Core NLP as Generative Prompting

Han He

Date: 2022-04-01 / 4:00 ~ 5:00 PM
Location: MSC E306 (https://emory.zoom.us/j/99364825782)


Abstract

Core NLP tasks have been dominated by task-specific encoders and decoders, limiting the design of any general frameworks. In this project, we challenge core NLP tasks with a new paradigm called generative prompting, which re-purposes a pre-trained seq2seq language model to answer many core NLP questions in human language (prompts). With constrained decoding, we further ensure the validity of the predictions produced in prompt generation. On several tasks, our single seq2seq model performs on par with other ad-hoc state-of-the-art models designed specifically for each task.

Link

Presentation