Cross-lingual Visual Pre-training for Multimodal Machine Translation

Abstract

Supplements for the paper entitled "Cross-lingual Visual Pre-training for Multimodal Machine Translation" which is accepted by the EACL'2021 conference. Further instructions on how to use these resources are explained at https://github.com/ImperialNLP/VTLM A tarball that contains a custom train, valid, test split of Conceptual Captions (CC) dataset. The included TSV files havean additional column containing automatic German translations of the original English captions. We only provide samples for which we could download the images and extract meaningful features. This amounts to ~3M out ouf ~3.3M original CC samples. A tarball of the exact object detector checkpoint used for feature extraction. A tarball with pre-extracted Multi30k dataset features.