KCLarity at SemEval-2026 Task 6: Encoder and Zero-Shot Approaches to Political Evasion Detection
| Authors | Archie Sage & Salvatore Greco |
| Year | 2026 |
| Field | NLP |
| arXiv | 2603.06552 |
| Download | |
| Categories | cs.CL |
Abstract
This paper describes the KCLarity team's participation in CLARITY, a shared task at SemEval 2026 on classifying ambiguity and evasion techniques in political discourse. We investigate two modelling formulations: (i) directly predicting the clarity label, and (ii) predicting the evasion label and deriving clarity through the task taxonomy hierarchy. We further explore several auxiliary training variants and evaluate decoder-only models in a zero-shot setting under the evasion-first formulation. Overall, the two formulations yield comparable performance. Among encoder-based models, RoBERTa-large achieves the strongest results on the public test set, while zero-shot GPT-5.2 generalises better on the hidden evaluation set.
Engineering Breakdown
Plain English
This paper presents the KCLarity team's approach to the CLARITY shared task at SemEval 2026, which focuses on detecting ambiguity and evasion techniques in political discourse. The team investigated two different modeling strategies: directly predicting clarity labels versus first predicting evasion techniques and then deriving clarity through a task taxonomy hierarchy. They evaluated multiple model architectures, from encoder-based models like RoBERTa-large to zero-shot decoder-only models like GPT-5.2, finding that both formulations achieved comparable performance on public test data. However, the zero-shot GPT-5.2 generalized better to a hidden evaluation set, suggesting that scaling and instruction-following capabilities provide advantages for out-of-distribution political discourse.
Core Technical Contribution
The core contribution is demonstrating that hierarchical label prediction—predicting intermediate evasion labels and deriving final clarity through task taxonomy—is competitive with direct end-to-end label prediction, despite adding an extra prediction step. The authors identify that the choice between these two formulations (direct vs. hierarchical) has minimal impact on performance, but the model class matters substantially: zero-shot large language models generalize better to held-out political discourse than fine-tuned encoder models. They also show that auxiliary training variants can be explored without significant performance degradation, opening design space for modular classification systems. The insight that zero-shot decoder-only models outperform fine-tuned smaller encoders on hidden test sets suggests that scale and pretraining diversity matter more than task-specific adaptation for evasion detection in political language.
How It Works
The system takes political discourse text as input and processes it through two alternative formulations. In the direct formulation, text is passed to a model (RoBERTa-large or GPT-5.2) that outputs a clarity label directly. In the hierarchical formulation, the same text first passes through a model trained to predict evasion technique labels (e.g., 'deflection', 'equivocation', 'false dilemma'), and then these predicted evasion labels are mapped through a predefined task taxonomy to derive the final clarity classification. For the zero-shot decoder-only approach, GPT-5.2 uses in-context learning—the model receives task descriptions and possibly examples in the prompt, then generates the label without any gradient-based training. Both paths use the same underlying transformer architecture but differ in supervision signal: the hierarchical path requires evasion labels during training, allowing the model to learn intermediate linguistic patterns before predicting clarity. The evaluation compares these approaches on both a public test set and a hidden evaluation set to measure both in-distribution and out-of-distribution generalization.
Production Impact
For teams building political discourse analysis systems, this work suggests that you don't need to choose between architectural simplicity (direct prediction) and explainability (hierarchical prediction)—both approaches perform equivalently. In practice, this means you could implement interpretable evasion-first pipelines without sacrificing accuracy, making it easier to audit decisions or debug failures in sensitive applications. However, the finding that zero-shot GPT-5.2 outperforms fine-tuned RoBERTa on hidden data carries significant cost and latency implications: calling a large language model API per document is orders of magnitude slower and more expensive than running a local RoBERTa-large inference. Organizations should conduct real-world latency and cost analysis before choosing the zero-shot path, especially at scale. The practical recommendation: start with fine-tuned RoBERTa-large for cost-sensitive production systems, but maintain the zero-shot GPT approach as a fallback or validation mechanism when encountering unusual discourse patterns, since its superior generalization suggests it will be more robust to domain drift.
Limitations and When Not to Use This
The paper does not address computational requirements or inference latency, which are critical for real-time content moderation systems—zero-shot GPT-5.2 likely requires API calls while RoBERTa allows local inference, a distinction not explored. The evaluation is limited to the CLARITY task taxonomy and may not generalize to other political discourse classification problems with different label sets or evasion types. The paper assumes that evasion techniques are predictable from text alone and doesn't incorporate conversational context, speaker history, or multimodal signals that might improve clarity detection in practice. There's no analysis of failure modes: when do the approaches disagree? When does hierarchical prediction fail despite correct evasion predictions? What types of political discourse (campaign ads, debate transcripts, social media) are hardest to classify, and do the two formulations fail differently? The claim that both formulations yield 'comparable performance' lacks detailed ablation studies showing which components matter most, making it difficult to understand why hierarchical prediction doesn't suffer from error propagation through the evasion intermediate step.
Research Context
This work builds on shared task traditions in NLP (SemEval) and extends prior work on detecting manipulative language, propaganda, and stance in political discourse. CLARITY 2026 represents an evolution of misinformation and evasion detection research, which historically focused on factuality or stance but increasingly targets linguistic evasion techniques—how politicians avoid direct answers. The hierarchical formulation connects to broader research in multitask learning and structured prediction, where intermediate tasks provide inductive bias for difficult target tasks. The comparison between fine-tuned encoder models and zero-shot LLMs reflects a significant industry trend: as decoder-only models scale, they increasingly outperform smaller task-specific models without any training, challenging the fine-tuning paradigm that dominated 2020-2024. This work opens research directions in understanding why hierarchical decomposition doesn't hurt performance (perhaps evasion and clarity are highly correlated?), and in building efficient zero-shot systems for politically sensitive content that still maintain the interpretability of intermediate predictions.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::