Using Natural Language Processing to aid in descriptive analysis of whisky

BENEFICIARIES: sensory researchers, companies and consumers

TOOLS USED: Multiple sensory science researchers hand-label (annotate) words within a set of reviews as descriptors or non-descriptors. To do this, we developed a novel interactive tagging tool to label words within the reveiws as descriptors or non-descriptors. We then developed a deep learning model architecture that uses the labeled data to train a model to perform this keyword extraction. We then tested the model on other labeled whisky data and obtained very high accuracy.

ISSUE: How to effectively extract whisky descriptors from reviews. A descriptor is a word that describes certain aspects of a whisky (taste), e.g., smoke or fruity. This is challenging due to the unique nature of descriptors used for describing whiskies. Being able to extract descriptors from whisky reviews with known characteristics, such as origin, age, and price, will allow a one to predict such characteristics for new whiskies when only a text description is provided.

WHAT WAS DONE TO ADDRESS THIS ISSUE: In the project we use thousands of whisky reviews and human labeled reviews to train a deep learning model to recognize a food descriptor. The result is the ability to identify, with high probability, what is and is not a descriptor within free-form text describing food.

POTENTIAL IMPACTS: To date, we are unaware of other research that has applied Natural Language Processing techniques to this problem space. We foresee the ability to perform predictive analysis for aspects of a whisky just based on its tasting notes (descriptors) and ideally create "food language" models for use by other researchers.