Technical Tools

Knowledge Hub Framework

We created a Virtual Machine, the ‘Knowledge Hub Framework’, which allows end-user applications to programmatically access a range of microservices. These tools assist with particular evidence verification and discovery tasks in relation to open source information, including object detection, face recognition and place name recognition. To preserve the confidentiality of evidence and to avoid the need for users to upload their evidence to the cloud, the ‘Virtual Machine’ format allows users to download the Knowledge Hub Framework to their local machine and run the tools from there. If you would like to test out the KHF, please contact us.

We are currently working on a cloud-based version of the Virtual Machine. To test an early prototype, please contact us.


As well as being housed on the VM, the FaceSearch tool also runs in Docker.
If you are interested in using FaceSearch in your human rights investigative work, please contact us.
Video installation guide and demo: link

Hate Speech Detection

The team developed social media post classifiers: machine learning-based models for classifying a social media post/message as containing ‘hate speech’, defined by the United Nations as ‘communication that attacks or uses pejorative or discriminatory language with reference to a person or a group on the basis of their religion, ethnicity, nationality, race, colour, descent, gender or other identity factor’. The classifiers obtained precision and recall of 96% and 87%, respectively, on a data set of ~25K tweets.

We are currently working on a follow-on project, which will package our NLP methods into a graphical web-based tool.

Using Natural Language Processing to identify the most relevant evidence

In a bespoke application designed in collaboration with our partners at Amnesty International, we developed a hybrid (rule and machine learning-based) model for recognising Arabic place names in video descriptions.

Using a combination of rules and machine learning, we have been able to address the challenging task of recognising Arabic place names, outperforming the state-of-the-art tools for the same task by at least 17 percentage points (in terms of F-score, the harmonic mean of precision and recall). This tool was tested by Amnesty International on an existing dataset.