Jvachier/meeting-notes-extractive-v1

Meeting Notes Extractive v1

Meeting transcripts in which every sentence carries two labels:

• importance (0/1): whether the sentence belongs in extractive meeting notes;
• zone: context, discussion, decision, action_item or closing.

Action-item sentences in the synthetic layer also carry character spans for the owner and the deadline.

Source code, seeds and the full build pipeline: https://github.com/jvachier/meeting-notes-extractive-corpus (built from commit b8587d90e23d).

Configurations

Config	Content
synthetic (default)	Sentences of the synthetic meetings (source = synthetic or synthetic_hn)
synthetic_notes	One row per synthetic meeting
ami	Sentences (dialogue acts) of 137 AMI scenario meetings, labelled from human annotation
ami_notes	One row per AMI meeting, with the human abstractive summary sections

from datasets import load_dataset

synthetic = load_dataset("Jvachier/meeting-notes-extractive-v1", "synthetic", revision="v1.1")
ami = load_dataset("Jvachier/meeting-notes-extractive-v1", "ami", revision="v1.1")

Statistics

Config	Split	Meetings	Sentences	importance = 1
synthetic	train	1,460	121,566	0.262
synthetic	validation	170	14,956	0.259
synthetic	test	170	14,854	0.249
ami	train	97	75,996	0.175
ami	validation	20	16,476	0.175
ami	test	20	15,876	0.186

How it was built

Synthetic layer. Claude Code (an AI coding assistant) wrote a bank of short labelled meeting segments once, by hand, following fixed briefs and labelling rules. Each segment has a paraphrase variant; a segment and its variants form a family. The bank also contains hard negatives: sentences that look like decisions or action items but are not (rejected proposals, tentative ideas, ownerless tasks, reversed decisions, decisions reported from elsewhere, hypotheticals, unanswered questions). A deterministic Python program then composed the segments into full meetings of five types (quick sync, standup, project review, stakeholder update, retrospective) across eight business domains. It maps speakers to participants, fills slots (names, projects, dates, figures) from lexicons of invented values, inserts backchannels, and renders an ASR-style copy of every sentence. No language model runs in the pipeline, and the program never creates or changes a label.

Splits are assigned per seed family (80/10/10), so no text from a test family appears in a training meeting.

AMI layer. Dialogue acts from the AMI Meeting Corpus manual annotations (v1.6.2), labelled from AMI's human extractive summaries and their links to the abstractive summary:

Dialogue act	importance	zone	confidence
Extracted, linked to a decisions summary sentence	1	decision	0.9
Extracted, linked to an actions summary sentence	1	action_item	0.9
Extracted, linked to a problems summary sentence	1	discussion	0.9
Extracted, linked to the abstract only, or not linked	1	discussion	0.8
Not extracted	0	discussion	0.8

AMI splits follow the standard scenario partition (test: ES2004, ES2014, IS1009, TS3003, TS3007; validation: ES2003, ES2011, IS1008, TS3004, TS3006).

Columns (sentences)

Column	Type	Description
note_id	string	Meeting id
source	string	synthetic, synthetic_hn or ami
meeting_type	string	Synthetic meeting type, or ami_design
domain	string	Synthetic domain, or product_design
split	string	train, validation, test
sentence_id	int32	0-based position in the meeting
speaker	string	Participant name, or AMI speaker letter
sentence_text	string	Clean text, cased and punctuated
sentence_text_asr	string	Lowercased, unpunctuated, light disfluencies (synthetic only)
importance	int8	0 or 1
confidence	float32	Label confidence (hard negatives 0.6–0.8; AMI 0.8–0.9)
zone	string	Sentence zone
segment_id, family_id	string	Seed provenance (fill-* for backchannels; null for AMI)
hard_negative	bool	Sentence comes from a hard-negative segment
token_count	int32	Whitespace word count
spans	list<struct<start, end, role>>	Character offsets of action-item owner / deadline

notes columns: note_id, source, meeting_type, split, n_sentences, n_words, n_participants, and for AMI summary_abstract, summary_decisions, summary_actions, summary_problems.

Known limitations

• Synthetic style. The synthetic meetings come from one author's seed bank and are cleaner and more regular than real speech. Evaluate on the AMI test split.
• Seed reuse. Each seed segment appears in about 23 meetings (maximum 72), with different names, values, speakers and order. Splits are by seed family, so test text never appears in training meetings.
• Topic coherence. A long synthetic meeting draws on segments from several domains, so it can move between unrelated topics more than a real meeting would.
• Slot values. Names, dates and figures come from lexicons; a figure can occasionally be implausible for its context.
• AMI domain. AMI meetings are role-played product-design meetings, with no context or closing annotation; zone metrics on AMI cover discussion, decision and action_item.
• Annotator agreement. AMI extractive summaries have modest agreement between annotators (κ ≈ 0.48), which the confidence column reflects.
• Backchannels and opening/closing small talk are labelled discussion, importance 0.

Licence and attribution

CC BY 4.0. Names, companies and products in the synthetic layer are invented or very common values. The AMI-derived layer is derived from the AMI Meeting Corpus (CC BY 4.0); please cite:

Carletta, J. et al. (2005). The AMI Meeting Corpus: A Pre-announcement. Machine Learning for Multimodal Interaction (MLMI 2005), LNCS 3869, Springer.

Provenance: version 1.1.0, generation seed 1729, commit b8587d90e23d830375525a51273626ddc9e1efea.