Greenwashing

The client for this project wanted to detect Greenwashing from text extracted from websites (note that details are scarce in order to satisfy an NDA).

The dataset was collected by an off-the-shelf automated sraping tool. This meant that the data collected comprised of many languages and a lot of the entries were gibberish. To clean the dataset, it was translated to English using Microsoft Azure language tools. To remove nonsense entries, I trained a bi-gram that analysed frequencies of letters appearing next to each other - this did a good job of removing some of the dirty data. I then employed a second package called 'Nostri' - this did a similar job and further cleaned up the dataset.

After some analysis to see our proof point distributions, we trained a model found on hugging face that had known good performance on sentiment analysis.

This was then handed over to the client for further development and work.