johanneskruse / nlp_qa_qg_app

Question Answering and Question Generation streamlit app

App for Question Answering & Question Generation

Welcome 🏆

This is a repository for a Streamlit App that enables you to play around with state-of-the-art NLP models in an easy user friendly environment. There are two main task, namely; Question Answering and Question Generation.

The default repository with all models are only compatible with Mac and Linux systems as AllenNLP models and some features are currently only supported by these systems (recommended systems). To run on Windows go to windows_setup and follow the instructions.

Getting started

It's super easy, we promise!

Preprare Envirorment

To install necesarry dependencies simple run

pip install -r requirements.txt

This will intall the packages:

• streamlit
• gitpython
• nltk
• allennlp (1.0.0)
• allennlp_models (1.0.0)
• transformers (3.0.0)
• langdetect

The question_generation is a git repository that has been added used for Question Generation. The functions.py script will check that it exists and if not download it.

All credits to patil-suraj for making an awesome end-to QG framework that mimics 🤗 transformers pipeline for easy inference (https://github.com/patil-suraj/question_generation).

Cloning the question_generation yourself:

git clone https://github.com/patil-suraj/question_generation.git

Running streamlit

Once streamlit and the other dependicies are installed you can simple run the app:

streamlit run app.py

The app will now run in a localhost server.

COLAB notebooks

Question Answering and Question Generation

The final streamlit is essentially a combination of two seperate COLAB prototypes solving Question Answering and Question Generation respectively. They have been merged into one streamlit app with user interface. The seperate notesbook can be found as well.

Steamlit app

To play around with all of models require some storage (~4-5gb) and computation power. To cope with this a COLAB notebook has been developed, which enables you to run the streamlit app through COLAB.

Note you will need to make a small modification, as you need to add your own ngrok to generate the localhost server (follow COLAB instructions).

Models

The app is essential a wrapper for a number of models that have been put together with a friendly user interface. All models are pre-trained and have not been fine-tuned further and are simple implemented using a user friendly API pipeline format.

Question Answering

For the Question Answering module four pre-trained models have been implemented:

1. ELMo-BiDAF (Trained on SQuAD)
2. BiDAG (Trained on SQuAD)
3. DistilBERT (distilbert-base-cased-distilled-squad)
4. BERT (bert-large-uncased-whole-word-masking-finetuned-squad)
5. Finetuned BERT multilingual base model [cased] (mrm8488/bert-multi-cased-finetuned-xquadv1)

Where 1) and 2) are AllenNLP models and 3), 4), and 5) are HuggingFace models.

Question Generation

The Question Generation is a pre-trained model with three pipeline tasks:

1. question-generation: for single task question generation models
2. multitask-qa-qg: for multi-task qa,qg models
3. e2e-qg: for end-to-end question generation

With the option of using a model sizes "small" or a "base" (https://huggingface.co/models). All models are easily implemented but only the end-to-end question generation is included in this demo (both the small and base model).

Store and Run models from local machine

It is possible to store the models on your local machine, to do so run

python run_locally/download_model_local.py 

This will allow you to download the models one-by-one. The models are stored in models. To run the models using the locally stored models replace screens.py and functions.py with the repective file from run_locally. Note this will be without the AllenNLP models.