Skip to content
oduwsdlPublic

About

A Benchmark for Detecting Social Media Content in TV News Archives

Resources

Stars

0 stars

Watchers

0 watching

Forks

Repository files navigation

Social2TV

A Benchmark for Detecting Social Media Content in TV News Archives

Social2TV is a manually curated multimodal gold-standard dataset for detecting social media references in U.S. cable television news. It covers 36 hours of prime-time programming from CNN, MSNBC, and Fox News across five dates (one per year, 2020–2024), with labels for:

  • Social media logos and post screenshots visible on TV screen
  • Verbal mentions of social media in closed captions

The underlying video and caption text remain in the Internet Archive TV News Archive. This repository releases annotations, timestamps, and episode identifiers that link back to those broadcasts.


Table of Contents


Quick links to dataset files

Logos and post screenshots

Positive visual annotations (frames that contain a social media logo and/or a post screenshot):

Channel File Direct download
CNN gold_standard_images_cnn.csv Download
Fox News gold_standard_images_foxnews.csv Download
MSNBC gold_standard_images_msnbc.csv Download

Folder: LogoAndPostScreenshots/

Closed captions

Gold-standard labels for every caption segment (without redistributing caption text):

Channel File Direct download
CNN gold_standard_cc_cnn.csv Download
Fox News gold_standard_cc_foxnews.csv Download
MSNBC gold_standard_cc_msnbc.csv Download

Folder: ClosedCaptions/


Overview

Social media content frequently appears in television news through platform logos, post screenshots, and verbal references, extending its reach beyond the original platform. These moments typically occupy only a small fraction of airtime, yet they matter because they show when posts, platforms, and online activity are selected, contextualized, and amplified for TV audiences.

Social2TV addresses the lack of a shared evaluation resource for this sparse, multimodal phenomenon. Beyond the benchmark itself, it demonstrates a reusable workflow for enriching television archives with searchable, research-ready metadata.

Contributions

  1. A multimodal gold standard covering 36 hours of CNN, MSNBC, and Fox News from five selected dates (2020–2024).
  2. An extensible annotation workflow adaptable to new dates, channels, platforms, countries, and research domains.
  3. Open evaluation resources for detecting visual and caption-based social media references (see Related code).

Repository structure

Social2TV/
├── LogoAndPostScreenshots/          # Visual gold standard (logos + post screenshots)
│   ├── gold_standard_images_cnn.csv
│   ├── gold_standard_images_foxnews.csv
│   └── gold_standard_images_msnbc.csv
├── ClosedCaptions/                  # Closed-caption gold standard (no caption text)
│   ├── gold_standard_cc_cnn.csv
│   ├── gold_standard_cc_foxnews.csv
│   └── gold_standard_cc_msnbc.csv
└── LICENSE

Corpus

Property Value
Networks CNN, Fox News, MSNBC
Window U.S. prime time, 8:00–11:00 p.m. Eastern Time
Selected dates (ET) 2020-03-13, 2021-01-09, 2022-04-24, 2023-01-21, 2024-01-31
Episodes 36 one-hour news programs (15 Fox News, 12 MSNBC, 9 CNN)
Frames reviewed ~130,000 one-second frames
Visual positives released 1,445 annotated frames with logo and/or screenshot
Caption segments 59,982 labeled segments

Dates were chosen via GDELT Television Explorer peak-mention searches to enrich for social media references. Social2TV is a positive-enriched benchmark for developing and evaluating detectors—not a probability sample for estimating overall prevalence across TV news.

Archive episode IDs use UTC timestamps (prime-time Eastern often falls on the following UTC calendar day). Example: March 13, 2020, 8–9 p.m. ET → CNNW_20200314_030000_....

Included episodes

Fox News (15)
Date (ET) 8–9 PM 9–10 PM 10–11 PM
2020-03-13 Fox News at Night With Shannon Bream Tucker Carlson Tonight Hannity
2021-01-09 Watters World Justice With Judge Jeanine The Greg Gutfeld Show
2022-04-24 Life, Liberty & Levin The Next Revolution With Steve Hilton Sunday Night in America With Trey Gowdy
2023-01-21 One Nation With Brian Kilmeade Unfiltered With Dan Bongino Lawrence Jones Cross Country
2024-01-31 Fox News at Night The Five Jesse Watters Primetime
MSNBC (12)
Date (ET) 8–9 PM 9–10 PM 10–11 PM
2020-03-13 The 11th Hour With Brian Williams The Rachel Maddow Show The Last Word With Lawrence O'Donnell
2021-01-09 The Week With Joshua Johnson The Week With Joshua Johnson — (Dateline Extra excluded)
2022-04-24 The Mehdi Hasan Show Ayman — (Dateline excluded)
2023-01-21 American Voices With Alicia Menendez Ayman — (Dateline excluded)
2024-01-31 The 11th Hour With Stephanie Ruhle Alex Wagner Tonight The Last Word With Lawrence O'Donnell
CNN (9)
Date (ET) 8–9 PM 9–10 PM 10–11 PM
2020-03-13 CNN Tonight With Don Lemon Anderson Cooper 360 Cuomo Prime Time
2021-01-09 — (documentary excluded) CNN Newsroom Live CNN Newsroom Live
2022-04-24 — (documentary excluded) — (documentary excluded) CNN Newsroom Live
2023-01-21 — (special excluded) — (special excluded) — (special excluded)
2024-01-31 Laura Coates Live Anderson Cooper 360 The Source With Kaitlan Collins

Nine documentaries/specials outside regular prime-time news were excluded from the final 36-episode corpus.


