Y
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
lumost
2669 days ago
What did you use for the parallel data? the paraphrases are much better than an auto-encoder.
link