Integer Linear Programming For Coreference Resolution

In this chapter, we introduce Integer Linear Programming (ILP) and review some of its best performing applications to coreference resolution in the literature. We develop some intuitions for how to pose ILPs based on learned-models and to how expert knowledge can be encoded as constraints that the learned models must then respect. We describe some of the difficulties encountered during both the development of an ILP and its deployment as well as how to deal with them. Finally, we see how ILP can create an environment in which independently learned models share knowledge for their mutual benefit.Most of the top results on coreference resolution over the last few years were achieved using an ILP formulation, and we provide a snapshot of these results. Conceptually, and from an engineering perspective, the ILP formulation is very simple and it provides system designers with a lot of flexibility in incorporating knowledge. Indeed, this is where we believe future research should focus.