SensoLangue

Automation of tasting comments provided by consumers

Project leader in Changins: Pascale Deneulin

Partner: Haute Ecole Arc

Fundings of Changins part: HES-SO

2016 – 2018

 

Research valorization:

Oral presentation of the results at the 13th Pangborn Sensory Science Symposium in July 2019 in Edinburgh, Scottland:

REBENAQUE P., GHORBEL H., ALBERTETTI F., VAN GYSEL L., DANTHE E., DENEULIN P., Does automated analysis of open comments from consumers allow us to get relevant results to understand their preference? (PDF)

 

Summary

Consumer opinion is increasingly of interest to agri-food companies (cheaper and more representative than an expert panel). However, the automatic analysis of comments remains little explored and is now done manually. This multidisciplinary project aims to find a process of automation integrating difficulties such as the treatment of sentences such as "very acidic but not bitter wine" or "wine with good acidity and refreshing bitterness". The automatic treatment of negations and the recognition of the polarity of the terms according to their environment appear as the main challenges in order to propose a useful tool in the field of marketing and sensory analysis.

Several stages were developped in this project:

1. A large number of tasting comments from professionals and consumers were collected via the Mondovino (Coop) platform and various databases.

2. Two wine ontologies with more than 2,225 distinct terms in French and 425 distinct terms in English were created to prioritize and standardize wine terms and concepts.

3. Natural language methods have been applied and developed using several deep machine learning algorithms to analyze the information contained in the comments collected.

4. The approach developed was trained and tested on more than 1000 tasting comments collected during consumer tests.

The optimization of machine learning algorithms applied to natural language processing (NLP) showed a clear improvement in the extraction of the concepts contained in the tasting comments, results supported by classical machine learning metrics. Similarly, the creation of a consistent wine ontology and its continuous implementation allow both to improve the automatic extraction and to reduce the variability of a manual processing.

If today the ontology of the wine is consequent in French, the other one in English would deserve to be completed due to lack of material available at the end of this project. Although specific to the wine sector, its construction process and use will be transposable to any other area affecting consumer opinion (eg tourism, agri-food). The valorization of this project will continue in the coming months with the publication of articles. A demonstration platform has been made available online to enhance the results.