home edit page issue tracker

This page pertains to UD version 2.

Guidelines Changes

This pages summarizes the history of notable changes to the universal annotation guidelines.

Significant changes are classified as:

(Note: Many minor clarifications are not listed.)

Changes in UDv2

Updates to UDv2 will NOT alter the inventory of basic top-level dependency relations, UPOS tags, etc. But some updates have been necessary to clarify how they should be applied to particular linguistic phenomena, as well as formal constraints to be enforced by validation.

  #   Date
Decided
Effective
Release
Type Title
14 2024-Apr 2.14 CLARIFICATION List Item Markers
13 2023-Dec 2.14 CLARIFICATION Foreign Expressions and Code-Switching
12 2023-Nov 2.14 CLARIFICATION Multiword Expressions
11 2023-Nov 2.14 CLARIFICATION Semi-mandatory Relation Subtypes
10 2023-Sep 2.13 CLARIFICATION Adpositional Predicative Complements
9 2023-Sep 2.13 CLARIFICATION Sufficiency and Excess Constructions
8 2023-Mar 2.12 AMENDMENT, VALIDATOR Typos in Multiword Tokens
7 2023-Jan 2.12 AMENDMENT Sole iobj
6 2022-Dec 2.12 AMENDMENT Nested flat
5 2022-May 2.10 AMENDMENT, VALIDATOR Multiple Subjects
4 2022-May 2.10 AMENDMENT Optional Depictives
3 2022-Feb 2.10 AMENDMENT Reported Speech
2 2022-Jan 2.10 AMENDMENT, VALIDATOR Typos and goeswith
1 2021-Dec 2.10 CLARIFICATION, VALIDATOR Deverbal Connectives

List Item Markers

The discourse page has been updated to include enumerators introducing a list item. (Some corpora were using nummod for this purpose, but it is inconsistent with the nummod definition.) Bullets are (still) PUNCT and thus must attach as punct. (issue 1027)

Foreign Expressions and Code-Switching

A new page on Foreign Expressions and Code-Switching consolidates the options for analyzing multilingual content (issue 1001).

Multiword Expressions

The wording in the guidelines around “multiword expressions” has been updated to be more precise in the syntax overview and on the flat, fixed, and compound pages (issue 989). The examples of flat have been revised and restructured around new prototypes (issue 974), emphasizing that the treatment of many idiosyncratic expressions involving names and numbers should be determined based on language-specific criteria. The wording recommending the subtype flat:foreign has been changed to merely acknowledge that some treebanks use it.

Semi-mandatory Relation Subtypes

While relation subtypes are technically optional, the guidelines state that certain subtypes are “semi-mandatory”, i.e., languages are strongly encouraged to use them if relevant. Previously this set was open-ended. The guidelines now commit to a closed set of subtypes that are semi-mandatory (issue 990).

Adpositional Predicative Complements

Examples of predicative complements marked with as etc. have been added to the xcomp guidelines.

Sufficiency and Excess Constructions

Guidelines for constructions like so X that Y (so dusty that I sneezed) and too X to Y (too slowly to finish on time) have been added.

Typos in Multiword Tokens

The rule that multiword tokens must not have features in the FEATS column has been relaxed to permit Typo=Yes, as this is a surface feature and not every typo can be localized to a single syntactic word. However, if a multiword token bears Typo=Yes, the feature cannot simultaneously appear on any of its component words. See discussion of typos.

Sole iobj

The restriction that iobj can only be used if there is also an overt obj in the same clause is hereby removed. It is preferable not to have a relation whose label depends on the presence or absence of a sister phrase when its interpretation remains the same (like them in tell them vs. tell them a story vs. tell them that the party is canceled). See iobj for examples (note that specific criteria for identifying indirect objects will vary by language).

Nested flat

In May 2021, the flat guidelines were revised to clarify that nested expressions were allowed under a flat dependent, so long as the nested expression was not itself a flat expression. This constraint is now lifted (in fact, it was never enforced): if a flat name, for example, has a part which is itself a flat name, it may be indicated with a nested flat expression. The only constraint on flat edges is that they must point forward in the sentence: the first word of the expression is always the technical head, though linguistically speaking, the expression is considered headless.

Multiple Subjects

In general, UD prohibits multiple subjects (i.e. a word may have at most one nsubj or csubj dependent), and enforcing this in validation is a useful way to catch errors. However, a clause may serve as the predicate in a copular construction (e.g. The problem is that we already paid), posing a problem for this constraint. Until now, the guidelines carved out an exception for such cases: the copula of the outer clause would be promoted to head its subject and the predicate of the inner clause would attach to it as ccomp (as explained in v1 guidelines). But this yielded an odd interpretation of some copulas as transitive and offered no solution for zero copula constructions. A change was necessary.

The new policy—a product of extensive deliberation—is that the predicate of the inner clause can have multiple subject dependents. The subject(s) of the non-innermost clause(s) can be subtyped with :outer to signify nesting: nsubj:outer, csubj:outer. The :outer subtype, like all subtypes, is (at least for now) technically optional. Therefore, as an alternative, the validator will allow a treebank’s maintainers to manually verify that any instances of multiple subjects are correct.

Note that using :outer just for subjects does not fully disambiguate the compositional structure: for example, cop, aux, mark, advmod, and obl dependents of the predicate may belong to either the inner or outer clause. In order to avoid a proliferation of subtyped relations, the trees in the new guidelines do not apply the :outer label to anything other than subjects. Treebanks are, of course, welcome to innovate in their use of subtypes and/or MISC attributes.

Optional Depictives

Reanalyzed optional depictives as adverbial (advcl) rather than adnominal (acl), given that the predicand may not always be overt in the sentence, and even when it is overt it doesn’t form a nominal phrase with the depictive. The secondary predication can instead be expressed via an enhanced dependency, similar to control. (A precise naming recommendation for the enhanced edge is deferred for further discussion.)

Reported Speech

Revised the policy regarding reported speech: the quoted material attaches as ccomp to the speech verb regardless of order and punctuation; parataxis should be used only if the quotation is interrupted.

Typos and goeswith

Updates to the policy on typos to clarify treatment of goeswith:

Deverbal Connectives

Deverbal connectives may be tagged as VERB while attaching as case or mark. Documented at ADP.

UDv1 and transition to UDv2

Discussions

Data Releases