Designing of Prompts for Hate Speech Recognition with In-Context Learning

In-context learning is a recent paradigm in natural language understanding, where a pretrained large language model (LLM) directly performs a new task without any update to its parameters by taking a test instance, new task description and a few training examples (e.g. input-label pairs) as its input. However, performance has been shown to strongly depend on the task description and selected training examples (both together termed as prompts here). In this paper, we use GPT-3 as the LLM, hate speech recognition as the new task, and we investigate how to design effective prompts for better performance. Our preliminary experimental results show that: (1) substantial number of input-labels pairs are necessary for good performance (2) informative task descriptions can further boost performance by ingesting our prior knowledge as inference guidance.