Revisiting paraphrase question generator using pairwise discriminator

In this paper, we propose a method for obtaining sentence-level embeddings. While the problem of word-level embeddings is very well studied, novel This obtained by simple in context solving paraphrase generation task. If use sequential encoder-decoder model generating paraphrase, would like generated to be semantically close original sentence. One way ensure adding constraints true and unrelated candidate sentence far. ensured using pair-wise discriminator that shares weights with encoder. trained suitable loss function. Our function penalizes embedding distances from being too large. used combination network. We also validate our evaluating sentiment analysis The proposed results semantic provide competitive on task standard dataset. These are shown statistically significant.