Dependency Parsing
Table of contents
Description
The dependency parsing module builds a tree structure of words from the input sentence, which represents the syntactic dependency relations between words. The resulting tree representations, which follow the Universal Dependencies formalism, are useful in many downstream applications. In Stanza, dependency parsing is performed by the DepparseProcessor, and can be invoked with the name depparse.
| Name | Annotator class name | Requirement | Generated Annotation | Description |
|---|---|---|---|---|
| depparse | DepparseProcessor | tokenize, mwt, pos, lemma | Determines the syntactic head of each word in a sentence and the dependency relation between the two words that are accessible through Word’s head and deprel attributes. | Provides an accurate syntactic dependency parsing analysis. |
Options
| Option name | Type | Default | Description |
|---|---|---|---|
| depparse_batch_size | int | 5000 | When annotating, this argument specifies the maximum number of words to process as a minibatch for efficient processing. Caveat: the larger this number is, the more working memory is required (main RAM or GPU RAM, depending on the computating device). This parameter should be set larger than the number of words in the longest sentence in your input document, or you might run into unexpected behaviors. |
| depparse_pretagged | bool | False | Assume the document is tokenized and pretagged. Only run dependency parsing on the document. |
Example Usage
Running the DepparseProcessor requires the TokenizeProcessor, MWTProcessor, POSProcessor, and LemmaProcessor. After all these processors have been run, each Sentence in the output would have been parsed into Universal Dependencies (version 2) structure, where the head index of each Word can be accessed by the property head, and the dependency relation between the words deprel. Note that the head index starts at 1 for actual words, and is 0 only when the word itself is the root of the tree. This index should be offset by 1 when looking for the govenor word in the sentence.
Accessing Syntactic Dependency Information
Here is an example of parsing a sentence and accessing syntactic parse information from each word:
import stanza
nlp = stanza.Pipeline(lang='fr', processors='tokenize,mwt,pos,lemma,depparse')
doc = nlp('Nous avons atteint la fin du sentier.')
print(*[f'id: {word.id}\tword: {word.text}\thead id: {word.head}\thead: {sent.words[word.head-1].text if word.head > 0 else "root"}\tdeprel: {word.deprel}' for sent in doc.sentences for word in sent.words], sep='\n')
As can be seen in the output, the syntactic head of the word Nous is atteint, and the dependency relation between the two words is nsubj (Nous is a nominal subject for atteint).
id: 1 word: Nous head id: 3 head: atteint deprel: nsubj
id: 2 word: avons head id: 3 head: atteint deprel: aux:tense
id: 3 word: atteint head id: 0 head: root deprel: root
id: 4 word: la head id: 5 head: fin deprel: det
id: 5 word: fin head id: 3 head: atteint deprel: obj
id: 6 word: de head id: 8 head: sentier deprel: case
id: 7 word: le head id: 8 head: sentier deprel: det
id: 8 word: sentier head id: 5 head: fin deprel: nmod
id: 9 word: . head id: 3 head: atteint deprel: punct
Start with Pretagged Document
Normally, the depparse processor depends on tokenize, mwt, pos, and lemma processors. However, in cases you wish to use your own tokenization, multi-word token expansion, POS tagging and lemmatization, you can skip the restriction and pass the pretagged document (with upos, xpos, feats, lemma) by setting depparse_pretagged to True.
Here is an example of dependency parsing with pretokenized and pretagged document:
import stanza
from stanza.models.common.doc import Document
nlp = stanza.Pipeline(lang='en', processors='depparse', depparse_pretagged=True)
pretagged_doc = Document([[{'id': 1, 'text': 'Test', 'lemma': 'Test', 'upos': 'NOUN', 'xpos': 'NN', 'feats': 'Number=Sing'}, {'id': 2, 'text': 'sentence', 'lemma': 'sentence', 'upos': 'NOUN', 'xpos': 'NN', 'feats': 'Number=Sing'}, {'id': 3, 'text': '.', 'lemma': '.', 'upos': 'PUNCT', 'xpos': '.'}]])
doc = nlp(pretagged_doc)
Training-Only Options
Most training-only options are documented in the argument parser of the dependency parser.