2022

Scattertext for dichotomous text variables

Advanced visualization for contrastive text analysis using Scattertext, applied to reputation and OSINT corpora at Eni.

A visual analytics workflow built on top of Scattertext to surface terms, n-grams and phrases that discriminate between two contrasting text populations — for instance, coverage in friendly vs hostile outlets, or pre- vs post-event mentions.

Integrated into the Eni reputation stack alongside topic models and transformer classifiers, the technique gave analysts a fast, interpretable way to explain why two corpora differ, before committing to a heavier modelling step.

The approach proved valuable for communicating findings to non-technical stakeholders: instead of a single aggregate score, analysts could point to the actual language driving a dichotomy and defend editorial or strategic conclusions from it.