Update README.md

README.md

@@ -31,6 +31,20 @@ and 0.720 in macro-F1 on a Croatian test set (440 instances from ParlaMint-HR 4.
 An additional evaluation on smaller samples from Czech ParlaMint-CZ, Bulgarian ParlaMint-BG and Ukrainian ParlaMint-UA datasets shows 
 that the model achieves macro-F1 scores of 0.736, 0.75 and 0.805 on these three test datasets, respectively.
 
+For end use scenarios, we recommend filtering out predictions based on the model's prediction confidence.
+
+When the model was applied to the ParlaMint datasets, we annotated instances that were predicted with confidence below 0.60 as "Mix".
+
+With this approach, we annotate as Mix:
+- 8.6% of instances in the English test set
+- 11.4% of instances in the Croatian test set
+
+Performance of the model on the remaining instances (all instances not annotated as "Mix"):
+
+|    |   micro-F1 |   macro-F1 |   accuracy |
+|:---|-----------:|-----------:|-----------:|
+| EN |   0.838 |    0.838 |   0.838 |
+| HR |   0.724 |   0.726 |   0.724 |
 
 ## Use