Using Surface and Semantic Features for Detecting Early Signs of Self-Harm in Social Media Postings

This paper describes the University of Hildesheim submission to the CLEF eRisk 2020 shared task on detecting early signs of self-harm in social media posts. We introduce four systems that apply different methods trying to address this task and a fifth ensemble system that combines the four other systems. The first four systems make use of features of different types, such as time intervals between posts, the sentiment and semantics of the writings by using bag-of-words vectors and contextualized word embeddings in a neural network approach. The results show that while all our systems achieve a high recall, the focus of future work should be further improvement of the precision. All systems and the ensemble model achieve a comparable performance of Flatency values in the range of 0.367 to 0.424.