Transforming dependency parses into ternary expressions for enhanced indexing and matching

Abstract

Advancements in dependency parsing allow machines to quickly and accurately analyze natural language sentences; however, these parses often require non-trivial manipulation to be useful for many applications. This thesis describes Astroparse, a system for producing ternary expression (subject–relation–object triple) parses by building on existing third-party dependency parsers. I present a design which uses a previously-studied training-example framework with additional augmentations to expand its parsing abilities. I analyze some ways that dependency parse representations fail to capture important relationships in sentences and present algorithms to recover ternary expressions despite those failures. I evaluate my system by examining its outputted ternary expressions manually as well as by qualitatively analyzing its learned transformations. On sentences from high-quality articles in Wikipedia, Astroparse achieves an average precision of up to 93.4% and an estimated recall of about 88.1%, and recovers an average of 35.3% more relations than raw dependency parses alone. My system is also flexible to changes in the underlying dependency parsers and produces human-readable explanations for each ternary expression it produces.