SetFitABSA: High-Performance Sentiment Analysis with Small Amounts of Data

SetFitABSA, jointly developed by Intel Labs and Hugging Face, is an efficient Few-shot learning framework that detects sentiment toward specific aspects within text.

Compared to existing LLM-based approaches, it has the following key advantages.

• No prompts needed: Without complex prompt engineering, it generates rich embeddings from just a small amount of labeled data, reducing variability in results.
• Fast training speed: Training is very fast and easy using a simple data format, without the need for special tagging tools.

The process consists of a total of 3 steps.

1. Aspect candidate extraction: Uses spaCy to extract nouns and noun phrases from sentences.
2. Aspect/non-aspect classification: Uses SetFit to classify whether the extracted candidates are actual aspects.
3. Sentiment polarity assignment: Assigns positive or negative sentiment to each classified aspect.

Experimental results showed that SetFitABSA performed on par with, or even better than, large generative models such as Llama2 and T5 in low-data settings.