search

greenelab / snorkeling

Extracting biomedical relationships from literature with Snorkel 🏊
Other
59 stars 17 forks source link

Using Label Functions Designed for One Relation to Predict Another Relation #78

Closed danich1 closed 5 years ago

danich1 commented 5 years ago

This pull request is designed to upload the results of my latest experiment, where I aimed to see if one relation type can aid in predicting another relation type. I.E Can compound treats disease label functions classify sentences that infer Disease associating a Gene relationship? How about compound binds gene label functions predicting sentences that infer disease associates a gene relationship? etc...

To answer these questions the approach here was to randomly add label function information to the baseline model (a distantly supervised classifier using relation specific database to label sentences) and estimate if performance improved or worsened.

There are a lot of files being added here; however, the only files that need to be reviews are the single_task_label_function_experiment.ipynb notebooks. There is a notebook for each relation type. Let me know what you think. If you get a chance take a look at the images within the figures folder. It will provide a global picture on what I had accomplished.

cgreene commented 5 years ago

The title is transfer learning, but I'm not sure if that's quite the right term. Maybe label transfer?

It's really pretty amazing how robust things are in AUC terms to moving labels around. It's also interesting how much having some text-based LF helps, somewhat (apparently) regardless of what exact LF it is. Cool!