A Qualitative Comparison of CoQA, SQuAD 2.0 and QuAC
share
In this work, we compare three datasets which build on the paradigm defined in SQuAD for question answering: SQuAD 2.0, QuAC, and CoQA. We compare these three datasets along several of their new features: (1) unanswerable questions, (2) multi-turn interactions, and (3) abstractive answers. We show that the datasets provide complementary coverage of the first two aspects, but weak coverage of the third. Because of the datasets' structural similarity, a single extractive model can be easily adapted to any of the datasets. We show that this model can improve baseline results on both SQuAD 2.0 and CoQA. Despite the core similarity between the datasets, models trained on one dataset are ineffective on another dataset, but we do find moderate performance improvement through pretraining. To encourage evaluation of methods on all of these datasets, we release code for conversion between them.
READ FULL TEXT
