Resolving Pronouns Robustly: Plumbing the Depths of Shallowness

In this paper, we explore various shallow techniques for salience-based pronoun resolution. We use simple noun chunking at the syntactic analysis stage and extract grammatical function information by pattern matching. Identifying subjects and objects is critical to salience calculations. We report that this important subject-object distinction can be made reliably with our shallow approach. We then explore a range of shallow semantic inference procedures (making use of only the text under consideration and lexical knowledge bases like WordNet) that significantly improve third-person pronoun resolution across a range of genre. We show how the same framework can also resolve relative pronouns with high accuracy. This is a result that might be relevant to other NLP applications like parsing.