Learning Visually Grounded Sentence Representations
This work addresses the challenge of enhancing sentence representations for NLP tasks by incorporating visual grounding, showing incremental improvements over existing text-only approaches.
The paper tackles the problem of learning sentence representations grounded in visual data by training models on an image captioning corpus to predict image features from captions, resulting in a grounded sentence encoder that improves performance on COCO caption and image retrieval and transfers well to various NLP tasks with better results than text-only models.
We introduce a variety of models, trained on a supervised image captioning corpus to predict the image features for a given caption, to perform sentence representation grounding. We train a grounded sentence encoder that achieves good performance on COCO caption and image retrieval and subsequently show that this encoder can successfully be transferred to various NLP tasks, with improved performance over text-only models. Lastly, we analyze the contribution of grounding, and show that word embeddings learned by this system outperform non-grounded ones.