Combining ILP and MLN for Coreference Resolution

Abstract

Coreference resolution is a very important problem for many NLP applications. Most existing methods for coreference resolution make use of attribute-value features over pairs of noun phrases, which can't adequately describe the coreference conditions and properties between noun phrases. In this paper, we present a new approach to coreference resolution by combining Inductive Logic Programming (ILP) and Markov Logic Network (MLN), which excels such existing approaches as just considering inductive logic or probabilistic reasoning respectively. The ILP technique is used to capture the relationships among coreferential mentions based on first-order rules. With MLN's powerful representational ability, the previous findings are easily assimilated into MLN. Moreover, we can add specific rules about coreference resolution into MLN. After MLN's learning and inference, whether two mentions are coreferential is decided from the global view. Evaluations on the ACE data set show that our method is promising for the coreference resolution task.