Sentiment-based masked language modeling for improving sentence-level valence–arousal prediction