Datasets:

SlayerLab
/

polish-dynaword

The Sejm and Senate part of IPI PAN's Polish Parliamentary Corpus (PPC), rebuilt from its speaker-labelled TEI dump: plenary and committee sittings limited to statements by official speakers (MPs and senators, members of the government, heads of constitutional organs, Sejm and Senate staff; see "Speaker allowlist"), and Sejm interpellations, questions and the government's replies kept whole: 329,863 records / 1,504,127,614 cl100k_base tokens / 4,098,061,466 characters (43,728 sittings, 286,135 interpellation records), 1919-2025. The Sejm from term 9 (2019-) is not here: sejm_api, sejm_committee_transcripts and sejm_interpellations cover it.

Origin and license

Origin: Polish Parliamentary Corpus, Institute of Computer Science, Polish Academy of Sciences (IPI PAN), https://clip.ipipan.waw.pl/PPC: the per-term TEI dump of 2025-05-21, https://kdp.ipipan.waw.pl/static/ppcdump-tei/<term>-tei.zip, 28 zips, 26,409,547,075 bytes, downloaded 2026-10-07 and pinned by size and SHA-256 in src/fetch_parliamentary.py. The texts are the stenographic reports, committee transcripts and interpellations published by the Chancelleries of the Sejm and the Senate (each TEI header names the publisher). Each record's ppc_path in parliamentary.attribution.jsonl pins its document in the dump.
Credit: the speaker turns, speaker labels, roles and dates that this source is cut and filtered by come from IPI PAN's TEI markup. Cite Maciej Ogrodniczuk (2018), "Polish Parliamentary Corpus", Proceedings of the LREC 2018 Workshop ParlaCLARIN: Creating and Using Parliamentary Corpora. PPC's CC BY licence covers its annotations; none of the linguistic annotation layers is shipped, and the speaker labels in the attribution sidecar are credited here.
License: official parliamentary materials are excluded from copyright by art. 4(2) of the Polish Copyright Act; reuse is governed by the Polish Open Data and Reuse of PSI Act. IPI PAN's PPC page states that the parliamentary data is public domain. Asked for its basis, IPI PAN answered (e-mail, 2026-10-07) that it is public information made available by statute (the Open Data Act and the Access to Public Information Act), with no agreement with the Chancelleries. Asked whether the documents it received from the Sejm Library (Biblioteka Sejmowa) came with conditions or limits on further use, IPI PAN answered (e-mail, 2026-10-08) that they were handed over without further restrictions. Record license: public-domain (official documents). No Creative Commons grant is claimed for the text. Not lawyer-signed-off; whether statements of guests at sittings are official documents under art. 4(2) is open (polish-dynaword-v2#54), so they are left out.

What the texts are

Genre: stenographic reports of plenary sittings; transcripts and bulletins of committee sittings; interpellations and questions ("zapytania") of MPs with the government's replies.
Register: parliamentary Polish as recorded and edited by the stenographers - speeches, debate with interruptions, chair announcements, read-out documents - and, in the interpellations, formal letters to and from ministers. The texts of 1919-1989 were digitised from the printed reports (the dump does not say how; each header names an .odt original): samples read clean, interwar spelling is kept ("zależnem"), and some speaker labels carry typos ("Poeł", "Marszalek"). Not machine-translated.

Time span:

Period	House	Kind	Records	Characters	Dates
1919-1939	Sejm	plenary	1,021	156,362,280	1919-02-10 - 1939-09-02
1919-1939	Senat	plenary	323	43,495,751	1922-11-28 - 1939-09-02
1944-1946	Krajowa Rada Narodowa	plenary	6	1,093,152	1944-12-31 - 1946-09-23
1947-1989	Sejm	plenary	460	111,637,467	1947-02-04 - 1989-05-30
1947-1989	Sejm	committee	2,432	62,298,363	1965-09-17 - 1989-05-29
1989-1991	Sejm	plenary	79	47,102,915	1989-07-04 - 1991-10-25
1989-1991	Senat	plenary	61	18,104,530	1989-07-04 - 1991-10-24
1991-2019	Sejm	plenary	2,093	927,132,601	1991-11-25 - 2019-10-16
1991-2019	Sejm	committee	31,360	1,177,647,251	1993-05-23 - 2019-10-16 (one sitting dated 1988, see caveats)
1997-2019	Sejm	interpellations	286,135	1,063,165,695	terms 3-7 undated (1997-10-20 - 2015-11-11 by term); term 8 2015-11-17 - 2019-04-04
1991-2025	Senat	plenary	1,375	346,591,211	1991-11-26 - 2025-04-24
2015-2025	Senat	committee	4,518	143,430,250	2015-11-13 - 2025-04-24

