brjoey/ecb-press-conferences

ECB Monetary Policy Press Conference Transcripts

Full transcripts of every ECB Governing Council monetary policy press conference from June 1998 to the present, structured paragraph-by-paragraph with an introductory statement / Q&A split and journalist-vs-ECB-official speaker labels.

Source: European Central Bank (ecb.europa.eu) — original text © European Central Bank, Frankfurt am Main, Germany. Reproduction permitted provided the ECB is cited as the source. See ECB copyright notice. Coverage: June 1998 – present (294 conferences as of June 2026) Update frequency: Automatically updated after each Governing Council meeting (8 times/year)

Dataset structure

Column	Type	Description
url	string	Source URL on ecb.europa.eu
date	string	ISO date YYYY-MM-DD
location	string	City where the conference was held
participants	string	ECB President and Vice-President (full names and titles)
president	string	ECB President name, e.g. "Christine Lagarde". Empty for the handful of non-standard briefings chaired by a board member.
vice_president	string	ECB Vice-President name, e.g. "Luis de Guindos". Also filled when a Vice-President chaired in the President's absence.
has_qa	bool	Whether a Q&A session is present (9 early conferences have no Q&A)
intro_paragraphs	list[str]	Body paragraphs of the introductory statement (headings excluded)
intro_headings	list[dict]	Section headings: {text, after_paragraph_index} — after_paragraph_index is the index of the last body paragraph before this heading (−1 if the heading precedes all body paragraphs). Present from 2021 onwards.
intro_full_text	string	Full introductory statement as plain text (body paragraphs only)
n_is_paragraphs	int	Number of IS body paragraphs
qa_turns	list[dict]	All Q&A turns: {text, role} where role is question, answer, or unknown
qa_questions	list[str]	Journalist questions only
qa_answers	list[str]	ECB official answers only
qa_full_text	string	Full Q&A session as plain text
n_qa_turns	int	Total number of Q&A turns

Speaker classification

Each Q&A paragraph carries a role label — "question" (journalist or reporter) or "answer" (ECB President or Vice-President).

Pipeline

Period	Method
Pre-2013 (approx. 140 conferences)	Regex on explicit "Question:" / speaker-name prefixes present in original HTML
2013–present	Kimi K2 LLM via HF Inference API (zero-shot, full-transcript context, two-pass verify)

Validation status

Pre-2013 (regex): Labels are derived directly from explicit speaker markers in the source HTML.

2013–2023 (LLM, Draghi and early Lagarde era): Validated against a manually-annotated data set of 100 press conferences. Some labeling errors were found and manually corrected.

2024–June 2026 (Lagarde era, automatic): All 20 conferences from January 2024 through June 2026 were also inspected after the LLM classification.

Future updates beyond June 2026: Classified automatically by the same LLM pipeline. Not manually inspected. The models seems to perform better in the entire Lagarde era (2021–2026) and indicates that future updates will be reasonable accurate.

Despite the validation, residual errors may exist. If you find a mislabeled paragraph, please open a discussion on the dataset community tab.

Usage

Load the dataset

from datasets import load_dataset

ds = load_dataset("brjoey/ecb-press-conferences", split="train")

# Introductory statement
is_texts = ds["intro_full_text"]    # list[str] — one string per conference
is_paras = ds["intro_paragraphs"]   # list[list[str]] — paragraph-level

# Q&A — all turns with speaker labels
all_turns = [
    {"date": row["date"], "role": t["role"], "text": t["text"]}
    for row in ds
    for t in row["qa_turns"]
]

# Journalist questions / ECB answers flat
all_questions = [q for conf in ds["qa_questions"] for q in conf]
all_answers   = [a for conf in ds["qa_answers"]   for a in conf]

Working with IS section headings

Modern conferences (from 2021 onwards) structure the introductory statement under named sections such as Economic activity, Inflation, and Monetary policy stance. These are stored in intro_headings (present from 2021 onwards) so you can work with body paragraphs alone or reconstruct the full labelled structure:

from datasets import load_dataset

ds = load_dataset("brjoey/ecb-press-conferences", split="train")

# Reconstruct IS with inline section labels
def is_with_headings(row: dict) -> list[str]:
    paras    = row["intro_paragraphs"]
    headings = {h["after_paragraph_index"]: h["text"] for h in row["intro_headings"]}
    result   = []
    if -1 in headings:
        result.append(f"## {headings[-1]}")
    for i, para in enumerate(paras):
        result.append(para)
        if i in headings:
            result.append(f"## {headings[i]}")
    return result

# Latest conference
print("

".join(is_with_headings(ds[0])))

# Drop headings and work with body paragraphs only (default behaviour)
body_only = ds["intro_paragraphs"]   # list[list[str]]

Related ECB data

The ECB also publishes a speeches dataset containing all speeches by ECB Executive Board members (plus some pre-ECB European Monetary Institute speeches) as a single monthly-updated CSV:

• Page: https://www.ecb.europa.eu/press/key/html/downloads.en.html
• Direct download: https://www.ecb.europa.eu/press/key/shared/data/all_ECB_speeches.csv
• Format: UTF-8 CSV with | separator; columns: date, speakers, title, subtitle, full text
• Coverage: EMI era – present; updated monthly

Issues and contributions

Found a labeling error or a missing conference? Please open a discussion on the dataset community tab.

License

The transcript texts are © European Central Bank, Frankfurt am Main, Germany. Reproduction is permitted provided the ECB is cited as the source. See the ECB copyright notice.