Consolidating the UD Annotation for Armenian

The current release of Universal Dependencies database (v.2.16, 15.05.2025) includes treebanks of three literary varieties of Armenian, Modern Eastern Armenian (“UD Armenian BSUT” and "UD Armenian ArmTDP"), Modern Western Armenian ("UD Western Armenian ArmTDP"), and Classical Armenian ("UD Classical Armenian CAVaL"). Although these varieties have many shared morphosyntactic features, they also show significant differences conditioned by diachronic and/or dialectal divergence. Besides, the treebanks of the modern varieties, on the one hand, and the one of Classical Armenian, on the other hand, have been developed by different teams of annotators, who do not always follow the same approach to the UD annotation guidelines. These two factors result in only partial compatibility of annotation across the Armenian treebanks.

The paper offers a systematic revision of discrepancies in the tagsets and principles of their application. The goal of this analysis is to identify which differences of annotation can be harmonized, and which are conditioned by genuinely dissimilar grammatical features.

The presentation will include a complete chart of correspondences between the tagsets for parts of speech and grammatical features. The major differences will be illustrated by cases studies, two of which are signalled below.

1) The Modern Eastern Armenian (MEA) treebanks offer context-dependent interpretations for values of some morphological features, whereas the Classical Armenian (CA) one follows the morphological principle, which requires to apply the same tag for each instance of a given morphological category. For example, the feature "Animacy" of the MEA treebanks has two values "Hum" (human) and "Nhum" (non-human), which are assigned to nouns based on the semantic interpretation of participants that they express. MEA does have the morphosyntactic expression of the contrast between human and non-human direct objects, which are flagged by the dative and nominative-accusative cases, respectively (Dum-Tragut 2009: 60-63); this grammatical information can be retrieved by a combination of tags "Case=Dat" + "obj" and "Case=Nom" + "obj", respectively. However, the MEA treebanks apply the values of the "Animacy" feature without regard to this morphsyntactic pattern, so that one finds combinations like "Animacy=Nhum|Case=Dat" + "obj". By contrast, the CA treebank assigns the values "Anim" (animate) and "Inan" (inanimate) of the "Animacy" feature to pronouns and deternimers that have parallel sets of forms corresponding to these values. Similar issues concern the features "Aspect" and "Voice". 

2) In the MEA treebanks, nouns with enclitic articles and verbs with proclitic negation particples are treated as single-token words carrying relevant features. By contrast, in the CA treebank, these structures are annotated as groups of separate tokens spelled without a space, and relevant features are distributed among the constituents.

The discussion of results can be helpful for developing hybrid multi-variant parsing models based on UD (see Vidal-Gorène et al. 2024) as well as for developing UD treebanks for other varieties of Armenian such as Middle Armenian or modern Armenian dialects.