Hacker News Reader

Seems to do a consistent job making comparable sentences. Curious what tech is underneath the UI.

It uses a deep sequence to sequence model, and the drop-down thesaurus are phrase embeddings trained on large corpuses like Reddit and Wikipedia.

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