Automatic Anaphora Resolution: Limits, Impediments, and Ways Forward

Abstract

The talk will discuss both the limits of and impediments to automatic anaphora resolution and will provide suggestions as to how to overcome some of these hurdles. To start with, anaphora resolution will be introduced as a daunting NLP task requiring the employment of different sources of knowledge. As recent research focusing on robust and knowledge-poor solutions will be summarised, as well as the latest trends, it will be argued that it is the absence of sophisticated semantic and realworld knowledge that imposes limits on all current algorithms. Next, the talk will explain that the lack of reliable pre-processing tools to feed correct inputs into resolution algorithms is another impediment which imposes a limit on the success rate of automatic anaphora resolution systems. It is argued that the performance of the best anaphora resolution algorithms can drop by over 20% if fully automatic pre-processing is undertaken. Comparative results and conclusions will be presented from the evaluation of the different versions of MARS, a fully automatic anaphora resolution system based on Mitkov's knowledge-poor approach to pronoun resolution, as well as from the evaluation of other (well-known) systems developed or re-implemented by members of the Research Group in Computational Linguistics at the University of Wolver-hampton. Outstanding evaluation issues will also be highlighted as a further impediment to measuring progress in anaphora resolution and a new 'evaluation workbench' will be proposed as a suitable environment for comprehensive evaluation of anaphora resolution algorithms. In addition, the issue of annotated corpora and how it is addressed at Wolverhampton will be discussed. Finally, ways forward in enhancing the performance of current algorithms will be suggested.