Abstract
Uniform Meaning Representation (UMR) is a document-level semantic formalism that extends the Abstract Meaning Representation (AMR) with an added focus on multilingual compatibility and cross-sentence relations such as coreference, modal dependencies and temporal dependencies. Yet, this added expressivity comes at a cost of greater structural complexity that renders the UMR parsing task more difficult.
To cope with this challenge that is further compounded by the limited availability of annotations, this work introduces a modular framework for UMR parse generation that decomposes the task into smaller, tractable subtasks, establishing baseline performance. This subtask-aware approach (1) eliminates the need for UMR annotations for training and (2) facilitates task-specific optimization—including principled aggregation of sub-model predictions—to improve overall parsing performance.
Within this framework, the dissertation focuses on improving the performance of two core subtasks: modal dependency parsing and temporal dependency parsing. Improvements in these subtasks, along with more effective aggregation strategies that better integrate information across sub-models, consequently yield overall gains in UMR parsing performance relative to the original baseline system.