MigrationNarrate: A Dataset for Detection of Migration Narratives in YouTube Videos
MigrationNarrate is a dataset for studying migration narratives in UK-related YouTube videos. It contains 1,115 human-annotated videos labelled using a hierarchical taxonomy of migration narratives, together with a larger set of 4,425 automatically filtered unlabelled videos.
The taxonomy consists of 12 super-narratives and 53 fine-grained narrative labels. Of these, 51 fine-grained narratives are represented in the human-annotated dataset.
The dataset was introduced in:
Fatima Haouari, Carolina Scarton, and Kalina Bontcheva. MigrationNarrate: A Dataset for Detection of Migration Narratives in YouTube Videos. Accepted to the Main Conference of EMNLP 2026. arXiv:2608.20984
MigrationNarrate was created to support research on the detection and analysis of migration narratives in video-based online discourse.
The original collection contains short YouTube videos related to migration in the UK, published between 1 January 2024 and 30 September 2025. Videos longer than three minutes were excluded.
The collection and filtering pipeline resulted in:
| Dataset | Videos | Unique channels | Avg. duration |
|---|---|---|---|
| Filtered corpus | 5,540 | 2,081 | 90.46 sec |
| Human-annotated dataset | 1,115 | 684 | 79.73 sec |
| Unlabelled subset | 4,425 | — | — |
The 1,115 annotated videos are divided into:
| Split | Videos |
|---|---|
| Train | 774 |
| Dev | 117 |
| Test | 224 |
| Total | 1,115 |
MigrationNarrate/
│
├── data/
│ ├── labelled_data.tsv
│ └── unlabelled_data.tsv
│
├── data_splits/
│ ├── train.tsv
│ ├── dev.tsv
│ └── test.tsv
│
├── taxonomy/
│ └── migration_narratives_taxonomy.tsv
│
├── scripts/
│ └── extract_transcripts.py
│
├── requirements.txt
└── README.md
Contains the identifiers and human annotations for the 1,115 annotated YouTube videos.
| Field | Description |
|---|---|
videoId |
YouTube video identifier |
super_narrative |
Human-annotated super-narrative |
narrative |
Human-annotated fine-grained narrative |
The annotation None indicates that, after human annotation/adjudication, no migration narrative was explicitly stated or strongly implied in the video.
Contains the identifiers of the 4,425 migration-related videos retained by the automatic filtering pipeline but not selected for human annotation.
These data can be useful for, for example, weakly supervised, semi-supervised, or further manual annotation experiments.
Contains the hierarchical migration narrative taxonomy used for annotation, including the 12 super-narratives and 53 narrative labels and their definitions.
The taxonomy is based on the migration narrative taxonomy introduced in the European Commission Joint Research Centre (JRC) report Navigating migration narratives: Research insights and strategies for effective communication.
Videos were collected using two complementary strategies:
Channel- and playlist-based collection. Videos were collected from migration-related and broader political/news channels and playlists using the YouTube Data API v3.
Search-based collection. Additional videos were retrieved using migration-related search phrases in order to increase the diversity of sources represented in the corpus.
The initial collection was subsequently filtered to retain videos:
- published between 1 January 2024 and 30 September 2025;
- with a duration of three minutes or less;
- identified as migration-related using semantic similarity filtering.
- Pre-annotated with GPT5.1 to select a diverse sample.
After this process, 5,540 videos remained in the filtered corpus.
A diverse subset of 1,115 videos was then selected for human annotation.
For full details of the collection, filtering, sampling, and annotation procedures, please refer to the paper.
MigrationNarrate uses a hierarchical annotation scheme.
Each video is first assigned a super-narrative and then, where applicable, a corresponding fine-grained narrative.
The annotation taxonomy contains:
- 12 super-narratives
- 53 fine-grained narratives
The human-annotated dataset contains instances belonging to 51 of the 53 fine-grained narrative categories, in addition to examples labelled None.
Each annotation batch was independently annotated by two annotators, with disagreements resolved through adjudication.
We do not redistribute the original YouTube videos or audio files.
Instead, the dataset provides YouTube video identifiers that can be used to locate videos that remain publicly available on YouTube.
Please note that YouTube videos may be removed, or become unavailable over time. Consequently, it may not always be possible to access all videos included in the original dataset.
We do not redistribute the generated video transcripts.
The transcripts used in the paper were generated locally from the video files using faster-whisper (large-v3).
Requirements
Install the required Python packages with:
pip install -r requirements.txt
The requirements.txt file contains the Python dependencies required by the transcription script, including faster-whisper and tqdm.
The script also requires FFmpeg to be installed and available on the system path.
Generating Transcripts
We provide a script for generating transcripts from locally available video files:
python scripts/extract_transcripts.py videos/The videos/ directory should contain the videos as MP4 files, with each file named using its corresponding YouTube video ID:
videos/
├── VIDEO_ID_1.mp4
├── VIDEO_ID_2.mp4
└── VIDEO_ID_3.mp4
For example:
videos/abc123.mp4
For each video, the script generates a .txt transcript with the same video ID:
videos/abc123.mp4
videos/abc123.txt
The script automatically detects the language of each video and generates the transcript using faster-whisper large-v3.
Please note that reconstructed transcripts may differ slightly from those used in the original experiments depending on software/model versions and the availability or version of the underlying video.
If you use MigrationNarrate in your research, please cite:
@article{haouari2026migrationnarrate,
title = {MigrationNarrate: A Dataset for Detection of Migration Narratives in YouTube Videos},
author = {Haouari, Fatima and Scarton, Carolina and Bontcheva, Kalina},
journal = {arXiv preprint arXiv:2608.20984},
year = {2026}
}The paper was accepted to the Main Conference of EMNLP 2026.
The MigrationNarrate dataset and annotations are released under the CC BY-NC-SA 4.0 license for non-commercial research use.
For questions about the dataset, please open an issue in this repository or contact the authors of the paper.