| Captions | |
| ======== | |
| The dev and test caption files follow the naming scheme: | |
| "img_<LANGUAGE>-en.tok.lc.filtered.(dev|test)" | |
| <LANGUAGE> is a two-letter code (de = German, fr = French, | |
| ru = Russian). | |
| The format of the parallel image-caption pairs is: | |
| "imagename<tab>foreign_caption ||| english_caption" | |
| Dev and Test Images | |
| =================== | |
| The images have to be downloaded manually. We provide a simple | |
| bash-script for downloading, which makes use of the following | |
| perl-script: | |
| https://commons.wikimedia.org/wiki/User:AzaToth/wikimgrab.pl | |
| Simply enter the "images" directory and call the script | |
| "download_dev_test.sh" from there, e.g.: | |
| $ cd images | |
| $ source ./download_dev_test.sh | |
| Downloading the Full Corpus | |
| =========================== | |
| We also provide a script for downloading the full corpus, i. e. | |
| (monolingual and multilingual) image annotations and their | |
| associated images directly from Wikimedia commons. Note that | |
| this may take a very long time (depending on your local bandwidth) | |
| and that we recommend distributing the task onto serveral machines. | |
| The script can be found in the "full_corpus" directory. | |
| We have compiled an index of 3.816.940 images on Wikimedia Commons. | |
| To download, for example, the first 20.000 images from the full | |
| corpus (and their annotatins) do the following: | |
| $ cd full_corpus | |
| $ ./download_full_corpus.sh 0 20000 | |
| The images will be placed in sequentially numbered directories, | |
| one directory containg a maximum of 10.000 images. | |
| The annotations will be written to files named XX.captions, where XX | |
| is a language identifier (en.cpations contains English captions, etc.) | |
| The files are tab-seperated in the following format: | |
| Corpus index <tab> Filename <tab> Annotated caption (including HTML "div" | |
| with metadata) | |
| Please note that some images might have been deleted since the | |
| index was created. They will be skipped. | |