In this repository, we provide files that contain information about the processed BBC News Corpus and its extracted relations. The uploaded data is made anonymous. Therefore, we also provide scripts for producing those files as well as reconstructing the original extracted relations. If you make use of these datasets please consider citing the publication:
R. Knaebel and M. Stede. "Semi-Supervised Tri-Training for Explicit Discourse Argument Expansion", Proc. LREC 2020 [PDF] [BibTeX]
For corpus preparation, we refer to the make_corpus.py
script.
It gets the path to one of the downloaded raw BBC corpora and writes all information into one json file.
The format is comparable to the CoNLL2016 format of the shared task.
Corpus links:
- bbc: http://mlg.ucd.ie/files/datasets/bbc-fulltext.zip
- bbcsport: http://mlg.ucd.ie/files/datasets/bbcsport-fulltext.zip
python3 make_corpus.py CORPUS_PATH JSON_PATH.json
For removing textual information, we use the dehydrate.py
script.
It returns a flattened json structure that contains only TokenList information and the corresponding document id.
python3 dehydrate.py RELATIONS_PATH > RELATION_ID.json
For back conversion, we use the hydrate.py
script.
It combines the extracted TokenLists with the corpus file and thus reconstructs the original extraction.
python3 hydrate.py JSON_PATH.json RELATION_ID.json > RELATION_FULL.json