New Technology Speeds Up Fight Against Online Harassment

Hate speech endangers people and cohesion. Quick screening of hate speech is mandatory for social networks in Europe. Graphic: Adobe

With the CS-2 system from Cerebras, AI models for the recognition of hate speech on social media platforms can be trained faster  and offensive texts can be detected more effectively. This is shown by a study from the LRZ that compares different AI accelerator systems.

Identifying hate speech and removing it from the digital world: This is a task that social media platforms and online media need to master as quickly as possible. Artificial intelligence (AI), in particular and pre-trained large language models (LLMs), are helping them to do this. Computing power provides additional advantage - especially clusters equipped with graphical processing units (GPUs) or AI systems such as the CS-2 from Cerebras Systems, whose chip, the Wafer Scale Engine 2, was specifically designed to meet the needs of LLMs. In a study, researchers at the Leibniz Supercomputing Centre (LRZ) compared the performance and effort of different AI technologies when implementing and fine-tuning language models: “Compared to classic training setups, the specialised AI accelerator from Cerebras speeds up training times by a factor of four,” reports Dr Michael Hoffmann, a specialist in big data and AI at the LRZ. “However, the Cerebras system is very new, so considerable effort is required for preparation or compilation, for example, and only a limited number of language models can be transferred to the CS-2 system.”

Testing and comparing different AI systems

For the total of 12 test series, a V100 GPU system competed against CS-2 systems at the LRZ and the Edinburgh Parallel Computing Centre (epcc), which differ from the implemented AI models and programmes. The monolingual and multilingual models BERT, mBert and XLM-RoBERTa were run on all three resources and trained using their own classifiers so that they were able to filter out hate speech from several thousand German, Italian, Spanish and English Twitter posts and other publicly accessible texts. "CS-2 does not directly make it easier to recognize hate speech," Hoffmann continues. "However, the systems accelerate the training and fine-tuning of language models that recognize hate speech." Online platforms or communities can use such specialized systems to adapt and optimize their own detection models or classifiers more quickly, although their performance depends on the models and training data used.

Results for hate speech research

In the experiments, the monolingual language models also outperformed the multilingual ones, delivering more accurate results. And, as expected, the performance differences between the two CS-2 systems were small. Although they are more complex to prepare than traditional GPU clusters, they run faster in a factor of four. The results, which were presented at the IEEE International Conference on Tools with Artificial Intelligence (ICTAI) in the US at the end of October, should be of interest to online communities and media, as well as researchers and specialists developing AI models for hate speech detection or multilingual content moderation. In addition to gaining practical experience, the LRZ researchers have also established further collaborations in the field of hate speech research. (vs/LRZ)

M. Hoffmann, J. John, N. Hammer: Exploring the Suitability of the Cerebras Wafer Scale Engine for the Fast Prototyping of a Multilingual Hate Speech Detection System. DOI 10.1109/ICTAI62512.2024.00048 Publication is planned at web presence of IEEE International Conference on Tools with Artificial Intelligence (ICTAI)