Open Research for Online Safety
Insights from Rewire’s CEO and Co-Founder Bertie Vidgen
A Starter Kit for Open Research
Keeping people safe online is a complex issue. That is why open research plays a critical role for online safety: it promotes transparency and supports trustworthy evidence. Being open in research can mean different things depending on who you speak to, but two key steps are to make data and models publicly available.
In this post, we want to provide a starter kit for open research in the online safety space, focusing on open data and open models. Read on below for our selection of top resources.
How to Access Social Media Data
YouTube has recently created a new academic access programme. Otherwise, you can collect comments from the YouTube Data API or you can pay for data.
TikTok has also recently announced plans to give some researchers access through their virtual Transparency Centre.
Reddit has its own API, but Push Shift offers an amazing alternative, with more data and historical insights.
Gab data can be accessed from (at least) three good sources. First, go to Push Shift. Second, go to the GabLeaks dataset, which you need to request access to. Third, use the data dumps from Zannettou et al.’s 2018 paper, if you want something to start using immediately.
Telegram data is best taken from… Push Shift! They have a dedicated API, documented with a full academic paper.
Where to Find Labelled Datasets
For datasets on hate speech and abuse you can use hatespeechdata.com, developed by ITU Copenhagen and Turing (and us!).
For extremism and terrorism research, check out Vox-Pol’s list of datasets and resources, Jihadology, the Terrorism Content Analysis Platform from Tech Against Terrorism, and datasets from C-REX.
And otherwise, you can always use the Data Search service from Google.
At Rewire, we’re working to open source as many of our datasets as we can. We were recently commissioned by Meta to develop a new dataset and AI models for better detection of sexist content—shared freely at SemEval 2023.
What Classification Tools to Use
There’s a growing crop of safety tech companies, like Rewire, who are building AI tools for content moderation.
Perspective API from Google Jigsaw provides classifications for 18 attributes, and works in 11 languages. Despite some criticism, Perspective has become an academic standard and now processes over 500 million entries each day.
The Rewire API provides classification for hate, abuse and profanity. It’s currently available in English, but will soon be released multilingually. We offer a Social Good programme that gives researchers access to 1 million calls per month for free.
Hugging Face is an open-source repository of AI models, datasets and evaluation frameworks. Anyone can upload and you’ll find models for classifying hate speech, suicide, bullying and more.
What’s Next?
Transparency is the most powerful lever in trust and safety. Real progress will only come from systemic change, which means regulators and governments putting more pressure on the platforms. But as we’ve hopefully shown, there are many tools out there that you can get started with straight away.
If you like what we’re doing here at Rewire, you can sign up for free trial access to our API. And if you have any questions about our work, get in touch—we’d love to hear from you!
