SG-UniBuc-NLP at SemEval-2026 Task 6: Multi-Head RoBERTa with Chunking for Long-Context Evasion Detection

2026-04-29

AI Summary

This study addresses the challenge of detecting evasive responses in political interviews by proposing a long-context modeling approach that combines overlapping sliding window chunking with multi-task learning, effectively circumventing the sequence length limitations of standard Transformers. The model employs a RoBERTa-large encoder and aggregates chunk-level representations through element-wise max pooling. To enhance robustness, predictions are integrated across seven stratified cross-validation folds. Evaluated on SemEval-2026 Task 6, the method achieved 11th place in both subtasks—coarse-grained clarity ternary classification and fine-grained evasion strategy nine-way classification—with Macro-F1 scores of 0.80 and 0.51, respectively, demonstrating the effectiveness of the proposed framework for analyzing complex political discourse.

Abstract

We describe our system for SemEval-2026 Task 6 (CLARITY: Unmasking Political Question Evasions), which classifies English political interview responses by coarse-grained clarity (3-way) and fine-grained evasion strategy (9-way). Since responses frequently exceed the 512-token limit of standard Transformer encoders, we apply an overlapping sliding-window chunking strategy with element-wise Max-Pooling aggregation over chunk representations. A shared RoBERTa-large encoder supplies two task-specific heads trained jointly via a multi-task objective, with inference-time ensembling over 7-fold stratified cross-validation. Our system achieves a Macro-F1 of 0.80 on Subtask 1 and 0.51 on Subtask 2, ranking 11th in both subtasks.

Problem

• evasion detection
• political interviews
• long-context
• clarity classification
• RoBERTa

Innovation

• sliding-window chunking
• multi-task RoBERTa
• evasion detection
• long-context modeling
• max-pooling aggregation