Title: An Open-Source, Universal Dependencies-Compatible Workflow for Parsing Katharevousa Greek Parliamentary Records

Abstract:

Despite the critical role of Katharevousa Greek in legal, administrative, and parliamentary archives, contemporary Natural Language Processing (NLP) tools have largely failed to address its specific needs. This paper introduces a reproducible framework designed to construct and assess a Universal Dependencies-style parsing resource based on parliamentary questions from Greece’s early post-junta era. Our approach integrates a series of rigorous steps: OCR-aware text reconstruction, LLM-assisted annotation bound by strict schema constraints, automated validation processes, deterministic CoNLL-U snapshotting, a fixed data split for evaluation, and a comparative analysis across different model families.

The resulting reference corpus, which has been automatically validated and frozen, comprises 1,697 sentences. These are divided into a training set of 1,357 sentences and a held-out test set of 340 sentences. Under a uniform scoring protocol, we benchmarked several systems, including off-the-shelf parsers for Greek and Ancient Greek, a feature-based parser, mBERT, XLM-R, and custom Stanza models.

Our results highlight a significant register mismatch in existing off-the-shelf systems. For instance, the strongest external baseline, spaCy’s Greek parser, achieved only a Labelled Attachment Score (LAS) of 0.4183. In contrast, the top-performing structural model, an XLM-R variant, attained an UPOS accuracy of 0.8893, a dependency-relation F1 score of 0.7250, an Unlabeled Attachment Score (UAS) of 0.6098, and a LAS of 0.5162. This represents an absolute LAS improvement of 0.0980 over the best external baseline. Additionally, the feature-based model proved competitive in UPOS and relation labeling, suggesting that transparent lexical-context features remain relevant at this scale of data.

Beyond quantitative metrics, this work offers an auditable methodology for converting challenging historical parliamentary OCR data into reusable syntactic NLP infrastructure. To ensure transparency and reproducibility, we release the entire pipeline—including code, annotation schemas, the frozen reference annotations, the fixed train/test split, and per-model benchmark reports—as an open-access companion to this publication.

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