Since August, a team from Darmstadt University of Applied Sciences has been working with the research project BoTox (bot and context detection in the area of hate comments) on a system that uses artificial intelligence to find so-called “hate speech” on the Internet. BoTox is intended to automatically detect when social media postings and communication histories contain aspects such as insults, incitement to hatred or threats of violence.
Advertisement
Twelve criminal offenses
“We identified twelve different criminal offenses,” explains computer linguist Melanie Siegel. The planned software should not only be able to automatically identify statements that fall under this category using machine learning and artificial intelligence (AI), but also evaluate them, pre-classify them and forward them to the HessenGegenHetze reporting center.
Digital forensic expert Dirk Labudde from the Fresenius University of Applied Sciences in Idstein contributes legal knowledge and basics. At the same time, BoTox should be able to distinguish whether a bot or a human has written a hate comment. In times of generative AI like ChatGPT, this is not an easy task, the professor knows. Ethical guardrails of chatbots can be overcome with the right questions and tasks.
The team also wants to use context analyzes to find out whether it makes more sense not to feed trolls or to consciously criticize toxic statements. In order to train the software to automatically recognize hate speech and all these aspects, large amounts of data are required. To do this, the scientists use interfaces from Telegram, Facebook and YouTube.
In contrast to the predecessor project DeTox, The project, which runs until 2025, is funded by the Hessian Ministry of the Interior with 292,000 euros.
Fundamentally sensitive training
However, the team continues to use existing training data, including from Twitter. The researcher explains that more current topics such as the corona pandemic or the war in Ukraine are not yet reflected in this. “But we are working on making the data portable.” For some people it's easy: “For Chancellor Merkel, for example, the choice of words was similar to that used today for traffic lights.” Otherwise it is usually about anti-Semitism, Holocaust denial, xenophobia, migration, racism or discrimination against minorities.
The demands on the training material are high, reports Siegel. Hate speech is posted significantly more often by men than by women. The data sets should reflect such realities. The basis for evaluation is also important. Three student assistants classified the comments according to whether they were an extreme opinion, an acceptable insult or a criminally relevant statement. A comparison will then be made to see whether the ratings for humans and machines were similar.
(vbr)