Automated Metamorphic Testing using Transitive Relations for Specializing Stance Detection Models

Alisa Arno, Futoshi Iwama, Mikio Takeuchi

Abstract: In machine-learning-based natural language processing, methods with high accuracy have been proposed for stance detection tasks. However, when they are applied to specific domains, they are often inaccurate due to domain-specific expressions. We propose an automated metamorphic testing method using transitive relations for creating training data that specializes stance detection in a specific domain. By specializing IBM Debater’s stance detection in currency exchange domain, we confirmed our proposed method can improve the accuracy of judging the currency exchange-related sentences.