UCLA

An end-to-end pipeline of text mining over consumer reviews on Amazon under the search phrase "face mask" was built for the product designers, engineers and manufacturers to identify "sweet spots" for product engineering and manufacturing optimization. Compared with current Natural Language Processing (NLP) and in particular Aspect Sentiment Clas- sification (ASC) approaches, this research achieved state-of-the-art (SOTA) classification accuracy in 31 classes of aspects (0.83 accuracy) and 3 classes of sentiments (0.91 accu- racy), with a small (<1,500) training dataset. The SweetSpotter pipeline took in raw review texts scraped from Amazon, split them into minimal semantic units, fine-tuned on bert- base-uncased transformer with training dataset labeled by a human expert, classified nearly 400,000 text units, and delivered insights on the most impactful and meaningful features to improve the product in terms of user experience. It turns out that consumers care most about fit, least about look, on face masks.