Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

COOL: Comprehensive Knowledge Enhanced Prompt Learning for Domain Adaptive Few-shot Fake News Detection

Yi Ouyang, Peng Wu, Li Pan

TL;DR

COOL tackles domain-adaptive few-shot fake news detection by enriching prompt-based PLMs with timely external knowledge. It introduces a comprehensive knowledge extraction module that harvests both structured KG relations and unstructured descriptions, filtered via signed correlation-aware attention to capture positive and negative correlations with news. This knowledge is injected through a hybrid prompt template combining prefix soft prompts and postfix hard prompts, while adversarial perturbations and cross-domain contrastive learning encourage domain-invariant news-knowledge interactions. Experiments on Snopes as the source and Politifact/CoAID as targets show COOL outperforms state-of-the-art baselines, confirming the value of integrating comprehensive knowledge with adversarial-contrastive prompt learning in few-shot settings.

Abstract

Most Fake News Detection (FND) methods often struggle with data scarcity for emerging news domain. Recently, prompt learning based on Pre-trained Language Models (PLM) has emerged as a promising approach in domain adaptive few-shot learning, since it greatly reduces the need for labeled data by bridging the gap between pre-training and downstream task. Furthermore, external knowledge is also helpful in verifying emerging news, as emerging news often involves timely knowledge that may not be contained in the PLM's outdated prior knowledge. To this end, we propose COOL, a Comprehensive knOwledge enhanced prOmpt Learning method for domain adaptive few-shot FND. Specifically, we propose a comprehensive knowledge extraction module to extract both structured and unstructured knowledge that are positively or negatively correlated with news from external sources, and adopt an adversarial contrastive enhanced hybrid prompt learning strategy to model the domain-invariant news-knowledge interaction pattern for FND. Experimental results demonstrate the superiority of COOL over various state-of-the-arts.

COOL: Comprehensive Knowledge Enhanced Prompt Learning for Domain Adaptive Few-shot Fake News Detection

TL;DR

COOL tackles domain-adaptive few-shot fake news detection by enriching prompt-based PLMs with timely external knowledge. It introduces a comprehensive knowledge extraction module that harvests both structured KG relations and unstructured descriptions, filtered via signed correlation-aware attention to capture positive and negative correlations with news. This knowledge is injected through a hybrid prompt template combining prefix soft prompts and postfix hard prompts, while adversarial perturbations and cross-domain contrastive learning encourage domain-invariant news-knowledge interactions. Experiments on Snopes as the source and Politifact/CoAID as targets show COOL outperforms state-of-the-art baselines, confirming the value of integrating comprehensive knowledge with adversarial-contrastive prompt learning in few-shot settings.

Abstract

Most Fake News Detection (FND) methods often struggle with data scarcity for emerging news domain. Recently, prompt learning based on Pre-trained Language Models (PLM) has emerged as a promising approach in domain adaptive few-shot learning, since it greatly reduces the need for labeled data by bridging the gap between pre-training and downstream task. Furthermore, external knowledge is also helpful in verifying emerging news, as emerging news often involves timely knowledge that may not be contained in the PLM's outdated prior knowledge. To this end, we propose COOL, a Comprehensive knOwledge enhanced prOmpt Learning method for domain adaptive few-shot FND. Specifically, we propose a comprehensive knowledge extraction module to extract both structured and unstructured knowledge that are positively or negatively correlated with news from external sources, and adopt an adversarial contrastive enhanced hybrid prompt learning strategy to model the domain-invariant news-knowledge interaction pattern for FND. Experimental results demonstrate the superiority of COOL over various state-of-the-arts.
Paper Structure (22 sections, 15 equations, 5 figures, 2 tables)

This paper contains 22 sections, 15 equations, 5 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related Works
  3. Problem Statement
  4. Methodologies
  5. Comprehensive Knowledge Extraction
  6. Heterogeneous Knowledge Retrieval
  7. Signed Correlation-aware Attention
  8. Hybrid Prompt Learning
  9. Hybrid Prompt Template
  10. Veracity Prediction
  11. Adversarial Contrastive Training
  12. Experiments
  13. Experiment Setup
  14. Main Results
  15. Abaltion Study
  16. ...and 7 more sections

Figures (5)

  • Figure 1: A piece of real-world fake news about the prevention of Covid-19.
  • Figure 2: Architecture of the proposed COOL model.
  • Figure 3: Ablation studies in comprehensive knowledge extraction module on CoAID, where “w/o KG” and “w/o KC” means removing structured knowledge retriever and unstructured knowledge retriever, respectively. “w/o CK” means removing the entire comprehensive knowledge extraction module. "KG-mean" means replacing the modulation mechanism with mean pooling in structured knowledge retriever. “w/o Pos” and “w/o Neg” means removing N-$\mathrm{E^{+}}$ and N-$\mathrm{E^{-}}$ attention, respectively.
  • Figure 4: Ablation studies in hybrid prompt learning module on Politifact, where “w/o Prefix” and “w/o Postfix” means removing the prefix soft prompt and the postfix hard prompt, respectively. “w/o AD” and “w/o CL” means removing adversarial augmentation and contrastive training, respectively. “w/o ACT” means eliminates the adversarial contrastive training strategy.
  • Figure 5: A real-world case from CoAID showing how COOL extracts comprehensive knowledge from external sources.