The benchmarks section lists all benchmarks using a given dataset or any of
its variants. We use variants to distinguish between results evaluated on
slightly different versions of the same dataset. For example, ImageNet 32⨉32
and ImageNet 64⨉64 are variants of the ImageNet dataset.
iSarcasm is a dataset of tweets, each labelled as either sarcastic or non_sarcastic. Each sarcastic tweet is further labelled for one of the following types of ironic speech:
sarcasm: tweets that contradict the state of affairs and are critical towards an addressee;
irony: tweets that contradict the state of affairs but are not obviously critical towards an addressee;
satire: tweets that appear to support an addressee, but contain underlying disagreement and mocking;
understatement: tweets that undermine the importance of the state of affairs they refer to;
overstatement: tweets that describe the state of affairs in obviously exaggerated terms;
rhetorical question: tweets that include a question whose invited inference (implicature) is obviously contradicting the state of affairs.
For each sarastic tweet, there's also:
an explanation, in English sentences, as to why it is sarcastic, and
a rephrase that conveys the same meaning non-sarcastically. Both have been provided by the author of the tweet.
iSarcasm contains 4,484 tweets, out of which 777 are labelled as sarcastic and 3,707 as non-sarcastic. You'll find two files, isarcasm_train.csv and isarcasm_test.csv, each containing 80% and 20% of the examples chosen at random, respectively. Each line in a file has the format tweet_id,sarcasm_label,sarcasm_type, where sarcasm_type are only defined for sarcastic tweets, as specified above.