Knowledge graphs¶
Knowledge graphs represent factual information as networks of entities (nodes) and relations/predicates (edges). Large-scale examples include Wikipedia (via DBpedia), Freebase, Wikidata, and YAGO. Knowledge graphs encode structured factual knowledge that can be queried, reasoned over, and analyzed. In the misinformation domain, knowledge graphs serve as ground-truth reference bases for fact checking and as data sources for training detection systems.
Knowledge graph properties relevant to misinformation work: - Completeness: Coverage varies; well-known entities have rich infoboxes while obscure entities may be sparse - Reliability: Crowdsourced graphs like Wikipedia exhibit high overall reliability despite vandalism and bias - Temporal dynamics: Facts evolve; entities are added/removed and relations change - Ambiguity: Polysemy and entity linkage challenges affect reasoning
Applications in misinformation research¶
Fact verification: Using paths and connectivity patterns to assess claim truthfulness. Example: checking if "X is a member of Y" by examining entity connectivity in the graph.
Knowledge base augmentation: Inferring missing facts from the graph structure to improve coverage and support detection systems.
Reasoning and inference: Applying logic-based and neural reasoning methods to combine multiple facts and verify complex claims.
Bias and knowledge measurement: Analyzing what facts are present/absent in graphs to understand coverage bias.
Key papers in this wiki¶
- Shiralkar et al. (2017) — Finding Streams in Knowledge Graphs to Support Fact Checking — Network-flow approach treating knowledge graphs as flow networks; uses relational similarity and node specificity to compute fact-checking truth scores; discovers interpretable patterns and relevant supporting facts
- Computational fact checking from knowledge networks (2015) — Demonstrates that graph topology and semantic proximity metrics enable effective fact checking
- FEVER: A large-scale dataset for fact extraction and verification (2018) — Evidence-based approach combined with retrieval from structured knowledge
- Mayank, Sharma & Sharma (2021) — DEAP-FAKED: Uses Wikidata knowledge graph embeddings via ComplEx to enhance fake news detection from article titles alone; maps named entities to KG and demonstrates KG embeddings outperform text-based entity encoding.
- Dun et al. (2021) — KAN: Knowledge-aware Attention Network for Fake News Detection — Incorporates entities and their knowledge graph contexts (neighbors) using entity linking to Wikidata, applies N-E and N-E2C attention mechanisms to weight entity importance, achieving 7.4% F1 improvement on PolitiFact.
Related topics¶
- Computational fact checking — Using knowledge graphs for verification
- Semantic similarity and relatedness — Measuring relatedness in knowledge graphs
- Graph Neural Networks — Learning on graph-structured data