Inductive Entity Representations from Text via Link Prediction

This research paper addresses the challenge of incomplete Knowledge Graphs (KGs) by proposing a method to learn vector representations of entities using textual descriptions. The authors introduce a holistic evaluation protocol for entity representations learned through a link prediction objective, focusing on inductive settings where entities were not seen during training. The study evaluates an architecture based on pretrained language models, demonstrating strong generalization capabilities. Results indicate that this approach outperforms state-of-the-art methods, achieving a 22% average improvement in Mean Reciprocal Rank (MRR) for link prediction. Furthermore, the learned representations transfer effectively to other tasks without fine-tuning, showing a 16% accuracy improvement in entity classification and up to an 8.8% increase in NDCG@10 for information retrieval tasks. The findings suggest that these representations possess greater generalization properties than previously evaluated, making them valuable for various web applications such as recommender systems and metadata annotation without the need for retraining.