IndoVSE-dataset — Indonesian Video Salient Entity Dataset

IndoVSE-dataset is the first Indonesian benchmark dataset for named entity salience detection in news videos. Each row represents a named entity extracted from an Indonesian news video, annotated with binary salience labels and four temporal features derived from word-level timestamps.

Dataset Summary

Statistic	Value
Total entity instances	27,115
Unique videos	573
Salient entities	2,193 (8.1%)
Non-salient entities	24,922 (91.9%)
Video source	YouTube (Metro TV, Kompas TV, tvOne)
Language	Bahasa Indonesia
Domains	Politik, Ekonomi, Kesehatan, Pendidikan
Inter-annotator agreement	Cohen's κ = 0.931
License	CC BY 4.0

Dataset Structure

Fields

Field	Type	Description
vid_id	string	YouTube video ID
entity_text	string	Surface form of named entity
entity_label	string	Entity type (see below)
freq_norm	float	Normalized mention frequency [0, 1]
burstiness	float	Temporal burstiness score (-1, 1)
first_appear	float	Normalized first appearance position [0, 1]
coverage	float	Fraction of video duration spanned [0, 1]
salient	int	Binary salience label: 1 = salient, 0 = not

Entity Labels

Label	Description	Count
PER	Person	4,335
ORG	Organization	4,013
PRD	Product	3,147
NOR	Norm/Rule	2,702
CRD	Cardinal number	1,832
PRC	Percentage	1,675
GPE	Geopolitical entity	1,494
EVT	Event	1,414
DAT	Date	1,254
MON	Money	1,109
QTY	Quantity	917
LAW	Law/regulation	794
ORD	Ordinal	751
LOC	Location	744
REG	Religion	633
Others	WOA, FAC, TIM, LAN	301

Temporal Features

Each entity's temporal features are derived from word-level timestamps produced by OpenAI Whisper:

Mention Frequency (freq_norm) Normalized mention count: freq / max_freq_in_video. Higher = mentioned more often relative to other entities in the same video.

Temporal Burstiness (burstiness) B = (σ_τ − μ_τ) / (σ_τ + μ_τ) where τ is inter-mention intervals. B > 0 = clustered mentions (bursty). B < 0 = evenly spaced. Set to 0 for single-mention entities.

First Appearance (first_appear) t₁ / D — normalized position of first mention. Lower = introduced earlier in video.

Coverage Spread (coverage) (t_k − t₁) / D — fraction of video duration spanned by mentions. Higher = discussed throughout the video.

Data Collection

Videos were collected from three major Indonesian news YouTube channels: Metro TV, Kompas TV, and tvOne, covering news from 2024 across four domains:

Domain	Videos
Politik	~200
Ekonomi	~150
Kesehatan	~120
Pendidikan	~103

Processing Pipeline

1. Audio extraction — ffmpeg
2. Speech-to-text — OpenAI Whisper large-v3 (language='id', word_timestamps=True)
3. NER — cahya/bert-base-indonesian-NER with aggregation_strategy='simple'
4. Temporal feature computation — derived from word-level timestamps
5. Human annotation — binary salience label per entity per video

Annotation Protocol

Annotators were shown the video title, domain, and entity list with temporal features. An entity was labeled salient (1) if its removal would substantially change a viewer's understanding of the video's primary topic. Otherwise labeled non-salient (0).

Inter-annotator agreement was measured using Cohen's Kappa (κ = 0.931), indicating near-perfect agreement.

Usage

from datasets import load_dataset

ds = load_dataset("galihkjaya/IndoVSE-dataset")
print(ds[0])
# {
#   'vid_id': '61o1g-mxCmo',
#   'entity_text': 'Pemilu 2024',
#   'entity_label': 'EVT',
#   'freq_norm': 0.5455,
#   'burstiness': 0.0981,
#   'first_appear': 0.0439,
#   'coverage': 0.2188,
#   'salient': 1
# }

Model

The IndoVSE salience classifier trained on this dataset is available at: galihkjaya/IndoVSE

Citation

@misc{indovse2025,
  title     = {IndoVSE: Temporal Entity Salience Detection for Indonesian News Videos},
  author    = {[Galih Kusuma Wijaya]},
  year      = {2026},
  publisher = {HuggingFace},
  url       = {https://huggingface.co/datasets/galihkjaya/IndoVSE-dataset}
}

License

This dataset is released under CC BY 4.0.