HATE COVID-19.PT: detecting hate speech in social media

A few weeks ago, the team that brought the FCT-funded HATE COVID-19.PT project to life met at INESC-ID. Since May 2021, HATE COVID-19.PT has developed methods for semi-automatically creating a large-scale Portuguese annotated corpus covering online hate speech, while exploring how the information in that annotated corpus can support hate speech detection, allowing users to visualize the metrics extracted from data.

Coordinated by Paula Cristina Quaresma da Fonseca Carvalho, researcher within the INESC-ID Information and Decision Support Systems (IDSS) Research Area, and with the participation of three other IDSS researchers (Mário Jorge Costa Gaspar da SilvaPaula Cristina Quaresma da Fonseca Carvalho and Danielle Caled Vieira) and two researchers from the Human Language Technologies (HLT)Research Area (Ricardo Daniel Santos Faro Marques Ribeiro and Fernando Manuel Marques Batista), HATE COVID-19.PT saw its results presented and discussed in a one-day workshop that took place on 17 May.

We spoke with Paula Carvalho about this exciting project and the May workshop:

How did the project come about and what were its initial objectives?

This research project was one of the applications approved for funding by the Fundação para a Ciência e a Tecnologia (FCT), within the scope of the call for special support for projects on hate speech. The idea of ​​submitting an application for this call made perfect sense at the time, since some of the researchers that make up the HATE COVID-19.PT team were involved in projects, more directly or indirectly, related to hate speech (namely the analysis and detection of misinformation in the media and in social networks). In addition to the linguistic and computational challenges that the recognition of hate speech presents, this phenomenon, increasingly present in social networks, constitutes a clear violation of human rights, as well as one of the greatest threats to social cohesion and democratic societies. From the point of view of natural language processing, it is crucial to create resources that allow, on the one hand to understand the linguistic and discursive materialization of this phenomenon and, on the other, to support the creation of learning models that allow its automatic recognition. However, with regard specifically to the Portuguese language, the existing resources are scarce and the geographic and temporal dimensions, crucial for the characterization of the evolution of online hate speech in the Portuguese context (also one of the objectives of this project), are not normally taken into account.

In a simplified way, the initial objectives of this project consisted in the creation of solid resources that allowed for the investigation of the main linguistic and discursive strategies underlying the materialization of direct and indirect hate speech, taking into account the different target groups studied, in the Portuguese context, as well as how to assess the impact of the pandemic on the evolution of online hate speech in Portugal. In addition, we proposed to explore semi-supervised learning approaches to address the limitations associated with manual annotation approaches. The manually and automatically annotated data would then be used to develop learning models that allow for the recognition of online hate speech in Portuguese.

What was the purpose of the May 17th workshop, and what is your perception of how the workshop went?

The workshop we organized had two main objectives: to share the main research results achieved during the project with the community and to involve civil society in their discussion. In our view, the dialogue between researchers and society, sometimes underestimated by the scientific community, is fundamental.

In the initial phase of this project we created three focus groups, made up of members of communities that are frequently the target of hate speech, offline and online, in Portugal, namely the Afro-descendant community, the Roma community and the LGBTQ+ community. This work, of a qualitative nature, was crucial to better understand this phenomenon and its impact on these communities. Therefore, for the project team, it made perfect sense to once again involve these communities in the workshop, not just as guests, but as participants with an active voice. Additionally, Paula Cardoso (founder of Afrolink), Hélder Bértolo (chairman of the board of Opus Diversidades) and Vanessa Lopes (currently an intern journalist at the Público and activist for the rights of the Roma Community) participated in the round table dedicated to the discussion of the results of this project.

We were also able to include the intervention of Portuguese specialists in the project who have been approaching the topic of hate speech in the scope of their research, namely Marisa Torres Silva, Professor at Nova FCSH and researcher at ICNOVA, and Rita Guerra, Professor at ISCTE and researcher at CIS-ISCTE.

Personally, I believe that it was a very important event, allowing for the sharing of knowledge, in a multidisciplinary perspective, while also highlighting stimulating and enriching interventions by the audience.

Now that the project has come to an end, what seems to be the biggest conclusions of the project and what impact did they have?

As I mentioned earlier, one of the main objectives that we defined within the scope of this project was the creation of linguistic resources, in particular annotated corpora (that is, data collections), which are scarce for Portuguese, in order to understand and characterize this phenomenon and ultimately contribute to its automatic recognition. This objective was accomplished and, at this moment, we are able to say that we have created perhaps the largest collection of data (on hate speech) annotated for Portuguese, considering several dimensions of the problem that are typically little explored or explored in an unsystematic way in the literature. In particular, around 200,000 comments were manually recorded, focusing on three target groups (Afro-descendant community, Roma and LGBTQ+) and two social networks (Twitter and YouTube). The annotation system includes, for example, the distinction between hate speech, offensive speech and counter-narratives, as well as the main rhetorical and discursive strategies used in the analyzed comments.

It should also be noted that the annotation process involved both individuals from the target communities and individuals who do not belong to any vulnerable or historically marginalized group. Although we cannot and should not make generalizations (taking into account the size and characteristics of the sample) there are some interesting conclusions that we can be drawn from the studies carried out. For example, we observed that indirect hate speech (implicit or covert) is as frequent or more frequent on social networks than direct hate speech. Indirect hate speech is often anchored in superficial and fallacious argumentation strategies, including the appeal to fear, the call to action (e.g., the appeal to vote in far-right parties); moreover, this type of discourse is also materialized through rhetorical figures such as irony or sarcasm. With particular reference to a study of a qualitative nature carried out in the initial phase of this project, which involved the creation of focus groups with members of the targeted communities, we were able to conclude that indirect hate speech is considered more harmful than direct hate speech by target groups, even when this type of discourse is expressed through praise or humor. These are, therefore, strategies that seek to normalize and perpetuate stereotypes associated with target groups, often present in comments identified as containing hate speech. We also observed that, according to the data we analyzed, namely on Twitter, it is not possible to establish a direct relationship between the pandemic and the increase in hate speech on social networks in Portugal. In fact, the highest peaks of online hate speech recorded during that period seem to be closely related to events, in the international and national paradigm, that involve the targeted communities (e.g., the murder of G. Floyd and the Marega case).

The results achieved in this project are therefore fundamental to support the development of models for automatic detection of hate speech in Portuguese. At the moment, the linguistic resources created, pioneers for the analysis of hate speech in Portuguese, are being explored within the scope of the master’s work of two research fellows, hired within the scope of this project, who will present their master’s dissertations, coming soon to Instituto Superior Técnico.

Fonte: https://www.inesc-id.pt/hate-covid-19-pt-detecting-hate-speech-in-social-media/