hakai-vulpes/congreso-opinion-publica

Parliamentary speech and public opinion in Spain (2020–2026)

Dataset Summary

Every speech turn in the plenary sessions of the Spanish Congreso de los Diputados (XIV and XV legislatures), with its cleaned transcript, speaker, parliamentary group, party and role (government / opposition), aligned month by month with public opinion (CIS barometers), registered unemployment (SEPE) and inflation (INE CPI).

• 42,075 speech turns from 458 plenary session days, 2020-01-05 → 2026-09-23, about 22.6 M tokens of Spanish.
• Monthly context: 75 CIS barometers, 117 months of unemployment, 105 months of CPI (14 series).
• Every row carries its provenance: source URL, download date and SHA-256 of the original file.

The data are in Spanish, and so are the column names and category values, except the common fields id, timestamp and text of intervenciones; this card explains them in English.

Purpose. The dataset was built to test whether the way deputies speak (figures, readability, questions, orientation to the future) changes with the economy and public opinion, beyond what they talk about. It was created by hakai-vulpes and Maverz-0 as a course project for Descubrimiento de Conocimiento en Datos Complejos (Grado en Ciencia de Datos e Inteligencia Artificial, Universidad Politécnica de Madrid). It received no specific funding. Each instance is a speech turn (text and metadata) or a monthly observation of a context series.

How to use

from datasets import load_dataset

turns = load_dataset("hakai-vulpes/congreso-opinion-publica", "intervenciones", split="full")
monthly = load_dataset("hakai-vulpes/congreso-opinion-publica", "alineacion_mensual", split="full")

import pandas as pd

turns = pd.read_parquet("hf://datasets/hakai-vulpes/congreso-opinion-publica/data/intervenciones/intervenciones.parquet")

Keys. id ({legislatura}-{num_diario}-{orden}) links intervenciones and metricas_texto; mes (YYYY-MM) links the monthly tables; grupo_id links intervenciones and grupos.

Dataset Details

Tables

Each table is a separate config (subset) with a single split, full. There is no train/test partition: how to split depends on the analysis (by legislature, by date, by speaker…), and the columns to do it are included.

Config	Grain	Rows	Columns	Content
intervenciones	speech turn	42,075	28	Every speech turn in the plenary sessions of the Congreso de los Diputados (XIV and XV legislatures): cleaned transcript, speaker, parliamentary group, party, role, agenda item and provenance.
metricas_texto	speech turn	42,075	10	Model-free text metrics for each turn in intervenciones (regular expressions and counts, reproducible).
alineacion_mensual	month × role × session type	310	42	Monthly aggregates of the text metrics by speaker role, joined with public opinion (CIS), unemployment (SEPE) and inflation (INE) for the same month and with 1- and 2-month lags.
cis_mensual	month	75	18	National aggregates of the CIS monthly barometer (public-opinion survey). Weighted with the CIS weight when available (from mid-2020). Only aggregates: the microdata are not redistributed.
sepe_mensual	month × province	117	14	Registered unemployment (SEPE), monthly, by economic sector. This release has the national total only (provincia_ine = ES).
ipc_mensual	month × ECOICOP group	1,457	9	Consumer Price Index (INE table 76125, base 2025, ECOICOP v2): general index and 13 groups.
grupos	group × legislature	19	4	Master table of parliamentary groups per legislature.
gobiernos	government period	3	5	Periods of each government and the groups supporting it; used to assign rol. Caretaker periods are assigned to the outgoing government.

