This paper compares the effectiveness of observed feature models and latent feature models in knowledge base completion tasks. The authors find that a simple observed features model outperforms state-of-the-art latent feature models on two benchmark datasets, FB15K and WN18. They also construct a more challenging dataset derived from FB15K and evaluate the impact of textual mentions from a web-scale corpus. The results show that the observed features model is most effective at capturing information for entity pairs with textual relations, while a combination of both model types performs best. The paper highlights the importance of direct relations and textual links in knowledge base completion and suggests that combining observed and latent features can significantly improve performance.This paper compares the effectiveness of observed feature models and latent feature models in knowledge base completion tasks. The authors find that a simple observed features model outperforms state-of-the-art latent feature models on two benchmark datasets, FB15K and WN18. They also construct a more challenging dataset derived from FB15K and evaluate the impact of textual mentions from a web-scale corpus. The results show that the observed features model is most effective at capturing information for entity pairs with textual relations, while a combination of both model types performs best. The paper highlights the importance of direct relations and textual links in knowledge base completion and suggests that combining observed and latent features can significantly improve performance.