The dataset is formatted as follows:
[task-name-folder]/ # constituency_parsing, coreference_resolution, data-to-text_generation, entity_linking, face_alignment, face_detection, hypernym_discovery, natural_language_inference
├── [article-counter-folder]/ # ranges from 0 to N-1 if we annotated N articles per task
│ ├── [articlename]-Grobid-out.txt # plaintext output from the [Grobid parser](https://github.com/kermitt2/grobid)
│ ├── [articlename]-Stanza-out.txt # plaintext preprocessed output from [Stanza](https://github.com/stanfordnlp/stanza)
│ └── ... # if N articles were annotated, this repeats (N-1) more times
└── ...
The submission will have be organized per the following directory structure:
[task-name-folder]/ # constituency_parsing, coreference_resolution, data-to-text_generation, entity_linking, face_alignment, face_detection, hypernym_discovery, natural_language_inference
├── [article-counter-folder]/ # ranges from 0 to N-1 if we annotated N articles per task
│ ├── sentences.txt # annotated contribution sentences in the file identified by the sentence number in the preprocessed data with counter starting at 1
│ └── entities.txt # annotated phrases from the contribution sentences where the phrase spans are specified by their first and last token numbers with the token counter starting at 1
│ └── triples/ # the folder containing information unit triples one per line
│ │ └── research-problem.txt # `research problem` triples (one research problem statement per line)
│ │ └── model.txt # `model` triples (one statement per line)
│ │ └── ... # there are 12 information units in all and each article may be annotated by 3 or 6
│ └── ... # if N articles were annotated, this repeats (N-1) more times
└── ... # if K tasks were selected overall, this repeats (K-1) more times
A valid sample submission for this phase can be downloaded here