A Personalised Approach to Lexical Complexity

Abstract

This thesis considers the task of lexical simplification: the simplification of word content within a text. Lexical simplification has been shown to positively impact the readability of a text and improve reader understanding and information retention. However, the process of simplifying word content is often modelled as a one-size-fits-all problem. To better model lexical simplification, it is important to understand how the perspective of word complexity varies for different audiences. The first contribution of the thesis explores lexical simplification from the application perspective by introducing a recursive and contextually-aware algorithm for the task. The modular nature of the system provides scope to personalise the simplification process. However, to personalise simplification, insights are needed on how best to model word complexity for differing audiences. To understand this, an investigation into how the perspective of word complexity differs for audiences is presented, including the extent to which annotators of varying proficiencies and language backgrounds hold the same notion of word complexity. Having shown that word complexity is highly subjective, the thesis presents a study demonstrating the utility of personalised word complexity models. In the study, per-individual models of word complexity are trained in real-time using active learning. The final contribution of this thesis is a discussion of the ethical implications when embedding such simplification models into real-world applications.