new! 🦜🏝️
Left-Branching Transformers Excel at Right-Branching Languages: Data Shapes Word Order Preferences in Language Models
Varvara Arzt, Allan Hanbury, and Terra Blevins
To appear in Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing (EMNLP 2026), main conference
Word order preferences in LMs are data-driven, not architectural. Comparing decoder-only LMs across 192 artificial languages and typologically diverse natural ones, the same architecture shows a left-branching preference on artificial languages but an emerging right-branching SVO advantage on natural ones as training data scales, and that advantage tracks resource level and data quality rather than word order itself. This implies LLM adoption may gradually narrow word order diversity.
Paper
Code
AnnoHID: LLM-Assisted Annotation Framework for Low-Resource Medical Texts
Annisa Maulida Ningtyas, Guntur Budi Herwanto, Yunita Sari, Rifki Afina Putri, Filip Kovacevic, Alaa El-Ebshihy, Varvara Arzt, and Florina Piroi
ACL 2026, System Demonstrations, pages 683–691, San Diego, California, USA
A human-in-the-loop annotation framework for medical text in low-resource languages. On Indonesian clinical data, LLM pre-annotation raised inter-annotator agreement (κ = 0.76 for NER) and human review lifted F1 from 0.39 to 0.45.
Paper
Relation Extraction or Pattern Matching? Unravelling the Generalisation Limits of Language Models for Biographical RE
Varvara Arzt, Allan Hanbury, Michael Wiegand, Gábor Recski, and Terra Blevins
IJCNLP–AACL 2025 (main conference), pages 1463–1484, Mumbai, India
RE models pattern-match rather than generalise. Cross-dataset experiments show relation extraction models fail to transfer even within similar domains, and that higher in-dataset performance often signals overfitting to dataset artefacts. Data quality, not lexical similarity, drives robust transfer, and zero-shot baselines occasionally beat every cross-dataset result.
Paper
Beyond the Numbers: Transparency in Relation Extraction Benchmark Creation and Leaderboards
Varvara Arzt and Allan Hanbury
2nd GenBench Workshop on Generalisation (Benchmarking) in NLP @ EMNLP 2024, pages 120–130, Miami, Florida, USA
RE benchmarks and leaderboards obscure real progress. An audit of widely used benchmarks (TACRED, NYT) documenting severe class imbalance, noisy labels and undocumented construction decisions, and showing that aggregate F1 leaderboards hide per-class failure, with recommendations applicable well beyond RE.
Paper
TU Wien at SemEval-2024 Task 6: Unifying Model-Agnostic and Model-Aware Techniques for Hallucination Detection
Varvara Arzt, Mohammad Mahdi Azarbeik, Ilya Lasy, Tilman Kerl, and Gábor Recski
18th International Workshop on Semantic Evaluation (SemEval-2024), pages 1183–1196, Mexico City, Mexico
An ensemble unifying model-agnostic and model-aware techniques for hallucination detection, placing 3rd of all systems in the model-aware track (80.6% accuracy).
Paper
Global PIQA: Evaluating Commonsense Reasoning Across 100+ Languages and Cultures
Tyler A. Chang, Catherine Arnett, et al. (350+ contributors, including Varvara Arzt)
arXiv:2510.24081, 2025
A participatory commonsense benchmark hand-built by over 350 researchers across 100+ languages, associated with the 5th Multilingual Representation Learning (MRL) workshop; I contributed the Albanian subset. State-of-the-art LLMs do well in aggregate but show accuracy gaps of up to 68% on lower-resource languages.
Paper
Analysis of Word Order Biases in Language Models: A Controlled Investigation Using Artificial and Natural Languages
Varvara Arzt
49th Austrian Linguistics Conference (Österreichische Linguistiktagung, ÖLT 2025), University of Klagenfurt, Austria, December 2025
Information Extraction from German Medieval Charters Abstracts
Sandy Aoun, Varvara Arzt, Daniel Luger, and Georg Vogeler
DH 2024, 19th International Conference of the Alliance of Digital Humanities Organizations, Washington, D.C., USA, August 2024
Exploring the Narratives of Rebetiko via Corpus-Based Analysis
Varvara Arzt
Storytelling, DARIAH Annual Event 2022, Athens, Greece, May 2022
Abstract
Published as V. A. Diveeva; name changed to Arzt in June 2019.
Markirovaniye aktantov dvuchmestnich predikatov v albanskom yazike [Actant marking of two-place predicates in Albanian]
Diveeva, V. A.
Book chapter in S. S. Say (ed.), Valentniye klassi dvuchmestnich predikatov v raznostrukturnich yazikach [Valency classes of two-place predicates], pages 72–88. ILS RAS, Saint Petersburg, Russia, 2018
Paper
Four conference talks on Albanian argument marking, labile verbs and experiencer constructions
International Conferences for Philology Students, Saint Petersburg State University (2012–2014), and the Institute of Linguistic Studies, Russian Academy of Sciences (2014)
