Abstract
The goal of Natural Language Understanding (NLU) is to develop systems that have the capability to understand the meaning of human language. Whether an event happened or not, and when did an event happen, are two important questions to ask in NLU. Event factuality prediction, and temporal relation extraction are the tasks that aim to answer those two questions in Natural Language Processing (NLP). To enable machine to understand the factuality and temporal information in texts, two components are needed: large-scale, high quality datasets annotated with machine understandable representations for event factuality and temporal relations, and systems that can take raw texts as input and extract factuality and temporal information with high accuracy. In this thesis, we address challenges from those two aspects.
We model the temporal relation extraction task as temporal dependency parsing, and model the event factuality prediction task as modal dependency parsing (TDP and MDP). Firstly, we propose a temporal annotation scheme called temporal dependency graphs which extend previous research on temporal dependency trees. We present the construction of a corpus of 500 Wikinews articles annotated with temporal dependency graphs (TDGs). We argue that temporal dependency graphs, built on previous research on narrative times and temporal anaphora, provide a representation scheme that achieves a good balance between completeness and practicality in temporal annotation. Next, we build a large data set annotated with modal dependency structures via crowdsourcing, and demonstrated the quality of this data set with a careful evaluation of each aspect of the annotation. Thirdly, we develop temporal and modal dependency parsers with a ranking architecture. The parsers extract events and conceivers (time expressions in TDP), and parse the dependency structures. To further improve the parsing accuracy, we build temporal and modal dependency parsers that are based on language model priming.