QA Tasks that generate appropriate answers to a given question have seen a lot of performance gains due to recent deep learning technologies. The well-known SQuAD is one such task.
However, as models are trained for each task, there is a problem that only the task can respond. In particular, it is considered difficult to combine multiple models into one because the desired shape and question answer format are slightly different for each task.
UnifiedQA, unveiled by AllenAI, is to solve this problem, taking an approach to learning a single model covering 20 QA datasets including SQuAD, NarrativeQA, ARC-challenge, and more. In particular, each QA Task has a slightly different purpose when context is given. For example, if it is necessary to extract facts, if you need a summary of the situation, if you choose one from several options, if you decide whether or not. UnifiedQA has the advantage of being able to respond to these four purposes with one model. The following is a related paper.
such as extractive span selection, multiple choice, etc. This has led to
format-specialized models, and even to an implicit division in the QA
community. We argue that such boundaries are artificial and perhaps
unnecessary, gi…
The code and models trained on T5 and BART are also published on the following github link:
