Content Based Fake News Detection Using Knowledge Graphs

Jeff Z. Pan; Siyana Pavlova; Chenxi Li; Xiaoxi Liu; Yangmei Li; Jinshuo Liu

doi:10.1007/978-3-030-00671-6_39

In brief

First work to apply knowledge graph embedding for content-based fake news detection. Proposes TransE-based models (single and binary variants) trained on triples extracted from news articles, combined with open knowledge graphs like DBpedia. Hybrid approach achieves F1 scores above 0.80 on benchmark datasets.

Summary figure for Content Based Fake News Detection Using Knowledge Graphs

Contributions

First application of knowledge graph embedding and incomplete/imprecise knowledge graphs to content-based fake news detection
Single TransE model training on extracted triples with average and max bias classification methods
Binary TransE (B-TransE) model pairing negative and positive models to handle triples with large biases
Hybrid approach using early fusion of feature vectors from multiple models
Evaluation on Kaggle Fake News and BBC News datasets showing improvements over baseline

Method

The approach extracts entity-relation triples from news article content using OpenIE, then encodes them via knowledge graph embedding. Three types of knowledge graphs are constructed: one from article base (fake or true), one from open knowledge graph (DBpedia), and one from the union. TransE encodes triples as vectors and computes bias scores to classify whether a news item is true or false.

For single TransE models, two bias functions are used: average bias across all triples (Avg Bias) and maximum bias from any triple (Max Bias). The binary approach trains separate models on true and false articles, computing classification functions that compare biases across both models. A hybrid approach concatenates max bias vectors from multiple models and uses SVM with different kernels (polynomial, linear, RBF) for final classification.

Results

Single TransE with TML-D4 (DBpedia-based model with max bias) achieves 0.79 F1 score. Binary TransE with FML+TML-D4 achieves 0.83 F1. The hybrid approach with RBF kernel reaches 0.81 accuracy on combined feature vectors. TransE-based models consistently outperform the Knowledge Stream baseline on the same datasets.

Notes

The paper is novel in combining content-based detection (which focuses on article text) with knowledge graph embeddings. The main limitation is dependency on high-quality triple extraction; OpenIE extraction quality affects downstream classification. The work assumes relevant knowledge can be encoded as entity-relation triples, which may miss nuanced linguistic or pragmatic features of misinformation. Future work should combine with style-based approaches for complementary signals.