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

• Paper: Coming soon
• Website: Coming soon
• Contact: heechan@kaist.ac.kr,

🌊 TIDES is an in-the-wild, longitudinal, multi-party collaboration dialogue dataset collected by tracking Korean university student teams over the course of a semester. TIDES includes transcripts and chat logs from meetings conducted by participants as they worked on real course projects throughout the semester, along with satisfaction surveys. Rather than simply providing raw dialogue, we offer emergent role, utterance type, and team development annotation layers to capture the development of social dynamics in multi-party collaboration settings.

Speakers are anonymized (PERSON_A, PERSON_B, …). The corpus includes seven primarily Korean-speaking teams (Team_1, Team_2, Team_3, Team_6, Team_7, Team_9, and Team_10) and five primarily English-speaking teams (Team_4, Team_5, Team_8, Team_11, and Team_12). All transcripts provide English text (Some participants speak in Korean seldomly!); Korean-source teams additionally include the original utterances in korean_text.

Scale	Count
Teams	12 (Team_1 … Team_12)
Meeting transcript files	104
Utterance segments	~75,971
Chat exports	11 teams (Team_3 has no chat file)
Approx. size	~27 MB

Some meetings are split across multiple files (…_Pt1.json, …_Pt2.json, …). Metadata lists meetings by date (88 dated meetings), so the number of JSON files can exceed the number of metadata meeting entries.

---

Directory layout

TIDES/
├── README.md
├── metadata/
│   ├── meeting_metadata.json      # per-team project blurb, meeting topics, speaker map
│   ├── satisfaction_metadata.json # per-meeting satisfaction survey aggregates
│   └── meeting_roles.csv          # per-participant role ratings per meeting
└── Team_N/
    ├── TeamN_YYYYMMDD.json        # meeting transcript (+ optional "_PtK" suffix)
    ├── TeamN_YYYYMMDD_PtK.json
    └── chat/
        └── TeamN_chat_anonymized.{txt,csv,xlsx}

---

Meeting transcripts (Team_N/*.json)

Each file is a single JSON object:

{
  "segments": [ /* ordered utterances */ ]
}

Segment schema

Field	Type	Required	Description
start	number	yes	Utterance start time (seconds)
end	number	yes	Utterance end time (seconds)
speaker	string	yes	Anonymized speaker id (PERSON_A, …)
text	string	yes	Utterance text (English)
utterance_type	string	yes	Dialogue-act / interaction label (see below)
annotation_source	string	yes	"gemma" (model) or "human" (gold)
korean_text	string	no	Original Korean utterance when retained
qwen_utterance_type	string	no	Secondary model (Qwen) label

Coverage notes

• annotation_source: mostly gemma (70k); human (5.7k). Human segments usually omit korean_text and qwen_utterance_type.
• korean_text: present on a majority of segments, not all.
• qwen_utterance_type: present on most gemma segments; absent on typical human segments.

Example segment

{
      "start": 76.56,
      "end": 80.3,
      "speaker": "PERSON_B",
      "text": "It's a bit different from what we had in mind.",
      "utterance_type": "Linking Solutions",
      "annotation_source": "gemma",
      "korean_text": " 근데 저희가 생각하는 이미지랑 좀 다르세요.",
      "qwen_utterance_type": "Linking Problems"
}

utterance_type vocabulary

Based on modified version of act4teams-SHORT (Klünder et al., 2020)

Label	~Count
Giving Information	26810
Active listening	9946
Linking Solutions	9260
Naming Solutions	4917
Other / Neutral	4703
Structuring	4643
Proactivity	3912
Naming Problems	3157
Social / Humor	2510
Linking Problems	2233
Cooperation	1812
Knowledge Transfer	878
Task/Process Negative	642
Social Negative	465
Linking & Connecting	83

---

Metadata

metadata/meeting_metadata.json

Top-level keys: Team_1 … Team_12.

Team_N:
  team_number: int
  project: string                    # short project / context description (may be Korean)
  speaker_id_to_pseudonym: {         # PERSON_* → English given-name pseudonym
    "PERSON_A": "Alex",
    ...
  }
  meetings: [
    {
      date: "YYYY-MM-DD",
      main_topics: [string, ...],    # participant- or annotator-written topic summaries
      participants: ["PERSON_A", ...],
      team_development_stage: string # optional; e.g. Forming, Storming, Norming, Performing, Adjourning
                                     # (occasionally a comma-combined value)
    },
    ...
  ]

Team development stage (Based on Tuckman's stages of group development (1965)): Forming, Storming, Norming, Performing, Adjourning.

metadata/satisfaction_metadata.json

Top-level keys: Team_1 … Team_12.

Team_N:
  team_number: int
  meetings: [
    {
      date: "YYYY-MM-DD",
      satisfaction: [
        {
          question: string,   # bilingual prompt text
          average: number | null,
          scores: [number, ...]   # may be empty when not collected
        },
        ...
      ]
    },
    ...
  ]

metadata/meeting_roles.csv

One row per participant × meeting.

Column	Description
Team_Number	Team id (1–12)
Meeting_Sequence	Meeting index within the team
Meeting_Date	YYYY-MM-DD
Participant_Name	PERSON_* id
Dominance_Average	Dominance score
Sociability_Average	Sociability score
Task_Orientation_Average	Task-orientation score
assigned_role	Nearest role label (see below)
role_distance	Distance to assigned role centroid

assigned_role values (Based on TRIAD model (Driskell et al., 2017)): Attention Seeker, Coordinator, Critic, Evaluator, Follower, Negative, Power Seeker, Problem Solver, Social, Task Completer, Task Motivator, Team Leader, Teamwork Support.

---

Team chat (Team_N/chat/)

Asynchronous team messaging exports (KakaoTalk-style), with participant names replaced by English pseudonyms consistent with speaker_id_to_pseudonym where applicable.

Team	File	Format
1, 2, 5, 6, 7, 8, 9	TeamN_chat_anonymized.txt	Plain-text chat export
10, 11, 12	TeamN_chat_anonymized.csv	CSV
4	TeamN_chat_anonymized.xlsx	Excel
3	—	Missing in this release

---

Anonymization & ethics notes

• Meeting speakers use PERSON_* ids; metadata maps them to English pseudonyms for readability.
• Chat logs use the same pseudonym style; residual indirect identifiers may still exist in free text—review before redistribution if your release policy requires it.
• Survey free-text / topic fields may contain Korean descriptions of course projects.

---

Citation

Add citation / license / paper link here before publishing to Hugging Face.