Step Semantics: Representations for State Changes in Natural Language

The dynamics of the world is often bound up in processes. These include continuous processes, such as flows and motion, and discrete processes, such as count and break. Things that occur in the world can often be described at multiple levels of detail, using combinations of continuous and discrete processes, and it is important to be able to shift among levels of detail as needed for communication and understanding. This paper describes step semantics, a framework that draws upon prior work in qualitative reasoning and discrete action representations to provide a set of representation conventions for processes described in natural language, independent of a particular task or dataset. We explore its potential in two ways: Analyses of recipes with complex temporal structure and learning from AI2’s ProPara dataset.