search

greenelab / snorkeling

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

Error Analysis and Few Other Results #89

Closed danich1 closed 5 years ago

danich1 commented 5 years ago

This pr contains three modules:

  1. a random label function that emits positive and negative labels at a specified frequency.
  2. Edge predictions reporting the test set only
  3. Model calibration tables for each edge type.

No need to review code unless you want to.

danich1 commented 5 years ago

Ah. For that figure I am incorporating a single label function onto the baseline model. This single label function randomly generates labels (pos or neg) at a certain frequency (x-axis). For example this mentioned function could emit negative labels for 25% of the training set or for 50% of the training set. I did this experiment to understand why the Generative model performance increases for one label function and then drastically decreases.