Multi-Head RoBERTa with Chunking for Political Evasion Detection

Researchers from the National University of Singapore and the University of Bucharest created a system for SemEval-2026 Task 6, titled CLARITY: Unmasking Political Question Evasions. This task focuses on categorizing English political interview replies based on coarse-grained clarity (3-way) and fine-grained evasion strategy (9-way). To manage responses that exceed the 512-token limit of standard Transformer encoders, they utilized an overlapping sliding-window chunking method along with element-wise Max-Pooling aggregation for chunk representations. A shared RoBERTa-large encoder features two task-specific heads that are trained together using a multi-task objective, with 7-fold stratified cross-validation applied during inference-time ensembling. Their system achieved a Macro-F1 score of 0.80 for Subtask 1 and 0.51 for Subtask 2, placing 11th in both categories.