Interpellation records: 110,162 interpellations (251,313,419 characters), 108,433 replies to them (610,085,424), 34,997 questions (60,724,007) and 32,543 replies to questions (141,042,845). 20,447 distinct speaker labels are kept.

Collection and normalization

Every step lives in src/fetch_parliamentary.py; the shard is reproducible from the pinned dump (--download fetches it and refuses a zip whose size or SHA-256 differs).

Scope: every term of the dump except 2019-2023/sejm/ and 2023-2027/sejm/ (5,940 documents), which the sejm_* sources cover; the Senate of 2019-2025 is kept. The Sejm part ends with the sitting of 2019-10-16; sejm_api checks that its own material starts after the end of term 8 (PARLIAMENTARY_SEJM_END = 2019-11-11 in src/sejm_api_common.py).
Records: one record per sitting document (a sitting day, as the dump splits it) or per interpellation, question or reply. A sitting's text is its kept utterances (<u>) joined by newlines, without speaker labels; a dropped turn leaves no marker. The attribution sidecar lists per record the kept speaker labels (speakers), kept_characters and dropped_characters, and the header's system, house, type, term_no and date. An interpellation's text is its paragraphs; its header names the author (279,628 records carry an author; sittings carry none). The dump runs co-authors together in one <author> ("Tomasz GarbowskiTadeusz TomaszewskiBogusław Wontor"); author keeps the string as given.
Dates: created is the sitting or letter date. The dump writes 175 sittings over several days as one date string ("1938-11-28 i 29", "1952-07-31 i 08-01"); they get the range "first, last". The 213,652 interpellation records of terms 3-7 carry no date in the dump and get their term's range (e.g. "1997-10-20, 2001-10-18"); the sidecar's date keeps the dump's string (empty for those).

Speaker allowlist (sittings only, plenary and committee): every speaker in a sitting's header is classified once, by its role followed by a placeholder name ("Jan Maria Kowalski"), against the anchored rules SPEAKER_RULES of src/fetch_sejm_committee_transcripts.py (the sejm_committee_transcripts allowlist) and then PPC_RULES in src/fetch_parliamentary.py. A turn is kept only if its speaker matches a rule; a speaker no rule recognises is dropped (fail closed). A person's own name that is not a 1-3 token name ("Piotr van der Coghen", "Jan Bury s. Antoniego", "Artur Płokszto z Litwy") is kept only if it is, or starts with, the name of an MP, senator or marshal of the same term zip (the chairs' member check, below; 349 persons kept, 673 dropped). PPC's removal of party tags sometimes leaves a cut tag as one more surname part ("(niez"; 1,850 persons), which is ignored. PPC strips the party tags that the base allowlist keys MPs on, so it matches only 3,804 characters of bare mp turns; the PPC rules close that gap. Kept, by rule (characters of utterance text, turns; parliamentary.speaker-roles.csv lists every role with its decision):

