PaperlayerView exact revision on arXiv

Source-linked AI summary

Explainable Fake News Detection With Large Language Model via Defense Among Competing Wisdom

Bo Wang, Jing Ma, Hongzhan Lin, Zhiwei Yang, Ruichao Yang, Yuan Tian, Yi Chang

arXiv:2405.03371v3cs.CL

TL;DR

Fake-news detectors often provide no justification, while debunked-report explanations can be delayed and crowd-majority explanations can inherit bias. L-Defense extracts evidence from competing parties, prompts an LLM to generate justifications for possible veracities, and models a defense between them. The authors report state-of-the-art detection results on two real-world benchmarks and high-quality justifications.

  • Problem

    Black-box detection, delayed debunking, and potentially biased crowd-majority explanations leave a need for timely, clear justifications alongside veracity predictions.

    Source: §1 · pp. 1–2

  • Method

    L-Defense extracts salient evidence from two competing parties, prompts an LLM to generate justifications for possible veracities, and models a defense among those justifications to determine veracity.

    Source: §1 · p. 2

  • Results

    The authors report state-of-the-art results on two real-world explainable fake-news datasets and high-quality explanations.

    Source: §1 · p. 2, §5 · p. 8

  • Takeaways & Limitations

    Comparing the informativeness and soundness of competing parties offers a basis for judging veracity without relying only on the quantity of crowd opinions.

    Source: §1 · p. 2

Abstract

from arXiv · show

Most fake news detection methods learn latent feature representations based on neural networks, which makes them black boxes to classify a piece of news without giving any justification. Existing explainable systems generate veracity justifications from investigative journalism, which suffer from debunking delayed and low efficiency. Recent studies simply assume that the justification is equivalent to the majority opinions expressed in the wisdom of crowds. However, the opinions typically contain some inaccurate or biased information since the wisdom of crowds is uncensored. To detect fake news from a sea of diverse, crowded and even competing narratives, in this paper, we propose a novel defense-based explainable fake news detection framework. Specifically, we first propose an evidence extraction module to split the wisdom of crowds into two competing parties and respectively detect salient evidences. To gain concise insights from evidences, we then design a prompt-based module that utilizes a large language model to generate justifications by inferring reasons towards two possible veracities. Finally, we propose a defense-based inference module to determine veracity via modeling the defense among these justifications. Extensive experiments conducted on two real-world benchmarks demonstrate that our proposed method outperforms state-of-the-art baselines in terms of fake news detection and provides high-quality justifications.

1 INTRODUCTION

The introduction argues that black-box detectors, delayed investigative journalism, and uncensored crowd opinions leave a need for timely, credible explanations. L-Defense compares evidence from competing parties and uses LLM-generated justifications to determine veracity.

1 INTRODUCTION

  • Unverified crowd reports can contain biased or inaccurate opinions, so treating majority opinion as justification risks misleading veracity predictions.Earlier explanation approaches also rely on labor-intensive debunked reports, while attention-based methods identify influential text without providing intuitive, credible explanations.Source: §1 · pp. 1–2
  • The proposed comparison treats evidence quality—not crowd size—as a signal: informativeness and soundness differences between parties indicate veracity.In the Figure 1 example, one refuting report gives detailed evidence while supporting reports offer brief or echoing claims.Source: §1 · p. 2, Figure 1 · p. 2
  • L-Defense extracts salient evidence for two competing parties, prompts an LLM to summarize reasons for each possible veracity, then models a defense between those justifications.The final explanation is adaptively selected from the generated justifications.Source: §1 · p. 2
  • The authors report state-of-the-art results on two real-world explainable fake-news datasets and high-quality explanations.They state that the model derives explanations comparable to human experts without debunked-report supervision.Source: §1 · p. 2

2 RELATED WORK

Prior explainable fake-news detection work ranges from highlighting salient content to generating explanations from debunked reports or crowd opinions. LLM-based fake-news work explores detection and reasoning, but the cited approaches lack easily comprehensible explanations.

2.1 Explainable Fake News Detection

  • Earlier EFND methods highlight salient phrases, attributes, users, or sentences, while other approaches summarize or learn from debunked reports; crowd-majority methods risk bias from unverified reports.The authors position splitting crowd wisdom into competing parties and comparing it through a defense-like strategy as their response.Source: §2.1 · p. 2

2.2 Large Language Model in Fake News Detection

  • Prior LLM fake-news approaches use prompting, rationale distillation, or LoRA tuning, but the authors say they lack easily comprehensible explanations.The authors also note that externally retrieved knowledge does not eliminate the majority-bias problem, and contrast their use of LLMs to generate competing justifications.Source: §2.2 · p. 3

3 PROPOSED APPROACH

L-Defense combines competing-evidence extraction, LLM-based reasoning, and defense-based inference to predict a claim’s veracity and provide an explanation. The approach develops these components through evidence ranking, veracity-oriented rationale generation, and a final prediction stage.

