Request Correction

Beyond OCR: Structural Segmentation and Speaker Attribution in Historical Italian Parliamentary Debates

Historical parliamentary debates are essential for longitudinal political and linguistic research, yet much early material remains available only as scanned images. In the Italian context, proceedings from 1848–1996 lack large-scale, structurally annotated, machine-readable representations. This paper addresses the challenge of transforming historical Italian parliamentary debates into structured corpora by moving beyond plain Optical Character Recognition (OCR) toward functional block segmentation and speaker attribution. We present detailed annotation guidelines and a manually annotated dataset of 300 randomly sampled pages. Two approaches are compared: (i) direct multimodal Large Language Model (LLM) annotation and (ii) a modular pipeline combining OCR with LLM-based structural reconstruction under zero-shot and few-shot prompting. Evaluation on a held-out test set shows that separating transcription from structural reasoning improves performance, with few-shot prompting yielding the most reliable results. The study demonstrates the feasibility of integrating LLM-based reasoning into historical parliamentary digitisation workflows.