Rule	Keeps	Characters	Turns
`mp_no_party`	"Poseł/Posłanka NAME", with "Sprawozdawca", "Sekretarz", "Wiceprzewodniczący"; "Sekretarz poseł NAME"; the typo "Poeł"	1,694,953,227	4,477,657
`state_secretary`	secretaries and undersecretaries of state in a ministry or the Chancellery of the Prime Minister (base rule)	325,593,817	614,984
`senator`	"Senator NAME" (base rule)	315,926,447	883,499
`marshal_no_house`	the chamber's own marshal and deputy marshals without "Sejmu"/"Senatu", "Marszałek Senior"; never a voivodeship marshal	190,540,640	2,407,101
`mp_committee_chair`	committee chairs who are MPs (base rule)	158,002,628	1,038,553
`government_member`	the Prime Minister, deputy PMs, ministers (base rule)	87,399,387	175,562
`committee_chair_named`	a committee chair or deputy chair named in the label, only if the name is an MP's, senator's or marshal's of that term (see below)	46,499,132	439,802
`government_historic`	"Prezydent Ministrów" (1919-21), "Prezes Rady Ministrów", "Wicepremier", ministers under historic names ("Minister Skarbu", "Minister Wyznań Religijnych i Oświecenia Publicznego", "Minister Przemysłu Chemicznego"), "Kierownik Ministerstwa ...", "Minister - Szef Urzędu Rady Ministrów"	41,519,929	78,012
`state_secretary_historic`	secretaries of state of historic ministries and of the Urząd Rady Ministrów (the Prime Minister's office before KPRM)	36,177,142	72,609
`constitutional_organ_head`	RPO, RPD, Prezes NIK, NBP, TK, NSA, Pierwszy Prezes SN, KRRiT, KRS, Prokurator Generalny (base rule)	34,642,977	55,742
`legislative_petitions_staff`	the Sejm's "Przedstawiciel Biura Legislacyjnego", Senate legislators and experts of the Biuro Legislacyjne, petitions staff of the Senate	32,427,046	97,234
`sejm_legislator`	Biuro Legislacyjne legislators (base rule)	20,914,344	73,630
`senator_no_party`	"Senator/Senatorka NAME" with "Sekretarz", "Sprawozdawca", "RP"	10,344,087	43,480
`stage_direction`	the stenographers' bracketed directions ("(Oklaski)", "(Początek posiedzenia o godz. 10)")	8,281,651	471,173
`mp_surname_1919_1939`	"P. Nowicki" in the Sejm of 1919-39 (not before an office)	6,505,299	12,467
`sejm_senate_office_head`	heads of the Sejm and Senate offices (base rule)	5,579,703	13,761
`sejm_bureau_expert`	BAS and BEOS experts (base rule)	3,373,664	5,918
`sejm_senate_marshal`	"Marszałek Sejmu/Senatu" (base rule)	2,502,494	9,317
`senator_surname_1919_1939`	"S. Karnicka" in the Senate of 1922-39	1,792,371	2,567
`komentarz_continuation`	an unbracketed stenographer's line continuing a kept speaker in the same turn block	548,873	28,611
`sejm_committee_secretary`	"Sekretarz Komisji ... NAME" (base rule)	543,439	8,392
`mp`	MP labels with a party tag (base rule)	3,804	13

Evidence for the PPC rules (also in the code comments and on polish-dynaword-v2#54): "P. X" and "S. X" are announced as MPs and senators in the same texts ("Głos ma poseł Nowicki" before "P. Nowicki", "Głos ma s. Karnicka"); "P." before an office in 1919-39 is "Pan" and is cut before the label is classified (12 persons), and "P." in the Senate (17k characters) stays out. A marshal without "Sejmu"/"Senatu" in a sitting of the Sejm or Senate is that chamber's marshal; labels naming a voivodeship, sejmik or board are excluded. A chair's name, and any name that is not a 1-3 token name, is checked against the MPs, senators and marshals of the same term zip (headers of the dropped 2019-2023 Sejm included). Case, punctuation, soft hyphens and one-letter tokens are ignored, two parts run together are split ("JanuszPalikot"), and a name that starts with a member's counts ("Jan Bury s. Antoniego"). A member's name of 4 tokens or with a particle ("Piotr van der Coghen") is taken only from the header of a plenary sitting. Chairs who are not members, such as a foreign or ministry committee's chair, are dropped (876 persons, 6,691,887 characters). The historic ministry words were taken from the label census of the whole dump (2026-10-07).

