Overview

Get a quick overview of the TextReveal Alerts workflow to learn more about we use AI to provide ESG controversy data.

At the beginning of the NLP pipeline, we leverage Language Models followed by deep learning models. These two steps help us compute sentiment, NER, etc. We then use similarity analysis based on keywords or full descriptions, decision trees, and finally, a second layer of Large Language Models – Generative AI (for final annotation).

These last steps of the process enable us to obtain the ESG classification, Intensity Score, and other final metrics. We also use clustering models to link multiple articles corresponding to the same event for a specific day, reducing the number of outputs.

We also have an annotation process to further fine-tune indicators. This includes manually identifying true/false positives/negatives, misclassified categories, incorrectly identified companies, and erroneous intensity scores. These files are then used to continuously improve the outputs, with a human-in-the-loop approach.

This iterative process of annotation, analysis, and model refinement helps us improve the classification of events and enhance the accuracy of the information. We work closely with our clients on this process and take client feedback into account.