Stanza Cheat Sheet

1Install & Pipeline1.x

pip install stanza import stanza★
Pulls in PyTorch. Python 3.9+.
stanza.download("en")★
Download the language's models once (cached in ~/stanza_resources). Do this before building a Pipeline for that language.
nlp = stanza.Pipeline("en", processors="tokenize,pos,lemma,depparse,ner")★
The neural pipeline. Omit processors for the default full set; list only what you need to run faster.
nlp = stanza.Pipeline("en", processors="tokenize", use_gpu=True)
Processors have dependencies (pos needs tokenize; lemma/depparse need pos) — Stanza enforces the order.

2The Documentnested structure

doc = nlp("Barack Obama was born in Hawaii.")★
Returns a Document. Pass a list of strings for batch processing.
for sent in doc.sentences: for word in sent.words: ...★
The hierarchy: Document → sentences → words (and tokens, which can differ for multi-word tokens).
doc.text · len(doc.sentences) · doc.num_words
Top-level accessors. tokens vs words differ in languages with multi-word tokens (e.g. French du → de le).

3Tokens, POS, lemmaword attributes

word.text · word.lemma · word.upos · word.xpos★
upos = universal POS (NOUN, VERB...), xpos = treebank-specific tag, lemma = base form.
word.feats # morphology: "Number=Sing|Tense=Past"★
Universal morphological features as a string — number, tense, case, person, etc.
for w in sent.words: print(w.text, w.upos, w.lemma)★
The typical extraction loop. Each word also has id, start_char/end_char offsets.

4Dependency parsesyntactic structure

word.head · word.deprel★
head = the id of this word's syntactic parent (0 = root); deprel = the relation (nsubj, obj, amod...).
sent.print_dependencies() · head_word = sent.words[word.head - 1]★
Word ids are 1-based, so index into sent.words with head - 1. A head of 0 marks the sentence root.
[(w.text, w.deprel, sent.words[w.head-1].text) for w in sent.words if w.head]
Build (dependent, relation, head) triples — the basis of relation/attribute extraction.

7Pipeline optionsspeed & input

stanza.Pipeline("en", tokenize_pretokenized=True)★
Feed already-tokenized text (space-separated, or a list of lists) — skips the tokenizer.
Pipeline("en", use_gpu=True, batch_size=64)
GPU + batching for throughput. Build the pipeline once and reuse it (loading models is the slow part).
Pipeline("en", package="default_accurate", download_method=None)
Choose a model package (accuracy vs speed); download_method=None to skip the auto-download check.

8Output & CoreNLPexport & Java

doc.to_dict() · stanza.utils.conll.CoNLL.write_doc2conll(doc, "out.conllu")★
Serialize to a list of dicts or CoNLL-U format (the UD standard).
from stanza.server import CoreNLPClient with CoreNLPClient(annotators=["tokenize","coref"]) as client: ...
Access the Java CoreNLP server for annotators Stanza's neural pipeline lacks (e.g. coreference). Needs Java + the CoreNLP jars.

!Common gotchasread before shipping

download() before Pipeline()
Building a Pipeline for a language whose models aren't downloaded triggers a download (or fails offline). Do stanza.download first.
build the Pipeline once
Loading neural models is expensive — construct nlp at startup and reuse it, never per-request.
word ids are 1-based; head 0 = root
Index sent.words[word.head - 1]; forgetting the offset points at the wrong word.
processor dependencies
lemma/depparse require pos, which requires tokenize. Request the whole chain or Stanza errors.
accuracy over speed
Stanza is neural & thorough, not the fastest. For high-throughput production, spaCy is often the better tradeoff.

Stanza cheat sheet