Back to all models
text-classification mask_token: <mask>
Query this model
🔥 This model is currently loaded and running on the Inference API. ⚠️ This model could not be loaded by the inference API. ⚠️ This model can be loaded on the Inference API on-demand.
JSON Output
API endpoint
								\$
curl -X POST \
-H "Authorization: Bearer YOUR_ORG_OR_USER_API_TOKEN" \
-H "Content-Type: application/json" \
-d '"json encoded string"' \

Share

last 30 days

pytorch

tf

#### Contributed by

How to use this model directly from the 🤗/transformers library:


from transformers import AutoTokenizer, AutoModelForSequenceClassification


Update on GitHub

## Model Description

This model takes facebook/bart-large-mnli and fine-tunes it on Yahoo Answers topic classification. It can be used to predict whether a topic label can be assigned to a given sequence, whether or not the label has been seen before.

You can play with an interactive demo of this zero-shot technique with this model, as well as the non-finetuned facebook/bart-large-mnli, here.

## Inteded Usage

This model was fine-tuned on topic classification and will perform best at zero-shot topic classification. Use hypothesis_template="This text is about {}." as this is the template used during fine-tuning.

For settings other than topic classification, you can use any model pre-trained on MNLI such as facebook/bart-large-mnli or roberta-large-mnli with the same code as written below.

#### With the zero-shot classification pipeline

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

from transformers import pipeline

sequence_to_classify = "Who are you voting for in 2020?"
candidate_labels = ["Europe", "public health", "politics", "elections"]
hypothesis_template = "This text is about {}."
nlp(sequence_to_classify, candidate_labels, multi_class=True, hypothesis_template=hypothesis_template)

#### With manual PyTorch

# pose sequence as a NLI premise and label as a hypothesis
from transformers import BartForSequenceClassification, BartTokenizer

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

# run through model pre-trained on MNLI
x = tokenizer.encode(premise, hypothesis, return_tensors='pt',
max_length=tokenizer.max_len,
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
prob_label_is_true = probs[:,1]
The model is a pre-trained MNLI classifier further fine-tuned on Yahoo Answers topic classification in the manner originally described in Yin et al. 2019 and this blog post. That is, each sequence is fed to the pre-trained NLI model in place of the premise and each candidate label as the hypothesis, formatted like so: This text is about {class name}. For each example in the training set, a true and a randomly-selected false label hypothesis are fed to the model which must predict which labels are valid and which are false.
This model was evaluated with the label-weighted F1 of the seen and unseen labels. That is, for each example the model must predict from one of the 10 corpus labels. The F1 is reported for the labels seen during training as well as the labels unseen during training. We found an F1 score of .68 and .72 for the unseen and seen labels, respectively. In order to adjust for the in-vs-out of distribution labels, we subtract a fixed amount of 30% from the normalized probabilities of the seen labels, as described in Yin et al. 2019 and our blog post.