Question Answering vs Machine Reading Comprehension (QA vs MRC)
I want to say two words about this theme because at first, I had some difficulties to understand that these two tasks are not the same thing. In my rescue came the paper A survey on Machine Reading Comprehension: Tasks, Evaluation Metrics and Benchmark Dataset, which has the aim to resume and explain the Machine Reading Comphension’s situation at the end of 2020. It is a good point to start in the fields of QA and MRC: it talks about everything that is needed to start with the right foot.
Anyway, here I want to talk about only one thing: how Question and Answering and Machine Reading Comprehension are linked. Let’s show an amazing image present in the paper.
[Note: I won’t talk about the Computer Vision part present in the image.]
As you can see QA and MRC are two different things but, sometimes, they converge. Just to resume the image:
- MRC tasks are a group of tasks which to solve them you need the ability to read some text
- QA tasks are a group of tasks which to solve them you have to answer a question
QA tasks can be solved in different ways, as we saw before (Retriever-Generator, Retriever-Extractor, Generator, and other). Sometimes a QA task is solved with a MRC technique, this is the reason for the convergence.
Retriever-Extractor and Retriever-Generator, for example, are solutions that fall in this case because they required the ability to read a text. Instead, the solution that is made up only of the Generator solve a QA problem without requiring this skill (so without using a MRC solution).
Just to be clear: of course, not all the MRC solutions are made up to solve QA problems. MRC is the ability to comprehend a text by reading it. This ability can be tested with different tasks, which can also be funded in other NLP problems. If while I am solving a QA task I find a MRC problem, I of course use the MRC solutions that fit well to solve it.
Another interesting thing that I found interesting in that article is the interesting classification of MRC tasks showed below.
Usually is not used this classification, it is a new proposal of the authors (remember, it was published at the end of 2020) that I find really interesting.
If you want to know more, or you didn’t understand well what I said before (I could probably say it better), I recommend to read the paper, you will not regret it.
Here I let you also some articles:
And some papers:
Enjoy ☺.
