Guidelines Changes
This pages summarizes the history of notable changes to the universal annotation guidelines.
Significant changes are classified as:
- AMENDMENT: change that negates a policy articulated in an earlier version of the guidelines
- CLARIFICATION: resolution of an important ambiguity in the guidelines (that doesn’t negate a previous policy)
- VALIDATOR: important change to the requirements checked by the official UD validator
(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 |
|---|---|---|---|---|
| 15 | 2024-Nov | 2.16 | CLARIFICATION, VALIDATOR | ExtPos for Fixed Expressions |
| 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 |
ExtPos for Fixed Expressions
In order to clarify the holistic status of expressions connected by the fixed relation, it is strongly recommended to specify ExtPos (in the FEATS column) on the first word of the expression. The validator has been updated to trigger a warning if ExtPos is missing for a fixed expression. Some checks that previously made exceptions for words in fixed expressions no longer do so; adding ExtPos may be necessary to pass these checks.
Worthy of note: (1) ExtPos (“external part of speech”) has already been in use across several treebanks. It is not always limited just to fixed expressions, and treebanks should feel free to apply it in other ways. At this time, we only specify as a universal matter that it is recommended for fixed expressions. (2) The fixed relation is not intended for all multiword expressions that are morphosyntactically fossilized. fixed should apply only to grammaticized expressions that appear internally anomalous, or whose external syntactic behavior diverges in important ways from the historical internal structure—especially for expressions functioning as ADP, CCONJ, DET, SCONJ, or PRON. (issue 1037)
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:
- Typo
=Yesis now required on the first part of a goeswith structure (previously stated to be unnecessary) - Morphological features and lemmas now prohibited on tokens attaching as goeswith; UPOS required to be X
- Further explanation of goeswith structures, including interaction with multiword tokens
Deverbal Connectives
Deverbal connectives may be tagged as VERB while attaching as case or mark. Documented at ADP.