Neural Fusion and Feature-Rich Modeling for L1-Aware Vocabulary Difficulty Prediction

Research Question and Approach

This study addresses the BEA 2026 shared task on L1-aware vocabulary difficulty prediction, which aims to estimate how difficult specific English words are for learners based on their native language (L1). The authors model this as a regression task, training separate systems for Spanish, German, and Chinese speakers. Their approach utilizes a neural fusion architecture that combines deep contextual representations from multilingual encoders (such as BGE-M3, multilingual E5, and LaBSE) with a rich set of engineered features. These features capture lexical frequency, surface form, retrieval evidence, semantic alignment, cognate similarity, and masked-language-model (MLM) predictability.

The researchers found that neural fusion consistently outperforms official closed-track baselines, achieving substantial reductions in Root Mean Square Error (RMSE) across all three languages. A key finding is that sentence-embedding-oriented models (e.g., BGE-M3) are significantly more effective than standard token-level encoders like mBERT or XLM-R. Feature analysis reveals that while frequency is the most stable and consistent predictor across languages, other features—such as semantic-domain alignment, cognate similarity, and contextual predictability—provide essential, complementary signals that improve model performance.

This work demonstrates that integrating structured, linguistically motivated features with modern neural encoders is a highly effective strategy for tasks where domain-specific knowledge (like L1-specific difficulty) is required. By moving beyond simple text-only models, the authors provide a robust framework for educational NLP applications, such as automated text simplification and personalized language learning, where accurately gauging the difficulty of vocabulary for specific learner groups is critical.