Portuguese project allows automatic detection of hate speech on the Internet

The rise of online hate speech in Portugal has prompted academic researchers to create a model to automatically identify this type of post using artificial intelligence.

Presenting the findings of the Knowhate project, coordinator Rita Guerra explained that the growth of online hate speech has increased, with more and more cases of online accusations and insults being translated into concrete actions in the real world.

“Hate speech has serious consequences for victims and there is a correlation between online and offline phenomena,” said the ISCTE researcher, stressing that this trend has increased with “peaks of violence and attacks against different communities,” mainly Roma, racial groups and LGBTQIA+ groups.

During the discussion entitled “Online Hate Speech in Portugal: What is the Present and What is the Future?” the researcher presented a prototype project for the automatic detection of this type of post, available at: https://huggingface.co/knowhate .

The prototype uses an annotated dataset of 24,739 comments from 88 YouTube videos and 29,846 posts from 2,775 conversations on the former social network Twitter, assessed by annotators trained in social psychology and linguistics.

Each user can add a phrase and ask the system to evaluate it to determine whether it is hate speech.

Hate speech can be indirect or implicit, often using stereotypes, irony or insults, making it difficult for automatic identification systems to work. Only with the help of artificial intelligence and the loading of large amounts of data can new phrases be identified by comparison.

This project began in 2021, during the pandemic, when a “significant increase in hate speech” was detected, explained Rita Guerra, stressing that the initiative has the support of several organizations such as SOS Racismo, Casa do Brasil, Ilga or the Commission for Equality and Against Racial Discrimination (CICDR).

The data “shows a rise in online hate speech against minorities and groups” and, at the same time, a “silent majority, people bombarded with thousands of comments,” for whom this type of speech becomes “normal” in a “desensitization phenomenon.”

The project’s ultimate goal is to “develop a set of tools and resources that can raise awareness of hate speech,” so the project also includes more traditional campaigns like videos and “podcasts,” Rita Guerra explained.

Paula Carvalho from the Institute of Systems and Computer Engineering (Inesc) explained that creating these language models was very complex, for example to allow the “nuances” between hate speech and simple irony to be identified.

“Direct hate speech is often associated with stereotypes” and “dehumanization of the other,” while “indirect hate speech is associated with the denial of hatred and role reversal” using “verbal irony.”

On the other hand, “counter-hate speech” uses “similar language,” he explained.

For the researcher, “context turns out to be even more important for detecting indirect hate speech,” since only then “can one understand what is actually meant.”

Sara Soares of Ilga, an association that advocates for LGBTQIA+ people, explained that activists have realized that “hate speech has increased significantly” during the pandemic.

In modern society, “hate speech plays a much greater role than being directed against a specific person. It is a message crime that does not create discomfort for that specific person, but expulsion from public space” for the entire community.

“Hate speech is leaving the network and penetrating our space, what we organize,” and “this is something that worries us more and more,” said the director of Ilga.