Jan-Christoph Kalo, Benno Kruit, Stefan Schlobach.

Automatic knowledge graph construction, using supervised relation extraction from text, has become the state-of-the-art to create large-scale repositories of background knowl- edge for various applications. Recent advances in machine learning and Natural Language Processing (NLP), in particular the advent of the large language models, have improved the performance of relation extraction systems significantly. Traditional leaderboard style benchmark settings show very high performance, suggesting that these models can be em- ployed in practical applications. Our analysis shows that in reality, though, the extraction quality varies drastically from one relation to another, with unacceptable performance for certain types of relations. To better understand this behaviour, we perform a seman- tic error analysis on a popular distantly supervised benchmark dataset, using ontological meta-relations to describe various error categories, which shows that relations that are confused by state-of-the-art systems are often semantically closely related, e.g., they are inverses of each other, in subproperty relations, or share the same domain and range. Such an extensive semantic error analysis allows us to understand the strengths and weaknesses of extraction models in a semantic way and to provide some practical recommendations to improve the quality of relation extraction in the future.

Citation

@inproceedings{
kalo2022understanding,
title={Understanding Distantly Supervised Relation Extraction through Semantic Error Analysis},
author={Jan-Christoph Kalo and Benno Kruit and Stefan Schlobach},
booktitle={4th Conference on Automated Knowledge Base Construction},
year={2022}
}