SilkPeak at SemEval-2026 Task 6: When Politicians Dodge — Unmasking Evasion in Political Interviews through Joint Multi-Task Transformer Learning
Year: 2026
Type: conference-paper
Abstract: We describe our system for SemEval-2026 Task 6 (CLARITY), a shared task centred on recognising evasive communication in political interviews.Two subtasks are involved: determining whether a politician's answer is a Clear Reply, an Ambivalent response, or a Clear Non-Reply (Subtask 1), and assigning one of nine fine-grained evasion strategy labels to the same exchange (Subtask 2).Our approach treats both subtasks as a single joint problem.A DeBERTa-v3-Large encoder is shared across both tasks, with the question-answer pair fed as one concatenated sequence; two independent linear heads then produce the respective predictions.Gradients from both cross-entropy losses flow back through the shared encoder simultaneously, so learning signals from the evasion taxonomy directly inform clarity-level decisions and vice versa.On the official evaluation set our system scores 0.76 macro F1 on Task 1, up from 0.58 with a standard singletask DeBERTa-Base.We also test hierarchical bi-encoding, weighted layer pooling, K-fold ensembling, and LoRA-tuned LLaMA-3-8B, and find that the joint discriminative approach beats all of them, often by wide margins. 1
Authors: Amruth Tetakali, Lavanya Tetakali
Institution: Indian Institute of Technology Roorkee
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7166823585