search

nyu-mll / jiant-v1-legacy

The jiant toolkit for general-purpose text understanding models
MIT License
21 stars 9 forks source link

Multitask learning for Sequence Labeling Task #962

Open jeswan opened 4 years ago

jeswan commented 4 years ago

Issue by antgr Sunday Nov 17, 2019 at 15:07 GMT Originally opened as https://github.com/nyu-mll/jiant/issues/962

Hi, I would like to apply multitask learning, for two tasks, one of which is a Sequence Labeling Task. If this information clarifies the things better, the other task could be a NLI task. Do you have an example of how I could create a proof of concept for that goal? Thanks!

jeswan commented 4 years ago

Comment by antgr Wednesday Nov 20, 2019 at 10:29 GMT

@pruksmhc Any thoughts on that? Is this supported from the library? Any pointers for that goal?

jeswan commented 4 years ago

Comment by antgr Saturday Nov 30, 2019 at 07:23 GMT

@sleepinyourhat Prof, any feedback on that?

jeswan commented 4 years ago

Comment by pruksmhc Thursday Dec 12, 2019 at 18:30 GMT

Hi! Sorry for the delay, we do have a task called TaggingTask in jiant/models.py (look at CCG task for reference of how to implement a sequence labelling task), as well as an NLI task (look at MNLI or any of the other NLI task for example). I would suggest following the tutorial on how to do multitask learning between the two tasks.

jeswan commented 4 years ago

Comment by antgr Sunday Dec 15, 2019 at 10:11 GMT

Thanks! Could I use torchtext.datasets, for a task ? (the data format is always .tsv?)