Heuristic algorithm for zero subject detection in Polish

Abstract

This article describes a heuristic approach to zero subject detection in Polish. It focuses on the zero subject detection as a crucial step in end-to-end coreference resolution. The zero subject verbs are recognized using a set of manually created rules utilizing information from different sources, including: a dependency parser, a shallow relational parser and a valence dictionary. The rules were developed and evaluated on the Polish Coreference Corpus. The experimental results show that the presented method significantly outperforms the only machine learning-based alternative for Polish, i.e., MentionDetector. We also discuss and evaluate the importance of zero subject detection for existing coreference resolution tools for Polish.