nativqa-framework 0.1.1

Framework for collecting multilingual culturally-aligned natural queries.

NativQA Framework: A Framework for Collecting Multilingual Culturally-Aligned Natural Queries

NativQA Framework is a powerful open-source toolkit designed to effortlessly build large-scale, culturally and regionally aligned question-answering (QA) datasets in native languages. By leveraging user-defined seed queries and real-time search engine results, NativQA captures location-specific, everyday information to generate natural, multilingual QA data. Ideal for evaluating and fine-tuning large language models (LLMs), the framework bridges the gap in region-specific QA resources, empowering inclusive AI development across diverse linguistic and cultural landscapes.

🌍 Key Features

Developing the NativQA Framework is an ongoing effort, and the framework will continue to grow and improve over time. Currently, it offers the following features:

• Supports Google, Yahoo, and Bing search engines for collecting "People also ask" questions and answers.
• Accepts seed queries in CSV or TSV format.
• Open-source and community-driven.
• Supports image collection for visual question answering tasks.
• Multilingual support for diverse language coverage.

Here is a quick overview of NativQA Framework:

Quick Start

1. Clone the repository git clone https://gitlab.com/nativqa/nativqa-framework.git
2. Navigate to the nativqa-framework directory: cd nativqa-framework
3. Install the requirements: pip install -r requirements.txt
4. Put your SerpAPI api key in envs/api_key.env
5. Run the program!

For example, to run the program using example seed queries:

python -m nativqa --engine google --search_type text --input_file data/test_query.csv --country_code qa --location "Doha, Qatar" --env envs/api_key.env --n_iter 3

which uses a sample seed query file

Parameters

• --engine Search engine to use for collect QA. Currently supports only Google, Bing, and Yahoo.
• --search_type Type of search either text, image, or video. [Currently supports only Google and Bing for image search and Google for video/audio search.]
• --input_file seed query file should be CSV/TSV
• --country_code Parameter defines the country to use for the Google search. The country code supported by Google.
• --location Parameter defines from where you want the search to originate.
• --multiple_countries Parameter defines one or multiple countries to limit the search to. For example, countryQA|countryBD
• --env API key file.
• --output output directory location. Defaults ./results/
• --n_iter number of search iteration to perform.

Outputs Format

• The framework will create a directory using input filename under the given output directory.
  • dataset directory has the final QA pair file with the same name of input file.
  • iteration_{n} directory contains output folder of each iteration and input queries of each iteration.
    • output contains the output of each iteration consisting of related_search.tsv, original_response.json, summary.jsonl, related_question.json
  • completed_queries.txt List of completed searched queries

Query collection

Template based Query Generation

python scripts/template2seeds.py --template_file templates/arabic_template.csv --output_file templates/test.csv --location "قطر"

LLM based Query Generation

To add

QA Validation

Domain Reliability Check (DRC)

Manually verified domains list file are located in domain/annotated_domains.csv.

To verify the answer source reliability: (the input file will be dataset file generated from nativqa framework.)

python scripts/check_domain_reliability.py --input_file <dataset_directory>/input_filename.csv --output_file <output_directory>/output_filename.csv

Note that we only support csv/tsv file for domain reliability task. We aim to extend other file types in future.

LLM based Annotation

python scripts/GPT_4o_labeling.py --input_file results/text/arabic_qa/dataset.json --env_path envs/gpt4-api-key.env --output_dir results/text/arabic_qa/GPT4o_labeling/ --location "Qatar"

Licence

The NativQA Framework is licensed under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International (CC BY-NC-SA 4.0).

You are free to share and adapt the framework for non-commercial purposes, provided that:

• You give appropriate credit.
• You indicate if changes were made.
• You distribute your contributions under the same license.

Citation

Please cite our papers when referring to this framework:

@article{Alam2025nativqa,
  title={NativQA Framework: A Framework for Collecting Multilingual Culturally-Aligned Natural Queries},
  author={Alam, Firoj and Hasan, Md Arid and Laskar, Sahinur Rahman and Kutlu, Mucahid and Chowdhury, Shammur Absar},
  journal={arXiv},
  year={2024}
}