QA Datasets. The NLP community has been striving to propose Question Answering (QA) datasets that fall into three categories: reading comprehension QA, cloze-style QA and span-based QA, all of which are studied enthusiastically. Reading comprehension QA requires the model to pick an answer from an available pool of answers after comprehending an evidence passage, similar to multiple choice questions. Important and relevant information must be learned attentively and accurately in order to predict the correct answer. Following is a list of datasets proposed with this format. MCTest is an open-domain dataset comprising short fictional stories [24]. RACE is a large dataset compiled from English assessments for 12-18 years old students [13]. TQA gives passages from middle school science lessons and textbooks [12]. SciQ gives passages from science exams collected via crowdsourcing [31]. DREAM gives multiparty dialogue passages from English-as-a-foreign-language exams [27]. The second is for cloze-style QA, for which the model fills in the blanks that obliterate certain contents in sentences describing the evidence passages. This task is challenging because it requires both context understanding surrounding the blank and a general but comprehensive understanding toward the evidence passage, since normally the prediction happens within a summary of the passage. Such style is important because it is also a popular way to test people’s English skills and reading comprehension in a passage. CNN/Daily Mail targets on entities in bullet points summarizing articles from CNN and Daily News [7]. Children’s Book Test focuses on named entities, nouns, verbs, and prepositions in passages from children’s books [8]. BookTest is similar to Children’s Book Test but 60 times larger [1]. Who-did-What gives description sentences and evidence passages extracted from news articles in English Gigaword Corpus [19]. Finally, span-based QA is a task in which the model finds the answer contents as spans in the evidence passages.This task is the hardest because giving an answer span resembles humans’ way of answering questions. It also challenges the level of understanding of the evidence passages the greatest. bAbI aims to reinforce learning on event types and infer a sequence of event descriptions [32]. WikiQA [33] and SQuAD [22] use Wikipedia, whereas NewsQA [28] use CNN articles as evidence passages. XX XXXXX gives questions involving zero to multiple answer contents from web documents [18]. Trivi...
