Skip to content
conftrace_
2026 EMNLP EMNLP 2026

Beyond Over-Simplification and Over-Personalization: Lexical Complexity Prediction for Leveled Reading

AbstractLexical Complexity Prediction (LCP) in real-world scenarios, particularly in language education, is often limited by the trade-off between overly simplified general models and unscalable personalized ones. To address this, we introduce LCP for Leveled Reading (LCP-LR), a task aiming to operationalize group-level personalization by identifying complex words relative to specific proficiency levels, thereby balancing pedagogical guidance and scalability. We construct the first large-scale benchmark dataset for this task, comprising 100,917 instance annotations across 1,401 leveled texts. Furthermore, we propose Dual-Prior Adversarial Learning (DPAL), a framework that incorporates word- and text-level difficulty priors alongside adversarial debiasing to mitigate the model’s over-reliance on simplistic absolute difficulty shortcuts. Experimental results show that DPAL significantly outperforms both conventional pre-trained baselines and large language models, improving the Macro of the best baseline (XLNet) from 74.19% to 77.81%. Our evaluations also reveal that large language models underperform in zero- and few-shot settings without fine-tuning or Chain-of-Thought prompting. Overall, this work promotes the practical deployment of LCP and lays a crucial technical foundation for dynamic text adaptation within leveled reading.

More from EMNLP 2026

All EMNLP 2026 papers →