1
$\begingroup$

We are using Google BERT for question and answering. We are using vanialla bert-base-uncased as well as squad trained checkpoints.

The answers from BERT are very short and crisp. For example, if we ask describe a chatbot, then it will simply return, takes input from user and replies ... Though the answer actually is complete paragraph.

  • Will training BERT on long paragraphs can solve this problem? If yes, any idea from where we can get such a QnA dataset, as it is not possible to create a huge QnA dataset manually.
  • Is there any other tweak which can be done at some BERT layer, so that it starts understanding a long answer?
  • Or is there any other framework or system already have solved this kind of problem, by maybe integrating with some other neural network technique, or by using a pipeline of multiple components?
$\endgroup$
1
$\begingroup$

You can check quora for your QnA dataset.

As far as long answers are concerned it may not be because of BERT as SQuADhas very crisp and short answers for most of its questions.

BERT already does understand long term dependencies in your data(one of reasons for using deeply bidirectional tranformer).

| improve this answer | |
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.