TPC: One Investigation Reveals the Exploration of Offensive Language

The Twin-Problem-Channel (TPC) method leverages few-shot learning to identify offensive language within a multi-turn dialogue QA framework. By integrating natural language instructions with corresponding sentences, it aids models in annotating and comprehending text. Instruction fine-tuning is applied to the T5 model, allowing it to extract more information from the sentence, enhancing its ability to detect offensive content. Trained on large datasets such as Hasoc, OLID, and social media data from Facebook and Twitter, T5 outperforms BERT and GPT-3 in aggressive language recognition. The instruction fine-tuning approach improves model efficiency and enables task adaptation without retraining, demonstrating strong generalization to unseen tasks. Experimental results show that the TPC method outperforms several baseline models, achieving new state-of-the-art performance on the OLID, TRAC, and Hasoc datasets.