ScandiNLI - Natural Language Inference model for Scandinavian Languages

This model is a fine-tuned version of NbAiLab/nb-bert-large for Natural Language Inference in Danish, Norwegian Bokmål and Swedish.

We have released three models for Scandinavian NLI, of different sizes:

• alexandrainst/scandi-nli-large-v2
• alexandrainst/scandi-nli-large (this)
• alexandrainst/scandi-nli-base
• alexandrainst/scandi-nli-small

A demo of the large-v2 model can be found in this Hugging Face Space - check it out!

The performance and model size of each of them can be found in the Performance section below.

Quick start

You can use this model in your scripts as follows:

>>> from transformers import pipeline
>>> classifier = pipeline(
...     "zero-shot-classification",
...     model="alexandrainst/scandi-nli-large",
... )
>>> classifier(
...     "Mexicansk bokser advarer Messi - 'Du skal bede til gud, om at jeg ikke finder dig'",
...     candidate_labels=['sundhed', 'politik', 'sport', 'religion'],
...     hypothesis_template="Dette eksempel handler om {}",
... )
{'sequence': "Mexicansk bokser advarer Messi - 'Du skal bede til gud, om at jeg ikke finder dig'",
 'labels': ['sport', 'religion', 'politik', 'sundhed'],
 'scores': [0.6134647727012634,
  0.30309760570526123,
  0.05021871626377106,
  0.03321893885731697]}

Performance

We assess the models both on their aggregate Scandinavian performance, as well as their language-specific Danish, Swedish and Norwegian Bokmål performance.

In all cases, we report Matthew's Correlation Coefficient (MCC), macro-average F1-score as well as accuracy.

Scandinavian Evaluation

The Scandinavian scores are the average of the Danish, Swedish and Norwegian scores, which can be found in the sections below.

Danish Evaluation

We use a test split of the DanFEVER dataset to evaluate the Danish performance of the models.

The test split is generated using this gist.

Swedish Evaluation

We use the test split of the machine translated version of the MultiNLI dataset to evaluate the Swedish performance of the models.

We acknowledge that not evaluating on a gold standard dataset is not ideal, but unfortunately we are not aware of any NLI datasets in Swedish.

Norwegian Evaluation

We use the test split of the machine translated version of the MultiNLI dataset to evaluate the Norwegian performance of the models.

We acknowledge that not evaluating on a gold standard dataset is not ideal, but unfortunately we are not aware of any NLI datasets in Norwegian.

Training procedure

It has been fine-tuned on a dataset composed of DanFEVER as well as machine translated versions of MultiNLI and CommitmentBank into all three languages, and machine translated versions of FEVER and Adversarial NLI into Swedish.

The training split of DanFEVER is generated using this gist.

The three languages are sampled equally during training, and they're validated on validation splits of DanFEVER and machine translated versions of MultiNLI for Swedish and Norwegian Bokmål, sampled equally.

Check out the Github repository for the code used to train the ScandiNLI models, and the full training logs can be found in this Weights and Biases report.

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 2e-05
• train_batch_size: 2
• eval_batch_size: 2
• seed: 4242
• gradient_accumulation_steps: 16
• total_train_batch_size: 32
• optimizer: Adam with betas=(0.9, 0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 500
• max_steps: 50,000