Data dictionary

intervenciones: speech turn, 42,075 rows

Column	Type	Null	Description
id	string	0 %	Stable identifier {legislatura}-{num_diario}-{orden}.
legislatura	string	0 %	Legislature: XIV (Dec 2019 – Aug 2023) or XV (Aug 2023 –).
num_diario	int64	0 %	Issue number of the Diario de Sesiones (DSCD) within the legislature.
num_sesion	int64	2.9 %	Plenary session number printed on the cover; null when not printed.
timestamp	date	0 %	Session date, when the turn was delivered and published. The Diario de Sesiones gives no time of day per turn.
mes	string	0 %	Session month, YYYY-MM: join key with the monthly tables.
tipo_sesion	string	0 %	Pleno (plenary). Committee sessions are supported by the pipeline but not part of this release.
comision	string	100.0 %	Committee name. Always null in this release (plenary only).
tipo_iniciativa	string	44.4 %	Type of initiative under debate, from the case-number prefix (pregunta oral pleno, interpelación urgente, proyecto de ley, real decreto-ley…); otro for unmapped prefixes; null for turns outside any agenda item (oaths, votes, points of order).
expediente	string	44.4 %	Congress case number (NNN/NNNNNN) of the agenda item.
punto_orden_dia	string	0.5 %	Title of the agenda item, as printed.
orden	int64	0 %	Position of the turn within the issue (1-based, counted before the length filter).
orador	string	0 %	Speaker as printed: surnames for deputies, name for government members and guests.
titulo_orador	string	0 %	Speaker heading as printed (e.g. MINISTRA DE HACIENDA).
cargo	string	0 %	Speaker type from the heading: diputado (deputy), gobierno (government member), presidencia (Speaker and Bureau), otro (guests: investiture candidates, regional parliaments, Ombudsman…).
grupo_id	string	28.9 %	Parliamentary group (key of grupos), matched by surname within the legislature. Null for non-deputies and for missing or ambiguous matches.
formacion	string	28.9 %	Party or electoral list of the deputy (congreso.es), same matching rule as grupo_id.
circunscripcion	string	28.9 %	Electoral district (province) of the deputy, same matching rule.
provincia_ine	string	28.9 %	INE province code (2 digits) of the electoral district.
rol	string	0 %	gobierno (government members, and deputies of the groups supporting the government on that date, see gobiernos), oposición (opposition), presidencia (Speaker and Bureau), otro (guests, and deputies whose group could not be matched).
text	string	0 %	Cleaned transcript: page headers removed, hyphenation undone, stage directions removed (counted in n_aplausos, n_rumores, n_protestas).
n_tokens	int64	0 %	Whitespace-separated tokens in intervenciones.text.
n_aplausos	int64	0 %	(Aplausos) (applause) annotations in the turn.
n_rumores	int64	0 %	(Rumores) (murmurs) annotations in the turn.
n_protestas	int64	0 %	(Protestas) (protests) annotations in the turn.
url_pdf	string	0 %	URL of the source PDF.
fecha_descarga	date	0 %	Date the source PDF was downloaded.
hash_pdf	string	0 %	SHA-256 of the source PDF.

metricas_texto: speech turn, 42,075 rows

Column	Type	Null	Description
id	string	0 %	Key of intervenciones (id).
n_tokens	int64	0 %	Whitespace-separated tokens in intervenciones.text.
n_frases	int64	0 %	Sentences, split at ., ! and ? (thousand/decimal separators ignored).
long_media_frase	float64	0 %	Mean sentence length, in tokens.
legibilidad_szigriszt	float64	0 %	Szigriszt-Pazos readability: 206.835 − 62.3·syllables/words − words/sentences. Heuristic syllabification.
densidad_cifras	float64	0 %	Numbers (including %, €, millones…) per 1,000 tokens.
prop_preguntas	float64	0 %	Share of sentences that are questions (0–1).
prop_primera_persona	float64	0 %	First-person markers (yo, nosotros, nuestro, verbs in -mos…) per 1,000 tokens.
orientacion_temporal	float64	0 %	(Future markers − past markers) per 1,000 tokens. Positive = future-oriented.
version_pipeline	string	0 %	Version of the metric definitions (v0: regex-based, no models).

