LLM selection and application settings to support transparency and reproducibility.The statement covers experimental models, application scenarios, and hyperparameter configurations.Source: D Statement on the Use of LLMs · p. 20

D.1 Model Selection for Core Modules

  • Qwen2.5-VL-7B-Instruct supports SSS sample cleaning, evidence screening and stance assessment, while Qwen2.5-VL-7B-Instruct and InternVL2.5-8B serve as LLM baselines.The LLM baselines receive news text, image, and retrieved evidence snippets in a locally deployable zero-shot setting.Source: §D.1 · p. 20

D.2 Global Generation Parameters

  • LLM calls use temperature 0.3, top-p 0.85, and a 64-token maximum generation length.The configuration is intended to stabilize output format and reduce API inference latency.Source: §D.2 · p. 20

E.1 Experimental Parameter Settings

The experiments compare RAEGNet with internal-content, external-knowledge, and LLM-based baselines. Across three datasets, RAEGNet reports stronger results, with event-level retrieval and broader post-hoc evidence offering further gains.

E.1 Experimental Parameter Settings

  • The experimental-parameter appendix points to Table 15 for the detailed settings.The supplied table caption identifies it as the detailed experimental settings table.Source: §E.1 · p. 20, Table 15 · p. 20

E.2 Baseline Models

  • The baseline set includes methods using internal multimodal content, external knowledge, and LLMs.Examples include CAFE and MRML, KEHGNN-FD and ERIC-FND, and direct multimodal models or GLPN-LLM.Source: §E.2 · pp. 20–21

E.3 Prompt for w/ MLLM Baselines

  • Direct multimodal LLM baselines receive a fixed zero-shot prompt and output only a binary authenticity label.The prompt jointly supplies the news text, image, and retrieved evidence snippets.Source: §E.3 · p. 21

F Complete Experimental Results

  • RAEGNet-E uses entity-level evidence, RAEGNet-B uses leakage-controlled pre-news event evidence, and RAEGNet-A combines pre- and post-news event evidence.The variants represent a controlled retrieval-granularity comparison, early detection, and post-hoc fact-checking, respectively.Source: F Complete Experimental Results · p. 21

F.1 Overall Performance

  • RAEGNet outperforms task-specific, external-knowledge-enhanced, and direct LLM-based baselines across Weibo-21, Fakeddit, and SSS.The comparison also finds event-level evidence more effective than entity-level evidence under the same architecture, with broader leakage-controlled evidence adding gains in the post-hoc setting.Source: §F.1 · p. 21

F.2 Ablation Study

The analyses test how evidence relations, harm supervision, and information sources contribute to authenticity and harm prediction. They report benefits from relation-aware modeling, event-level retrieval, and the harm-aware branch, including under harm-matched evaluation.

F.2 Ablation Study

  • Removing the evidence graph, edge weights, or edge types weakens RAEGNet; removing text, image, or evidence also reduces performance.The ablations support the contribution of relation-aware evidence modeling and all three information sources.Source: §F.2 · pp. 21–22, Table 20 · p. 24, Table 21 · p. 24

F.3 Performance Across Harm Levels

F.4 Conditional-Harm Prediction

  • The RAEGNet harm branch matches human harm annotations better than a text-only predictor.The authors attribute this result to useful signals captured by the multimodal representation and learned harm-risk branch.Source: §F.4 · p. 22, Table 26 · p. 26

F.5 Harm-Matched Evaluation

  • The full model remains better than the ablated variant when real and fake samples are balanced within each harm level.This matched evaluation reduces the possibility that performance comes merely from marginal correlations between harm and authenticity labels.Source: §F.5 · p. 22, Table 27 · p. 26

F.6 Analysis of Retrieval and Model Contributions

  • Under a shared source pool, filtering process, and evidence budget, event-level retrieval beats entity-level retrieval, while RAEGNet beats simple evidence fusion.The controlled comparisons separate retrieval-granularity gains from the gains of relation-aware graph modeling.Source: §F.6 · p. 22, Table 28 · p. 26

F.7 Event-Level Evidence for Other w/ EK Models

  • Replacing external evidence in knowledge-enhanced baselines with ELERF evidence consistently improves them, especially entity-oriented methods.The authors report this as evidence of ELERF’s usefulness beyond its pairing with RAEGNet.Source: §F.7 · p. 22, Table 29 · p. 27

Related papers

Loading 2609.36902v1…