ACloudCenter/ami-addressee

AMI Addressee (stage-1 extraction)

Addressee-labeled dialogue acts extracted from the AMI Meeting Corpus manual annotations v1.6.2 (groups.inf.ed.ac.uk/ami/download/, CC BY 4.0) for training/evaluating addressee detection ("who is this utterance addressed to") as a candidate proposer for multi-party floor control.

Scale — important caveat

The manual annotations mark addressee only on a 22-meeting subset (76 speaker files). Across all 139 meetings / 117,915 dialogue acts, exactly 8,876 dacts carry addressee labels; the remaining 109K have no addressee markup. Downstream papers reporting "8.9K AMI DAs with addressee" (e.g. Malik et al.) refer to this same subset.

Schema

• data/records.jsonl — the 8,876 labeled dialogue acts.
• data/records_all.jsonl — all 110,795 dialogue acts that resolve to words (labeled or not; addressee_raw is null and labeled is false for unlabeled ones). Use this for building conversational-context windows: the labeled subset is only ~8% of the dialogue flow, and windows need the backchannels/fragments in between.
• data/participants.json — meeting -> sorted participant channels; all 139 meetings are 4-participant.
• Per-record fields: meeting_id, speaker, da_id, da_type, addressee_raw, other, unexplained, text, start_time, end_time, n_words, labeled.

Label distribution (top of data/stats.json)

Multi-speaker addressees dominate (~6,190 of 8,876, e.g. D,C,B), single-speaker ~3,600. This differs from the Jovanović addressing-behaviour subset (61.7% individual / 34.2% group) — treat cross-subset comparability with care.

Provenance

build_ami_addressee.py in this repo reproduces data/* from the official zip.