📄 Improving Surface-syntactic Universal Dependencies (SUD): MWEs and deep syntactic features

Summary

This paper refines SUD (Surface-syntactic Universal Dependencies), an annotation scheme that closely parallels UD and can be automatically converted to it. Unlike UD, which links content words and treats function words as markers, SUD selects heads by distributional criteria, so functional words often become heads. It also defines relation labels by which elements can substitute for each other in a position, not by the dependent's POS. The authors present an updated, hierarchical relation inventory organized around a few broad functions. They separate surface-syntactic relations from an orthogonal layer of deep-syntactic features, marked with "@" on labels, which captures diathesis, control, expletives, and light verbs. For multiword expressions, they propose a feature giving the POS of the whole expression, so that MWEs can receive a regular internal analysis. They also describe how conversion between UD and SUD works in both directions, which is reversible in most cases. This work matters because SUD gives communities that object to some UD design choices a linguistically motivated alternative that remains compatible with the UD ecosystem.

Key points

• Seventeen UD relations, including nsubj, obj, xcomp, ccomp, amod, advcl, aux, cop, case and mark, are collapsed into three core SUD relations: subj, comp and mod. These have subtypes such as comp:obj, comp:obl, comp:pred, comp:aux and comp:cleft, with udep and unk serving as underspecified fallbacks.
• The deep-syntactic features are @x (a dependent whose subject is not realized locally, replacing UD's xcomp logic), @pass, @caus, @agent, @expl, @lvc and @fixed.
• A new feature, ExtPOS (with a companion Type feature), records the external POS of idioms, titles and "grafts." The authors suggest UD adopt it as well, partly to simplify treebank validation.
• To keep conversion reversible, two new UD subtypes are introduced: nummod:det and parataxis:discourse.
• Conversion in both directions is implemented as GREW grammars. Resources are collected at surfacesyntacticud.github.io. SUD is already used for a Naija treebank, and French and Chinese treebanks are in progress.