I have been looking for BERT for many tasks. I would like to compare the performance to answer an FAQ, using BERT semantic similarity and BERT Q/A. However, I'm not sure it is a good idea to use semantic similarity for this task. If it is, do you think it is possible to find a dataset to fine-tune my algorithm?
Maybe the following article can help you:
They evaluate their model in localgovFAQ and StackExchange datasets.
You can do that and you'll propably find data, yet that depends on the kind of FAQ data you will apply it on. More importantly, what insight do you gain by comparing two BERT models?
Secondly, f you mean with semantic similarity that vector space embeddings are used, even for the retrieval/ranking and not just for re-ranking, then I can tell you that the performance still isn't SOTA. But you can simply use a such a neural semantic model for re-ranking.
We are working on that. So if you wanna know more, PM me.