It’s perfectly reasonable to label and train for better responses to real world questions. What’s bad though is that the system doesn’t seem to actually be lear

It’s perfectly reasonable to label and train for better responses to real world questions. What’s bad though is that the system doesn’t seem to actually be learning much, as demonstrated by the question about whether to use a bat or a match to start a fire - if the network can’t extrapolate trained answers onto slightly rephrased questions, it’s doing a bad job at understanding and learning.