Dropped, by reason (characters of utterance text):

Reason	Characters	What it is
`speaker:unmatched`	601,219,281	a named speaker whose role matches no rule (117,972 persons)
`speaker:no_role`	84,057,911	a name without a role in the header (12,222 persons): MP or guest cannot be told
`komentarz_colon`	24,479,841	a stenographer's line with ": ", i.e. an interjection ("Głos z sali: ...")
`komentarz_unattributed`	9,757,058	an unbracketed stenographer's line with no speaker before it in its turn block
`speaker:role_only_unmatched`	9,322,256	a role without a name that matches no rule (8,525 persons)
`inline_speaker`	8,416,268	text after a speaker line inside a turn whose label is not kept (see caveats)
`speaker:chair_not_member`	6,691,887	a named chair who is not an MP or senator of that term
`speaker:bad_name`	3,074,332	a name that is not 1-3 tokens and not a member's (673 persons: ambassadors, witnesses, "Dr NAME, Uniwersytet ...", foreign MPs "Poseł Artur Płokszto z Litwy")
`speaker:free_unmatched`	1,782,824	a free-text header label that no rule keeps
`komentarz_continuation_dropped`	1,548,523	a stenographer's line continuing a dropped speaker
`inline_label_line`	352,435	the speaker lines inside turns themselves

The allowlist keeps 3,024,072,101 of 3,774,774,717 characters of utterance text (80.1%). The largest dropped roles: the role-less names above (84.06M), bare "Podsekretarz stanu" (27.39M, 65,361 turns) and "Sekretarz stanu" (6.70M) without a ministry - the President's Chancellery uses the same titles, so they cannot be told apart and fail closed -, "Dyrektor" (16.82M), "Wiceminister" (11.80M, 31,147 turns; not in the allowlist), "Prezes" (9.49M), "Wiceprezes NIK" (5.32M), "Sprawozdawca" without "poseł" (4.64M), "Ekspert Komisji" (3.07M), candidates for ambassador (2.91M), representatives of citizens' legislative initiatives (2.35M), the secretaries of state of the President's Chancellery (2.21M) and witnesses of investigative committees (0.91M). Also out: the RCL, ministry departments, heads of agencies other than the constitutional organs, MEPs and the heads of state ("Naczelnik Państwa", "Prezydent RP").

Stage directions (the "komentarz" speaker): a bracketed direction is kept; any other komentarz line is kept only as the continuation of a kept speaker in the same turn block; a line with ": " (an interjection from the floor) is always dropped.
Speaker lines inside turns: a <u> that ends with ":" and reads as a speaker label switches the speaker until the next speaker change. Such a label is kept only if it matches a rule other than committee_chair_named and either names a speaker that the zip's headers keep or puts a title ("ob.", "ks.", "gen.") before the name; this guards against roles that swallow words of a name ("Minister Spraw Zagranicznych Ukrainy Borys Tarasiuk", "Marszałek Polski Józef Piłsudski"). 8,910 such lines: 315 kept by name, 62 by title, 333 refused by the guard, 8,200 matching no rule.
PII (which scrubber ran): the pattern pass redact() of src/fetch_sejm_committee_transcripts.py - e-mail -> [ADRES_EMAIL], checksum-valid PESEL, NIP, REGON (9 and 14 digits) and IBAN candidates -> [DANE_NUMERY], labelled phone/fax -> [DANE_KONTAKTOWE]. Counted in the shipped shard: 11,626 redactions (6,075 [DANE_KONTAKTOWE], 5,119 [ADRES_EMAIL], 432 [DANE_NUMERY]). This is not SlayerLab/NERGAL: interpellations quote citizens' cases by name, so the shard must be re-run through NERGAL before it is admitted to a release.
Short documents: 440 sittings keep no turn and 164 interpellation records have no text; 37 sittings and 583 interpellation records have under 200 characters after the filter and the PII pass. All are dropped; a dropped sitting's turns stay counted under their rule above.
Encoding: some records are Windows-1250 text that was read as ISO-8859-2, so "ą" shows as "š" and "ś", "ź" as control characters ("Nawišzujšc" for "Nawiązując"). Such a record is re-decoded when no line of it has a correct "ą", "ś" or "ź"; otherwise only its lines without one are re-decoded, and the stray control characters left next to correct text are mapped to the letters they stand for (Poto\x88nik -> "Potočnik"; _C1_MAP in the code) or, for a stray 0x98, removed. Kept records repaired (encoding_repair in the attribution sidecar): 1,170 re-decoded whole, 1,162 of them Sejm interpellation records of 2007-2011, and 26 repaired by line. A record that cannot be re-decoded fails the build. No control character of U+0080-U+009F is left in the shard. Not detected: 7 lines in 5 records whose other lines are correct and which have no control character ("Dane dotyczšce powiatu lubartowskiego przedstawiajš się następujšco:"; 199701-sjm-ioxxx-01425-01, -04177-01, -04543-01, 199701-sjm-zoxxx-02436-01, 200711-sjm-ioxxx-05532-01); a rule on "š" alone would also change correct names ("Babiš", "Miloš").
Dedup: records whose text is identical after the encoding repair and the PII pass are kept once: the first in MANIFEST order of the term zips, then document order. 18,612 records in 2,143 groups are dropped, all but 7 of them interpellation records: mostly an MP's interpellation or question sent to several ministers, each a record of its own, and a ministry's reply filed under each question it answers (the largest group is one reply under 538 questions). 18,543 of the dropped copies have the same author as the kept one. parliamentary.decisions.jsonl lists each dropped id with the duplicate_of id kept in its place. No near-dedup and none against other sources. This source is the reference set that src/clean_sejm_api_2011_2022.py deduplicates sejm_api_2011_2022 against.

