README.md

Question Generation Integrating Knowledge Basis

General Instructions

Please run the following python code:

import stanza
stanza.download('en') 
import nltk
nltk.download('stopwords')
nltk.download('wordnet')
nltk.download('averaged_perceptron_tagger')

How to run QPGen

What you will need to provide

You will need to provide a JSON file containing your whole dataset (train+dev+test) with the following schema:

{
    "facts": List[String],
    "base_question": String,
    "target_question": String
}

Every string should be tokenized and lower cased. An example showing how to go about creating this file can be found in data_processing.data_generator.generate_repeat_q_squad_raw.

Data Processing Step

Next, you will need to run models.repeat_q in preprocessing mode, passing in argument the path to the JSON file mentioned above. This will create a vocabulary file and optionally an embedding matrix file for you.

Training

You can now train the model using models.repeat_q in training mode. Please refer to the arguments' descriptions for more information by running:

python -m models.repeat_q --help

Knowledge Graph API

Google Knowledge Graph

Please store you api key in a file ".gkg_api_key" located at the root directory

Instructions for the NQG model (Seq2Seq)

To pre-process SG DQG data, you'll need to run python -m spacy download en_core_web_sm prior to doing anything.

To run anything related to the NQG model, you'll want to use the script models/seq2seq.py.

Train

Command: seq2seq.py train

Description: Trains the model using data located at /data/processed/nqg. This directory shall contain two subdirectories "dev" and "train". The content of these directories shall follow the format used by the original NQG team: https://res.qyzhou.me/redistribute.zip, even though the original dataset can be any of your liking and the NER and POS features can be modified as well to use any convention/tool.

Options:

--vocab_size num : prune the vocabulary of the dataset to the required number of words "num". Optional; Default value: 20000

Generating data

Command: seq2seq.py generate_data

Description: Generates the necessary data from the raw SQuAD dataset to train the NQG model on. The SQuAD 1.1 data files shall be stored at /data/squad_dataset.

Make predictions

Command: Use translate.py from the NQG repository with any .pt file storing your trained model.

Make predictions (NQG+ SQuAD dev set)

Command: seq2seq.py beam_search

Description: Makes predictions for the SQuAD dev set (see data format from section train).

Options:

--model_path path : path to a .pt trained model file.

Instructions for SG-DQG

• Follow instructions 1 from: https://github.com/yanghoonkim/NQG_ASs2s
• Have the GloVe embedding .txt file in /models/pretrained/ and run models/NQG_ASs2s/data/process_embedding.py