Hacker News new | ask | show | jobs
by saternius 2670 days ago
It uses a deep sequence to sequence model, and the drop-down thesaurus are phrase embeddings trained on large corpuses like Reddit and Wikipedia.
1 comments

What did you use for the parallel data? the paraphrases are much better than an auto-encoder.