Feedback utterances such as ‘yeah’, ‘mhm’,and ‘okay’, convey different communicative functions depending on their prosodic realizations, as well as the conversational context in which they are produced. In this paper, we investigate the performance of different models and features for classifying the communicative function of short feedback tokens in American English dialog. We experiment with a combination of lexical and prosodic features extracted from the feedback utterance, as well as context features from the preceding utterance of the interlocutor. Given the limited amount of training data, we explore the use of a pre-trained large language model (GPT-3) to encode contextual information, as well as SimCSE sentence embeddings. The results show that good performance can be achieved with only SimCSE and lexical features, while the best performance is achieved by solely fine-tuning GPT-3, even if it does not have access to any prosodic features.