Funded by Innosuisse: An algorithm that detects hate speech
Hateful comments severely affect public debate on the internet. The social science research project “Stop Hate Speech” is investigating how such comments can be detected using algorithms, and which counter strategies can curb them.
News platforms and social media provide online forums for comments. But not every comment that is posted is constructive and objective, let alone written respectfully. Some are even insulting, or hateful; they seek to intimidate, frighten, or devalue certain people or groups of people.
Such insulting or damaging comments are known as hate speech. Hate speech can harm public discourse by deterring people from expressing their viewpoint in a comment section or participating in a discussion. “For example, if female politicians constantly have to deal with misogynistic comments, it may discourage women from entering politics in the first place,” says Dominik Hangartner, ETH Professor of Public Policy and Head of the Public Policy Group.
Hangartner’s research group and the University of Zurich’s external page Digital Democracy Lab have teamed up with alliance F, the transpartisan Federation of Swiss Women’s Associations, to investigate how machine learning can be used to detect online hate speech, and which counterspeech strategies are most effective in reducing it.
Digital dog to detect hate speech
alliance F has launched the Stophatespeech.ch platform to reduce online hate speech. The aim of the project is to automatically detect online hate speech with an algorithm named “BotDog”, and to empirically evaluate which kind of counterspeech strategies are most effective in improving the quality of discussions.
To teach the Bot Dog how to find hate speech, rather than just any comments, the researchers train it with numerous examples. These are drawn from the comment sections of Swiss news platforms such as Blick.ch and 20 Minuten. To date, about 12,500 comments – in both German and French – have already been classified.
Because there is no universally accepted definition of hate speech, and because different people assess comments differently, the research team works with different groups who classify the comments according to whether they contain hate speech or not. This ensures that the classification does not solely reflect the opinions of a single group of people.
So far, almost 60,000 classifications have been made, and more are being added continuously; the more training data the Bot Dog is fed, the more accurate its search on the web will be.
What’s the best way to tackle hate speech?
In addition, the research team seeks to find out how to respond effectively to negative comments. They study eight different strategies – ranging from fact-based corrections of false statements, moral appeals and warnings, to humorous, sarcastic or even positive responses.
“We currently know very little about which counterspeech strategies are effective in reducing the various types of hate speech,” says Selina Kurer, Lab Manager in Hangartner’s group. “As soon as we understand that, alliance F and media companies can use the knowledge to counter hate speech in online comment sections.”
The close collaboration with alliance F and the Swiss media companies Ringier (Blick) and TX Group (20 Minuten) ensures that ETH’s and UZH’s research findings translate directly into practice.
Plenty of potential for Innosuisse projects
Stop Hate Speech has received funding from Innosuisse, the Swiss innovation funding agency, since November 2020, as an “innovation project with implementing partners”. This may seem surprising, given that innovation funding is typically associated with collaborations between the STEM disciplines and industrial groups or small and medium-sized enterprises. For the D-GESS department, Stop Hate Speech is the very first Innosuisse project. “In fact, Innosuisse also funds projects with a non-profit organisation as implementing partner,” says Jan Zimmermann, Industry Relations Manager at ETH Zurich.
Of course, innovation projects with non-profit partners must also generate a benefit. While an industry research partnership develops a product or process innovation for a market, in the case of projects with non-profit partners, projects with a measurable societal impact or public good are eligible for funding. The benefit of Stop Hate Speech lies in enabling a constructive, hate-free public debate in which everybody feels welcome to participate.
The ETH Industry Relations team has been advising ETH researchers on Innosuisse funding opportunities and new funding programmes such as Swiss Innovation Power, Flagship Initiative and NTN Innovation booster since 2020. It is seeking to reverse a prevailing trend: although Innosuisse has grown strongly since replacing its predecessor CTI in 2018, applications at ETH have not kept pace (see Table 1).
“Whereas the funding volume used to be around 15 million francs per year, in 2018 and 2019 the number of Innosuisse projects at ETH Zurich dropped by more than half,” says Zimmermann. “Which is why we have expanded our range of support for ETH researchers. Stop Hate Speech and the recent increase in applications show that this support is starting to bear fruit.”
Table 1: Number of approved projects by ETH researchers at Innosuisse/KTI since 2016.