From Physical to Social Commonsense: Natural Language and the Natural World

Abstract

Along with the meteoric rise of computation-hungry models, NLP research has also produced new handcrafted datasets. These datasets allow us to study problems that are difficult by web scraping alone. We can use such data to evaluate and extend machine learning models into new areas. One area of natural interest is work that connects NLP to the outside world. This dissertation describes four projects that present such datasets and computational models. Each project attempts to situate NLP in a context broader than text alone. As a common thread throughout, we make use of commonsense knowledge, either explicitly or implicitly. The first half of the dissertation covers two projects, Verb Physics and Social Chemistry, which contain explicit representations of commonsense knowledge. Respectively, they capture physical commonsense (e.g., that my house is bigger than I am) and social commonsense (e.g., that it's rude for my roommate to run the blender at 5am). The second half studies language production and evaluation. In this half, commonsense implicitly informs the work. Neural Naturalist addresses language generation from image comparisons. Scarecrow focuses on evaluating text generated by large language models. In the conclusion, we urge the field to embrace communication—not merely natural language—and thereby extend the richness of groundings we consider.