Hate speech online: detection methodologies between algorithmic and qualitative evaluations. A case study on antiSemitism on Twitter

Research on the forms of contemporary hatred (Siegel, 2020; Santerini, 2021), and in particular studies on the changes that have taken place on the social Web (Pasta, 2018, 2019), agree that this phenomenon requires a multidisciplinary approach. At an international level, the field of Hate Studies, which combines the legal and IT fields with the humanities (sociological, pedagogical, anthropological, philosophical, linguistic, semiotic) and the interests of scholars, researchers, politicians, communication experts, human rights, NGO leaders, is marked by a significant number of research aimed at automating detection processes and creating an algorithm capable of identifying online hatred. The corpus is almost always taken from Twitter, since among the main social networks it is the only one with easy access to data automatically through APIs, i.e. application programming interfaces. In this field of research there is a tension between human-non-human and technology-human action, with the tendency to limit interventions to artificial intelligence to the detriment of more interpretative approaches. At the macro level, we can identify two groups among international studies. The first includes searches that use only machine learning methods, while the second includes studies that combine automatic search with human classification (Pasta, 2021; 2023). The contribution presents an analysis that combines socio-educational approach and automatic computer processing. This methodology is applied to various target groups and aims, alongside detection, at a more in-depth study of its characteristics, in order to design coherent educational interventions. This case deals with the classification of antiSemitic hate speech on Twitter, in Italian from 1st March 2019 to 28th February 2023. The question is whether there are monthly spikes in antiSemitic hatred, and the research is carried out through temporal analyses of samples manually classified by experts, and later is specified which rhetoric and forms of hatred are prevalent. The methodology used falls under the techniques of social network analysis (SNA). The data were collected using the open-source Python library GetOldTweets3, which allows to obtain tweets via query search. With the search string that combined the presence of a lemma identifying the target group with (AND) a reference to elements typical of antiSemitism according to the literature, all the tweets published in the two years were extracted. Subsequently, following the technique of simple random sampling without repetition, a sample consisting of 100 tweets per month was selected, thus obtaining a sample dataset of 4800 total posts (Gareth et al., 2017). The latter was manually classified by industry experts (“annotators”), who determined whether the tweet contained hate or not. In case there was a hate content, they assigned the rhetoric and the corresponding form of antiSemitism, according to the Working Definition of the International Holocaust Remembrance Alliance (IHRA). The former were derived from a psycho-social analysis and historical-literary on linguistic forms of hostility and already tested for other target groups by the same interdisciplinary team (insults, derision/irony, exclusion/separation, prejudice, dehumanization, humiliation/contempt, fear, competition, incitement/violence). After returning the main results (however the contribution focuses on the methodological approach), the last step is to submit the results to a confusion matrix, i.e. a tool for analyzing the errors made by a machine learning model (Gareth, Witten, Tibshirani, 2017). All the texts classified by the annotators are thus also evaluated by an algorithm capable of establishing whether the tweet contains hate, after applying a series of typical Natural Language Processing (NLP) procedures to “clean” the texts, such as the removal of superfluous characters, the conversion of text to lowercase and the removal of stopwords (Bird, Klein, Loper, 2009). For the “alternative” classification, the dictionary of negative wordsfrom anotherscientific research in Italian is used. It will be highlighted how the algorithm that takes into account only the roots of the words does not perform well in identifying hateful content, with a high degree of difference in the evaluation on the same tweets between the manual annotation process, also called tagging, which requires collaboration with experts, and that is produced by the algorithm created with the dictionary of negative words of another methodology. From the case study, connected to the debate in the Hate Studies, emerges the need to contextualize the signs (words, images, memes…) in context. The high error rate produced by the algorithm confirms that semantic analysis alone is not sufficient to be able to correctly automate such a complex process. It is necessary to possess a knowledge of reality and therefore of the context in which it is found, as well as reflect on the different means used in the mechanisms of othering, i.e. on a set of dynamics, processes, structures, including linguistic ones, which dialectically group subjects into a “us” and “them” in groups presented as homogeneous and alternative to each other (Powell & Menendian, 2018), and in perspectivation strategies of polarization us vs. them (Graumann, Kallmeyer, 2002). Algorithmic logic, as Bruner (1996) has shown, deals with already encoded information, the meaning of which is established in advance; computational logic is interested in stimuli and responses, not in the meaning to be attributed to things, processes information, while those who do culture and education interpret and produce meaning: an operation full of ambiguity and above all sensitive to the context. Therefore, from the case study, it emerges that algorithms that limit themselves to identifying insults, or that only detect the presence of hate words, are not enough, while it is necessary to continue experimenting with research that integrates the two phases of human and automatic classification, applying interdisciplinary approaches and start from an in-depth knowledge of the manifestations of the phenomena at the center of the investigation.