Improving Adversarial Data Collection by Supporting Annotators: Lessons from GAHD, a German Hate Speech Dataset

TL;DR

GAHD tackles biases in hate speech data by applying four rounds of dynamic adversarial data collection (DADC) with explicit annotator-support strategies to generate diverse German adversarial examples. The resulting 10,996-example dataset yields substantial robustness gains for the German hate speech detector, with macro $F_1$ improvements of about $18$–$20$ percentage points when training on all rounds, and the gains are strongest when mixing strategies. The paper also analyzes annotator disagreements, reports inter-annotator agreement, and benchmarks large language models and content moderation APIs, with GPT-4 achieving the best performance. These findings offer practical guidance for designing efficient, robust data collection pipelines and improving hate-speech detectors in German.

Abstract

Hate speech detection models are only as good as the data they are trained on. Datasets sourced from social media suffer from systematic gaps and biases, leading to unreliable models with simplistic decision boundaries. Adversarial datasets, collected by exploiting model weaknesses, promise to fix this problem. However, adversarial data collection can be slow and costly, and individual annotators have limited creativity. In this paper, we introduce GAHD, a new German Adversarial Hate speech Dataset comprising ca.\ 11k examples. During data collection, we explore new strategies for supporting annotators, to create more diverse adversarial examples more efficiently and provide a manual analysis of annotator disagreements for each strategy. Our experiments show that the resulting dataset is challenging even for state-of-the-art hate speech detection models, and that training on GAHD clearly improves model robustness. Further, we find that mixing multiple support strategies is most advantageous. We make GAHD publicly available at https://github.com/jagol/gahd.

Figures (8)

• Figure 1: We use four rounds of dynamic adversarial data collectionkiela-etal-2021-dynabench to improve a German hate speech classifier. We start with a target model trained on existing datasets. Then, in each round (R1-R4), annotators try to trick the target model using a different method. After each round, we train a new target model including the new adversarial examples.
• Figure 2: DADC workflow for R2, where we let annotators validate model tricking translations of English adversarial examples.
• Figure 3: Workflow of R3, where we task annotators with validating model tricking newspaper sentences.
• Figure 4: Workflow of R4, where we let annotators create contrastive examples to challenging entries from previous rounds.
• Figure 5: An overview of the most important topics in GAHD. We generate the topics via clustering and use GPT-3.5 to obtain cluster descriptions. Section \ref{['subsection:full-dataset']} describes the procedure.