Train models on joint dataset and predict on Businesswire data

[leonhardhennig]

Train models on joint dataset of #54 and predict on Businesswire data (after NER validation by crowd workers)

[phucdev]

Until we get the NER validated Businesswire data we will use our RE model to predict on the Businesswire data with NER labels from our NER model ensemble (majority label vote), resulting in a weakly supervised dataset

[phucdev]

Our current best model that was trained on the joint data had the following evaluation results:

f1 = 0.8844128308597224 precision = 0.8759653163527977 recall = 0.8930248618784531

• [x] Evaluate the model on the tacrev test split to compare to 0.77 f1 with a bert uncased model that was only trained on tacrev data

[phucdev]

The same model tested on the converted tacrev dataset achieved the following results:

f1 = 0.7601943798859075 precision = 0.7175907459114479 recall = 0.8081761006289309