Detecting hate in digital environments
Reference number | |
Coordinator | TOTALFÖRSVARETS FORSKNINGSINSTITUT, FOI - Avdelningen för Försvars- och säkerhetssystem |
Funding from Vinnova | SEK 413 500 |
Project duration | April 2019 - January 2020 |
Status | Completed |
Venture | Challenge-Driven Innovation – Stage 1 initiation |
Call | Challenge-Driven Innovation stage 1 initiation – 2019 |
Purpose and goal
The goal of the project has been to develop a digital tool that automatically recognizes hate speech in Swedish and that helps to reduce hate speech in digital environments. An initial model has been developed but more tests and more annotated data are needed before a model with satisfactory performance is obtained.
Expected results and effects
Within the the project, we have increased our understanding of digital hate. Among other things, we have launched the website ´hatomaten´ where anyone could submit hateful texts to contribute to the development of our hate detection model. The aim with ´hatomaten´ was also to inform the initiative against online hate. Initial tests to understand our hate detection model´s ability to detect hate have been conducted by the Police Authority. The ambition is that the police will eventually be able to use similar techniques in their work with identifying threats.
Planned approach and implementation
FOI and Uppsala Universitet have worked on developing machine learning models that have been trained with annotated data to recognize hate. The police have tested one model and given feedback. In order to collect data that can be used to develop better models, the webpage hatomaten.se was also launched, where the public had the opportunity to assist the development of machine learning models by contributing with training data.