Analyses of open-ended questions by renormalized associativities and textual networks: A study of perception of minerality in wine

- Use of semi-automatic pre-treatments for term lemmatization and construction of contingency table.
- Construction of a remormalized textual network from the text-document contingency table.
- Comparison of Ncut and modularity clustering.
- Construction of meaningful, well-defined terms communities.

Verbal tasks are increasingly used in food science, but still often suffer from time-consuming manual preprocessing procedures. Also, traditional visualization techniques are not always successful at clearly revealing the structure of term co-occurrences. The present study proposes a few statistical innovations in the analysis of textual data resulting from an open-ended survey on minerality perception, without tasting phase. First, we use dedicated, amenable software aimed at producing term lemmatization and construction of contingency table, enabling minimum manual verification and correction. Furthermore, co-occurrences are treated as a textual weighted network, which can be further iterated and renormalized in a flexible way, filtering out rare terms and their associations. In addition, visualization and clustering techniques, initially developed in social networks studies, reveal meaningful and well-defined terms communities, corresponding to distinct conceptions of minerality. Results are exclusively based upon statistical methods, without resorting to semantic nor linguistic considerations. Altogether, they demonstrate the polysemy and ill-definiteness of the concept of minerality among wine professionals.