AI & Society
One neural architecture performed satisfactorily at detecting online hate speech in English, Italian, and German

In a 2020 paper in ACM Transactions on Internet Technology, researchers reported that one neural architecture performed satisfactorily at detecting hateful or aggressive posts on social media in English, Italian, and German. A neural architecture is a machine-learning model made of layers of simple processing units. The authors tested the system on three freely available datasets, one for each language, and compared how different model designs and text features contributed to the result.
How did the researchers test the system?
They used three freely available datasets for hate speech detection on social media, one each in English, Italian, and German. The study compared three recurrent network designs: long short-term memory, gated recurrent unit, and a bidirectional version of long short-term memory. Recurrent networks process a sequence of words in order. The researchers also examined features such as n-grams, which are short sequences of words or characters, social-network-specific features, emotion word lists, emojis, and word embeddings, which map words to numerical vectors.
What did the study compare?
The paper describes an in-depth analysis of how much each design and feature set contributed across the three languages. The abstract reports that the overall system performed in a satisfactory way, but it does not say which component mattered most.
What should readers know?
Which languages did the study cover?
The system was evaluated on English, Italian, and German.
Did the model perform equally across all three languages?
The abstract says the model performed in a satisfactory way across the three languages. It does not break the results down by language.
What kind of data did the study use?
The study used three freely available datasets of hate speech on social media. The abstract does not describe the platforms or date ranges of those datasets.
Limits and what's next
The study was a dataset-based evaluation on three freely available social-media datasets, one each in English, Italian, and German. The abstract frames this as one study and reports no independent replication. It does not report results for languages beyond English, Italian, and German, and it does not state planned next steps.
Source: Michele Corazza et al., “A Multilingual Evaluation for Online Hate Speech Detection,” ACM Transactions on Internet Technology, 2020.