---
dataset_info:
  features:
  - name: meeting_id
    dtype: string
  - name: abstractive
    struct:
    - name: abstract
      list: string
    - name: actions
      list: string
    - name: decisions
      list: string
    - name: problems
      list: string
  - name: participants_summary
    struct:
    - name: A
      struct:
      - name: abstract
        list: string
      - name: actions
        list: string
      - name: decisions
        list: string
      - name: problems
        list: string
    - name: B
      struct:
      - name: abstract
        list: string
      - name: actions
        list: string
      - name: decisions
        list: string
      - name: problems
        list: string
    - name: C
      struct:
      - name: abstract
        list: string
      - name: actions
        list: string
      - name: decisions
        list: string
      - name: problems
        list: string
    - name: D
      struct:
      - name: abstract
        list: string
      - name: actions
        list: string
      - name: decisions
        list: string
      - name: problems
        list: string
  - name: topics
    list: string
  - name: key_speakers
    list:
    - name: speaker_id
      dtype: string
    - name: speech
      dtype: string
    - name: name_entities
      list:
      - name: text
        dtype: string
      - name: type
        dtype: string
  - name: utterance
    list:
    - name: speaker
      dtype: string
    - name: start_time
      dtype: float64
    - name: end_time
      dtype: float64
    - name: text
      dtype: string
  splits:
  - name: train
    num_bytes: 9648986
    num_examples: 139
  download_size: 9666002
  dataset_size: 9648986
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
---

# Dataset Card for AMI_annotation

## Dataset Details

### Dataset Description
**AMI_annotation** is a structurally parsed and annotated multi-modal meeting corpus derived from the AMI Meeting Corpus manual annotations. It features detailed meeting-level metadata, abstractive meeting summaries, individual participant summaries mapped by roles/speakers, segmented timestamps, speaker-specific text transcriptions, integrated Named Entity Recognition (NER), and discussion topic vectors.

- **Curated by:** Rabindra Nath Nandi
- **Language(s) (NLP):** English
- **License:** [More Information Needed]

## Dataset Structure

### Data Instances
Each row in the dataset represents a unique meeting instance indexed by its `meeting_id`. 

### Data Fields
* `meeting_id` (string): Unique identifier for the meeting session (e.g., `ES2002a`).
* `abstractive` (struct): Meeting-level abstractive summaries containing:
  * `abstract` (list of strings): High-level meeting summary sentences.
  * `actions` (list of strings): Stated action items and directives.
  * `decisions` (list of strings): Concluded decisions made by the team.
  * `problems` (list of strings): Open issues or problems discussed.
* `participants_summary` (struct): Role-specific summaries for participants (`A`, `B`, `C`, `D`), tracking individual perspectives on abstract, actions, decisions, and problems.
* `topics` (list of strings): Chronological meeting agenda and discussion topics.
* `key_speakers` (list): Speaker profile records containing:
  * `speaker_id` (string): Identifier code for the speaker.
  * `speech` (string): Consolidated full-text monologue or contributions.
  * `name_entities` (list): Extracted entities mapped to types (`text`, `type`).
* `utterance` (list): Time-aligned utterance blocks containing:
  * `speaker` (string): Speaker label.
  * `start_time` (float64): Utterance start timestamp in seconds.
  * `end_time` (float64): Utterance end timestamp in seconds.
  * `text` (string): Transcribed text segment.

## Uses

### Direct Use
Designed for training and evaluating Audio Large Language Models (AudioLLMs), automatic meeting summarization, multi-party dialogue analysis, and named entity recognition tasks.

## Dataset Creation

### Curation Rationale
Processed using automated XML parsers to structure legacy AMI corpus annotations into clean, standardized JSON table rows compatible with the Hugging Face `datasets` ecosystem.

## Bias, Risks, and Limitations
Corpus reflects natural multi-party meeting dynamics, containing colloquial speech patterns, interruptions, and subjective participant summaries.