This project was a collaborative effort between the University of Toronto and IBM’s Center for Advanced Studies (CAS) to extend prior work in speech-act theory into the realm of deep learning.
Background
Speech-act theory, originating from Austin (1962) and Searle (1969), classifies utterances not just by their content but by the action they perform: requests, commitments, assertions, declarations. Carvalho & Cohen (2005) adapted this framework to email, defining “email speech acts” — illocutionary points specific to workplace communication.
Approach
We built taxonomic word embeddings for semantic relatedness, leveraging the hierarchical structure of speech-act categories to learn representations that capture both syntactic and pragmatic similarity between email utterances.
The resulting classifier could identify the intent behind an email — whether it was a request for action, a commitment, a question, or an informational statement — enabling downstream automation of task extraction and delegation.
Impact
Our work culminated in the launch of IBM Watson Orchestrate, a natural-language-powered productivity assistant that won a CES 2022 Innovation Award.
Supervision
This research was supervised by Dr. Eric Yu at the University of Toronto.