home edit page issue tracker
This page pertains to UD version 2.

Introduction

Universal Dependencies (UD) is a project that is developing cross-linguistically consistent treebank annotation for many languages, with the goal of facilitating multilingual parser development, cross-lingual learning, and parsing research from a language typology perspective. The annotation scheme is based on an evolution of (universal) Stanford dependencies (de Marneffe et al., 2006, 2008, 2014), Google universal part-of-speech tags (Petrov et al., 2012), and the Interset interlingua for morphosyntactic tagsets (Zeman, 2008). The general philosophy is to provide a universal inventory of categories and guidelines to facilitate consistent annotation of similar constructions across languages, while allowing language-specific extensions when necessary.

This is illustrated in the following parallel examples from English, Bulgarian, Czech and Swedish, where the main grammatical relations involving a passive verb, a nominal subject and an oblique agent are the same, but where the concrete grammatical realization varies.

# visual-style 4 2 nsubj:pass	color:blue
# visual-style 4 7 obl	color:blue
1	The	the	DET	_	Definite=Def|PronType=Art	2	det	_	_
2	dog	dog	NOUN	_	Gender=Neut|Number=Sing	4	nsubj:pass	_	_
3	was	be	AUX	_	Mood=Ind|Number=Sing|Tense=Past|VerbForm=Fin	4	aux:pass	_	_
4	chased	chase	VERB	_	Tense=Past|VerbForm=Part	0	ROOT	_	_
5	by	by	ADP	_	_	7	case	_	_
6	the	the	DET	_	Definite=Def|PronType=Art	7	det	_	_
7	cat	cat	NOUN	_	Gender=Neut|Number=Sing	4	obl	_	_
8	.	.	PUNCT	_	_	4	punct	_	_


# visual-style 3 1 nsubj:pass	color:blue
# visual-style 3 5 obl	color:blue
1	Кучето	куче	NOUN	_	Definite=Def|Gender=Neut|Number=Sing	3	nsubj:pass	_	_
2	се	се	PRON	_	Case=Acc|PronType=Prs|Reflex=Yes	3	expl:pass	_	_
3	преследваше	преследвам	VERB	_	Aspect=Imp|Mood=Ind|Number=Sing|Person=3|Tense=Past|VerbForm=Fin	0	root	_	_
4	от	от	ADP	_	_	5	case	_	_
5	котката	котка	NOUN	_	Definite=Def|Gender=Fem|Number=Sing	3	obl	_	_
6	.	.	PUNCT	_	_	3	punct	_	_


# visual-style 3 1 nsubj:pass	color:blue
# visual-style 3 4 obl	color:blue
1	Pes	pes	NOUN	_	Animacy=Anim|Case=Nom|Gender=Masc|Number=Sing	3	nsubj:pass	_	_
2	byl	být	AUX	_	Aspect=Imp|Gender=Masc|Number=Sing|Tense=Past|VerbForm=Part|Voice=Act	3	aux:pass	_	_
3	honěn	honit	VERB	_	Aspect=Imp|Gender=Masc|Number=Sing|VerbForm=Part|Voice=Pass	0	root	_	_
4	kočkou	kočka	NOUN	_	Case=Ins|Gender=Fem|Number=Sing	3	obl	_	_
5	.	.	PUNCT	_	_	3	punct	_	_


# visual-style 2 1 nsubj:pass	color:blue
# visual-style 2 4 obl	color:blue
1	Hunden	hund	NOUN	_	Definite=Def	2	nsubj:pass	_	_
2	jagades	jaga	VERB	_	Tense=Past|Voice=Pass	0	root	_	_
3	av	av	ADP	_	_	4	case	_	_
4	katten	katt	NOUN	_	Definite=Def	2	obl	_	_
5	.	.	PUNCT	_	_	2	punct	_	_

What is needed for UD to be successful?

The secret to understanding the design and current success of UD is to realize that the design is a very subtle compromise between approximately 6 things:

  1. UD needs to be satisfactory on linguistic analysis grounds for individual languages.
  2. UD needs to be good for linguistic typology, i.e., providing a suitable basis for bringing out cross-linguistic parallelism across languages and language families.
  3. UD must be suitable for rapid, consistent annotation by a human annotator.
  4. UD must be suitable for computer parsing with high accuracy.
  5. UD must be easily comprehended and used by a non-linguist, whether a language learner or an engineer with prosaic needs for language processing. We refer to this as seeking a habitable design, and it leads us to favor traditional grammar notions and terminology.
  6. UD must support well downstream language understanding tasks (relation extraction, reading comprehension, machine translation, …).

It’s easy to come up with a proposal that improves UD on one of these dimensions. The interesting and difficult part is to improve UD while remaining sensitive to all these dimensions.

History

The Stanford dependencies were originally developed in 2005 as a backend to the Stanford parser to help in Recognizing Textual Entailment systems, then eventually emerged as the de facto standard for dependency analysis of English, and have since been adapted to a number of different languages (Chang et al., 2009, Bosco et al., 2013, Haverinen et al., 2013, Seraji et al., 2013, Tsarfaty, 2013, Lipenkova and Souček 2014). The Google universal tag set grew out of the cross-linguistic error analysis based on the CoNLL-X shared task data by McDonald and Nivre (2007), was initially used for unsupervised part-of-speech tagging by Das and Petrov (2011), and has since been adopted as a widely used standard for mapping diverse tagsets to a common standard. The Interset (Zeman, 2008) started as a tool for conversion between morphosyntactic tagsets of multiple languages. It dates back to 2006 when it was used in the first experiments with cross-lingual delexicalized parser adaptation (Zeman and Resnik, 2008). It was later employed as the morphological layer in HamleDT (Zeman et al., 2014) – a project that brings treebanks of many languages under a common annotation scheme.

The first attempt to combine Stanford dependencies and Google universal tags into a universal annotation scheme was the Universal Dependency Treebank (UDT) project (McDonald et al., 2013), which released treebanks for 6 languages in 2013 and 11 languages in 2014, and the first proposal for incorporating morphology was made by Tsarfaty (2013). The second version of HamleDT (Rosa et al., 2014) provided Stanford/Google annotation for 30 languages in 2014. This was followed by the development of universal Stanford dependencies (USD) (de Marneffe et al., 2014). The new Universal Dependencies is the result of merging all these initiatives into a single coherent framework, based on universal Stanford dependencies, an extended version of the Google universal tagset, a revised subset of the Interset feature inventory, and a revised version of the CoNLL-X format (called CoNLL-U).

The first version of the new guidelines, released in October 2014, introduced a somewhat extended universal part-of-speech tag set. This set makes some distinctions that were missing in the original proposal, but were perceived to be of importance by many, and clarifies the definition of categories. As a result of this work, universal POS categories have substantive definitions and are not necessarily just equivalence classes of categories in underlying language-particular treebanks. Hence, work to convert to UD POS tags often requires context-sensitive rules, or some hand correction. The UD morphological features aim to provide a stripped down basic set of features which are most crucial for analysis and are widespread across languages. The dependency representation of UD evolves out of Stanford Dependencies (SD), which itself follows ideas of grammatical relations-focused description that can be found in many linguistic frameworks. That is, it is centrally organized around notions of subject, object, clausal complement, noun determiner, noun modifier, etc. The goal of the new universal version was to add or refine relations to better accommodate the grammatical structures of typologically different languages and to clean up some of the quirkier and more English-specific features of the original version. Hence, the new taxonomy has less relations than the original SD.

Project organization

UD is an open collaboration with many project members. The administrative structure is kept at a minimum and currently consists of the following:

List of contributors