Phd in COMPUTER SCIENCE

Thesis title: Enhancing Semantic Parsing in the Age of Pre-trained Language Models

In recent years, Semantic Parsing (SP) has seen significant advancements, from novel semantics formalisms to more sophisticated parsers which leverage the power of Pretrained Language Models (PLMs). However, the challenges around cross-linguality and computational cost hinder broader adoption by researchers. In the first part of the thesis, we highlight the current limitations of semantic formalisms – with a focus on Abstract Meaning Representation (AMR) – to be considered a truly interlingua-meaning representation (chapter 2). Consequently, we present the BabelNet Meaning Representation, a novel linguistic formalism that successfully scales across different languages (chapter 3). Then, we describe our efforts to build BMR 1.0, the first dataset annotated with the formalism that allows us to develop BMR parsers. As a result, we provide insights into how BMR outperforms previous formalisms as an interlingua representation (chapter 4). The second part provides an overview of the historical evolution of AMR parsing, starting with its foundations and moving towards the latest architectural innovations (chapter 5). Then, our focus moves towards our own contributions to the field, which include i) an efficient and adaptable framework for parsing (chapter 6), ii) the development of parsers with state-of-the-art performance (chapter 7), iii) the exploration of the cross-attention mechanism for simultaneous cross-lingual alignment generation during parsing (chapter 8), iv) the introduction of advanced ensemble frameworks (chapter 9), and v) the creation of a comprehensive broad-domain dataset for AMR parsing interconnected in different languages and other NLU tasks (chapter 10). Finally, we summarise our contributions and we give some insights into potential future directions in the constantly evolving field of Semantic Parsing.