Abstract
An important task in understanding the meaning of natural language text is to represent and understand the temporal information in the text. Time expressions, events that happened at some time points, and temporal relations between these time expressions and events are the three basic temporal information commonly present in texts. A well designed machine-readable temporal representation is crucial for representing and understanding these information efficiently. Most fundamental research on temporal information modeling has been representing temporal relations in a pair-wise manner -- the temporal relation between pairs of time expressions and/or events are explicitly and separately modeled. This stream of representations faces a few challenges. First, human annotation for this representation is laborious and on some level arbitrary. Second, computation on this representation is expensive and inefficient on scalability. Third, due to the nature of temporal transitivity,annotations (human or computational) harbor potential conflicts on temporal relations.\r \r In this dissertation, we introduce a new temporal representation to address these challenges -- the Temporal Dependency Tree (TDT) structure. A Temporal Dependency Tree represents temporal information in a text as a single dependency tree. Time expressions and events are represented as nodes on the tree, while temporal relations are represented as edges. A TDT explicitly models n temporal relations for a text with n time expressions and events, reducing human annotation labor, computation complexity, and temporal transitivity conflicts. As a proof-of-concept, we performed annotation experiments on the TDT representation to show stable and high inter-annotator agreements. To support further linguistic study on TDT and automatic system training, we built an expert-annotated TDT corpus (on two domains: news and narratives). One step closer to automatic temporal information modeling and understanding, we built a competitive Temporal Dependency Parser that parses time expressions and events in a text into a Temporal Dependency Tree structure. Finally, to collect larger amount of TDT data more efficiently, and further support the training of better temporal dependency parsers, we experimented with crowdsourcing approaches and built a TDT corpus with high agreements through crowdsourcing.