Logo image
SurreyCTS at BEA 2026 Shared Task 1: Semantic Funnelling and Entropy-based Multilingual Lexical Difficulty Prediction
   

SurreyCTS at BEA 2026 Shared Task 1: Semantic Funnelling and Entropy-based Multilingual Lexical Difficulty Prediction

Proceedings of the 21st Workshop on Innovative Use of NLP for Building Educational Applications (BEA 2026), pp.1016-1023
21st Workshop on Innovative Use of NLP for Building Educational Applications (BEA 2026), 21 (San Diego, California, United States)
The 64th Annual Meeting of the Association for Computational Linguistics (San Diego, California, USA, 03/07/2026–04/07/2026)
07/2026
Text accessibility Large Language Models Natural Language Processing
This paper describes the SurreyCTS submission 1 to the BEA 2026 shared task on lexical difficulty prediction, entered in the Open Track. Our approach progressed from base-line multilingual encoders to a hybrid Rem-BERT architecture with extensive feature engineering , combining semantic funnelling, lexical similarity features, attention-derived signals , and language-aware representations. On our internal production validation set, the best single models achieved RMSE 0.8122 (prod-H) and Pearson correlation 0.8968 (prod-G). A weighted ensemble of the five strongest systems, with weights proportional to inverse squared validation RMSE, was submitted as our final entry. On the official shared-task test set, the ensemble achieved RMSE 1.034, 0.945, and 0.861 for Spanish, German, and Chinese respectively, outperforming the open-track base-line in all three settings and placing fifth among open-track teams.

(3)

url
https://aclanthology.org/2026.bea-1.70/
Published (Version of record)
url
https://sig-edu.org/bea/2026
Event Website Workshop website
url
https://2026.aclweb.org/
Event Website Conference website
1
Logo image