Information Extraction using Non-consecutive Word Sequences

We address an important deficiency in existing machine learning approaches for information extraction from natural language texts. Existing techniques for information extraction employ rules that exploit properties of consecutive word sequences. We argue that sequences of non-consecutive words capturing long range contextual correlations are vital features for information extraction from natural language text. We propose an efficient method that extends the a-priori algorithm to mine frequently occurring non-consecutive word sequences from a given corpus. We also perform a simplistic aggregation of feature information across multiple mentions of an entity in a document to avoid independent classification of the multiple occurrences of the entity. Experiments on some standard data sets show substantial improvements over previously reported