Multi-Document Relationship Fusion via Constraints on Probabilistic Databases

Proceedings of NAACL HLT 2007, pages 332–339, Rochester, NY, April 2007. c⃝2007 Association for Computational Linguistics Multi-document Relationship Fusion via Constraints on Probabilistic Databases Gideon Mann Department of Computer Science University of Massachusetts Amherst, MA 01003 gideon.mann@gmail.com Abstract Previous multi-document relationship ex- traction and fusion research has focused on single relationships. Shifting the fo- cus to multiple relationships allows for the use of mutual constraints to aid extrac- tion. This paper presents a fusion method which uses a probabilistic database model to pick relationships which violate few constraints. This model allows improved performance on constructing corporate succession timelines from multiple doc- uments with respect to a multi-document fusion baseline. 1 Introduction Single document information extraction of named entities and relationships has received much atten- tion since the MUC evaluations1 in the mid-90s (Ap- pelt et al., 1993; Grishman and Sundheim, 1996). Recently, there has been increased interest in the extraction of named entities and relationships from multiple documents, since the redundancy of infor- mation across documents has been shown to be a powerful resource for obtaining high quality infor- mation even when the extractors have access to little or no training data (Etzioni et al., 2004; Hasegawa et al., 2004). Much of the recent work in multi- document relationship extraction has focused on the extraction of isolated relationships (Agichtein, 2005; Pasca et al., 2006), but often the goal, as in 1http://www.itl.nist.gov/iaui/894.02/related projects/muc/ single document tasks like MUC, is to extract a tem- plate or a relational database composed of related facts. With databases containing multiple relationships, the semantics of the database impose constraints on possible database configurations. This paper presents a statistical method which picks relation- ships which violate few constraints as measured by a probabilistic database model. The constraints are hard constraints, and robust estimates are achieved by accounting for the underlying extraction/fusion uncertainty. This method is applied to the problem of con- structing management succession timelines which have a rich set of semantic constraints. Using con- straints on probabilistic databases yields F-Measure improvements of 5 to 18 points on a per-relationship basis over a state-of-the-art multi-document extrac- tion/fusion baseline. The constraints proposed in this paper are used in a context of minimally super- vised information extractors and present an alterna- tive to costly manual annotation. 2 Semantic Constraints on Databases This paper considers management succession databases where each record has three fields: CEO’s name and the start and end years for thata person’s tenure as CEO (Table 1 Column 1). Each record is represented by three binary logical pred- icates: ceo(c,x), start(x,y1), end(x,y2), where c is a company, x is a CEO’s name, and y1 and y2 are years.2 2All of the relationships in this paper are defined to be bi- nary relationships. When extracting relationships of higher ar- 332 Explicit Implicit Logical Constraints Relationships Relationships (a partial list) ceo(c, x) precedes(x1,x2) ceo(c,x1), precedes(x1,x2) ⇔ ceo(c,x2), end(x1,y), start(x2, y) start(x, y) inoffice(x, y) start(x,y1), inoffice(x,y2), end(x, y3) ⇒ y1 ≤ y2 ≤ y3 end(x, y) predates(x1, x2) inoffice(x1, …

Content truncated to 500 words