Release 1.0 (March 11, 2021)

NB-Bert base model finetuned on Norwegian machine translated MNLI

Description

The most effective way of creating a good classifier is to finetune it for this specific task. However, in many cases this is simply impossible. Yin et al. has proposed a very clever way of using pre-trained MNLI model as a zero-shot sequence classifiers. The methods works by reformulating the question to an MNLI hypothesis. If we want to figure out if a text is about "sport", we simply state that "This text is about sport" ("Denne teksten handler om sport").

When the model is finetuned on the 400k large MNLI task, it is in many cases able to solve this classification tasks. There are no MNLI-set of this size in Norwegian but we have trained it on a machine translated version of the original MNLI-set.

Hugging Face zero-shot-classification pipeline

The easiest way to try this out is using the Hugging Face pipeline. Please note that you will improve the results by overriding the English hypothesis template.

from transformers import pipeline
classifier = pipeline("zero-shot-classification", model="NBAiLab/nb-bert-base-mnli")

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

sequence_to_classify = "Folkehelseinstituttets mest optimistiske anslag er at alle over 18 år er ferdigvaksinert innen midten av september."
candidate_labels = ["politikk, helse, sport, religion"]
hypothesis_template = "Denne teksten handler om {}."
classifier(sequence_to_classify, candidate_labels, hypothesis_template=hypothesis_template, multi_class=True)


#{'labels': ['travel', 'dancing', 'cooking'],
# 'scores': [0.9938651323318481, 0.0032737774308770895, 0.002861034357920289],
# 'sequence': 'one day I will see the world'}

More information

For more information on the model, see

https://github.com/NBAiLab/notram

Here you will also find a Colab explaining more in details how to use the zero-shot-classification pipeline.