COLM 2026 · Dataset

🌊 TIDES
A Longitudinal Bilingual Dataset for Modeling Multi-Party Social Dynamics

Heechan Lee*,1 Jeonggyu Kang*,1 Junho Myung1 Jaywoong Jeong1 Juho Kim1,2 Joseph Seering1 1 KAIST  ·  2 SkillBench * Equal contribution

We followed 12 student teams over one semester to capture how real-world collaboration and social dynamics develop over time.

LongitudinalIn-the-wildMulti-partyCollaborative
🤗 Dataset arXiv · soonOpenReview · soon
TEAM DYNAMICS OVER TIME
FORMING NORMING PERFORMING

Utterances, emergent roles, and team development stages are linked across each team’s semester.

12teams
88meetings
104transcripts
75,971utterances
7 KR / 5 ENteams by language
01

Motivation

Why another multi-party dialogue dataset?

Models have difficulty tracking subtle, changing social dynamics in multi-party conversation.

A high-quality dataset of group conversations is necessary for training and evaluating group conversation ability, but existing datasets fall short in three ways.

TIDES answers each limitation directly: authentic in-the-wild recordings, a full semester of longitudinal coverage, and roles that emerge from behavior.

Existing datasets · Limitation 01

Scripted or synthesized

TV/movie scripts, lab observation, or LLM-generated dialog — lacking the authenticity of real conversation.

TIDES: Authentic

In-the-wild recordings of real student teams working on actual course projects.

Existing datasets · Limitation 02

Short-term conversations

Usually a single session under one hour — long-term evolving social dynamics stay invisible.

TIDES: Longitudinal

The same 12 teams followed across a full semester of meetings and chats.

Existing datasets · Limitation 03

Fixed, assigned roles

Professor–student, parent–child — emergent roles like coordinator or critic never surface.

TIDES: Emergent roles

Roles grounded in observed behavior: dominance, sociability, and task orientation.

02

Dataset

Three annotation layers for studying social dynamics.

Real teams, followed over time.

We tracked 12 university student teams as they worked on authentic course projects over one semester. TIDES combines synchronous meetings and asynchronous communication with repeated team surveys.

75,971Utterances

Time-aligned, speaker-anonymized transcripts with English text and retained Korean source text.

SurveysSatisfaction & roles

Meeting satisfaction plus dominance, sociability, and task-orientation ratings.

11 teamsChat logs

Asynchronous team messages connected to the same semester-long collaborations.

KR + ENBilingual teams

Seven Korean-source and five English-source teams across varied project settings.

Annotation layers

Each layer is grounded in an established theory of teamwork and group development.

Emergent role

Participant roles based on dominance, sociability, and task orientation.

TRIAD model · Driskell et al., 201713 role labels
“ ”

Interaction type

Utterance-level labels describing how each turn contributes to teamwork.

Modified act4teams-SHORT · Klünder et al., 202015 interaction labels

Development stage

Meeting-level labels locating a team along its longitudinal development.

Group development · Tuckman, 19655 development stages
Bilingual data

All transcripts provide English text. Korean-source teams additionally retain the original Korean utterances when available.

7 Korean-source teams  ·  5 English-source teams
03

Dataset examples

Explore three excerpts from Korean- and English-source team meetings.

Each name identifies a unique person within a team; the same name appearing in different teams refers to different people.

Speaker / timeUtteranceInteraction type
English translations with Korean source text9 utterances
04

Citation

This is a placeholder and will be updated after publication.

Citation placeholder — full BibTeX will be updated after publication.
KAIST CSTL KIXLAB SkillBench