alineacion_mensual: month × role × session type, 310 rows

Column	Type	Null	Description
mes	string	0 %	Month, YYYY-MM.
rol	string	0 %	Speaker role, as in intervenciones.rol.
tipo_sesion	string	0 %	Session type, as in intervenciones.tipo_sesion.
n_intervenciones	int64	0 %	Turns aggregated in the row.
long_media_frase_media	float64	0 %	Monthly mean of long_media_frase.
legibilidad_szigriszt_media	float64	0 %	Monthly mean of legibilidad_szigriszt.
densidad_cifras_media	float64	0 %	Monthly mean of densidad_cifras.
prop_preguntas_media	float64	0 %	Monthly mean of prop_preguntas.
prop_primera_persona_media	float64	0 %	Monthly mean of prop_primera_persona.
orientacion_temporal_media	float64	0 %	Monthly mean of orientacion_temporal.
long_media_frase_p90	float64	0 %	Monthly 90th percentile of long_media_frase.
legibilidad_szigriszt_p90	float64	0 %	Monthly 90th percentile of legibilidad_szigriszt.
densidad_cifras_p90	float64	0 %	Monthly 90th percentile of densidad_cifras.
prop_preguntas_p90	float64	0 %	Monthly 90th percentile of prop_preguntas.
prop_primera_persona_p90	float64	0 %	Monthly 90th percentile of prop_primera_persona.
orientacion_temporal_p90	float64	0 %	Monthly 90th percentile of orientacion_temporal.
valoracion_presidente	float64	8.1 %	Mean rating (1–10) of the Prime Minister, Pedro Sánchez. Excludes doesn't know him, DK and NA.
valoracion_lider_pp	float64	8.1 %	Mean rating (1–10) of the PP leader (Pablo Casado, then Alberto Núñez Feijóo).
valoracion_lider_vox	float64	8.1 %	Mean rating (1–10) of the VOX leader, Santiago Abascal.
valoracion_lider_sumar_up	float64	8.1 %	Mean rating (1–10) of the Unidas Podemos / Sumar leader (Pablo Iglesias in XIV, Yolanda Díaz in XV).
sit_economica_mala_pct	float64	8.1 %	% rating Spain's economic situation as bad or very bad.
sit_politica_mala_pct	float64	96.5 %	% rating Spain's political situation as bad or very bad. Asked only in a few barometers: mostly null.
problema_1_paro_pct	float64	6.8 %	% naming unemployment as Spain's main problem (first mention).
problema_1_politica_pct	float64	6.8 %	% naming a politics-related category as Spain's main problem (first mention).
intencion_voto_psoe_pct	float64	9.4 %	Direct vote intention for PSOE, % of all interviews (not the CIS vote estimate).
intencion_voto_pp_pct	float64	9.4 %	Direct vote intention for PP, % of all interviews.
intencion_voto_vox_pct	float64	9.4 %	Direct vote intention for VOX, % of all interviews.
paro_registrado	int64	0 %	Registered unemployed, national total (SEPE).
paro_var_interanual	float64	0 %	Year-on-year change in registered unemployment, %.
ipc_interanual	float64	0 %	Year-on-year change in the general CPI, %.
valoracion_presidente_dif	float64	15.8 %	Month-on-month change in valoracion_presidente.
paro_var_interanual_lag1	float64	0 %	paro_var_interanual 1 month earlier.
ipc_interanual_lag1	float64	0 %	ipc_interanual 1 month earlier.
valoracion_presidente_lag1	float64	9.0 %	valoracion_presidente 1 month earlier.
valoracion_presidente_dif_lag1	float64	18.1 %	valoracion_presidente_dif 1 month earlier.
sit_economica_mala_pct_lag1	float64	10.3 %	sit_economica_mala_pct 1 month earlier.
paro_var_interanual_lag2	float64	0 %	paro_var_interanual 2 months earlier.
ipc_interanual_lag2	float64	0 %	ipc_interanual 2 months earlier.
valoracion_presidente_lag2	float64	9.0 %	valoracion_presidente 2 months earlier.
valoracion_presidente_dif_lag2	float64	18.1 %	valoracion_presidente_dif 2 months earlier.
sit_economica_mala_pct_lag2	float64	9.0 %	sit_economica_mala_pct 2 months earlier.
mes_del_anio	int64	0 %	Calendar month (1–12), for seasonality.

