Overview of HOPE-EXP at IberLEF 2026: Outcome-Oriented Expectation Analysis in Social Media

Resumen

This paper presents an overview of HOPE-EXP, a shared task organized at IberLEF 2026 on hope detection in social media as a multi-level, outcome-oriented structured-prediction problem. Given a bilingual (English/Spanish) post, systems had to jointly (A) assign one of six primary hope labels, (B) detect a multi-label set of trigger emotions, (C) extract up to three spans naming the expected or avoided outcome, and (D) classify the stance (Desired/Avoided) and actor (Self/Other/World-System/Unclear) of each span. The corpus comprises 4,857 training and 2,082 test posts from Reddit. Twenty-three participants registered and submitted 195 runs; eight systems obtained a leaderboard score, and six teams submitted a system-description paper. Approaches ranged from lexical baselines to fine-tuned transformers (XLM-RoBERTa) and parameter-efficient LLMs (QLoRA-tuned Qwen3.5-4B), with the best system reaching 0.785 overall. Primary label classification was comparatively tractable, while span extraction and actor classification proved the hardest components.