This dissertation studies implicit semantic roles – instances where a participant is not explicitly stated in the text, such as the arguments in “∅_you eat ∅_food yet?”. It focuses upon the task of resolution of these implicit roles – determining what these unstated arguments refer to.

This thesis proposes a typology of different kinds of these implicit roles, distinguished not by their syntactic behavior (which is very language-specific), but by their referential behavior. Implicit roles in some contexts act like pronouns, looking to recently mentioned referents; in other contexts, implicit roles can refer generically to “people in general”, or to a speaker or addressee. The first contribution of the thesis is to outline the range of these various interpretations seen for these implicit roles across different languages so that we might make apples-to-apples comparisons from language to language.

The second part of this thesis presents new corpora of English implicit semantic roles, and presents computational models trained upon those corpora to do implicit role resolution. This provides data to do the full task of resolving all unstated semantic roles in a document. On those new corpora, a set of implicit role resolution models are trained, showing that while this data is difficult, one can build wide-coverage systems which predict implicit semantic roles using the PropBank semantic role inventory.

These implicit role resolution models are used to illuminate characteristics currently being learned by implicit role resolution models, and to highlight issues that are still poorly represented. It is hoped that this thesis lays the groundwork for future work in implicit role resolution, and for addressing related linguistic questions which those models might enable.