3 PROPOSED APPROACH

  • The framework has three components: competing evidence extraction, prompt-based reasoning, and defense-based inference.The paper also discusses the framework’s generalization.Source: §3 · p. 3

3.1 Competing Evidence Extraction

  • The extraction module splits reports into sentences and ranks evidence separately for false and true veracity, using claim–evidence scores and veracity-specific attention weights.It trains with a soft objective based on temporary veracity labels and a three-category classification objective, combining the two losses; at inference, it selects top-k false and true evidence.Source: §3.1 · p. 3, §3.1.1 · p. 3, §3.1.2 · p. 4

3.1.1 Claim and Evidence Representation.

  • Claim and candidate-evidence representations are produced with a pretrained Transformer encoder, with pooling used to obtain contextualized sequence representations.Reports are divided into individual candidate sentences because a report can contain evidence for different veracities.Source: §3.1.1 · p. 3

3.1.2 Competing Veracity Scoring.

  • Claim–candidate pairs receive complementary false and true scores, while veracity-specific attention weights contribute to evidence ranking.The resulting false and true evidence scores are used to rank candidates.Source: §3.1.2 · p. 4

3.1.3 Training and Inference.

  • A soft KL objective supervises competing scores alongside temporary three-way classification; inference ranks candidates separately to select top-k false and true evidence.The extraction module combines the two training losses linearly.Source: §3.1.3 · p. 4

3.2 Prompt-based Reasoning with LLM

  • An LLM uses extracted evidence and a prior veracity label to generate streamlined rationales for both false and true possibilities.The two resulting explanations are described as having relative strengths in confidence that can facilitate detection.Source: §3.2 · p. 4

3.3 Defense-based Inference

  • The claim and both explanations are encoded together for veracity classification; the predicted label selects the final explanation, while a “half” prediction presents both.The concatenated texts are processed by a pretrained Transformer encoder, whose representation supports the classifier.Source: §3.3 · p. 5

3.4 Discussion about Generalization

  • The authors state that the three-stage architecture generalizes whether or not competing evidence is present.Source: §3.4 · p. 5

4 EXPERIMENT

The experiments assess veracity prediction and explanation quality on RAWFC and LIAR-RAW, then examine component contributions and case-study behavior.

4.1 Evaluations on Veracity Prediction

  • On RAWFC, both L-Defense variants improve at least 4% across all metrics over FactLLaMAknow; on LIAR-RAW, both outperform it in macF1 but trail in precision.The paper reports state-of-the-art or competitive performance on the two datasets.Source: §4.1 · pp. 5–6, Table 2 · p. 6

4.2 Evaluations on Explanation

  • L-Defense achieves superior or comparable explanation quality to Oracle, while ChatGPT evaluation trends for misleadingness and discrepancy partially align.Oracle uses the actual veracity label and is treated as a ceiling for explanation quality.Source: §4.2 · pp. 6–7, Table 3 · p. 7
  • CofCED explanations are incoherent and highly misleading; ChatGPTfull's relative performance shifts by dataset, and L-Defense variants trade misleadingness against other quality metrics.ChatGPTfull performs better than ChatGPTclaim on RAWFC's last three metrics but worse across all metrics on LIAR-RAW.Source: §4.2 · p. 7
  • Human evaluation rankings largely match ChatGPT's, though humans favor the ChatGPT L-Defense variant while ChatGPT favors LLaMA2; explanations also reduce annotator errors.Ten volunteers assessed explanations, and the confusion matrices compare judgments with claims alone versus claims plus L-Defense explanations.Source: §4.2 · p. 7, Figure 3 · p. 7, Table 4 · p. 7

4.3 Ablation Study

  • Removing extracted evidence, prior labels, generated explanations, or inference training degrades veracity performance, supporting the contribution of each component.Random evidence remains commendable but performs below the original model; removing inference training exposes a gap between LLM and fine-tuned SLM predictions.Source: §4.3 · pp. 7–8, Table 5 · p. 7

4.4 Case Study

  • In the RAWFC cases, competing evidence aligns with gold justifications, and L-Defense predicts correctly where CofCED misses the true aspect and misclassifies.The case study also reports that competing LLM-generated justifications support a correct defense-based decision.Source: §4.4 · p. 8, Table 6 · p. 8

5 CONCLUSION

The framework uses competing crowd wisdom to produce concise justifications and reports state-of-the-art explainable fake-news detection on two real-world benchmarks. Its majority-bias analysis finds lower average bias ratios for L-Defense than for CofCED.

