Flair — State-of-the-Art NLP Framework

Flair

Flair provides a very simple API for applying SOTA NLP models — including its own contextual string embeddings — to sequence labelling, classification, and entity extraction tasks in 12+ languages. It stacks seamlessly on top of Hugging Face Transformers.

Key Features

• SOTA NER: 93.9 F1 on CoNLL-2003 English with stacked embeddings
• POS tagging, chunking, dependency parsing, NEL
• Contextual string embeddings (character-level LM) + BERT/XLM-R stacking
• Text classification: sentiment, intent, topic
• Multilingual: 12+ languages, zero-shot cross-lingual transfer
• Training API: train custom sequence labellers and classifiers in ~10 lines

Quick Start

from flair.data import Sentence
from flair.models import SequenceTagger

tagger = SequenceTagger.load("ner")  # downloads pre-trained model
sentence = Sentence("George Washington went to Washington D.C.")
tagger.predict(sentence)

for entity in sentence.get_spans("ner"):
    print(entity)  # Span[0:2]: "George Washington" → [PER (0.9998)]

Install via ai-supply

npx ai-supply add flair-nlp-framework

Curated mirror of the open-source Flair (MIT). Get it from the source.