cis_mensual: month, 75 rows

Column	Type	Null	Description
mes	string	0 %	Fieldwork month, YYYY-MM. There is no barometer in August.
num_estudio	int64	0 %	CIS study number.
n_entrevistas	int64	0 %	Number of interviews.
valoracion_presidente	float64	1.3 %	Mean rating (1–10) of the Prime Minister, Pedro Sánchez. Excludes doesn't know him, DK and NA.
valoracion_lider_pp	float64	1.3 %	Mean rating (1–10) of the PP leader (Pablo Casado, then Alberto Núñez Feijóo).
valoracion_lider_vox	float64	1.3 %	Mean rating (1–10) of the VOX leader, Santiago Abascal.
valoracion_lider_sumar_up	float64	1.3 %	Mean rating (1–10) of the Unidas Podemos / Sumar leader (Pablo Iglesias in XIV, Yolanda Díaz in XV).
sit_economica_mala_pct	float64	1.3 %	% rating Spain's economic situation as bad or very bad.
sit_politica_mala_pct	float64	94.7 %	% rating Spain's political situation as bad or very bad. Asked only in a few barometers: mostly null.
problema_1_paro_pct	float64	0 %	% naming unemployment as Spain's main problem (first mention).
problema_1_politica_pct	float64	0 %	% naming a politics-related category as Spain's main problem (first mention).
intencion_voto_psoe_pct	float64	2.7 %	Direct vote intention for PSOE, % of all interviews (not the CIS vote estimate).
intencion_voto_pp_pct	float64	2.7 %	Direct vote intention for PP, % of all interviews.
intencion_voto_vox_pct	float64	2.7 %	Direct vote intention for VOX, % of all interviews.
fuente	string	0 %	Source institution.
url_origen	string	0 %	URL of the source file or API call.
fecha_descarga	date	0 %	Date the source file was downloaded.
hash_fichero_origen	string	0 %	SHA-256 of the downloaded source file.

sepe_mensual: month × province, 117 rows

Column	Type	Null	Description
mes	string	0 %	Month, YYYY-MM.
provincia_ine	string	0 %	INE province code; ES = national total.
parados	int64	0 %	Registered unemployed.
var_mensual	float64	0.9 %	Month-on-month change, % (recomputed from the series).
var_interanual	float64	10.3 %	Year-on-year change, % (recomputed from the series).
parados_agricultura	int64	0 %	Registered unemployed, agriculture.
parados_industria	int64	0 %	Registered unemployed, industry.
parados_construccion	int64	0 %	Registered unemployed, construction.
parados_servicios	int64	0 %	Registered unemployed, services.
parados_sin_empleo_anterior	int64	0 %	Registered unemployed with no previous job.
fuente	string	0 %	Source institution.
url_origen	string	0 %	URL of the source file or API call.
fecha_descarga	date	0 %	Date the source file was downloaded.
hash_fichero_origen	string	0 %	SHA-256 of the downloaded source file.

ipc_mensual: month × ECOICOP group, 1,457 rows

Column	Type	Null	Description
mes	string	0 %	Month, YYYY-MM.
grupo_ecoicop	string	0 %	General or ECOICOP v2 group name (in Spanish).
indice	float64	< 0.1 %	Index, base 2025 = 100. The INE publishes the latest index one month after the rates, so it may be null in the latest month.
var_mensual	float64	0 %	Month-on-month change, %.
var_interanual	float64	0 %	Year-on-year change, %.
fuente	string	0 %	Source institution.
url_origen	string	0 %	URL of the source file or API call.
fecha_descarga	date	0 %	Date the source file was downloaded.
hash_fichero_origen	string	0 %	SHA-256 of the downloaded source file.

grupos: group × legislature, 19 rows

Column	Type	Null	Description
grupo_id	string	0 %	Short group identifier, stable across legislatures.
nombre_oficial	string	0 %	Official group name.
legislatura	string	0 %	Legislature.
n_diputados	int64	0 %	Deputies who belonged to the group in that legislature (congreso.es, substitutes included).

gobiernos: government period, 3 rows

Column	Type	Null	Description
inicio	date	0 %	Start date.
fin	date	33.3 %	End date; null = ongoing.
presidente	string	0 %	Prime Minister.
grupos_gobierno	string	0 %	;-separated grupo_id of the groups in government.
nota	string	0 %	Notes.

