Analysis of Mastodon's Italian Messages: Networks, Topics, LLM and Machine Learning

This study examines Italian-language content on the social media platform Mastodon, employing some typical tools for textual analysis. A set of posts (known as ‘toots’ on Mastodon) related to the hashtag #intelligenzaartificiale from the past two years were collected. Co-occurrence networks between hashtags and cross-author referrals within the same toots were analyzed. Next, structural topic modeling was utilized to identify four topics and their most important keywords. A comparison was then made between the predictive ability of a set of words and that of a transformer using a classifier, resulting in similar findings. Furthermore, a SHAP analysis was conducted to demonstrate the impact of individual words on the classification model, providing an explanation of the contribution of individual features.