Leveraging Deep Learning models to assess the temporal validity of Emotional Text Mining procedures

The paper presents the training of a surrogate model to mimic the functioning of a text mining social profiling method, emotional Text Mining, applied on the Italian tweets containing #Covid-19 during the first lockdown. This surrogate model is based on stateof-the-art Deep learning models in the field of Natural language processing (Nlp). These models feature an architecture called Transformer, which adopts a mechanism of self-attention that weighs the significance of each part of the input data. The model was trained on the tweets and factors of an initial time period to predict the polarisation of a tweet according to each factor found using the ETM technique. The model was then applied to subsequent periods. Using appropriate metrics, it was then possible to quantify the need to re-run the technique from scratch on the following time periods, while also being able to assess which factors still remained relevant over time.