BERT Can See Out of the Box: On the Cross-modal Transferability of Text Representations

Pre-trained language models such as BERT have recently contributed to significant advances in Natural Language Processing tasks. Interestingly, while multilingual BERT models have demonstrated impressive results, recent works have shown how monolingual BERT can also be competitive in zero-shot cross-lingual settings. This suggests that the abstractions learned by these models can transfer across languages, even when trained on monolingual data. In this paper, we investigate whether such generalization potential applies to other modalities, such as vision: does BERT contain abstractions that generalize beyond text? We introduce BERT-gen, an architecture for text generation based on BERT, able to leverage on either mono- or multi- modal representations. The results reported under different configurations indicate a positive answer to our research question, and the proposed model obtains substantial improvements over the state-of-the-art on two established Visual Question Generation datasets.