Stanza

Stanza provides a fully integrated neural pipeline for 70+ languages, including tokenization, multi-word token expansion, part-of-speech tagging, lemmatization, morphological feature tagging, and dependency parsing. It wraps pre-trained models that achieve state-of-the-art or near-state-of-the-art performance on Universal Dependencies treebanks, making it a drop-in solution for cross-lingual text analysis without requiring language-specific feature engineering.

At its core, Stanza implements the Deep Biaffine Parser architecture introduced by Dozat and Manning. This graph-based parser uses:

• BiLSTM encoders to generate contextualized word representations
• Biaffine attention to score all possible head-dependent pairs simultaneously
• Maximum spanning tree decoding to find the globally optimal parse

The result is highly accurate labeled attachment scores (LAS) across diverse language families, handling both projective and non-projective structures.

Stanza is built natively on the Universal Dependencies (UD) framework, outputting annotations in standard CoNLL-U format. This ensures:

• Consistent part-of-speech tags and dependency relation labels across languages
• Seamless interoperability with other UD-compliant tools and treebanks
• Support for enhanced dependencies that capture implicit predicates and control relationships
• Easy integration into downstream tasks like relationship extraction and semantic role labeling

Beyond general-domain parsing, Stanza includes specialized biomedical and clinical NLP models trained on PubMed abstracts and MIMIC clinical notes. These models handle domain-specific tokenization, entity recognition, and syntactic analysis, enabling accurate processing of medical literature and electronic health records. The biomedical pipeline achieves significant accuracy improvements over general models when parsing complex clinical syntax and specialized terminology.

Stanza is designed for both research and production use with a clean Python API:

• Efficient batch processing with automatic GPU acceleration via PyTorch
• Memory-efficient model loading with options to keep models on disk
• Streaming document processing for large corpora
• Seamless integration with other NLP libraries and data science workflows

The library handles the full pipeline from raw text to structured annotations in a few lines of code.

Stanza's tokenizer handles complex linguistic phenomena including:

• Multi-word tokens common in languages like French (e.g., 'au' expanding to 'à le')
• Contractions in English and other languages
• Clitic splitting in Romance and Slavic languages
• Sentence segmentation with high accuracy across writing systems

This preprocessing ensures that downstream dependency parsing receives correctly segmented input, which is critical for accurate syntactic analysis in morphologically rich languages.