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

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, quick

Summary

The SurreyCTS system, developed for the BEA 2026 shared task on lexical difficulty prediction, integrates multilingual transformer encoders such as RemBERT and COMET. This approach also incorporates several engineered linguistic features, including semantic funnelling, lexical similarity metrics, attention-derived signals, and language-aware representations. The system achieved fifth place among open-track teams in the competition, demonstrating a robust methodology for assessing word difficulty. Notably, its weighted ensemble of the five strongest components surpassed the open-track baseline performance across all three target learner L1 groups: Spanish, German, and Chinese, highlighting its effectiveness in predicting lexical difficulty for diverse language backgrounds and its competitive standing in the task.

Key takeaway

For NLP Engineers developing multilingual language learning tools or content localization systems, consider integrating a hybrid approach combining transformer encoders like RemBERT or COMET with engineered linguistic features. Your systems can achieve competitive performance in lexical difficulty prediction, especially for diverse L1 learner groups. Employing weighted ensembles of your best models can further enhance robustness and accuracy, improving the user experience for language learners.

Key insights

SurreyCTS combines multilingual transformers with engineered linguistic features to effectively predict lexical difficulty across diverse L1 groups.

Principles

• Multilingual transformers enhance lexical difficulty prediction.
• Engineered linguistic features improve model performance.
• Ensemble weighting boosts system robustness.

Method

SurreyCTS combines RemBERT and COMET multilingual transformer encoders with engineered features like semantic funnelling, lexical similarity, and attention-derived signals. A weighted ensemble of the top five systems is then used for final prediction.

In practice

• Integrate RemBERT or COMET for multilingual NLP.
• Develop custom linguistic features for specific tasks.
• Apply weighted ensembles for robust model deployment.

Topics

• Lexical Difficulty Prediction
• Multilingual Transformers
• RemBERT
• COMET
• Linguistic Feature Engineering
• Ensemble Methods

Best for: Research Scientist, AI Scientist, NLP Engineer

Related on AIssential

• Equity with Efficiency: An Empirical Study of Tokenizers for Multilingual Large Language Models — Cross-lingual fairness and tokenization efficiency are not fundamentally at odds in multilingual LLMs.
• What Makes Words Hard? Sakura at BEA 2026 Shared Task on Vocabulary Difficulty Prediction — Soft-target loss fine-tuning significantly improves LLM and MLM performance in continuous value prediction tasks.
• How Far Can Classical NLP Go? From Bag-of-Words to Stacking on Spooky Author Identification — Classical NLP, with careful feature engineering and stacking, excels at stylistic authorship attribution.
• Extracting Problem and Method Sentence from Scientific Papers: A Context-enhanced Transformer Using Formulaic Expression Desensitization — Formulaic expression desensitization and context-enhanced transformers improve problem/method sentence extraction from scientific papers, outperforming baselines.
• Hybrid Human-LLM Corpus Construction and LLM Evaluation for the Caused-Motion Construction — LLMs struggle with complex linguistic constructions like CMC, necessitating novel corpus creation methods.
• A Beginner’s Guide to NLP Token Classification: NER, POS Tagging, Chunking, and BERT — Token classification, including NER, POS tagging, and chunking, is fundamental for machines to understand human language.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.