Google’s AI folks have made a new chatbot using a transformer-based architecture (but a network substantially bigger than full-size GPT2). Blog post; paper on arXiv. They claim it does much better than the state of the art (though I think everyone would agree that the state of the art is rather unimpressive) according to a human-evaluated metric they made up called “sensibleness and specificity average”, which means pretty much what you think it does, and apparently correlates with perplexity in the right sort of way.
Google’s AI folks have made a new chatbot using a transformer-based architecture (but a network substantially bigger than full-size GPT2). Blog post; paper on arXiv. They claim it does much better than the state of the art (though I think everyone would agree that the state of the art is rather unimpressive) according to a human-evaluated metric they made up called “sensibleness and specificity average”, which means pretty much what you think it does, and apparently correlates with perplexity in the right sort of way.