Analysis and Application of Language Models to Human-Generated Textual Content

Social networks are enormous sources of human-generated content. Users continuously create information, useful but hard to detect, extract, and categorize. Language Models (LMs) have always been among the most useful and used approaches to process textual data. Firstly designed as simple unigram models, they improved through the years until the recent release of BERT, a pre-trained Transformer-based model reaching state-of-the-art performances in many heterogeneous benchmark tasks, such as text classification and tagging. In this thesis, I apply LMs to textual content publicly shared on social media. I selected Twitter as the principal source of data for the performed experiments since its users mainly share short and noisy texts. My goal is to build models that generate meaningful representations of users encoding their syntactic and semantic features. Once appropriate embeddings are defined, I compute similarities between users to perform higher-level analyses. Tested tasks include the extraction of emerging knowledge, represented by users similar to a given set of well-known accounts, controversy detection, obtaining controversy scores for topics discussed online, community detection and characterization, clustering similar users and detecting outliers, and stance classification of users and tweets (e.g., political inclination, COVID-19 vaccines position). The obtained results suggest that publicly available data contains delicate information about users, and Language Models can now extract it, threatening users' privacy.