README.md

---
language:
  - en
  - vi
pretty_name: VQAv2 in Vietnamese
source-datasets:
  - VQAv2
tags:
  - VQAv2-vi
  - VQA
license: unknown
task_categories:
  - visual-question-answering
task_ids:
  - visual-question-answering
---
# VQAv2 in Vietnamese

This is Google-translated version of [VQAv2](https://visualqa.org/) in Vietnamese. The process of building Vietnamese version as follows:
- In `en/` folder,
  - Download `v2_OpenEnded_mscoco_train2014_questions.json` and `v2_mscoco_train2014_annotations.json` from [VQAv2](https://visualqa.org/).
  - Remove key `answers` of key `annotations` from `v2_mscoco_train2014_annotations.json`. I shall use key `multiple_choice_answer` of key `annotations` only. Let call the new file `v2_OpenEnded_mscoco_train2014_answers.json`
  - By using [set data structure](https://docs.python.org/3/tutorial/datastructures.html#sets), I generate `question_list.txt` and `answer_list.txt` of unique text. There are 152050 unique questions and 22531 unique answers from 443757 image-question-answer triplets.
- In `vi/` folder,
  - By translating two `en/.txt` files, I generate `answer_list.jsonl` and `question_list.jsonl`. In each of entry of each file, the key is the original english text, the value is the translated text in vietnamese.

To load Vietnamese version in your code, you need original English version. Then just use English text as key to retrieve Vietnamese value from `answer_list.jsonl` and `question_list`. I provide both English and Vietnamese version.

Please refer to [this code](https://github.com/dinhanhx/velvet/blob/main/scripts/apply_translate_vqav2.py) to apply translation.