Zero-Shot Classification
Transformers
PyTorch
Safetensors
bert
text-classification
Inference Endpoints
scandi-nli-large / README.md
saattrupdan's picture
Update README.md
47523da
|
raw
history blame
9.77 kB
metadata
pipeline_tag: zero-shot-classification
language:
  - da
  - 'no'
  - nb
  - sv
license: mit
datasets:
  - strombergnlp/danfever
  - KBLab/overlim
  - MoritzLaurer/multilingual-NLI-26lang-2mil7
widget:
  - example_title: Danish
    text: >-
      Mexicansk bokser advarer Messi - 'Du skal bede til gud, om at jeg ikke
      finder dig'
    candidate_labels: sundhed, politik, sport, religion
  - example_title: Norwegian
    text: >-
      Regjeringen i Russland hevder Norge fører en politikk som vil føre til
      opptrapping i Arktis og «den endelige ødeleggelsen av russisk-norske
      relasjoner».
    candidate_labels: helse, politikk, sport, religion
  - example_title: Swedish
    text:  luras kroppens immunförsvar att bota cancer
    candidate_labels: hälsa, politik, sport, religion
inference:
  parameters:
    hypothesis_template: Dette eksempel handler om {}

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:

A demo of the large 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.

Model MCC Macro-F1 Accuracy Number of Parameters
alexandrainst/scandi-nli-large (this) 73.70% 74.44% 83.91% 354M
MoritzLaurer/mDeBERTa-v3-base-xnli-multilingual-nli-2mil7 69.01% 71.99% 80.66% 279M
alexandrainst/scandi-nli-base 67.42% 71.54% 80.09% 178M
joeddav/xlm-roberta-large-xnli 64.17% 70.80% 77.29% 560M
MoritzLaurer/mDeBERTa-v3-base-mnli-xnli 63.94% 70.41% 77.23% 279M
NbAiLab/nb-bert-base-mnli 61.71% 68.36% 76.08% 178M
alexandrainst/scandi-nli-small 56.02% 65.30% 73.56% 22M

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.

Model MCC Macro-F1 Accuracy Number of Parameters
alexandrainst/scandi-nli-large (this) 73.80% 58.41% 86.98% 354M
MoritzLaurer/mDeBERTa-v3-base-xnli-multilingual-nli-2mil7 68.37% 57.10% 83.25% 279M
alexandrainst/scandi-nli-base 62.44% 55.00% 80.42% 178M
NbAiLab/nb-bert-base-mnli 56.92% 53.25% 76.39% 178M
MoritzLaurer/mDeBERTa-v3-base-mnli-xnli 52.79% 52.00% 72.35% 279M
joeddav/xlm-roberta-large-xnli 49.18% 50.31% 69.73% 560M
alexandrainst/scandi-nli-small 47.28% 48.88% 73.46% 22M

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.

Model MCC Macro-F1 Accuracy Number of Parameters
alexandrainst/scandi-nli-large (this) 76.69% 84.47% 84.38% 354M
joeddav/xlm-roberta-large-xnli 75.35% 83.42% 83.55% 560M
MoritzLaurer/mDeBERTa-v3-base-mnli-xnli 73.84% 82.46% 82.58% 279M
MoritzLaurer/mDeBERTa-v3-base-xnli-multilingual-nli-2mil7 73.32% 82.15% 82.08% 279M
alexandrainst/scandi-nli-base 72.29% 81.37% 81.51% 178M
NbAiLab/nb-bert-base-mnli 64.69% 76.40% 76.47% 178M
alexandrainst/scandi-nli-small 62.35% 74.79% 74.93% 22M

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.

Model MCC Macro-F1 Accuracy Number of Parameters
alexandrainst/scandi-nli-large (this) 70.61% 80.43% 80.36% 354M
joeddav/xlm-roberta-large-xnli 67.99% 78.68% 78.60% 560M
alexandrainst/scandi-nli-base 67.53% 78.24% 78.33% 178M
MoritzLaurer/mDeBERTa-v3-base-xnli-multilingual-nli-2mil7 65.33% 76.73% 76.65% 279M
MoritzLaurer/mDeBERTa-v3-base-mnli-xnli 65.18% 76.76% 76.77% 279M
NbAiLab/nb-bert-base-mnli 63.51% 75.42% 75.39% 178M
alexandrainst/scandi-nli-small 58.42% 72.22% 72.30% 22M

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