New Data-Driven Approaches to Text Simplification

Abstract

Many texts we encounter in our everyday lives are lexically and syntactically very complex. This makes them difficult to understand for people with intellectual or reading impairments, and difficult for various natural language processing systems to process. This motivated the need for text simplification (TS) which transforms texts into their simpler variants. Given that this is still a relatively new research area, many challenges are still remaining. The focus of this thesis is on better understanding the current problems in automatic text simplification (ATS) and proposing new data-driven approaches to solving them. We propose methods for learning sentence splitting and deletion decisions, built upon parallel corpora of original and manually simplified Spanish texts, which outperform the existing similar systems. Our experiments in adaptation of those methods to different text genres and target populations report promising results, thus offering one possible solution for dealing with the scarcity of parallel corpora for text simplification aimed at specific target populations, which is currently one of the main issues in ATS. The results of our extensive analysis of the phrase-based statistical machine translation (PB-SMT) approach to ATS reject the widespread assumption that the success of that approach largely depends on the size of the training and development datasets. They indicate more influential factors for the success of the PB-SMT approach to ATS, and reveal some important differences between cross-lingual MT and the monolingual v MT used in ATS. Our event-based system for simplifying news stories in English (EventSimplify) overcomes some of the main problems in ATS. It does not require a large number of handcrafted simplification rules nor parallel data, and it performs significant content reduction. The automatic and human evaluations conducted show that it produces grammatical text and increases readability, preserving and simplifying relevant content and reducing irrelevant content. Finally, this thesis addresses another important issue in TS which is how to automatically evaluate the performance of TS systems given that access to the target users might be difficult. Our experiments indicate that existing readability metrics can successfully be used for this task when enriched with human evaluation of grammaticality and preservation of meaning.