Answer like span extraction about question_generation HOT 1 CLOSED

Thanks @s20ss.
This is described in this section.

I did this using text-to-text approach. There are two ways we can do this

1. Provide the context as input and ask the model to generate ans spans separated by sep token.
   So for this example Python is a programming language. Created by Guido Van Rossum and first released in 1991.
   ans could be Python, a programming language, Guido Van Rossum and 1991. So we can process input as follows

input text: Python is a programming language. Created by Guido Van Rossum and first released in 1991.

target text: Python <sep> a programming language <sep> Guido Van Rossum <sep> 1991 <sep>

But there's one problem with this approach. What if a particular answer span is repeated more than once in the context and depending on where it occurs it could be an answer or not, so how can we make the model to understand the position of answer spans.

So the second approach is (which is used in this project) to highlight the sentence which contains answer spans and ask the model to generate answers only for that sentence. In SQuAD we know the start index of of the answers, so here's how it works

1. split the input into sentences.
2. If a sentence contains ans spans highlight that sentence and prepare a training example ,and the target will be answer spans in that sentence separated by <sep>

So for above example, we will have 2 input examples as there are two sentences and both of them contain answer spans

so first
input text: <hl> Python is a programming language. <hl> Created by Guido Van Rossum and first released in 1991.
target text: Python <sep> a programming language <sep>

and second
input text: Python is a programming language. <hl> Created by Guido Van Rossum and first released in 1991. <hl>
target text: Guido Van Rossum <sep> 1991 <sep>

the assumption is that, this should force the model to understand the position of the spans. And also for QG, ans spans are highlighted within the context, so we need to know the position of the answer spans. With this approach at inference time we can highlight each sentence, extract ans from that sentence then highlight those answers and ask the model to generate the question.

There's another way to do this, which is using BERT like models. Basically we can model the task of answer extraction as unconditional span extraction. This is described in this paper . I decided to not use this as I wanted to keep everything uniform and as simple as possible.

Hope this helps. ;)

from question_generation.