FarsTail: A Persian Natural Language Inference Dataset

Natural Language Inference (NLI), also called Textual Entailment, is an important task in NLP with the goal of determining the inference relationship between a premise p and a hypothesis h. It is a three-class problem where each pair (p, h) is assigned to one of these classes: "ENTAILMENT" if the hypothesis can be inferred from the premise, "CONTRADICTION" if the hypothesis contradicts the premise, and "NEUTRAL" if none of the above holds.
There are large datasets such as SNLI, MNLI, and SciTail for NLI in English, but there are few datasets for poor-data languages like Persian.
Persian (Farsi) language is a pluricentric language spoken by around 110 million people in countries like Iran, Afghanistan, and Tajikistan. Here, we present the first relatively large-scale Persian dataset for NLI task, called FarsTail. A total of 10,367 samples are generated from a collection of 3,539 multiple-choice questions. The train, validation, and test portions include 7,266, 1,537, and 1,564 instances, respectively. Please refer to the manuscript for more details.

Reading data

To read the raw data in Persian alphabet, use the following code:

train_data = pd.read_csv('data/Train-word.csv', sep='\t')
val_data = pd.read_csv('data/Val-word.csv', sep='\t')
test_data = pd.read_csv('data/Test-word.csv', sep='\t')

The train_data and val_data have three columns, premise, hypothesis, and label. The test_data has two more columns denoted as hard(hypothesis) and hard(overlap) which indicate whether or not each sample belongs to the hard subset based on the hypothesis-only and overlap-based biased models, respectively.

Non-Persian researchers can use the following code to read the indexed data:

with np.load('data/Indexed-FarsTail.npz', allow_pickle=True) as f:
    train_ind, val_ind, test_ind, dictionary = f['train_ind'], f['val_ind'], f['test_ind'], f['dictionary'].item()

The train_ind and val_ind are numpy arrays with the shape of (n, 3) where n is the number of samples in each set. Each entry in these arrays includes the tokenized, indexed version of the premise and hypothesis along with the respective label for one instance. The entries of test_ind variable have two more elements corresponding to the hard(hypothesis) and hard(overlap) columns, respectively. The dictionary variable maps the indexes to tokens.

Results

Here is test accuracies obtained by training some models on FarsTail training set. Please refer to the manuscript for more results.

Model	Test Accuracy	Hypothesis-only (Easy)	Hypothesis-only (Hard)	Overlap-based (Easy)	Overlap-based (Hard)
DecompAtt (word2vec)	0.6662	0.7341	0.5823	0.7633	0.5404
HBMP (word2vec)	0.6604	0.7618	0.5350	0.7565	0.5360
ESIM (fastText)	0.7116	0.7931	0.6109	0.8120	0.5815
mBERT	0.8338	0.8763	0.7811	0.8981	0.7504

Reference

@article{amirkhani2020farstail,
  title={FarsTail: A Persian Natural Language Inference Dataset},
  author={Hossein Amirkhani, Mohammad Azari Jafari, Azadeh Amirak, Zohreh Pourjafari, Soroush Faridan Jahromi, and Zeinab Kouhkan},
  journal={arXiv preprint arXiv:2009.08820},
  year={2020}
}