Inclusion argument and caveats

What this adds: a century of Polish parliamentary language - interwar, PRL and post-1989 plenary debate, Sejm committee work since 1965, the Senate since 1922, and 1.06B characters of interpellations and ministerial replies. 1.50B tokens.

Changes against the shard shipped up to v0.2.5 (SpeakLeash's unlabelled PPC_corpus: 324,622 records / 4,492,746,014 characters / 1,646,835,986 tokens, no authors): rebuilt from IPI PAN's labelled dump; guests and other non-official speakers removed from the sittings; 1919-1989 and the Senate of 2019-2025 added; the Sejm from 2019 left to the sejm_* sources; interpellation authors recorded; multi-day and undated records dated; misdecoded records repaired; exact copies dropped.

Caveats: the allowlist is the conservative reading of art. 4(2), and it reads the printed role label, not the person: a name slot accepts any 1-3 token name, and only committee chairs and longer or unusual names are checked against the members of the term. Officials whose name is damaged (OCR "Teresa MaIczewska", "Krzysztof Michałkiewicz:", the typo "Piotr Van der Koghen") fail that check: about 0.13M characters in a hand review of the 0.39M characters of such names whose role a rule keeps. Officials the rules miss are dropped with the guests: bare "Podsekretarz stanu"/"Sekretarz stanu" (34.09M characters), "Wiceminister" (11.80M) and the role-less names (84.06M, possibly including MPs; a member check like the chairs' could recover those). The speaker-line check inside turns errs towards dropping: on a character-weighted sample of 80 refused lines (seed 20261007), 47 were guests, 28 were prose or headings ending with a colon (the secretary's "Dekrety przedstawione przez Ministerstwo Skarbu:" in 1919, "Koreferat przedstawił poseł X (PZPR)") and 5 were officials whose label the rules miss, so about two fifths of the 8,416,268 characters dropped after such lines is official text lost. A guest's label that does not stand alone in its utterance is not detected and stays inside the previous turn (not measured). Dates: one term-3 committee sitting (199701-sjm-nusxx-00004-01) is dated 1988-07-02 in the dump and kept as dated; the 213,652 undated interpellation records carry their term's range, not a day. The dump's Sejm interpellations end on 2019-04-04 and the dump has none of term 9, so Sejm interpellations from April 2019 to the start of sejm_interpellations are in no source. Quotations of statutes, prints and citizens' letters are embedded and not filtered (near-dedup against dziennik_ustaw and a benchmark-overlap check remain targets). PII treatment is pattern-based (see above). Not lawyer-signed-off. Training benefit is an untested hypothesis.

Limitations: records are sitting documents, not turns; the order of kept turns is preserved, and ppc_path in parliamentary.attribution.jsonl links each record to the TEI document with the full sitting. Figures: parliamentary.stats.json.