Previous studies of predictive mechanisms in computer-mediated communication (CMC) suggested that native speakers (L1) rely on auditory cues and emotion in conversation processing. To understand how the prediction mechanism differs for non-native speakers (L2) in CMC, this study assessed how the loss of multi-modal cues affects word predictability in turn-taking, considering language background and social factors. L2 watched videos, listened to audio, or read a transcript of conversations, and predicted the same set of omitted words with different levels of predictability and semantic relatedness in different CMC. Results showed that, similar to the L1 study in He et al. (2025), higher response similarity but longer response time were observed in conditions with richer cues in L2 predictive processing. Semantic relatedness, self-emotion, attention, and language proficiency did not affect predictability. Participants reporting negative emotions and more limited L2 exposure demonstrated reduced prediction accuracy, particularly in cue-rich environments. These findings expand our understanding of L2 predictive processing in CMC by highlighting how multimodal cue integration operates differently for L2. The results have implications for developing communication technologies and language pedagogies tailored to L2 across various mediated communication contexts.