These days the biggest issue is getting a decent labelled dataset. Once you have that you just fine-tune a BERT-dervived transformer neural network. See https

These days the biggest issue is getting a decent labelled dataset. Once you have that you just fine-tune a BERT-dervived transformer neural network. See https://paperswithcode.com/task/sarcasm-detection/codeless https://paperswithcode.com/task/sarcasm-detection/codeless for some examples and datasets.