GWC2025 — Pavia

Abstract

This paper presents a PropBank annotation project for Turkish, focusing on core arguments in matrix clauses. The dataset comprises five different corpora with 25,580 sentences labeled for verbal predicates and core arguments. Using a semi-automatic approach, a dependency layer was leveraged to pre-assign some ARG0s and ARG1s, followed by manual corrections. This work feeds into the development of Abstract Meaning Representations (AMRs), enhancing Turkish NLP resources for semantic parsing and higher-level language tasks.

Methodology

• Predicate selection: identifying verbal predicates while excluding noun phrases and nominal clauses
• Dependency-based pre-labeling: automatically assigning ARG0/ARG1 from NSUBJ and OBJ relations, with manual correction for unaccusative and passive predicates
• Manual argument selection using the StarDust annotation interface, assigning frame-file roles from TRopBank v2.0

Datasets & Argument Distribution

ARG0 typically marks the agent or causer, while ARG1 denotes the patient or theme. ARG2 and ARG3 represent more specialized roles depending on the verb's semantics.