README.md
3.7 KB · 82 lines · markdown Raw
1 ---
2 license: mit
3 thumbnail: https://huggingface.co/front/thumbnails/facebook.png
4 pipeline_tag: zero-shot-classification
5 datasets:
6 - multi_nli
7 ---
8
9 # bart-large-mnli
10
11 This is the checkpoint for [bart-large](https://huggingface.co/facebook/bart-large) after being trained on the [MultiNLI (MNLI)](https://huggingface.co/datasets/multi_nli) dataset.
12
13 Additional information about this model:
14 - The [bart-large](https://huggingface.co/facebook/bart-large) model page
15 - [BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension
16 ](https://arxiv.org/abs/1910.13461)
17 - [BART fairseq implementation](https://github.com/pytorch/fairseq/tree/master/fairseq/models/bart)
18
19 ## NLI-based Zero Shot Text Classification
20
21 [Yin et al.](https://arxiv.org/abs/1909.00161) 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.
22
23 This method is surprisingly effective in many cases, particularly when used with larger pre-trained models like BART and Roberta. See [this blog post](https://joeddav.github.io/blog/2020/05/29/ZSL.html) 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.
24
25 #### With the zero-shot classification pipeline
26
27 The model can be loaded with the `zero-shot-classification` pipeline like so:
28
29 ```python
30 from transformers import pipeline
31 classifier = pipeline("zero-shot-classification",
32 model="facebook/bart-large-mnli")
33 ```
34
35 You can then use this pipeline to classify sequences into any of the class names you specify.
36
37 ```python
38 sequence_to_classify = "one day I will see the world"
39 candidate_labels = ['travel', 'cooking', 'dancing']
40 classifier(sequence_to_classify, candidate_labels)
41 #{'labels': ['travel', 'dancing', 'cooking'],
42 # 'scores': [0.9938651323318481, 0.0032737774308770895, 0.002861034357920289],
43 # 'sequence': 'one day I will see the world'}
44 ```
45
46 If more than one candidate label can be correct, pass `multi_label=True` to calculate each class independently:
47
48 ```python
49 candidate_labels = ['travel', 'cooking', 'dancing', 'exploration']
50 classifier(sequence_to_classify, candidate_labels, multi_label=True)
51 #{'labels': ['travel', 'exploration', 'dancing', 'cooking'],
52 # 'scores': [0.9945111274719238,
53 # 0.9383890628814697,
54 # 0.0057061901316046715,
55 # 0.0018193122232332826],
56 # 'sequence': 'one day I will see the world'}
57 ```
58
59
60 #### With manual PyTorch
61
62 ```python
63 # pose sequence as a NLI premise and label as a hypothesis
64 from transformers import AutoModelForSequenceClassification, AutoTokenizer
65 nli_model = AutoModelForSequenceClassification.from_pretrained('facebook/bart-large-mnli')
66 tokenizer = AutoTokenizer.from_pretrained('facebook/bart-large-mnli')
67
68 premise = sequence
69 hypothesis = f'This example is {label}.'
70
71 # run through model pre-trained on MNLI
72 x = tokenizer.encode(premise, hypothesis, return_tensors='pt',
73 truncation_strategy='only_first')
74 logits = nli_model(x.to(device))[0]
75
76 # we throw away "neutral" (dim 1) and take the probability of
77 # "entailment" (2) as the probability of the label being true
78 entail_contradiction_logits = logits[:,[0,2]]
79 probs = entail_contradiction_logits.softmax(dim=1)
80 prob_label_is_true = probs[:,1]
81 ```
82