ishiki-labs/multi-party-dialogue

Dataset Card for Multi-Party Dialogue

Dataset Details

Dataset Description

This dataset is designed for proactive decision prediction (SPEAK or REMAIN SILENT) in multi-party conversations: given a conversational context and a current utterance, the task is to predict whether a specific participant (the target speaker) will speak next (SPEAK) or remain silent (SILENT).

Each sample captures a decision point in a conversation — a moment where the target speaker either takes the floor or stays silent. The dataset is balanced (50/50 SPEAK/SILENT) and spans three domains: meeting transcripts (AMI, SPGI) and TV show dialogues (Friends).

• Curated by: Ishiki Labs
• Language(s): English
• License: Apache 2.0

Dataset Sources

• AMI Corpus: Meeting recordings and transcripts: AMI Corpus
• Friends Corpus: TV show transcripts from the sitcom Friends Friends Corpus
• SPGI2.0 Corpus: Earnings call transcripts from S&P Global SPGI2.0

Dataset Structure

Splits and Sizes

Dataset	Train	Val	Test	Total
AMI	11,270	1,408	1,410	14,088
Friends	10,289	1,286	1,287	12,862
SPGI	79,432	9,929	9,929	99,290

Fields

Field	Type	Description
decision_point_id	string	Unique identifier for the decision point
sequence_id	string	ID of the conversation sequence
meeting_id	string	Source meeting or episode ID
target_speaker	string	The speaker whose next-turn behavior is predicted
all_speakers	list[str]	All active participants in the conversation
turn_index	int	Index of the current turn in the sequence
turn_type	string	Whether the current turn is addressing or context
source	string	Annotation source (e.g. explicit, friends_explicit, spgi_heuristic)
is_explicit	bool	Whether the addressee is explicitly marked
inference_confidence	int	Confidence score (1–10) for the addressee inference
context_turns	list[dict]	Prior turns with speaker and text fields
current_turn	dict	The current utterance with speaker and text
addressees_in_current	list[str]	Speakers explicitly addressed in the current turn
target_is_addressed	bool	Whether the target speaker is directly addressed
target_spoke_next	bool	Ground truth: did the target speaker take the next turn?
decision	string	Label: SPEAK or SILENT
confidence	string	Confidence of the label (high, medium, low)
reason	string	Human-readable explanation for the label
category	string	Fine-grained label category (see below)
num_context_turns	int	Number of context turns provided
num_query_words	int	Word count of the current turn
current_turn_text_length	int	Character length of the current turn
is_filler	bool	Whether the current turn is a filler utterance

Label Categories

Category	Description
SPEAK_explicit (I1)	Target is directly addressed and speaks next
SPEAK_implicit (I2)	Target speaks next without being explicitly addressed
SILENT_no_ref (S1)	Target is not addressed and does not speak next
SILENT_ref (S2)	Target is referenced/addressed but does not speak next

Files per Split

Each split contains:

• {split}_samples.jsonl — the main samples
• {split}_samples_with_reasoning.jsonl — same samples with an added reasoning field explaining the label
• filtering_summary.json — summary statistics from dataset curation
• stage4_filtered_samples.jsonl — all filtered samples before train/val/test splitting

Dataset Creation

Curation Rationale

Multi-party conversations require participants (and models) to track who is being addressed and anticipate who will speak next. Existing dialogue datasets largely focus on two-party settings. This dataset was created to support research on proactive, context-aware response prediction in realistic multi-party settings with 2–7 participants.

Source Data

Data Collection and Processing

1. AMI Corpus — Meeting transcripts with explicit addressee annotations. Decision points were extracted at each turn where a speaker change could occur.
2. Friends Corpus — TV show scripts with scene-level speaker turns. Explicit addressee mentions (e.g., "Hey Joey, ...") were used to label addressee information.
3. SPGI Corpus — Earnings call transcripts. Addressees were inferred heuristically from conversation structure (e.g., question-answer pairs, moderator patterns).

Each decision point consists of a fixed-length context window, a current utterance, and a binary label indicating whether the target speaker took the next turn. The dataset was filtered to remove low-confidence samples and balanced to 50/50 SPEAK/SILENT via stratified sampling.

Who are the source data producers?

• AMI Corpus: Recorded and annotated by the AMI Consortium (University of Edinburgh and collaborators)
• Friends Corpus: Transcripts from the TV show Friends (NBC)
• SPGI Corpus: Earnings call transcripts from S&P Global Market Intelligence

Annotation Process

• AMI / Friends: Labels are derived from ground-truth speaker-turn sequences and explicit addressee annotations in the original corpora.
• SPGI: Labels are derived heuristically from turn-taking patterns; inference_confidence reflects annotation certainty.
• A reasoning field (in *_with_reasoning.jsonl files) provides a natural language explanation for each label, generated to support chain-of-thought training.

Personal and Sensitive Information

• AMI and SPGI data may contain real names of meeting/call participants. No additional anonymization has been applied beyond what is present in the original corpora.
• Friends data is fictional dialogue from a TV show.

Citation

If you use this corpus, please cite our work:

@misc{bhagtani2026speakstaysilentcontextaware,
  title={Speak or Stay Silent: Context-Aware Turn-Taking in Multi-Party Dialogue},
  author={Bhagtani, Kratika and Anand, Mrinal and Xu, Yu Chen and Yadav, Amit Kumar Singh},
  year={2026},
  archivePrefix={arXiv},
  url={https://arxiv.org/abs/2603.11409}
}