This paper introduces MULTIR, a novel approach for multi-instance learning with overlapping relations that combines a sentence-level extraction model with a simple, corpus-level component for aggregating individual facts. The model is designed to handle overlapping relations, which are common in real-world data, and enables accurate sentence-level predictions. It is computationally efficient, with inference reducing to weighted set cover, and performs well in both aggregate and sentence-level extraction tasks. The model is trained using weak supervision from a database, such as Freebase, and is tested on NY Times text. Experiments show that MULTIR outperforms previous approaches, including Riedel et al.'s method, in both aggregate and sentence-level extraction. The model is able to handle overlapping relations, which previous methods could not, and achieves higher precision and recall. It also demonstrates computational efficiency, with training and testing times significantly faster than previous methods. The paper also discusses related work, including weak supervision and multi-instance learning, and highlights the advantages of the proposed approach in handling overlapping relations and improving extraction accuracy.This paper introduces MULTIR, a novel approach for multi-instance learning with overlapping relations that combines a sentence-level extraction model with a simple, corpus-level component for aggregating individual facts. The model is designed to handle overlapping relations, which are common in real-world data, and enables accurate sentence-level predictions. It is computationally efficient, with inference reducing to weighted set cover, and performs well in both aggregate and sentence-level extraction tasks. The model is trained using weak supervision from a database, such as Freebase, and is tested on NY Times text. Experiments show that MULTIR outperforms previous approaches, including Riedel et al.'s method, in both aggregate and sentence-level extraction. The model is able to handle overlapping relations, which previous methods could not, and achieves higher precision and recall. It also demonstrates computational efficiency, with training and testing times significantly faster than previous methods. The paper also discusses related work, including weak supervision and multi-instance learning, and highlights the advantages of the proposed approach in handling overlapping relations and improving extraction accuracy.