After analyzing tens of thousands of comments and tweets, these are the preliminary results of the kNOwHATE project

Using a participatory approach, the project's partners suggested YouTube videos for analysis, which were then used to search for similar content using YouTube's recommendation system. Among the selection criteria, the kNOwHATE team highlights the fact that they are content by Portuguese authors with the potential to generate hate speech against target communities, with at least 100 comments and 1,000 views. As for X (formerly Twitter), a lexical approach was used to search for potentially relevant content and some selection criteria were adopted (e.g., the first tweet was geo-located in Portugal, publication occurred between 2021 and 2022).

In a first phase, the analysis of the thousands of comments analyzed and coded by a team of annotators focused on identifying the prevalence of hate speech, direct and indirect, offensive speech, and counter speech. Rita Guerra, the project coordinator, explains the differences between these types of speech: "Direct hate speech explicitly spreads, incites, promotes or justifies hatred, exclusion and/or violence/aggression against a social group or a person because of their group membership; it is explicitly prejudiced and inflammatory speech, which usually contains offensive, abusive or derogatory language. Indirect hate speech does the same thing, but in this case the message is not explicit and can be veiled or masked through a variety of strategies, such as humor, irony, or even stereotypical compliments. For these reasons, its meaning must be inferred. Counter-speech is a direct response to hate speech with the aim of fighting it."

The project's preliminary data indicates that in both social networks, indirect hate speech was more prevalent than direct hate speech in the different communities analyzed. As expected, direct and indirect hate speech had different characteristics and expressions. For example, direct hate speech was more often expressed through the mobilization of threat and the dehumanization of target communities, while indirect hate speech more often mobilized hate denial and role reversal strategies. The emotions present in the different types of hate speech were also analyzed: in direct hate speech, hate was the most frequent emotion, followed by anger; in indirect hate speech, the pattern was the opposite, with anger being the most prevalent emotion, followed by hate.

According to the research team, these results suggest the existence of multiple "hate speeches" with different psychosocial and linguistic characteristics, varying according to the communities targeted. For example, anger was more prevalent in hate speech against LGBTI+ communities, and hatred was more prevalent in speech against Roma/Gypsy and Migrant communities. Despite these nuances, there are common elements in hate speech against all these communities, such as the use of stereotypes and dehumanization.

 On a more positive note, all may not to be lost. Counter-speech was also analyzed on both networks, being infrequent on Youtube but standing out as the most prevalent discourse on Twitter/X. Overall, the results show a similar pattern on both networks, with the counter-speech characterized mostly using counter-stereotypes (e.g. "Have you ever tried to find out the true facts about the subsidies declared for the Roma community? I don't think so, but you have several websites that can prove your gigantic stupidity disguised as ethnic prejudice. If you need me to, I can send you some", empathy ("Usually instead of trying to explain what gender or sex is, I prefer to just go with the argument, let people be what makes them happy, if they're not hurting anyone" and reference to inclusive identities ("We are the human race and we have to know how to live together in this world. Instead of focusing on our differences, let's focus on what we have in common").

Overall, the results obtained from the annotation of tens of thousands of comments and tweets clearly suggest that understanding hate speech requires contextual, culturally sensitive approaches that consider the contextual and linguistic nuances that for now still elude algorithms. This data is therefore useful for informing the development of algorithms that are more sensitive to these contextual and cultural clues. Part of the project team has worked on computational models that can improve current detection algorithms. Rita Guerra is confident that this project will serve as a first step in improving the understanding of this complex phenomenon, which in turn could contribute to an improvement in the automatic detection of online hate speech in Portugal, as well as raising awareness about this phenomenon and the ways to combat it.

"These preliminary results highlight the complexity and variety of online hate speech. It is crucial to understand these dynamics in order to develop effective strategies to combat hate and promote a safer and more inclusive digital environment," explains the project coordinator. As well as offering an integrated and culturally sensitive view of the prevalence and nature of hate speech on social media, these results from the kNOwHATE project underline the importance of counter-discourse initiatives to mitigate the negative impacts on affected communities.

This and other data will be discussed in person at Iscte at the Final Conference on July 9, 2024. Find out more here