TBMM Parliamentary Proceedings Corpus with Speaker-Turn Segmentation, Party Linkage, and Validated Metadata, 1950–2023

A speech-level corpus of Turkish Grand National Assembly (TBMM) parliamentary proceedings covering 4 January 1950 through 23 April 2023 — Turkey's post-1950 parliamentary era, including periods of democratic competition, military interruption, and constituent assemblies. The corpus contains 2,418,313 speaker turns across five chambers (TBMM, Millet Meclisi, Senato, Danışma Meclisi, MGK).

Version 1.1 introduces a redesigned record-linkage pipeline. A failure-mode analysis of v1.0 revealed a candidate-generation ordering error: the matcher tested name uniqueness across the entire roster before applying the legislative-term filter, so MPs serving multiple terms were discarded as ambiguous before scoring. Three deterministic rescue passes correct this — exact name+term matching with province used only for disambiguation; title-stripping, initial-expansion and ordered-token-subsequence name variants; and historical-party completion. No fuzzy string matching is used.

Party linkage rises from 58.8% to 84.2% of eligible speaker turns (1,005,329 turns; identity linkage 84.25%, 1,006,117 turns). The analytical core grows from 136,118 to 233,435 floor speeches (82.9 million words, 6,547 unique linked speakers, 30 parties). Decade-level party linkage now reaches 93.5% (1990s), 96.0% (2000s), 92.3% (2010s) and 97.3% (2020s); the bicameral era (1960s–70s) remains lower because the MP roster does not cover the Senate.

A stratified manual validation of 180 newly linked turns yields a conservative precision estimate of 98.3% (177 confirmed correct, none confirmed incorrect, three unresolved pending roster verification). An independent automated release audit — covering file integrity, core-definition compliance, coverage, selection bias, text quality and linkage logic — ships with this version and reports zero failures. The audit also identified 161 core rows whose length is inflated by parsing artifacts (end-of-sitting annexes attached to the final speaker turn); these are preserved but flagged row-level in known_issue_rows.csv.

The corpus combines two sources. For 1950–2018 it builds on the Turkronicles OCR archive (Yazar et al. 2025, CC BY 4.0), adding speaker-turn segmentation, role classification, MP–party linkage, gender annotation and confidence tiers. For the 27th Legislative Term (2018–2023) it contributes an original extraction of 314,899 speaker turns from 537 official Word-format session transcripts published by TBMM.

Speech text, boundaries and all v1.0 linkage layers are unchanged; v1.0 remains archived under its own DOI for reproducibility. This release adds a stable speech_id, a self-contained METHODS document, a data dictionary, a match-layer dictionary, a party crosswalk, year-level linkage reports, row-level issue flags, the validation sample, and a SHA-256 manifest. Known limitations are documented in the README. Released under CC BY 4.0.