Abstract
We propose a stance classification framework based on hierarchically organized Differentiating Properties (DPs) -- linguistic features capturing opposing semantic attributes that define the poles of discourse polarization. DPs are automatically extracted from corpora via a multi-stage LLM pipeline and paired into oppositional variants, forming a bipolar semantic space. A linear Support Vector Machine (SVM) classifier operating on grouped DP embeddings performs stance classification by projecting target texts onto the polarization axis defined by opposing DP poles. Spearman rank correlation (ρ) and cosine similarity between pole embeddings define semantic convergence, serving as geometric diagnostic tools for assessing DP quality prior to classification. Experiments on three polarized corpora across two embedding models show that both metrics reliably predict quality of classification, as higher values correspond to higher F1 scores, while replacing high-ρ convergent pairs with low-ρ divergent pairs leads to a systematic performance degradation of up to 17%.
Paper Type
Short Paper
DOI
10.62036/ISD.2026.96
Semantic Convergence of Differentiating Properties as a Quality Predictor for Stance Classification
We propose a stance classification framework based on hierarchically organized Differentiating Properties (DPs) -- linguistic features capturing opposing semantic attributes that define the poles of discourse polarization. DPs are automatically extracted from corpora via a multi-stage LLM pipeline and paired into oppositional variants, forming a bipolar semantic space. A linear Support Vector Machine (SVM) classifier operating on grouped DP embeddings performs stance classification by projecting target texts onto the polarization axis defined by opposing DP poles. Spearman rank correlation (ρ) and cosine similarity between pole embeddings define semantic convergence, serving as geometric diagnostic tools for assessing DP quality prior to classification. Experiments on three polarized corpora across two embedding models show that both metrics reliably predict quality of classification, as higher values correspond to higher F1 scores, while replacing high-ρ convergent pairs with low-ρ divergent pairs leads to a systematic performance degradation of up to 17%.
Recommended Citation
Gierszewski, M., Szymański, J. & Duch, W.(2026). Semantic Convergence of Differentiating Properties as a Quality Predictor for Stance Classification. In M. Valenta, B. Mannová, R. Pergl, A. Przybylek, M. Lang, H. Linger, C. Schneider, N. Iivari, & E. Insfran (Eds.), Making ISD Sustainable: Reloaded with AI and Automation (ISD2026 Proceedings). Prague, Czech Republic: Czech Technical University in Prague. ISBN: 978-80-01-07585-2. https://doi.org/10.62036/ISD.2026.96