Text classification

createTextClassifier() labels a text with a fine-tuned classifier. The default checkpoint is DistilBERT SST-2, a sentiment model with POSITIVE and NEGATIVE labels (~65 MB).

import { createTextClassifier } from 'ngx-transformers';

readonly classifier = createTextClassifier(); // Xenova/distilbert-base-uncased-finetuned-sst-2-english

const [top] = await this.classifier.classify('This library makes on-device ML in Angular a joy.');
// { label: 'POSITIVE', score: 0.9997 }

const all = await this.classifier.classify(text, 2); // every label, best first

Try it: sentiment analysis story.

API

classify(text, topK = 1) resolves to ClassificationResult[] (label, score) sorted by score, best first. topK is forwarded to the pipeline; pass the number of labels to get the full distribution.

The handle is a PipelineHandle, so load(), dispose() and the signals work as described in Concepts. run(text, options) gives raw access to the pipeline with any Transformers.js option.

Other classifiers

Any text-classification checkpoint works; the labels come from the model:

Model	Labels
Xenova/bert-base-multilingual-uncased-sentiment	1 star to 5 stars, reviews in six languages
Xenova/toxic-bert	toxic, severe_toxic, obscene, threat, insult, identity_hate

readonly reviews = createTextClassifier({ model: 'Xenova/bert-base-multilingual-uncased-sentiment' });
const [stars] = await this.reviews.classify('Das Essen war hervorragend.'); // { label: '5 stars', ... }

For labels the model was not trained on, use zero-shot classification.