SetFitABSA: Few-Shot Aspect Based Sentiment Analysis using SetFit

Hugging Face Blog · 2 years ago

Intel Labs and Hugging Face introduced SetFitABSA, a framework for training aspect-based sentiment analysis models with minimal labeled data by using SetFit's embedding approach instead of large language models. SetFitABSA uses 220 million parameters total (two 110M models) yet matches or outperforms Llama 2 (7 billion parameters) and surpasses T5 and GPT2 in few-shot scenarios on SemEval14 benchmark datasets. The method eliminates the need for handcrafted prompts and specialized tagging tools, enabling faster model training and data labeling for detecting product or service aspects and their associated sentiment polarities in customer feedback.