Add QuestionGenerator

[brandenchan]

This PR creates a QuestionGenerator class that takes a document as input and generates questions as output. Since the default models used only seem to generate about 3 questions per passage, input text is split into 50 word chunks with 10 word split overlap. This is somewhat inefficient since T5's max seq len is significantly longer. But this allows us to handle variable length documents.

As a bonus, this also allows the DocumentStore to be used as an iterator

TODO:

• Documentation
• Design implementation into pipelines

For future:

• Consider how to speed up QG
• Consider a QuestGen implementation

[lalitpagaria]

Oh nice addition.

BTW this one can also be integrated https://github.com/ramsrigouthamg/Questgen.ai, yes but in separate PR if you feel beneficial.

cc: @ramsrigouthamg

[julian-risch]

I left some minor comments. The code looks good to me. However, before merging, please add some test cases for the new QuestionGenerator node. You could reuse the tutorial code for that.

[julian-risch]

curr_id = self.ids_iterator.pop(0) could be used here so that self.ids_iterator = self.ids_iterator[1:]is not necessary. Just a thought. Not a request for a change.

[brandenchan]

Yup good idea!

[julian-risch]

I understand that having standard pipelines, such as QuestionGenerationPipeline, RetrieverQuestionGenerationPipeline, and QuestionAnswerGenerationPipeline, make it easier to run a pipeline because there is no need to define the individual pipeline components. However, with more and more standard pipelines, how do users know which standard pipeline fits there needs? We already have DocumentSearchPipeline, ExtractiveQAPipeline, GenerativeQAPipeline , FAQPipeline, SearchSummarizationPipeline, and TranslationWrapperPipeline. Maybe we need an overview of all these pipelines in the documentation? And some guidelines when to define a new standard pipeline?

[julian-risch]

All other tutorials are numbered so let's rename this file to Tutorial13_Question_Generation.py.

[lalitpagaria]

it would be nice to suggest user what type of model are supported like T5, GPT or we can tell text generation models only used here.

[julian-risch]

Agreed. Maybe this link could help here as a comment: https://huggingface.co/models?filter=question-generation

[julian-risch]

I understand this is a generic solution in the BaseDocumentStore. Maybe it makes sense to implement a get_all_document_ids() method in the individual document stores (although it's more effort)? Or at least keep this in mind for future improvements? Otherwise this call will load a lot of data.
In ElasticSearchDocumentStore, you could actually make use of the existing generator:

haystack/haystack/document_store/elasticsearch.py

Line 601 in 2a90471

yield document

[brandenchan]

Yes I agree there are more efficient ways of implementing the Document Store as an iterator. I think that is outside the scope of this PR but we should think of reimplementing it, especially if this becomes an idiomatic way of accessing docs in the DocumentStore

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB

[cvgoudar]

@brandenchan : For the tutorial 13 to be complete can the haystack and Elasticsearch setup be added as part of the Tutorial. I tried in the notebook just by adding the following and the entire notebook worked end to end.

!pip install grpcio-tools==1.34.1
!pip install git+https://github.com/deepset-ai/haystack.git

# In Colab / No Docker environments: Start Elasticsearch from source
! wget https://artifacts.elastic.co/downloads/elasticsearch/elasticsearch-7.9.2-linux-x86_64.tar.gz -q
! tar -xzf elasticsearch-7.9.2-linux-x86_64.tar.gz
! chown -R daemon:daemon elasticsearch-7.9.2

import os
from subprocess import Popen, PIPE, STDOUT
es_server = Popen(['elasticsearch-7.9.2/bin/elasticsearch'],
                   stdout=PIPE, stderr=STDOUT,
                   preexec_fn=lambda: os.setuid(1)  # as daemon
                  )
# wait until ES has started
! sleep 30

[tholor]

Hey @cvgoudar , that's right. The current ES init uses docker and is therefore not working on colab. Do you want to create a PR?