# SuperGLUE [A Stickier Benchmark for General-Purpose Language Understanding Systems](https://w4ngatang.github.io/static/papers/superglue.pdf) GLUE原先的任務只包含sentence pair的分類,SuperGLUE增加了指代消集與QA # SuperGLUE任務: | 語料 | Train | Dev | Test | 任務 | 資料來源 | | ------- | ----- | ---- | ---- | ----------- | --- | | BoolQ | 9427 | 3270 | 3245 | QA |google queries, Wikipedia | | CB | 250 | 57 | 250 | NLI |various | | COPA | 400 | 100 | 500 | QA |blogs, photographt encyclopedia | | MultiRC | 5100 | 953 | 1800 | QA |various | | ReCoRD | 101k | 10k | 10k | QA |news(CNN, DailyMail) | | RTE | 2500 | 278 | 300 | NLI |news, Wikipedia | | WiC | 6000 | 638 | 1400 | WSD |WordNet, VerbNet, Wiktionary | | WSC | 554 | 104 | 146 | coreference |fiction books | *WSD(word sense disambiguation 單詞歧義消除) ## [BoolQ(Boolean Question)](https://arxiv.org/abs/1905.10044) input : Context + Question output : Answer(Yes/No) ## [CB(CommitmentBank)](https://semanticsarchive.net/Archive/Tg3ZGI2M/Marneffe.pdf) 判斷說話的人對某一事件的承認程度,"真實", "不真實", "不確定" input : 兩句的context+一句target組成 output : -3~3的分數 ## [COPA(Choice of Plausible Alternatives)](https://ict.usc.edu/pubs/Choice%20of%20Plausible%20Alternatives-%20An%20Evaluation%20of%20Commonsense%20Causal%20Reasoning.pdf) input : Premise, Question, Alternative 1, Alternative 2 output : Correct Alternative ## [MultiRC(Multi-Sentence Reading Comprehension)](https://cogcomp.seas.upenn.edu/page/publication_view/833) 多重選擇題 input : context, question, alternative output : answer ## [ReCoRD(Reading Comprehension with Commonsense Reasoning Dataset)](https://arxiv.org/abs/1810.12885) input : context, query(query會挖空格) output : 預測空格要填入什麼 ## [RTE(Recognizing Textual Entailment)](https://openreview.net/pdf?id=rJ4km2R5t7) GLUE的其中一項困難任務,所以被保留 RTE1 + RTE2 + RTE3 + RTE5 二分類任務 entailment/not_entailment 受益於transfer learning ## [WiC(Word-in-Context)](https://arxiv.org/abs/1808.09121) input : 兩段context + 兩句中重複出現的詞 output : 是否意義相同 ## [WSC(Winograd Schema Challenge)](https://cs.nyu.edu/faculty/davise/papers/WinogradSchemas/WS.html) 指代消集問題