Title: Interpreting Arc-Standard Dependency Derivations Through a Tree Lens

Abstract

Projective dependency trees undergoing arc-standard derivations can be viewed as the step-by-step assembly of lexicalized ordered trees featuring contiguous yields. In this framework, every transition—whether it is a \textsc{shift}, \textsc{leftarc}, or \textsc{rightarc}—triggers a specific, deterministic update to the tree structure. Consequently, the final ordered tree generated by the derivation uniquely identifies the dependency arcs established during the process.

This study demonstrates that such a contiguous ordered representation is not merely an arbitrary encoding scheme; a single-headed dependency tree possesses this property if and only if it is projective. Because the ordered structure is defined directly by the transition sequence rather than being derived from a pre-existing, completed dependency graph, the approach is fundamentally derivational rather than conversion-based. This perspective offers a tree-theoretic understanding of arc-standard parsing, wherein projective dependency derivations implicitly build ordered trees reminiscent of constituency parsing, which remain recoverable. For inputs that are non-projective, this interpretation can still be applied via pseudo-projective lifting and subsequent inverse decoding. A preliminary implementation study validates this model, showing that these mapped derivations can be successfully executed within an existing neural transition-based parser.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC