In the first phase of the project, between August 2022 and February 2023, we monitored hate speech and hurtful speech appearing online, affecting the Roma, LGBTQ+, Jewish, migrant and Muslim communities. We have created an almost completely automated system which is suitable for finding texts that most likely contain hate speech or offensive expressions based on keywords.
108 635 texts were manually downloaded, and then filtere through the keywords. As a result 11 354 were sorted into Excel tables, along with their sources and dates of publication. In the next step, the volunteers previously trained for the task classified these texts according to a predetermined system of criteria as hate speech or hurtful speech. The final result is a table that includes the texts which most likely contain hate speech or hurtful speech.
In addition to the data set, an artificial intelligence algorithm was created that can learn from human decisions. Similarly to a human, if it sees enough examples of hate speech and non-hate speech, it can tell the difference between with high accuracy.
You can read and download the full report here.