Fast and automated sensory analysis: Using natural language processing for descriptive lexicon development

Abstract

As sensory evaluation relies upon humans accurately communicating their sensory experience, the diverse and overlapping vocabulary of flavor descriptors remains a major challenge. The lexicon generation protocols used in methods like Descriptive Analysis are expensive and time-consuming, while the post-facto analyses of natural vocabulary in “quick and dirty” methods like Free Choice or Flash Profiling require considerable subjective decision-making on the part of the analyst. A potential alternative for producing lexicons and analyzing the sensory attributes of products in nonstandardized text can be found in Natural Language Processing (NLP). NLP tools allow for the analysis of larger volumes of free text with fewer subjective decisions. This paper describes the steps necessary to automatically collect, clean, and analyze existing product descriptions from the web. As a case study, online reviews of international whiskies from two prominent websites (2309 reviews from WhiskyCast and 4289 reviews from WhiskyAdvocate) were collected, preprocessed to only retain potentially-descriptive nouns, adjectives, and verbs, and then the final term list was grouped into a flavor wheel using Correspondence Analysis and Agglomerative Hierarchical Clustering. The wheel is compared to an existing Scotch flavor wheel. The ease of collecting nonstandardized descriptions of products and the improved speed of automated methods can facilitate collection of descriptive sensory data for products where no lexicon exists. This has the potential to speed up and standardize many of the bottlenecks in rapid descriptive methods and facilitate the collection and use of very large datasets of product descriptions.