Coverage

Legislature	Dates	Session days	Speech turns	Tokens
XIV	2020-01-05 → 2023-08-16	261	24,754	12.2 M
XV	2023-08-17 → 2026-09-23	197	17,321	10.4 M

Turns by role: oposición 19,036 · presidencia 11,557 · gobierno 10,985 · otro 497. 98.7% of the deputies' turns have a parliamentary group assigned.

This is the full population, not a sample: every plenary issue of the Diario de Sesiones published on congreso.es for both legislatures (471 issues), minus the turns under 30 tokens. Issue numbers missing on the server are not reconstructed (XIV: 1, 2, 19, 25, 83, 121, 136; XV: 23, 41, 137, 163). tipo_sesion is always Pleno in this release. There are no labels for subpopulations beyond the descriptive columns (legislature, role, group, party, constituency, initiative type).

How it was built

The pipeline is a set of idempotent Jupyter notebooks with four data layers (the code will be published later):

Collection. The authors downloaded all files between 2026-10-02 and 2026-10-04: PDFs from congreso.es, CIS microdata and SEPE files over HTTP, and the INE CPI and the congreso.es deputy register through their JSON APIs. No sampling and no manual annotation were involved. There was no ethics review: the sources are public records of public officials and published statistics, and only aggregates of the survey are kept.

Layer	Content	Published
raw	Files as downloaded: 471 PDF issues of the Diario de Sesiones, 75 CIS microdata files, SEPE and INE files, with a manifest (URL, date, SHA-256).	No: re-downloadable from the URLs in each row.
cleaned	One table per source: PDFs parsed into 81,839 speech turns with cleaned text; CIS microdata harmonised across studies; SEPE and INE series tidied.	No: CIS microdata cannot be redistributed.
processed	Speakers matched to groups, parties and roles; scope filter; text metrics; monthly aggregates; alignment.	No: working layer.
curated	This release: fixed schema, documented columns, Parquet only.	Yes.

Main steps:

1. Parsing. Each PDF is split into speech turns at the speaker headings (El señor …: / La señora …:), discarding cover, summary and page headers. The agenda item and case number are taken from the last heading in capitals before the turn.
2. Cleaning. Hyphenation is undone and stage directions in brackets (Aplausos, Rumores, Protestas…) are counted and removed.
3. Speakers. The transcripts do not say a deputy's group. It is taken from the congreso.es deputy register by matching surnames within the legislature; ambiguous matches are left empty rather than guessed. rol combines the speaker type with the groups in government on the session date (gobiernos).
4. Scope. Turns under 30 tokens (greetings, Gracias, presidenta, procedural turns) are dropped: 81,839 parsed → 42,075 published. They are many rows but a small share of the text. The Speaker's turns are kept (rol = presidencia) so they can be filtered.
5. Context series. CIS variables are matched across studies by their labels (variable names change between studies) and aggregated per study; SEPE changes are recomputed from the series; the INE CPI comes from the INE API.

