Aspect-based sentiment analysis has always been the tedious cousin of regular sentiment classification. Instead of just labeling a review as positive or negative, you’re tagging individual words — “screen” gets a thumbs up, “battery” gets a thumbs down — and that granular labeling is slow, expensive, and frankly boring for human annotators. Intel Labs and Hugging Face just made it dramatically less painful.
SetFitABSA is a three-stage framework built on SetFit, the few-shot sentence classification approach that’s been quietly outperforming much larger models on small datasets. The pipeline starts by using spaCy to pull out noun candidates from training sentences — because in product reviews, aspects are almost always nouns or noun compounds. Then a binary SetFit classifier decides whether each candidate is actually an aspect or just a random noun. The clever part: each candidate gets concatenated with its full sentence to create training instances, turning a token-level problem into a sentence-level one that SetFit already handles well. A second SetFit model then assigns sentiment polarity to the extracted aspects.
The results are genuinely surprising. SetFitABSA outperforms Llama2 and T5 in few-shot scenarios — models that are orders of magnitude larger. No prompts to hand-tune, no brittle in-context learning that falls apart when you rephrase the instruction. Just a handful of labeled examples and you’ve got a domain-specific ABSA model trained in minutes, not hours.
For teams drowning in customer feedback across different product lines, this changes the math. You don’t need a prompt engineer or a GPU cluster to get fine-grained sentiment insights. You need maybe 50 labeled sentences and a working knowledge of the SetFit library. The question now is whether this approach holds up as the number of aspects per sentence grows, or when aspects aren’t tidy noun compounds — “screen brightness” yes, but what about “easy to hold”? That’s the next test.
