Model with Sentence Transforrmers
Hi,
question - many Cross-Encoders take pairs of Question and Candidate Documents as input and deliver a single relevance score value from -10 ... +10 for each such pair (not related ... very relevant).
There is also a model wrapper CrossEncoder in the Sentence Transformers package, that I'd like to use.
I can use this CrossEncoder together with this model, but instead of single numbers i get pairs of numbers (Label0/Label1). Is Label1 the positive score (match) and i can use this as relevance score? Seems to work quite well...
Or should I somehow combine this two values? A bit more documentation would be great, because not everyone want to use the full Haystack/FARM-stack and just want to use the model itself, e.g. via a seperate Inference Server etc.
How is it trained, what does each of these 2 classification labels indicate?
Another question: How well does it work on mixed German / English?
Thx and best regards,
André