I asked the (at the time) SOTA author to replace its BERT implementation with XLnet. He accepted and won 0.5% of accuracy on constituency parsing https://github

I asked the (at the time) SOTA author to replace its BERT implementation with XLnet. He accepted and won 0.5% of accuracy on constituency parsing https://github.com/sebastianruder/NLP-progress/blob/master/english/constituency_parsing.md https://github.com/sebastianruder/NLP-progress/blob/master/e... which is actually a huge, it means 5% less errors. Which tasks remain to be tried? Most actually but an obvious one would be coreference resolution https://github.com/sebastianruder/NLP-progress/blob/master/english/coreference_resolution.md https://github.com/sebastianruder/NLP-progress/blob/master/e...