Multilingual Emotion Recognition: Discovering the Variations of Lexical Semantics between Languages

Xulang Zhang, Rui Mao, Erik Cambria

Abstract: The task of multilingual emotion recognition holds significant importance in cross-cultural communication and data mining. While prior research has concentrated on enhancing classification accuracy using state-of-the-art techniques, it has often overlooked a crucial linguistic aspect—the semantic disparities across different languages. This study aims to address this gap by introducing a novel method to identify lexical semantic variations in diverse languages. The detected semantic variation features are subsequently injected into a multilingual emotion recognition model to enhance its performance within a target language. Notably, existing multilingual pre-trained language models are likely biased toward English word meanings, leading to inaccurate emotion predictions in other languages due to the misinterpretation of semantics. Our proposed semantic variation injection method tackles this limitation, resulting in improved accuracy. These findings contribute to the ongoing development of robust and culturally sensitive emotion recognition systems, offering valuable insights for both the linguistics and computational linguistics communities engaged in multilingual research.