Flair

Generates contextualized word and document embeddings by stacking forward and backward language models (flair embeddings), capturing semantic meaning based on surrounding context rather than static word vectors. This approach combines character-level CNN encoders with LSTM layers to produce embeddings that adapt to polysemy and word sense variation, enabling superior performance on downstream NLP tasks compared to static embeddings.

Implements a SequenceTagger model combining BiLSTM (bidirectional LSTM) layers with Conditional Random Fields (CRF) for structured prediction on token sequences. The architecture processes embedded tokens through bidirectional recurrent layers to capture long-range dependencies, then applies CRF decoding to enforce valid tag sequences and output globally optimal predictions rather than independent token classifications.

Computes comprehensive evaluation metrics for different NLP tasks including precision, recall, F1-score per class, and task-specific metrics (entity-level F1 for NER, accuracy for classification). The evaluation system provides detailed error analysis including confusion matrices, per-class performance breakdowns, and prediction confidence distributions, enabling practitioners to understand model behavior and identify failure modes.

Provides biomedical-specific embeddings and pre-trained models for NER, relation extraction, and text classification on biomedical literature. The biomedical models are trained on PubMed abstracts and biomedical corpora, with embeddings that capture domain-specific terminology and entity types (proteins, genes, diseases, chemicals). This enables practitioners to apply state-of-the-art biomedical NLP without extensive domain-specific training data.

Enables training custom contextual embeddings (flair embeddings) from scratch or fine-tuning pre-trained embeddings on domain-specific text. The language model training uses forward and backward LSTM-based language models with character-level CNN encoders, optimized for predicting next/previous tokens. This approach allows practitioners to create domain-specific embeddings without requiring massive transformer models, enabling better performance on specialized domains with limited data.

Provides core data structures (Sentence, Token, Label, Span) that represent text and annotations in a unified format. Sentence objects contain Token objects with embeddings and predictions, Label objects store classification labels with confidence scores, and Span objects represent entity mentions with types and confidence. These structures enable seamless integration between text processing, embedding, and prediction components throughout Flair's pipeline.

Performs document-level text classification by aggregating token embeddings into a single document representation (via pooling or attention mechanisms), then passing through feed-forward neural networks with optional multi-layer architecture. The TextClassifier model supports both single-label and multi-label classification, with configurable loss functions (cross-entropy for single-label, binary cross-entropy for multi-label) and automatic handling of class imbalance through weighted sampling.

Extracts relations between entities by treating relation extraction as a pairwise classification problem: for each pair of entities in a sentence, the model predicts whether a relation exists and its type. The RelationExtractor uses entity-aware embeddings that concatenate token embeddings with entity type information, enabling the model to distinguish between different entity types and their interactions while maintaining awareness of entity boundaries through special markers.

Links named entities in text to entries in a knowledge base (e.g., Wikipedia) through a two-stage pipeline: candidate generation identifies potential knowledge base entries for each entity mention, then disambiguation ranks candidates using entity context embeddings and knowledge base information. The EntityLinker uses mention embeddings combined with entity type constraints to select the most likely knowledge base entry, supporting both exact matching and fuzzy matching strategies.

Enables zero-shot classification and tagging by leveraging label semantics and task descriptions without requiring labeled training data. The TARS (Task Aware Representation System) model uses a prompt-based approach where task descriptions and label definitions are encoded as embeddings, then compared against input text embeddings to predict labels. This approach allows the same model to handle different classification tasks by changing the prompt and label definitions without retraining.

Trains multiple NLP tasks simultaneously using a shared embedding and encoder layer with task-specific output heads, enabling knowledge transfer between related tasks. The multi-task architecture uses a single BiLSTM or transformer encoder that processes embeddings, then branches into separate task-specific layers (CRF for tagging, softmax for classification) that predict task-specific outputs. This approach improves generalization by leveraging task relationships and reducing overfitting on small datasets.

Integrates pre-trained transformer models (BERT, RoBERTa, DistilBERT, etc.) from Hugging Face as embedding providers or task-specific models, enabling fine-tuning on downstream NLP tasks. Flair wraps transformer models through a unified TransformerWordEmbeddings interface that handles tokenization, subword token aggregation, and embedding extraction, allowing transformers to be used interchangeably with Flair's native embeddings in any downstream task architecture.

Provides a unified Corpus abstraction for managing labeled NLP datasets, handling data loading, preprocessing, and train-validation-test splitting. The Corpus class automatically manages multiple Sentence objects with their annotations, supports various input formats (CoNLL, TSV, JSON), and provides utilities for dataset statistics, class distribution analysis, and stratified splitting to ensure balanced class representation across splits.

Provides a unified training loop that handles model optimization, loss computation, and evaluation across different NLP tasks. The ModelTrainer class manages training dynamics including learning rate scheduling, gradient clipping, early stopping, and checkpoint management. It supports task-specific loss functions (cross-entropy for classification, CRF loss for tagging, weighted loss for imbalanced data) and multiple optimization strategies (Adam, SGD with momentum, AdamW).