How to finetune on customized dataset?

Hi, thanks for this beautiful work!

I have a coreference dataset that has format Text, Pronoun, Pronoun-offset, A, A-offset, A-coref, B, B-offset, B-coref

Example:

Tom went out and bought a book. He found it very interesting.

Then, ‘Tom’ is the Pronoun, A is ‘He’, A-coref is True, B is ‘it’, B-coref is False, off-sets indicates each token’s position.

Can I use HMTL to finetune on such dataset? If so, how should I modify the json file, dataset reader? If not, how can I use HMTL to predict the conference pair of the Pronoun in each text? Like predict which word in the text corefers the Pronoun?

I appreciate your help!

[VictorSanh]

Hey!

Thanks for your interest!

I’ve experimenting using transfer learning to fine tune on another coreference dataset, so yes, it should be possible. I think you have two choices here: 1/ You convert your dataset format to a CoNLL20120-coreference format. It should be feasible if I understand correctly that the offsets are the token positions. Then it is pretty much just replacing the dataset paths in the config files and call fine_tune.py (and probably some minor parameters to add such as the pre-trained weights to fine-tune). 2/ You implement a DatasetReader (see hmtl/dataset_readers) that support your dataset format and modify the config file to call this specific dataset reader type (and then call fine_tune.py, it should be seamless once you add another dataset reader).

Victor

[VictorSanh]

I am not sure to fully understand what you dataset format is and more specifically what does A-coref=False and B-coref=False mean… From what I see, when there are no coreference links, there should not be any cluster id in a conll formatted file, and thus canonical_clusters in the dataset reader should be an empty list, and then gold_clusters should also be an empty list. But naturally, you want to train your model on these “negative” examples so it is able to detect when there are no coreference links.