📄 TOEBM at BEA 2026 Shared Task 1: Improving Lexical Difficulty Prediction with Context-Aligned Contrastive Learning and Ridge Ensembling

Summary

The paper tackles L1-aware lexical difficulty prediction for BEA 2026 Shared Task 1. The goal is to estimate how hard an English word is for learners with a particular first language. The authors argue that training with only a scalar regression target leaves the embedding space unstructured. In that setting, items of similar difficulty need not cluster, and cross-lingual alignment is not enforced.

Their framework, Context-Aligned Contrastive Regression, adds two auxiliary objectives to MSE regression:

• Cross-View Context Contrastive loss: pulls together encodings of the English target view and the full L1-aware input for the same item.
• Ordinal Soft Contrastive loss: weights pairs by a Gaussian kernel over difficulty-score distance.

Three multilingual encoders are trained independently with this objective, and their outputs are combined by a RidgeCV meta-regressor. The system is evaluated on Spanish, German and Chinese learner data.

The ensemble beats each single encoder and keeps errors flatter across difficulty levels. It also corrects systematic under- and over-estimation by individual models. Representation analysis indicates better cross-lingual alignment and an ordinal structure in the embedding space. This matters for building level-appropriate learning materials and vocabulary assessments.

Key points

• Data and labels: three L1 groups (ES, DE, CN), with targets given as continuous GLMM difficulty scores. Lower GLMM scores mean harder words.
• Base encoders: xlm-roberta-base, mdeberta-v3-base and mmBERT-base, stacked via ridge regression (RidgeCV).
• Error by difficulty bin: XLM-R shows strong score compression, with error rising sharply on easier words. mmBERT shows moderate error growth, and mDeBERTa-v3 is the most stable single model. The Ridge ensemble has the lowest or near-lowest MAE in almost all bins.
• Qualitative example: for ES "cruise" (GLMM 1.21), mmBERT predicts −0.50 while the ensemble predicts 1.23.
• Remaining failures: some items stay far off even after ensembling, such as "baking", "stake" and "umbrella". The authors attribute these errors to ambiguity or context dependence.
• Code: available at github.com/airlanggawicaksono/BEA2026TOEBM.