Detection of Privacy-Harming Social Media Posts in Italian

As many psychological and sociological study reveal, many people disclose too much privacy-harming information in social media in the form of text and multimedia posts, thus exposing themselves and other persons to several security risks. Consequently, many researchers have addressed this problem by investigating on the detection and analysis of the so-called self-disclosure behavior in social media and blogging platforms. Among the others, content sensitivity analysis has emerged as a promising research direction, but, so far, it has only focused on English text posts, although it is well-known that people tend to disclose mostly in their own native languages. Therefore, in this paper, we address this limitation by proposing a new text corpus of Italian posts that we have annotated following to the anonymity assumption. We then apply several language models based on transformers to classify them according to their sensitivity. Moreover, since Italian is a lower-resource language compared to English, we also apply some multilingual zero-shot transfer learning architectures trained on a rich and manually annotated English corpus and tested on the Italian one. We show experimentally that the approaches trained directly on the Italian corpus, still outperform multilingual ones trained on the English data and tested on Italian, although some of them exhibit promising prediction performances.