katanemo / bart-large-mnli

bart-large-mnli

This is the checkpoint for bart-large after being trained on the MultiNLI (MNLI) dataset.

Additional information about this model:

• The bart-large model page
• BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension
• BART fairseq implementation

NLI-based Zero Shot Text Classification

Yin et al. proposed a method for using pre-trained NLI models as a ready-made zero-shot sequence classifiers. The method works by posing the sequence to be classified as the NLI premise and to construct a hypothesis from each candidate label. For example, if we want to evaluate whether a sequence belongs to the class "politics", we could construct a hypothesis of This text is about politics. . The probabilities for entailment and contradiction are then converted to label probabilities.

This method is surprisingly effective in many cases, particularly when used with larger pre-trained models like BART and Roberta. See this blog post for a more expansive introduction to this and other zero shot methods, and see the code snippets below for examples of using this model for zero-shot classification both with Hugging Face's built-in pipeline and with native Transformers/PyTorch code.

With the zero-shot classification pipeline

The model can be loaded with the zero-shot-classification pipeline like so:

from transformers import pipeline
classifier = pipeline("zero-shot-classification",
                      model="facebook/bart-large-mnli")

You can then use this pipeline to classify sequences into any of the class names you specify.

sequence_to_classify = "one day I will see the world"
candidate_labels = ['travel', 'cooking', 'dancing']
classifier(sequence_to_classify, candidate_labels)
#{'labels': ['travel', 'dancing', 'cooking'],
# 'scores': [0.9938651323318481, 0.0032737774308770895, 0.002861034357920289],
# 'sequence': 'one day I will see the world'}

If more than one candidate label can be correct, pass multi_label=True to calculate each class independently:

candidate_labels = ['travel', 'cooking', 'dancing', 'exploration']
classifier(sequence_to_classify, candidate_labels, multi_label=True)
#{'labels': ['travel', 'exploration', 'dancing', 'cooking'],
# 'scores': [0.9945111274719238,
#  0.9383890628814697,
#  0.0057061901316046715,
#  0.0018193122232332826],
# 'sequence': 'one day I will see the world'}

With manual PyTorch

# pose sequence as a NLI premise and label as a hypothesis
from transformers import AutoModelForSequenceClassification, AutoTokenizer
nli_model = AutoModelForSequenceClassification.from_pretrained('facebook/bart-large-mnli')
tokenizer = AutoTokenizer.from_pretrained('facebook/bart-large-mnli')

premise = sequence
hypothesis = f'This example is {label}.'

# run through model pre-trained on MNLI
x = tokenizer.encode(premise, hypothesis, return_tensors='pt',
                     truncation_strategy='only_first')
logits = nli_model(x.to(device))[0]

# we throw away "neutral" (dim 1) and take the probability of
# "entailment" (2) as the probability of the label being true 
entail_contradiction_logits = logits[:,[0,2]]
probs = entail_contradiction_logits.softmax(dim=1)
prob_label_is_true = probs[:,1]