Re:LSTM trained on HN
by minimaxir on Monday August 07, 2017 @01:22PM
(#14949220)
As someone who's cited by the OP and coincidentally released a tool today that easily trains a LSTM on any dataset demoed with Hacker News data (https://github.com/minimaxir/textgenrnn), I have a couple comments:
1) My Get All Hacker News Submissions script is somewhat obsolete since all HN data is now on BigQuery (will add a note to the README today)
2) There is a massive, massive selection bias in terms of the quality of the selected generated texts. If you look at the sample output (https://github.com/danhon/deep-hackernews/blob/master/sample...), there is a large disparity in quality between the chosen submissions for the blog post and a typical generated submission. (this is expected, although the need for human curation damages the "Turing test" suggestions proposed by other commenters)
Compare with the sample output at similar temperature from my 128-cell LSTM network (https://github.com/minimaxir/textgenrnn/blob/master/outputs/...), which is several orders smaller (OP uses a 512-cell network), and many magnitudes lower amount of training data. (I am curious how long OP spent training the network.)
1) My Get All Hacker News Submissions script is somewhat obsolete since all HN data is now on BigQuery (will add a note to the README today)
2) There is a massive, massive selection bias in terms of the quality of the selected generated texts. If you look at the sample output (https://github.com/danhon/deep-hackernews/blob/master/sample...), there is a large disparity in quality between the chosen submissions for the blog post and a typical generated submission. (this is expected, although the need for human curation damages the "Turing test" suggestions proposed by other commenters)
Compare with the sample output at similar temperature from my 128-cell LSTM network (https://github.com/minimaxir/textgenrnn/blob/master/outputs/...), which is several orders smaller (OP uses a 512-cell network), and many magnitudes lower amount of training data. (I am curious how long OP spent training the network.)