Contextual stance classification using prompt engineering
A prompt-based method that uses the existing conversation thread to create natural language prompts for few-shot learning with minimal reliance on training samples is introduced, whose preliminary results suggest that prompt engineering may be a competitive alternative to supervised methods both in terms of accuracy and development costs for the task at hand.
Abstract
This paper introduces a prompt-based method for few-shot learning addressing, as an application example, contextual stance classification, that is, the task of determining the attitude expressed by a given statement within a conversation thread with multiple points of view towards another statement. More specifically, we envisaged a method that uses the existing conversation thread (i.e., messages that are part of the test data) to create natural language prompts for few-shot learning with minimal reliance on training samples, whose preliminary results suggest that prompt engineering may be a competitive alternative to supervised methods both in terms of accuracy and development costs for the task at hand.