REGULAR: A Framework for Relation-Guided Multi-Span Question Generation

Install Packages

• run

    pip install -r requirements.txt
  

Download Models and Data

• You may need to download Llama-1B-Instruct (https://huggingface.co/meta-llama/Llama-3.2-1B-Instruct) and UnifiedQA-T5-large (https://huggingface.co/allenai/unifiedqa-t5-large) from Huggingface and put them into the MODEL directory before starting constructing the MSQA dataset. The fine-tuned checkpoint will be available in the future update.

• You should also download data.zip from https://drive.google.com/file/d/1VS9vVGWE4AYDkgH0XhA5bZ29yNS1vgsI/view?usp=sharing, which contains both REGULAR datasets and other data for constructing MSQA dataset.

Introduction to this Repository

• This repository contains the code and data of the papre REGULAR: A Framework for Relation-Guided Multi-Span Question Generation. The contents of each subfolder or zipfile are shown below:

  |-DATA.zip            # You should unzip this file first                
    |- corpus           # Open-Source corpus for construct MSQA dataset
    |- MultiSpanQA_data # Dataset for training QA model used in step 4
    |- QG_train_data    # Dataset for training QG model used in step 3 
    |- REGULAR_DATASET  # Synthetic Dataset used in our experiment
  |- MODEL      # Pre-trained Models for construct MSQA dataset    
  |- SCRIPT     # Bash scripts
  |- SOURCE     # Code for step 1-5
    |- 1_convert_passage_to_graph   # Code for step 1  
    |- 2_extract_answers            # Code for step 2
    |- 3_generate_question          # Code for step 3
    |- 4_select_optimal_question    # Code for step 4
    |- 5_construct_msqa_dataset     # Code for step 5
  |- TMP        # Directory for temporary outputs

Enviroments

• python: 3.9.13
• CUDA Version: 11.7
• GPU: Tesla V100 32GB

Constructing MSQA Dataset

• run each bash file in the SCRIPTS directory. Each file in the in the SCRIPTS directory is named with number. Each bash file is named with a number at the beginning, which represents the order in which it should be run. Therefore, you only need to run the files starting with 1, 2, 3_1, 3_2, 4_1, 4_2, and 5 in sequence, like:

  # After clone this repository
  cd REGULAR
  sh SCRIPTS\1_convert_passage_to_graph.sh
  

• After running these bash files, you are excepted to obtain the final output directory in TMP/REGULAR_dataset, which contains train file and validation file for fine-tuning QA models.

Contact Us

• FEEL FREE to contact us if you have any questions or encounter issues while using the repository.

20251220 Update

• This article has been accepted by AACL-IJCNLP 2025 recently.
• We have uploaded the camera-ready version of the paper REGULAR.pdf.