Graph-Linguistic Fusion: Using Language Models for Wikidata Vandalism Detection

TL;DR

The paper tackles vandalism detection in Wikidata by unifying structured and textual edits into a text-based representation (Graph2Text) and applying a single multilingual transformer to assess edits. It follows with a production-friendly pipeline that combines an LM-based content analyzer and a CatBoost final classifier, achieving state-of-the-art performance (AUC of $0.924$) and better fairness than the ORES baseline while supporting CPU-only deployment. A large open dataset (~$4.84$ million revisions) and open-source code accompany the approach, enabling reproducibility and further research. The work demonstrates practical impact by reducing patrollers' workload and providing a scalable, multilingual vandalism detector suitable for real-world Wikidata maintenance and growth.

Abstract

We introduce a next-generation vandalism detection system for Wikidata, one of the largest open-source structured knowledge bases on the Web. Wikidata is highly complex: its items incorporate an ever-expanding universe of factual triples and multilingual texts. While edits can alter both structured and textual content, our approach converts all edits into a single space using a method we call Graph2Text. This allows for evaluating all content changes for potential vandalism using a single multilingual language model. This unified approach improves coverage and simplifies maintenance. Experiments demonstrate that our solution outperforms the current production system. Additionally, we are releasing the code under an open license along with a large dataset of various human-generated knowledge alterations, enabling further research.

Figures (9)

• Figure 1: Diagram with the most important parts of the Wikidata record.
• Figure 2: Example of a revision (ID: 593195479) vandalizing the Wikidata entry for Bulgaria. Original triple IDs are mapped to their corresponding English labels.
• Figure 3: Wikidata vandalism detection system schema.
• Figure 4: Text processing schema.
• Figure 5: Data splitting logic.