NetherlandsForensicInstitute/squad-nl-v2.0

SQuAD-NL v2.0 for Sentence Transformers

The SQuAD-NL v2.0 dataset (on Hugging Face: GroNLP/squad-nl-v2.0), modified for use in Sentence Transformers as a dataset of type "Pair with Similarity Score".

Score

We added an extra column score to the original dataset. The value of score is 1.0 if the question has an answer in the context (no matter where), and 0.0 if there are no answers in the context. The allows the evaluation of embedding models that aim to pair queries and document fragments. Please note that is a very hard task for embedding models, because SQuAD v2.0 was specifically designed to contain unanswerable questions adversarially written to look similar to answerable ones. Expect your models to perform poorly.

Translations

SQuAD-NL is translated from the original SQuAD and XQuAD English-language datasets. From the SQuAD-NL v2.0 Readme:

Split	Source	Procedure	English	Dutch
train	SQuAD-train-v2.0	Google Translate	130,319	130,319
dev	SQuAD-dev-v2.0 \ XQuAD	Google Translate	10,174	10,174
test	SQuAD-dev-v2.0 & XQuAD	Google Translate + Human	1,699	1,699

For testing Dutch sentence embedding models it is therefore recommended to only use the test split. Also it would be advisable to not train your model on the other splits, because you would train answering this specific style of questions into the model.

Example code using Sentence Transformers

import pprint

from datasets import load_dataset
from sentence_transformers import SentenceTransformer
from sentence_transformers.evaluation import EmbeddingSimilarityEvaluator, SimilarityFunction


eval_dataset = load_dataset('NetherlandsForensicInstitute/squad-nl-v2.0', split='test')

evaluator = EmbeddingSimilarityEvaluator(
     sentences1=eval_dataset['question'],
     sentences2=eval_dataset['context'],
     scores=eval_dataset['score'],
     main_similarity=SimilarityFunction.COSINE,
     name="squad_nl_v2.0_test",
 )

model = SentenceTransformer('NetherlandsForensicInstitute/robbert-2022-dutch-sentence-transformers')

results = evaluator(model)
pprint.pprint(results)

Original dataset

SQuAD-NL is a derivative of the SQuAD and XQuAD datasets, and their original CC BY-SA 4.0 licenses apply.

Code used to generate this dataset

code

import json

import requests
from datasets import Dataset, DatasetDict


def squad(url):
  response = requests.get(url)

  rows = json.loads(response.text)['data']

  for row in rows:
      yield {'question': row['question'],
             'context': row['context'],
             'score': 1.0 if row['answers']['text'] else 0.,
             'id': row['id'],
             'title': row['title'],
             'answers': row['answers']}

if __name__ == '__main__':
  url = 'https://github.com/wietsedv/NLP-NL/raw/refs/tags/squad-nl-v1.0/SQuAD-NL/nl/{split}-v2.0.json'

  dataset = DatasetDict({
      split: Dataset.from_generator(squad, gen_kwargs={'url': url.format(split=split)})
      for split in ('train', 'dev', 'test')
  })

  dataset.push_to_hub('NetherlandsForensicInstitute/squad-nl-v2.0')