Use a fine-tuned BERT for identification of tricky introductions

[MansMeg]

Now more length introductions can be missed in the current algorithms. One way would be to use a BERT model on a small trainings set of positive samples (complicated introductions), then use that to predict new possible introductions. based on these we would then add identified complicated introductions to a new training set and re-train/fine-tune the introductions and continue like that until we have found a couple of thousands of introductions.

[MansMeg]

This is now being done as a masters thesis.

[MansMeg]

It is a little too far away from my work. But they might find the same technique useful.