The paper says "To train embeddings, we use domain-matched data for each downstream task. For coreference we train on Wikipedia data, ..." in section 4.1 Datasets.
However, README.md says,
Oh sorry, the readme was out of date from our initial experiments, good catch! We use wikipedia as the paper says. I have updated and corrected the readme.
The paper says "To train embeddings, we use domain-matched data for each downstream task. For coreference we train on Wikipedia data, ..." in section 4.1 Datasets. However, README.md says,
Which is right?