Decompositional Semantics for Events, Participants, and Scripts in Text
MetadataShow full item record
This thesis presents a sequence of practical and conceptual developments in decompositional meaning representations for events, participants, and scripts in text under the framework of Universal Decompositional Semantics (UDS) (White et al., 2016a). Part I of the thesis focuses on the semantic representation of individual events and their participants. Chapter 3 examines the feasibility of deriving semantic representations of events from dependency syntax; we demonstrate that predicate- argument structure may be extracted from syntax, but other desirable semantic attributes are not directly discernible. Accordingly, we present in Chapters 4 and 5 state of the art models for predicting these semantic attributes from text. Chapter 4 presents a model for predicting semantic proto-role labels (SPRL), attributes of participants in events based on Dowty’s seminal theory of thematic proto-roles (Dowty, 1991). In Chapter 5 we present a model of event factuality prediction (EFP), the task of determining whether an event mentioned in text happened (according to the meaning of the text). Both chapters include extensive experiments on multi-task learning for improving performance on each semantic prediction task. Taken together, Chapters 3, 4, and 5 represent the development of individual components of a UDS parsing pipeline. In Part II of the thesis, we shift to modeling sequences of events, or scripts (Schank and Abelson, 1977). Chapter 7 presents a case study in script induction using a collection of restaurant narratives from an online blog to learn the canonical “Restaurant Script.” In Chapter 8, we introduce a simple discriminative neural model for script induction based on narrative chains (Chambers and Jurafsky, 2008) that outperforms prior methods. Because much existing work on narrative chains employs semantically impoverished representations of events, Chapter 9 draws on the contributions of Part I to learn narrative chains with semantically rich, decompositional event representations. Finally, in Chapter 10, we observe that corpus based approaches to script induction resemble the task of language modeling. We explore the broader question of the relationship between language modeling and acquisition of common-sense knowledge, and introduce an approach that combines language modeling and light human supervision to construct datasets for common-sense inference.