PaperlayerView exact revision on arXiv

Source-linked AI summary

End-to-End Multimodal Fact-Checking and Explanation Generation: A Challenging Dataset and Models

Barry Menglong Yao, Aditya Shah, Lichao Sun, Jin-Hee Cho, Lifu Huang

arXiv:2205.12487v2cs.CL

TL;DR

The paper addresses the limited ability of existing fact-checking systems to use multimodal evidence and generate explanations. It constructs Mocheg and evaluates an end-to-end pipeline for evidence retrieval, claim verification, and explanation generation, finding that performance across all three subtasks remains far from sufficient.

  • Problem

    Existing fact-checking studies often omit multimedia evidence or explanations, despite the need to justify truthfulness decisions using reliable information.

  • Method

    The paper constructs Mocheg and evaluates neural baselines for multimodal evidence retrieval, claim verification, and explanation generation.

  • Results

    Performance on all three Mocheg subtasks remains far from sufficient.

  • Takeaways & Limitations

    Mocheg provides a benchmark for end-to-end multimodal fact-checking and explanation generation.

  • Takeaways & Limitations

    Verification errors include cross-modality reasoning failures, and explanation generation is constrained by missing evidence, text-only evidence use, and weak logical organization.

Abstract

from arXiv · show

We propose end-to-end multimodal fact-checking and explanation generation, where the input is a claim and a large collection of web sources, including articles, images, videos, and tweets, and the goal is to assess the truthfulness of the claim by retrieving relevant evidence and predicting a truthfulness label (e.g., support, refute or not enough information), and to generate a statement to summarize and explain the reasoning and ruling process. To support this research, we construct Mocheg, a large-scale dataset consisting of 15,601 claims where each claim is annotated with a truthfulness label and a ruling statement, and 33,880 textual paragraphs and 12,112 images in total as evidence. To establish baseline performances on Mocheg, we experiment with several state-of-the-art neural architectures on the three pipelined subtasks: multimodal evidence retrieval, claim verification, and explanation generation, and demonstrate that the performance of the state-of-the-art end-to-end multimodal fact-checking does not provide satisfactory outcomes. To the best of our knowledge, we are the first to build the benchmark dataset and solutions for end-to-end multimodal fact-checking and explanation generation. The dataset, source code and model checkpoints are available at https://github.com/VT-NLP/Mocheg.

1 INTRODUCTION

The paper targets fact-checking systems that can use multimodal web evidence and explain their truthfulness decisions. It introduces Mocheg and baseline models for this end-to-end task, finding substantial room for improvement.

  • Misinformation complicates reliable online information seeking, while manually verifying claims requires substantial effort.
  • Existing fact-checking studies often ignore multimedia evidence, and some multimodal datasets use automatically generated labels or evidence without guaranteed human consistency.
  • The proposed task retrieves relevant sources, predicts claim truthfulness, and generates an explanation of the reasoning and ruling process.
  • Mocheg contains 15,601 claims with truthfulness labels, multimodal evidence, and ruling statements for benchmarking the task.
  • Baseline experiments indicate that end-to-end multimodal fact-checking and explanation generation remains substantially challenging.

2 RELATED WORK

Prior work covers multimodal misinformation detection and explainable fact-checking, but the paper distinguishes Mocheg by combining human-validated multimodal evidence with generated explanations.

  • Multimodal Fake News Detection and Fact-checking: Most earlier fake-news and fact-checking datasets are text-based, although newer work incorporates images and videos.
  • Multimodal Fake News Detection and Fact-checking: Multimodal methods commonly use cross-modality consistency or fused representations, while some predict truthfulness without explicit evidence.
  • Multimodal Fake News Detection and Fact-checking: The closest multimodal-evidence datasets automatically generate labels or evidence, whereas this work uses annotations from fact-checking journalists.
  • Explainable Fact-Checking: Explainable fact-checking approaches include direct evidence presentation, knowledge graphs, question proxies, and natural-language generation.
  • Explainable Fact-Checking: Directly presented evidence can be difficult to interpret because extracted sentences may lack logical connections.

3 DATASET CONSTRUCTION

Mocheg is built from journalist-verified Snopes and PolitiFact material, processed into human-labeled multimodal evidence, ruling explanations, and three end-to-end fact-checking tasks.

  • 3.1 Data Source: Snopes and PolitiFact provide claims, truthfulness labels, evidence references, and ruling articles written by journalists.
  • 3.2 Data Preprocessing: The dataset maps 68 original labels into Supported, Refuted, and NEI categories, removing claims assigned other labels.
  • 3.2 Data Preprocessing: Mocheg contains 15,601 claims, 33,880 text evidence paragraphs, and 12,112 image evidence items.
  • 3.2 Data Preprocessing: The collection includes 91,822 articles and 122,246 images, but only 30% of articles and 10% of images contain claim evidence.
  • 3.3 Task Definition: The benchmark defines multimodal evidence retrieval, claim verification, and explanation generation as its three subtasks.
  • 3.3 Task Definition: The framework combines text and image evidence retrieval with claim verification and explanation generation.
  • 3.4 Train / Dev / Test Split: The dataset is split into training, development, and test sets using 75%, 10%, and 15% proportions.

4 APPROACH

The framework addresses end-to-end multimodal fact-checking through evidence retrieval, claim verification, and explanation generation. It retrieves text and image evidence, fuses their stance representations to predict truthfulness, and generates ruling statements conditioned on claims, labels, and evidence.

