language:
- da
- 'no'
- nb
- sv
license: apache-2.0
datasets:
- strombergnlp/danfever
- KBLab/overlim
- MoritzLaurer/multilingual-NLI-26lang-2mil7
pipeline_tag: zero-shot-classification
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: Så luras kroppens immunförsvar att bota cancer
candidate_labels: hälsa, politik, sport, religion
inference:
parameters:
hypothesis_template: Dette eksempel handler om {}
base_model: NbAiLab/nb-bert-base
ScandiNLI - Natural Language Inference model for Scandinavian Languages
This model is a fine-tuned version of NbAiLab/nb-bert-base 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
- alexandrainst/scandi-nli-base (this)
- alexandrainst/scandi-nli-small
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-base",
... )
>>> 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', 'sundhed', 'politik'],
'scores': [0.724335789680481,
0.1176532730460167,
0.08848614990711212,
0.06952482461929321]}
Performance
We evaluate the models in Danish, Swedish and Norwegian Bokmål separately.
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 |
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 (this) |
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 |
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 (this) |
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 |
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 (this) |
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 |
70.61% | 80.43% | 80.36% | 354M |
joeddav/xlm-roberta-large-xnli |
67.99% | 78.68% | 78.60% | 560M |
alexandrainst/scandi-nli-base (this) |
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: 8
- eval_batch_size: 8
- seed: 4242
- gradient_accumulation_steps: 2
- 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