Style-transfer and Paraphrase: Looking for a Sensible Semantic Similarity Metric

Datsets used in AAAI's paper Style-transfer and Paraphrase: Looking for a Sensible Semantic Similarity Metric by Ivan Yamshchikov, Viacheslav Shibaev, Nikolay Khlebnikov, Alexey Tikhonov.

Repo contains following datasets:

• Yelp paraphrases - contains paraphrases for Yelp dataset

• Yelp random - contains random pairs from Yelp dataset

• Paraphrase - contains paraphrases from Paraphrase dataset

• Paraphrase random - contains random pairs from Paraphrase dataset

• Paralex paraphrases - contains paraphrases from Paralex dataset

• Yelp random - contains random pairs from Paralex dataset

• Bible paraphrases - contains paraphrases from Bible dataset

• Bible random - contains random pairs from Bible dataset

• GYAFC formal paraphrases - contains paraphrases from formal part of GYAFC dataset

• GYAFC formal random - contains random pairs from formal part of GYAFC dataset

• GYAFC informal paraphrases - contains paraphrases from informal part of GYAFC dataset

• GYAFC informal random - contains random pairs from informal part of GYAFC dataset

• GYAFC rewrites paraphrases - contains paraphrases from rewrites part of GYAFC dataset

• GYAFC rewrites random - contains random pairs from rewrites part of GYAFC dataset

All csv-files have following fields:

• text_1 - given text
• text_2 - random/real paraphrase of text_1
• label_1 - score of the first labeller
• label_2 - score of the second labeller
• label_3 - score of the third labeller
• avg_score - average labellers score

All scores are in the range from 1 to 5, where 1 - "Not Similar At All" and 5 - "Highly Similar"