Comparing methods for the syntactic simplification of sentences in information extraction

This article describes research aimed at improving the accuracy of an information extraction (IE) system by treating coordinate structures systematically. Commas, coordinating conjunctions, and adjacent comma–conjunction pairs are considered to be potential indicators of coordination in natural language. A recursive algorithm is implemented which converts sentences containing classified potential coordinators into sequences of simple sentences. Several approaches to the classification of potential coordinators are presented, one exploiting memory-based learning, another exploiting the publicly available Stanford parser, and a hybrid approach that classifies commas and conjunctions using the former system and comma–conjunction pairs using the latter. The article describes the initial set of features developed for exploitation by the memory-based classifier and presents optimization of that classifier. A baseline system is also described. The sentence simplification module was exploited by an IE system. With regard to the automatic classifiers that form the basis for simplification, comparative evaluation demonstrated that IE can be performed with greatest accuracy when exploiting the hybrid classifier. It also demonstrated that a simple baseline classifier induces improved accuracy when compared to systems that ignore the presence of coordinate structures in input sentences. The article presents an analysis of the errors made by the different sentence simplification modules and the IE system that exploits them. Directions for future research are suggested.