PaperlayerView exact revision on arXiv

Source-linked AI summary

Generalizing to the Future: Mitigating Entity Bias in Fake News Detection

Yongchun Zhu, Qiang Sheng, Juan Cao, Shuokai Li, Danding Wang, Fuzhen Zhuang

arXiv:2204.09484v1cs.CLcs.AIcs.LG

TL;DR

Fake-news detectors must generalize from past news to future news, but temporal changes in entity–veracity correlations pose a challenge. ENDEF separately models entity and content contributions during training and removes the direct entity effect at inference; the authors report effectiveness in offline and online experiments.

  • Problem

    Existing fake-news detection methods often assume IID training and test data, although future-news detection involves temporal shifts and changing entity–veracity correlations.

    Source: §1 · p. 2

  • Method

    ENDEF models entity and overall-content contributions separately during training, then removes the direct entity effect at inference.

    Source: §1 · p. 2, §3.2 · p. 3, §5 · p. 5

  • Results

    Offline and online experiments demonstrated ENDEF’s effectiveness, and five tested fake-news detectors performed better on future data with the framework.

    Source: §1 · p. 2, §5 · p. 5, §4.2.2 · p. 5

  • Takeaways & Limitations

    The authors present ENDEF as a practical approach to improving fake-news detectors’ generalization to future data.

    Source: §5 · p. 5

    The proposed framework is currently text-only; extending it to multimodal and social-graph-based detection is left for future work.

    Source: §5 · p. 5

Abstract

from arXiv · show

The wide dissemination of fake news is increasingly threatening both individuals and society. Fake news detection aims to train a model on the past news and detect fake news of the future. Though great efforts have been made, existing fake news detection methods overlooked the unintended entity bias in the real-world data, which seriously influences models' generalization ability to future data. For example, 97\% of news pieces in 2010-2017 containing the entity `Donald Trump' are real in our data, but the percentage falls down to merely 33\% in 2018. This would lead the model trained on the former set to hardly generalize to the latter, as it tends to predict news pieces about `Donald Trump' as real for lower training loss. In this paper, we propose an entity debiasing framework (\textbf{ENDEF}) which generalizes fake news detection models to the future data by mitigating entity bias from a cause-effect perspective. Based on the causal graph among entities, news contents, and news veracity, we separately model the contribution of each cause (entities and contents) during training. In the inference stage, we remove the direct effect of the entities to mitigate entity bias. Extensive offline experiments on the English and Chinese datasets demonstrate that the proposed framework can largely improve the performance of base fake news detectors, and online tests verify its superiority in practice. To the best of our knowledge, this is the first work to explicitly improve the generalization ability of fake news detection models to the future data. The code has been released at https://github.com/ICTMCG/ENDEF-SIGIR2022.

1 INTRODUCTION

Fake-news detectors must generalize from past news to future news, but temporal shifts and entity–veracity correlations challenge that goal. ENDEF addresses entity bias by modeling entity and content contributions separately and removing the direct entity effect at inference.

1 INTRODUCTION

  • A temporal split caused existing methods to lose about 10% performance compared with an ideal IID split.Most existing methods assume training and test news are sampled IID from the same period, unlike the future-data setting.Source: §1 · p. 2
  • The share of Donald Trump-related news labeled real fell from 97% in 2010–2017 to 33% in 2018, illustrating changing entity–veracity correlations.Models trained on earlier data may over-rely on an entity’s label association and generalize poorly when that association changes.Source: §1 · p. 2
  • ENDEF models entity contributions separately from overall news content, then removes entities’ direct effect for debiased prediction.The framework aims to retain more generalizable non-entity signals, such as writing style and emotions.Source: §1 · p. 2
  • The authors report improved future-data performance for five detectors and describe offline and online experiments demonstrating the framework’s effectiveness.They present the framework as deployable alongside different fake-news detection models.Source: §1 · p. 2

2 RELATED WORK

Prior fake-news detection work uses content features or social context, while model-debiasing research includes data-level, model-level, and causal approaches. ENDEF focuses on entity bias in textual detection and temporal generalization.

