Sentence Simplification for Text Processing

Abstract

Propositional density and syntactic complexity are two features of sentences which affect the ability of humans and machines to process them effectively. In this thesis, I present a new approach to automatic sentence simplification which processes sentences containing compound clauses and complex noun phrases (NPs) and converts them into sequences of simple sentences which contain fewer of these constituents and have reduced per sentence propositional density and syntactic complexity. My overall approach is iterative and relies on both machine learning and handcrafted rules. It implements a small set of sentence transformation schemes, each of which takes one sentence containing compound clauses or complex NPs and converts it one or two simplified sentences containing fewer of these constituents (Chapter 5). The iterative algorithm applies the schemes repeatedly and is able to simplify sentences which contain arbitrary numbers of compound clauses and complex NPs. The transformation schemes rely on automatic detection of these constituents, which may take a variety of forms in input sentences. In the thesis, I present two new shallow syntactic analysis methods which facilitate the detection process. The first of these identifies various explicit signs of syntactic complexity in input sentences and classifies them according to their specific syntactic linking and bounding functions. I present the annotated resources used to train and evaluate this sign tagger (Chapter 2) and the machine learning method used to implement it (Chapter 3). The second syntactic analysis method exploits the sign tagger and identifies the spans of compound clauses and complex NPs in input sentences. In Chapter 4 of the thesis, I describe the development and evaluation of a machine learning approach performing this task. This chapter also presents a new annotated dataset supporting this activity. In the thesis, I present two implementations of my approach to sentence simplification. One of these exploits handcrafted rule activation patterns to detect different parts of input sentences which are relevant to the simplification process. The other implementation uses my machine learning method to identify compound clauses and complex NPs for this purpose. Intrinsic evaluation of the two implementations is presented in Chapter 6 together with a comparison of their performance with several baseline systems. The evaluation includes comparisons of system output with human-produced simplifications, automated estimations of the readability of system output, and surveys of human opinions on the grammaticality, accessibility, and meaning of automatically produced simplifications. Chapter 7 presents extrinsic evaluation of the sentence simplification method exploiting handcrafted rule activation patterns. The extrinsic evaluation involves three NLP tasks: multidocument summarisation, semantic role labelling, and information extraction. Finally, in Chapter 8, conclusions are drawn and directions for future research considered.