5 CONCLUSION

  • The framework achieves state-of-the-art explainable fake-news detection on two real-world benchmarks.Source: §5 · p. 8
  • Evidence extraction splits competing wisdom, an LLM generates justifications for two possible veracities, and inference models the defense between them.Source: §5 · p. 8

A THE MEASURING OF MAJORITY BIAS

  • Majority bias is measured as the ratio of candidate sentences semantically similar to extracted evidence; higher ratios indicate greater bias.Source: A THE MEASURING OF MAJORITY BIAS · p. 10
  • 0.32 for CofCED and 0.22 for L-Defense were the respective average majority-bias ratios on RAWFC; human Informativeness evaluations also found CofCED evidence less informative.Source: A THE MEASURING OF MAJORITY BIAS · p. 10

B DETAILED DISCUSSION ABOUT GENERALIZATION

L-Defense extracts true- and false-oriented evidence, prompts an LLM with each party's evidence and a prior veracity label, then predicts by comparing the resulting explanations. The described cases include claims with or without competing evidence.

B DETAILED DISCUSSION ABOUT GENERALIZATION

C.1 Training Setup

Training uses SGD with Adam, warm-up, and learning-rate decay, while prompt design iteratively refines instructions for streamlined rationales. The prompt asks the LLM to reason from a veracity label and related sentences, which may contain noise.

C.1 Training Setup

  • Training uses mini-batch SGD with Adam, 10% warm-up, and linear learning-rate decay; LIAR-RAW labels are grouped into false, half-true, and true categories.Source: §C.1 · p. 10
  • Reasoning generation uses temperature 0.8, explanation evaluation uses temperature 0, and prediction results are the best values across ten runs.Source: §C.1 · p. 10

C.2 Prompt Designing

  • The prompt was iteratively refined using randomly selected training samples to elicit rationales that let readers assess veracity without additional background knowledge.Source: §C.2 · p. 11
  • The system prompt requests abductive reasoning from a veracity label and related sentences, allows for noisy evidence, and sets an explanation limit below 200 words.Source: §C.2 · p. 11

D.1 Validation of Assumption

The class-wise evaluations support the assumption that the truth-indicating party is more informative and sounder than its competitor. For explanation length comparisons, CofCED uses its top six sentences, while other LLM-based explanations are limited to 200 words.

D.1 Validation of Assumption

  • For false claims, false justifications evaluate better than true ones; for true claims, true-oriented justifications evaluate better.Source: §D.1 · p. 11
  • For half-true claims, true and false justifications yield competitive evaluation results.Source: §D.1 · p. 11

D.2 The Length of Explanations

  • CofCED explanations use the top six ranked sentences, while other LLM-based explanations are capped at 200 words for the length comparison.Source: §D.2 · p. 11
  • The explanation for L-Defense's half-true prediction is slightly longer because it combines two competing justifications.Source: §D.2 · p. 11

D.3 Ablation Study of Extraction Objective

The section examines each extraction-module objective’s contribution to veracity prediction on RAWFC. The accompanying captions identify hyperparameter and explanation-evaluation context.

D.3 Ablation Study of Extraction Objective

  • Table 10 reports an ablation study of the extraction-module objectives for veracity prediction on RAWFC.The cited passages state that the study assesses each objective’s contribution but provide no outcome values.Source: §D.3 · p. 11, Table 10 · p. 11
  • Table 7 presents the hyper-parameters.Source: Table 7 · p. 11
  • The accompanying evaluation captions describe ChatGPT Likert ratings of competing explanations and average explanation-token counts on RAWFC and LIAR-RAW.Source: Table 8 · p. 11, Table 9 · p. 11

D.4 More Case Studies

The case studies compare the informativeness and soundness of truth-side and competing explanations, while the authors identify evidence extraction as a source of errors. Future work targets extraction, prompting, and LLM selection.

D.4 More Case Studies

  • The first two Table 11 cases show higher informativeness and soundness for the truth side than the competing side, consistent with the authors’ assumption.Source: §D.4 · p. 11, Table 11 · p. 12
  • Most L-Defense errors on RAWFC’s test set are attributed to failures to extract accurate evidence sentences.Source: §D.4 · p. 11

E FUTURE WORK

  • Future work targets more accurate salient-evidence extraction, stronger prompting for fake-news reasoning, and adaptive LLM selection by claim.Source: E FUTURE WORK · p. 11
Loading 2405.03371v3…