Many interesting research problems, such as ontology alignment and collective classification, require probabilistic and collective inference over imprecise evidence. Existing approaches are typically ad-hoc and problem-specific, requiring significant effort to devise and implement and provide poor generalizability. In this paper, we introduce probabilistic similarity logic (PSL), a simple, yet powerful language for describing problems which require probabilistic reasoning about similarity where, in addition to reasoning probabilistically, we want to capture both logical constraints and imprecision. We prove that PSL inference is polynomial and describe an efficient implementation which uses an off-the-shelf convex optimization package. Lastly, we outline a wide range of application areas for PSL.

@InProceedings{broecheler:srl09,
  author       = "Broecheler, Matthias and Getoor, Lise",
  title        = "Probabilistic Similarity Logic",
  booktitle    = "International Workshop on Statistical Relational Learning (SRL'09)",
  year         = "2009",
}