Researchers built a system that mines product reviews for sensory details like color and scent to improve recommendations, then found the real driver of its gains is structured data extraction, not the sensory content itself.
The method, called ASER, fine-tunes a large language model to pull evidence-grounded attribute records, such as color: matte black or scent: vanilla, out of review text. Those records get distilled into a compact encoder that builds a frozen sensory profile for every item in a catalog. At recommendation time, the system keeps its core model frozen and learns only a lightweight scoring function on top of that profile, with the correction capped at a validation-tuned limit. Across five Amazon product categories and four backbone models, the approach improved two standard ranking metrics in all 20 test combinations, with average gains of 6.1% and 6.4%.
Here is the part a marketing deck would skip: a control test using the same pipeline but no sensory data matched most of the hit-rate improvement. The sensory vocabulary only earned its keep on ranking quality, and even then in eight of nine comparisons, not all. That is a useful dose of rigor in a field prone to crediting whatever sounds most sophisticated for a performance bump.
An audit found 94.8% of the extracted sensory records traced back to actual review text, which counts as a solid showing in a field that still too often treats 'the model said so' as an answer.