This paper compares the effectiveness of observed feature models and latent feature models on two knowledge base completion datasets, FB15K and WN18. The observed feature model outperforms the latent feature models on these datasets, possibly due to the unrealistic redundancy in the KB graphs. However, when a more challenging dataset derived from FB15K is used, the latent feature models outperform the observed feature models. When textual mentions are added to the dataset, the observed feature model becomes more powerful than the latent feature models, but a combination of both models performs better than either alone. The paper introduces a simple observed feature model that uses direct links between candidate entity pairs. It also presents two latent feature models: model E, which captures the compatibility between entities and relation positions, and DISTMULT, a bilinear model. The observed feature model uses binary features based on direct links and entity relation positions, while the latent feature models use continuous representations of entities and relations. The paper also explores the impact of textual mentions on knowledge base completion. It shows that textual mentions can significantly improve performance, especially for test cases with textual occurrences. The observed feature model benefits from textual mentions, and combining it with a latent feature model leads to better performance. The paper concludes that the presence of relations between candidate pairs can be a strong signal in some cases, and that textual links extracted from a large document collection can significantly improve performance. Combining observed and latent feature models is effective for capturing inferences among KB relations and direct cues from text. The paper also highlights the importance of balancing the weight of the loss incurred from textual versus KB relations in a dataset where training and test triples are not limited to those with textual mentions.This paper compares the effectiveness of observed feature models and latent feature models on two knowledge base completion datasets, FB15K and WN18. The observed feature model outperforms the latent feature models on these datasets, possibly due to the unrealistic redundancy in the KB graphs. However, when a more challenging dataset derived from FB15K is used, the latent feature models outperform the observed feature models. When textual mentions are added to the dataset, the observed feature model becomes more powerful than the latent feature models, but a combination of both models performs better than either alone. The paper introduces a simple observed feature model that uses direct links between candidate entity pairs. It also presents two latent feature models: model E, which captures the compatibility between entities and relation positions, and DISTMULT, a bilinear model. The observed feature model uses binary features based on direct links and entity relation positions, while the latent feature models use continuous representations of entities and relations. The paper also explores the impact of textual mentions on knowledge base completion. It shows that textual mentions can significantly improve performance, especially for test cases with textual occurrences. The observed feature model benefits from textual mentions, and combining it with a latent feature model leads to better performance. The paper concludes that the presence of relations between candidate pairs can be a strong signal in some cases, and that textual links extracted from a large document collection can significantly improve performance. Combining observed and latent feature models is effective for capturing inferences among KB relations and direct cues from text. The paper also highlights the importance of balancing the weight of the loss incurred from textual versus KB relations in a dataset where training and test triples are not limited to those with textual mentions.