CompCog: Large-scale, empirically based, publicly accessible database of argument structure to support experimental and computational research

Boston College, Chestnut Hill MA

Investigators

Abstract

A basic function of language is to say who did what to whom. Linguists have identified many of the ways English fits who, what, and whom into a sentence, but it is still unclear why different rules apply to different sentences. But the language system is far more complex than this example suggests. English has about 150 different ways of fitting who, what, and whom into a sentence. Consider the sentence "Agnes received the package from Bart" in which the subject of the sentence (Agnes) gets the package. In contrast, in "Agnes gave the package to Bart," Agnes is still the subject of the sentence but she doesn't get the package. We might also say "Agnes tore at the package" or "Agnes looked at the package" but "Agnes saw at the package" doesn't work. This variability is a problem for teaching language and for building more robust voice-enabled systems. In this project, the investigators organize a large team of citizen scientists to try to identify some of the rules of who, what, and whom in English by analyzing the grammar and meaning of over 6000 verbs. This project will develop and evaluate methods for harnessing the power of collaborations with citizen scientists, facilitating broader engagement of the public in science. This project focuses on characterizing the semantics of the 6340 verbs listed in VerbNet, the most complete compendium of verb argument structure. From VerbNet, the investigators will generate example sentences for every verb in every compatible argument structure. These sentences are posted onto a website (gameswithwords.org/VerbCorner) where volunteers can code these sentences for 100 semantic features that have been previously identified as likely being relevant to verb argument structure rules. The website incorporates a number of strategies in order to make volunteer participation more enjoyable and rewarding. As part of the project, the investigators will assess existing and new analytic models for determining when sufficient judgments have been collected for a given item. Given the increasing importance of crowd-sourcing, developing and assessing these models should provide dividends beyond the current project.