Visual annotations (logos & post screenshots)

Path: LogoAndPostScreenshots/

Each CSV lists frames that contain at least one visual social media signal (logo and/or post screenshot). Frame files follow:

{episodeID}-{second}.jpg

Example: CNNW_20200314_030000_CNN_Tonight_With_Don_Lemon-000136.jpg is the frame at second 136 of that episode and (in the gold standard) shows both a Twitter logo and a Twitter post screenshot.

Schema

Column Description
filename One-second frame filename (episodeID-second.jpg)
Social Media Logo Yes / No — whether a social media platform logo is visible
Social Media Logo Type Platform(s) for the logo (e.g., Twitter, TikTok, Twitter (X), Instagram)
Social Media Screenshot Yes / No — whether a social media post screenshot is visible
Social Media Screenshot Type Platform(s) for the screenshot (e.g., Twitter, Truth Social, No Platform)

Counts (released positive frames)

Channel Annotated frames Logo = Yes Screenshot = Yes File
CNN 229 154 221 csv
Fox News 768 728 188 csv
MSNBC 448 448 282 csv
Total 1,445 1,330 691

Frames without logos or screenshots are not listed; for full-episode evaluation, treat unlisted seconds in the 36 episodes as negatives.


Closed-caption annotations

Path: ClosedCaptions/

Each row is one closed-caption segment. Caption text is not redistributed. Labels and timing metadata are provided, plus a clip_url pointing into the Internet Archive so you can inspect the corresponding audiovisual context.

Schema

Column Description
date Broadcast date in the Archive identifier (UTC calendar date)
channel CNN, FOXNEWS, or MSNBC
episodeid TV News Archive episode identifier
clip_url 60-second Archive clip containing the caption midpoint
starttime / endtime Caption segment times (HH:MM:SS,mmm)
starttime_seconds / endtime_seconds Same bounds in seconds
Social Media Mention Yes (social media reference), No, or Ad (commercial break)
Social Media Platform Platform(s) mentioned when Social Media Mention is Yes (e.g., Twitter, Social Media, TikTok)

How clip_url is constructed

Archive.org serves programs as contiguous 60-second windows (0–60, 60–120, …). For each caption segment with midpoint (m = (t_{\mathrm{start}} + t_{\mathrm{end}}) / 2):

https://archive.org/details/{episodeid}/start/{T}/end/{T+60}

where (T = 60 \lfloor m / 60 \rfloor). Boundary ties prefer the window with greater overlap (earlier window on a remaining tie). Caption clocks can be slightly offset from video; treat clip_url as a retrieval pointer, not an exact cut.

Example clip:
https://archive.org/details/CNNW_20200314_030000_CNN_Tonight_With_Don_Lemon/start/420/end/480

Counts

Channel Segments Mention = Yes Mention = Ad File
CNN 14,764 83 2,028 csv
Fox News 22,561 121 3,257 csv
MSNBC 22,657 81 4,775 csv
Total 59,982 285 10,060

Accessing source broadcasts

Social2TV does not ship video frames or closed-caption text.

Need How
Watch a full episode https://archive.org/details/{episodeid}
Watch a time range https://archive.org/details/{episodeid}/start/{S}/end/{E}
One-second frames / bulk captions Contact the Internet Archive (see below)
Explore caption mention volume GDELT Television Explorer

For bulk research access beyond the public interface, contact roger@archive.org and info@archive.org with a description of the proposed research and intended use.


Related code


Data availability and usage restrictions

  • Annotations in this repository are released for research use under the MIT License.
  • Television media and closed-caption text are owned by rights holders / the Internet Archive and are not redistributed here.
  • Use episode IDs, timestamps, and clip_url values to retrieve corresponding material from the TV News Archive.
  • Do not treat this positive-enriched set as a prevalence sample of platforms or networks.

Acknowledgments

We thank Roger Macdonald (founder, Internet Archive TV News Archive), Mark Graham (Director, Internet Archive), and Kalev Leetaru (GDELT Project) for supporting this research and facilitating access to Television News Archive data.


License

Annotation files and documentation in this repository are released under the MIT License. Source television media and closed captions are not included in this release and remain subject to the copyright and usage terms of the Internet Archive.

About

A Benchmark for Detecting Social Media Content in TV News Archives

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages