Leveraging Text and Image BERTs for Few-Shot Classification in Cryptocurrency Tweets

Investors frequently turn to Twitter as a critical resource for gaining insights into cryptocurrency. The cryptocurrency market could inflate through influencing messages. This paper explores a few off-the-shelf Hugging Face's BERTs (Bidirectional Encoder Representations from Transformers) to decipher tweet interests in this context. Influential tweets, rich in text and multimodal information like images and videos, are worthy of a comprehensive understanding. It employed DistilBERT (a distilled version of BERT) for text features and Bootstrapping Language-Image Pre-training 2 (BILP-2) for image features. It also extracted linguistic representations for other multimodal data, e.g., videos. It explored 6 architectures to fuse multimodal features. It validated them on the PAN23 dataset created for Profiling Cryptocurrency Influencers with Few-shot Learning. One novel system shows a marked improvement in Macro F1 by +10.5% over the DistilBERT-only baseline.