2 RELATED WORK

  • Fake-news detection methods are broadly grouped into content-based approaches using content or factual resources and social-context approaches using user engagement.The paper places its work among textual content-based methods.Source: §2 · p. 2
  • Event-invariant features and emotional signals target generalization, but the paper says they do not explicitly address bias from a temporal perspective.The experiments are reported to show additional improvements when ENDEF is added to these methods.Source: §2 · p. 2
  • Model-debiasing approaches include data-level manipulations [38], model-level balancing [10] [11], and causal-learning methods [36] [39].ENDEF applies causal learning to entity bias in fake-news detection.Source: §2 · p. 2

3 METHODOLOGY

ENDEF extends a base detector by modeling entity and full-content contributions during training, then using the content-path prediction without the direct entity effect at inference. The method also uses token-level augmentation during training.

3.2 Model

  • ENDEF adds an entity-based model to the base detector, modeling the direct entity contribution alongside the effect of overall news content.The entity model takes only E as input to represent the direct E→Y effect, while the full-text detector represents the E→P→Y path; an auxiliary loss trains the entity model to capture entity bias and encourages less-biased information in the detector.Source: §3 · p. 2, §3.2 · p. 3

3.1 Problem Formulation

  • A news piece P contains tokens, including entities E, and has a ground-truth label y; the detector predicts whether the piece is fake or real.The entity set may contain people or locations.Source: §3.1 · pp. 2–3

3.3 Inference

  • At inference, ENDEF uses the content-path prediction σ(r_P), removing the direct entity effect from the aggregated prediction.The detector can use diverse base models.Source: §3.3 · p. 3

3.4 Data Augmentation

  • Training randomly applies either token-level drop or mask augmentation to words or entities, with probability p.One of the two augmentation policies is randomly selected for each sample.Source: §3.4 · p. 3

4 EXPERIMENTS

The experiments evaluate ENDEF on future-data offline benchmarks, an online detection system, and analyses of its predictions. Across these settings, the reported results favor the framework, while the analysis links entity-based signals to debiased predictions.

4.1 Experimental Settings.

  • Evaluation uses temporal splits on Weibo and GossipCop, with the most recent 40% of news randomly included in test and validation sets and five content-based detectors; metrics include AUC, accuracy, macro-F1, class-specific F1, and spAUCFPR≤10%.Weibo training spans 2010–2017, with test and validation samples posted in 2018; GossipCop training spans 2000–2017.Source: §4.1 · pp. 3–4

4.2.3 Analysis (RQ3).

  • 0.58 versus 0.49 average fake-news prediction scores for BERT-Emo w/ ours and BERT-Emo, respectively, accompany higher-confidence fake predictions in two cases.The analysis says the entity-based model captures biased information, enabling the framework to make debiased predictions.Source: Figure 3 · p. 5, Figure 4 · p. 5, §4.2.3 · p. 5

4.2.1 Offline Experiments (RQ1).

  • Most base models improve on most metrics with ENDEF on future data; Table 3 compares each base model with and without the framework.The improvement is reported as larger on Weibo than on GossipCop; the authors attribute the difference to the shorter average Weibo articles (120 versus 606).Source: Table 3 · p. 4, §4.2.1 · p. 5

4.2.2 Online Experiments (RQ2).

  • ENDEF improves base models in the highly skewed online setting, where real-to-fake items occur at roughly 40:1; Table 4 reports relative improvements over baselines.Absolute online performance is not reported because of business-rule restrictions, and BERT-Emo w/ ours was deployed in the system.Source: Table 4 · p. 4, §4.2.2 · p. 5

5 CONCLUSION AND FUTURE WORK

ENDEF targets future-data generalization by mitigating entity bias through a causal framework that models entity effects during training and removes them at inference. The authors report offline and online effectiveness, while identifying several directions for extending the work.

5 CONCLUSION AND FUTURE WORK

  • ENDEF models entities’ direct contribution to news veracity during training and removes that effect at inference to mitigate entity bias.The framework uses a causal graph of entities, news contents, and veracity to improve base detectors’ generalization to future data.Source: §5 · p. 5
  • Offline and online experiments demonstrated the effectiveness of ENDEF.Source: §5 · p. 5
  • Future work includes adapting unbiased models to future news, studying features shared across periods, and extending text-only detection to multimodal and social-graph settings.Source: §5 · p. 5

Related papers

Loading 2204.09484v1…