Can I change the 'n_pointers' dynamically?

[ShengleiH]

Hello, thanks to your awesome job. I want to adapt it to my work. Can I change the number of pointers dynamically, such as for the first sentence, there are two answers that could answer the first question and for the second sentence, there are three answers that could answer the second question?

[Kenta426]

The original implementation of Pointer Network can compute variable-length probability-distributions but the number of outputs must remain constant. In our implementation, we are assuming an "answer" is a range of words in the source texts, thus the number of outputs is always 2, pointing at the beginning and the end of "answers".

[ShengleiH]

@Kenta426 Thank you for your reply. But I really need to vary the number of outputs.
For example, in sentence:
"Washington D.C. is the capital city of America and Beijing is the capital city of China."
the question is "which cities are the capital cities?"
then there are two answers in this sentence: "Washington D.C." and "Beijing".
But for another sentence:
"Obama is the 44th president of the United States."
the question is "Who is the president?"
then there is only one answer in this sentence: "Obama".

As you can see, in different sentences, there are different number of answers. Thus the number of pointers should be variable.

And I thinks one approach to solve this problem is to set the number of pointers to the 'maximum number' just as the encoder does and then mask the maximum outputs to the real length of outputs to calculate the real loss.