TFCS-IELTS: A Three-Tier Cascaded Scoring System for Offline IELTS Writing Assessment

Abstract

Automated IELTS Writing assessment requires reliable scoring under privacy and resource constraints. We propose TFCS-IELTS, a Three-Tier Cascaded Scoring architecture that escalates computation only when earlier stages are uncertain. Stage 1 uses Term Frequency-Inverse Document Frequency (TF-IDF) with Ridge regression and k-nearest-neighbor variance estimation; Stage 2 uses a fine-tuned DistilBERT regressor; Stage 3 invokes a local Qwen2.5-3B model conditioned by retrieved IELTS reference essays. Experiments are conducted on the Hugging Face chillies/IELTS-writing-task-2-evaluation corpus, cleaning 10,324 raw essays to 8,722 valid samples split into 6,782 train, 1,454 validation, and 486 test essays. The lightweight cascade reaches a Mean Absolute Error (MAE) of 0.9033 and a Quadratic Weighted Kappa (QWK) of 0.4659 while invoking the LLM for only 5.3% of test essays with 0% parsing failures. The max-accuracy cascade achieves MAE 0.8909, outperforming both TF-IDF Ridge (0.9599) and fine-tuned DistilBERT (0.9105). Results demonstrate that retrieval-augmented LLM scoring can be selectively integrated into a practical offline pipeline without sacrificing efficiency or reliability.

Related publications

• Neural Network with Fine-Tuned BERT for IELTS Writing EvaluationIEEE International Conference on Smart Information Systems and Technologies · 2025 · Citations: 3 · Shared topic: Natural Language Processing Techniques
• CulturalRiddles: A Multicultural Benchmark of RiddlesarXiv · 2026 · Shared topic: Natural Language Processing Techniques
• Aligning Cultural Behaviors in Arabic Language Models via PreferencesarXiv · 2026 · Shared topic: Natural Language Processing Techniques
• Codewithsayanjib/slm-systematic-review: v1.1Zenodo · 2026 · Shared topic: Natural Language Processing Techniques