2023

Measuring corporate reputation with BERT-based transformers

Fine-tuned BERT-family transformers to score corporate reputation signals across very large multilingual news and social corpora, presented at Machine Learning Prague 2023.

A production-grade reputation measurement system built on fine-tuned BERT-family transformers, applied to multilingual news and social media at scale during my time at Eni. The models classify content along multiple reputation dimensions (sentiment, stance, topical salience) and feed downstream KPIs used by communications and public-relations stakeholders.

Beyond modelling, the project required careful attention to label design, inter-annotator agreement and calibration across domains and languages — without which aggregate “reputation scores” are indistinguishable from noise.

Presented as a poster at Machine Learning Prague 2023 (Measuring Corporate Reputation with BERT-based Transformers) and used internally to drive reputation reporting for a multi-billion-euro brand portfolio.