  • Framework overview: The baseline framework contains three components corresponding to evidence retrieval, claim verification, and explanation generation.The approach is designed as an end-to-end multimodal fact-checking and explanation-generation framework.
  • Evidence retrieval: Text retrieval ranks document sentences using SBERT similarity and a BERT-based re-ranker, selecting the top-K evidence sentences.The re-ranking model is pretrained on MS MARCO Passage Ranking.
  • Evidence retrieval: Image retrieval uses CLIP to compute claim-image cosine similarities, ranks the image corpus, and selects the top-K candidate images.CLIP is fine-tuned with the same InfoNCE loss used for text evidence retrieval.
  • Claim verification: Claim verification pairs each evidence item with the claim, applies cross-attention for stance detection, and fuses updated and original claim representations using subtraction and multiplication.The resulting stance representations are aggregated separately across text and image evidence.
  • Claim verification: Mean-pooled text and image stance representations are concatenated to predict the truthfulness label with a cross-entropy objective.CLIP parameters are fixed during training while the other parameters are tuned.
  • Explanation generation: Explanation generation feeds the claim, truthfulness label, and text evidence into BART to generate a ruling statement, then uses a truthfulness reward for reinforcement learning.The reward is based on the difference between confidence in the correct label and confidence in wrong labels.
  • Assumption: The approach assumes evidence annotated by Politifact and Snopes is reliable when combining evidence stance for truthfulness prediction.The authors leave checking evidence trustworthiness for future work.

5 EXPERIMENTS

Experiments evaluate multimodal evidence retrieval, claim verification, and explanation generation on Mocheg using neural baselines and ablations. Results show that retrieval and explanation remain difficult, evidence quality matters, and machine verification trails human agreement.

  • Evidence Retrieval: Evidence retrieval is low for both text and images, requiring retrieval of roughly two text pieces per claim from 2,792,639 sentences.Semantic matching can miss evidence that contradicts a claim when that evidence has lower similarity than retrieved passages.
  • Claim Verification: Majority Label and Average Similarity perform close to random, while Pre-CoFactv2 and SpotFakePlus underperform the proposed approach.These comparisons indicate that Mocheg lacks simple label-distribution or claim-evidence semantic shortcuts.
  • Claim Verification: Gold and system evidence are compared in claim-verification ablations, alongside text-only, image-only, and no-evidence settings.Without evidence, models can still exceed the majority baseline on claims containing obvious clues or common-sense violations.
  • Claim Verification: Image evidence can add crucial information, such as confirming that a Boeing B-17E bomber was found in a jungle.The approach also returns top-5 images even for claims without associated image evidence, which introduces noise.
  • Claim Verification: Human annotators reached Fleiss κ scores of 0.67 with gold evidence, 0.59 with system evidence, and 0.42 without evidence, leaving a machine–human performance gap.The study sampled 50 claims and counted a human prediction as correct only when both annotators agreed on the true label.
  • Explanation Generation: Evidence retrieval affects explanation generation more than claim verification, while low LEAD-3 and ORACLE performance indicates that the explanations are highly abstractive.Concatenated evidence may contain necessary information but lacks coherent logical connections for human interpretation.

6 REMAINING CHALLENGES

The paper identifies persistent errors in multimodal claim verification and explanation generation, including cross-modal, cross-document, visual, arithmetic, commonsense, and coherence challenges.

  • Verification challenges: 30% of verification errors involve deep cross-modality reasoning and evidence fusion.Text and image evidence can provide complementary information, requiring models to connect both modalities.
  • Verification challenges: 30% of verification errors require reasoning across multiple textual evidence pieces or sentences.The model may need to combine separate statements to compare a claim against a historical benchmark.
  • Verification challenges: 6% of wrongly predicted claims require deep visual understanding of charts, tables, or maps.For example, verifying a drug-overdose claim requires extracting numerical information from an image.
  • Verification challenges: Mathematical calculation accounts for 4% of errors, commonsense reasoning for 8%, and partially supported or refuted claims for 12%.Some claims combine a correct component with an incorrect one, requiring more than a single-label evidence match.
  • Explanation-generation challenges: Evidence truncation can lose information because pre-trained models encode or decode only limited sequence lengths.Some evidence and ruling statements exceed the maximum length and are truncated.
  • Explanation-generation challenges: Missing evidence and weak logical organization remain challenges for explanation generation.Some ruling-statement information is absent from the collected sources, while evidence sentences may be difficult to organize coherently.

7 CONCLUSION

The paper introduces Mocheg as a benchmark for end-to-end multimodal fact-checking and explanation generation and finds that all three subtasks remain far from sufficient performance.

  • Dataset and findings: Mocheg contains 15,601 claims, 33,880 text evidence items, 12,112 image evidence items, and explainable statements.The dataset supports multimodal evidence retrieval, claim verification, and explanation generation.
  • Dataset and findings: Baseline results show that performance on multimodal evidence retrieval, claim verification, and explanation generation remains far from enough.The paper identifies advanced methods, deep visual understanding, and open-domain fact-checking as future directions.

8 ETHICAL STATEMENT

The ethical statement describes dataset licensing, restricted tweet sharing, privacy considerations, potentially offensive claims, and the use of fact-checking websites as sources.

  • Licensing and data sharing: The dataset is licensed under CC BY 4.0, while the associated crawler and baseline code use the Apache License 2.0.The dataset contains 2,916 tweets, for which only Twitter IDs and crawling scripts are shared under Twitter developer terms.
  • Privacy and content: The dataset does not use features or labels involving sensitive personally identifiable information.The authors note that internet claims may include offensive content.
  • Source selection: The authors use Politifact and Snopes because their journalists manually check claims and write ruling articles based on varied sources and expertise.The stated process includes contacting the statement originator and consulting government reports, academic studies, and other data.
Loading 2205.12487v2…