REBECCA at eRisk 2024: Search for Symptoms of Depression Using Sentence Embeddings and Prompt-Based Filtering

Abstract Depression is a complex mental health disorder characterized by persistent feelings of sadness, hopelessness, and a lack of interest or pleasure in daily activities. It significantly affects an individual’s well-being, impairing their ability to work, socialize with others and be creative. Social media is used by billions of people globally who interact and generate an abundance of posts and texts. Analysis of social interaction data offers opportunities to gain valuable insights into people’s mental health and potentially take supportive action. eRisk 2024 focuses on the challenge of early risk detection on the Internet and has established a number of tasks for this reason. We participated in Task 1: Search for symptoms of depression. The aim of this task is to rank user sentences in terms of 21 symptoms of depression. This paper presents our approach combining ranking sentences using cosine similarity and Transformer embeddings and refining our results with the use of a Large Language Model (LLM). Our LLM-refined approach was among the best performing ones among the 29 runs submitted by the 9 participating teams.