Abstract
In this thesis, we explore foundational issues related to how paraphrases are handled in NLP and how they can be applied to enhance NLP tasks. Firstly, we show that key problems exist, such as the underspecification of paraphrases, leading to issues like the low generalisation ability of commonly used paraphrase datasets. As such, we investigate mitigation strategies and propose new approaches to better understand and work with paraphrases in NLP. Next, we introduce new metrics to better characterise paraphrase diversity. We then propose a new ontology, Semantic Paraphrase Types (SPT), to guide how different types of paraphrases and non-paraphrases are defined and handled in NLP tasks. Based on this, we created a new dataset and benchmark task for fine-grained paraphrase classification, which also includes an explainability task to further address issues with explainability and subjectivity in the real world. Lastly, we also demonstrate the applications of our research on a real-world task, Banking77, showing that it can be used to characterise and adjust the behaviour of a paraphrase generation model, informing how we can apply the model to perform data augmentation to improve low-resource text classification performance, and how it can be applied towards test case generation to improve testing processes in NLP.