Hate speech detection is a crucial endeavor in maintaining the safety of online spaces, but the effectiveness of supervised approaches hinges mainly on the availability of annotated data. Prior research has explored the utility of natural language inference (NLI) models for zero-shot hate speech detection (ZSHSD), which leverages the capacity of these models to learn semantic relationships and adapt to downstream task without relying on large annotated datasets. NLI models assess if a premise sentence logically entails a hypothesis sentence, relying on precise hypothesis design to achieve adequate downstream task performance. Existing frameworks that use NLI model for ZSHSD rely on multiple inferences with different hypothesis to extract characteristics to achieve desirable outcomes. In light of the challenges surrounding ZSHSD and the method of relying on discrete tokens to design hypotheses, we aim to optimize and identify ideal intermediate representations by applying p-tuning techniques. On HateCheck, a fully tuned hypothesis led to a 18.8% point (pp) accuracy improvement, compared to a discrete designed hypothesis. Our work surpassed prior work by achieving a 5.6 pp accuracy enhancement, outperforming previous approaches that required multiple inferences. Also, the optimized tokens uncover relations to broader aspects of hate speech, offering insights for hypothesis design.
