This line in the documentation states that label_mapping is an accepted parameter of TokenClassificationEvaluator.compute(), when it in fact is not.

I suspect this line was absent-mindedly copied over from the TextClassificationEvaluator documentation (that class's compute() method actually does accept this parameter).

In my opinion this would be a useful feature to implement (easily allowing users to evaluate models on datasets that don't use exactly the same labels), but at the very least this line should be removed from the documentation as long as the method doesn't support this parameter.