Sentiment in spoken interaction is not a fixed property of utterances but a dynamic and multimodally constructed phenomenon. This study investigates how sentiment polarity is differently interpreted across modalities -- text, audio, silent video, and full video -- within Chinese screen dialogue, using a corpus in which each modality has been independently annotated. We show that judgments of affective stance frequently diverge, and these divergences shed light on how emotion and evaluation are shaped and perceived across semiotic channels.