Events

NLP and Text-as-Data Speaker Series: Improving model controllability with input and output alterations

Speaker: Emily Pitler

Location: 60 Fifth Avenue, Room 7th Floor Open Space

Date: Thursday, September 22, 2022

Language models have enabled exciting advances in language understanding accuracy. Tasks with dynamic output spaces introduce some new and interesting challenges that go beyond accuracy measurements. In settings such as semantic parsing, the output space may change from instance to instance and the model may need to adapt without retraining. This talk will describe some techniques for altering the inputs and/or the outputs to a language model that can improve the adaptability and controllability of an overall system, with examples drawn primarily from semantic parsing.