Universal Dependencies Joakim Nivre Uppsala University

Universal Dependencies Background: Treebank annotation schemes vary across languages Hard to compare results across languages [Nivre et al. 2007] Hard to evaluate cross-lingual learning [McDonald et al. 2013] Hard to build multilingual systems Universal Dependencies (http://universaldependencies.github.io/docs/): Stanford universal dependencies [de Marneffe et al. 2014] Google universal part-of-speech tags [Petrov et al. 2012] Interset morphological features [Zeman 2008] First guidelines released Oct 1, 2014 First 10 treebanks released Jan 15, 2015

Universal Dependencies Syntactic words – explicit splitting of clitics and contractions Universal part-of-speech tags + morphological features Dependency tree + augmented dependencies (not shown)

Goals Cross-linguistically consistent grammatical annotation Support multilingual NLP and linguistic research Build on common usage and existing de-facto standards Complement – not replace – language-specific schemes Open community effort – anyone can contribute

Guiding Principles Maximize parallelism Don't annotate the same thing in different ways Don't make different things look the same Don't annotate things that are not there Languages select from a universal pool of categories Allow language-specific extensions

Design Principles Dependency Lexicalism Recoverability Widely used in practical NLP systems Available in treebanks for many languages Lexicalism Basic annotation units are words – syntactic words Words have morphological properties Words enter into syntactic relations Recoverability Transparent mapping from input text to word segmentation

Morphological Annotation Le La DET Definite=Def Gender=Masc Number=Sing chat NOUN chasse chasser VERB Mood=Ind Person=3 les le chiens chien Number=Plur . PUNCT Lemma represent the semantic content of a word Part-of-speech tag represent its grammatical class Features represent lexical and grammatical properties of the lemma or the particular word form

Syntactic Annotattion Content words are related by dependency relations Function words attach to the content word they modify Punctuation attach to head of phrase or clause

CoNLL-U Format 1 Le DET _ 2 det chat NOUN 3 nsubj boit boire VERB Root ID FORM LEMMA UPOSTAG XPOSTAG FEATS HEAD DEPREL DEPS MISC 1 Le DET _ 2 det chat NOUN 3 nsubj boit boire VERB Root 4-5 du 4 de De ADP 6 case 5 le DEP lait Lait obj SpaceAfter=no 7 . PUNCT Punct

Dependency Structure English Swedish Keeping content words as heads promotes parallelism Function words often correlate with morphology

Dependency Relations [de Marneffe et al. 2014] Taxonomy of 42 universal grammatical relations, broadly supported across many languages in language typology Language specific subtypes can be added

Morphology: POS Open class words Closed class words Other ADJ ADV INTJ NOUN PROPN VERB ADP AUX CONJ DET NUM PART PRON SCONJ PUNCT SYM X Taxonomy of 17 universal part-of-speech tags, based on the Google Universal Tagset [Petrov et al. 2012]

Morphology: Universal Features Standardized inventory of morphological features, based on the Interset system [Zeman 2008] Lexical features Inflectional features Nominal* Verbal* PronType Gender VerbForm NumType Animacy Mood Poss Number Tense Reflex Case Aspect Foreign Definite Voice Abbr Degree Evident Polarity Person Polite

Morphology: Examples la Definite=Def|Gender=Fem|Number=Sing|PronType=Art hanno Mood=Ind|Number=Plur|Person=3|Tense=Pres|VerbForm=Fin fatto Gender=Masc|Number=Sing|Tense=Past|VerbForm=Part casa Gender=Fem|Number=Sing