This paper addresses the challenge of relation extraction without explicit labeled text by proposing a novel approach that relaxes the distant supervision assumption. The authors argue that the assumption that every sentence mentioning two related entities expresses the given relation is often violated, especially when the knowledge base is external to the text being processed. They introduce a factor graph model that captures both the prediction of relations between entities and the prediction of which sentences express these relations. This model is trained using constraint-driven semi-supervision, where the assumption is relaxed to "at least one sentence" expressing the relation. The approach is evaluated on the New York Times corpus using Freebase as the external knowledge base, achieving a 31% error reduction compared to a state-of-the-art distant supervision method. The paper also discusses related work and provides a detailed experimental setup, including held-out and manual evaluation, to validate the effectiveness of the proposed method.This paper addresses the challenge of relation extraction without explicit labeled text by proposing a novel approach that relaxes the distant supervision assumption. The authors argue that the assumption that every sentence mentioning two related entities expresses the given relation is often violated, especially when the knowledge base is external to the text being processed. They introduce a factor graph model that captures both the prediction of relations between entities and the prediction of which sentences express these relations. This model is trained using constraint-driven semi-supervision, where the assumption is relaxed to "at least one sentence" expressing the relation. The approach is evaluated on the New York Times corpus using Freebase as the external knowledge base, achieving a 31% error reduction compared to a state-of-the-art distant supervision method. The paper also discusses related work and provides a detailed experimental setup, including held-out and manual evaluation, to validate the effectiveness of the proposed method.