Seeing Flavors

The way foods and beverages taste—their set of flavors—is extremely important to most people. Scientists who develop and study food products must use tasting panels of human volunteers to evaluate the flavors of foods. These panels create popular tools—like the “wine wheel” used by countless wine lovers and professionals—and use them to make important processing and economic decisions. However, human panels are expensive, time-consuming, and limited in the number of products they can evaluate. We think it is possible to use Natural Language Processing (NLP) techniques on digital reviews of food products in order to develop useful flavor tools without a human panel. The Seeing Flavors project aims to take advantage of more than 6500 online whiskey reviews—a dataset orders of magnitude larger than those usually available in food science—in order to develop practical and interactive new types of “flavor wheels” for whiskeys.

Since the Seeing Flavors project was initiated in June 2019 we have worked to develop tools to implement our overall project goal. We have produced several concrete results and intermediate products. First, we have submitted the first paper that has been supported by this research for publication in Food Quality & Preference. Second, we have developed a prototype software application for use by humans-in-the-loop to identify sensory descriptors from our word corpus—this is a key step in being able to train a deep-learning tool to identify these same terms without human intervention. Third, we have started wireframe prototyping for the interface that we will deploy for our new, interactive flavor wheels. Of course, these three products are a snapshot and subset of the work done so far, but we believe they are illustrative of our work so far and where we are in the project.

Midterm Report

The following products have been developed at the midpoint of our project.  Appropriate examples/drafts/screenshots are included in the same numbered order as below in the subfolder of the report directory labeled “Products So Far”.

1. Submission of a methods paper documenting and reporting our work so far to the journal Food Quality & Preference. 
   
2. We have worked with Jonathan Bradley, Head of Studios and Innovative Technologies in the Newman Library data group to develop a JavaScript program to allow easy human-in-the-loop data cleaning of text.  Briefly, the key problem to be addressed is that there is no current machine-learning tool that “knows” what sensory descriptors (“sweet”, “cinnamon”, “green”) are, but this is an easy task for humans.  We have created an interactive wordcloud tool that takes arbitrary lists of text and allows humans to “clean” it by selecting non-descriptor words.  The tool will continue to read in more words from the dataset, allowing easy cleaning of text data for this project.  In fact, we believe this tool will have broader applicability beyond this project and will be valuable for sensory-science workflows in general, which we will explore in the second half of this project and beyond. 
3. We have begun the process of wireframing our interface for the interactive flavor wheels to be generated as the final product.  These will help guide our analysis work and determine priority.