Limitations and biases

• Plenary only: committees are not included in this release.
• PDF extraction errors are possible (headings split across lines, unusual layouts). The parser was checked against raw heading counts on sample issues.
• Group from the current register: congreso.es lists each deputy's current group, so a deputy who changed group during the legislature appears with the latest one.
• Speech in co-official languages (Catalan, Basque, Galician) appears as transcribed in the Diario.
• Unequal weight of roles: the government and the Speaker take many more turns than any single group.
• CIS gaps: there is no barometer in August; leader ratings are missing in September 2023, vote intention in June–July 2023, and the political-situation question is asked only in a few barometers.
• Extreme metric values: some turns have no sentence punctuation (the Speaker reading vote results or roll calls, lists of names read in the chamber), so long_media_frase reaches several hundred words and legibilidad_szigriszt, which is unbounded for long sentences, goes far below zero. They are real values, kept as they are; filter rol = presidencia to remove most of them.
• Residual noise: a few hundred lowercase parentheticals remain in the text; most are the speaker's own (sic, nueva, names addressed), a small tail are stage directions the rules did not catch.
• Correlation, not causation: the monthly alignment supports exploring associations; it does not identify causal effects.

Personal data and responsible use

The only people named are public officials acting in their public role (deputies, government members), whose interventions are published by the Congress itself. Processing rests on public interest (GDPR art. 6.1.e). Only aggregates of the CIS survey are included, so there is no data about survey respondents. The data must not be used to profile individuals or to generate speeches attributed to real deputies.

The speeches express political opinions, a special category under GDPR art. 9. They are opinions that public officials made public themselves in a public session (art. 9.2.e). The text is a record of political debate and contains heated and sometimes offensive language: insults between members, references to terrorism, violence and named victims. It is reproduced as published and not filtered.

The tables are self-contained. The only external resources are the source URLs in each row (url_pdf, url_origen), useful for checking a value against the original, not needed to use the data.

Licence and attribution

This dataset is released under CC BY 4.0. When you use it, cite this dataset and the original sources:

Source	Data	Terms (checked 2026-10-04)
Congreso de los Diputados	Diario de Sesiones (Pleno), deputy register	Reuse allowed if the content is not altered or misrepresented, citing the source and the update date
Centro de Investigaciones Sociológicas (CIS)	Monthly barometers (aggregates only)	Reuse and transformation allowed under Ley 37/2007, citing the source; microdata not redistributed here
Servicio Público de Empleo Estatal (SEPE)	Registered unemployment	Reuse and transformation allowed under Ley 37/2007, citing the source
Instituto Nacional de Estadística (INE)	CPI, table 76125	CC BY 4.0

The speech text is reproduced verbatim except for extraction normalisation (spacing, hyphenation, stage directions removed and counted apart); the authoritative version of each session is the signed PDF linked in url_pdf. The text metrics and monthly aggregates are the authors' own derived work. None of the sources endorses this dataset, and derived values must not be presented as official figures.

Authors

hakai-vulpes, Maverz-0.

Citation

DOI: 10.57967/hf/10759

@misc{congreso_opinion_publica_v1,
  title  = {Parliamentary speech and public opinion in Spain (2020--2026)},
  author = {hakai-vulpes and Maverz-0},
  year   = {2026},
  url    = {https://huggingface.co/datasets/hakai-vulpes/congreso-opinion-publica},
  doi    = {10.57967/hf/10759},
  note   = {Version v1}
}

Cite the sources as well:

• Congreso de los Diputados, Diario de Sesiones del Congreso de los Diputados. Pleno y Diputación Permanente, XIV and XV legislatures (session date in each row).
• Origen de los datos: Centro de Investigaciones Sociológicas (study numbers in cis_mensual.num_estudio).
• Origen de los datos: SEPE.
• Elaboración propia con datos extraídos del sitio web del INE: www.ine.es.

Acknowledgements

To the Congreso de los Diputados, the CIS, the SEPE and the INE for publishing their data openly. No funding was received.

Maintenance

The authors maintain the dataset. Questions, errata and contributions go through the Community tab of this repository. Fixes are released as new versions listed in the changelog; earlier versions stay available as revisions of the repository (revision= in load_dataset). There is no fixed update schedule; new sessions may be added in later versions.

Changelog

• v1 (2026-10-04): id_intervencion, fecha_sesion and texto renamed to id, timestamp and text (id_intervencion also in metricas_texto); DOI added. 52 MB of Parquet in total.
• v0 (2026-10-04): first release, DOI 10.57967/hf/10758.