Ng, Vincent2007-04-042007-04-042003-12-23http://techreports.library.cornell.edu:8081/Dienst/UI/1.0/Display/cul.cis/TR2003-1918https://hdl.handle.net/1813/5630State-of-the-art coreference resolution systems are mostly knowledge-based systems that operate by relying on a set of hand-crafted coreference resolution heuristics. Recently, however, machine learning approaches have been shown to be a promising way to build coreference resolution systems that are more robust than their knowledge-based counterparts. Nevertheless, there are several key issues in existing machine learning approaches to the problem that are either not explored or being overlooked, potentially leading to a deterioration of system performance. This document examines each of these issues in detail and suggests potential solutions.614245 bytesapplication/postscripten-UScomputer sciencetechnical reporttechnical report