~/divyam/stories lexigrade-text-readability-classifier
the lab · models and research
LexiGrade: A Text Readability Classifier
A calibrated NLP pipeline that combines LightGBM, sentence-transformer embeddings, and classical readability indices to classify text by reading level.
- when
- April 2026
- role
- ML Engineer
- stack
- LightGBMall-MiniLM-L6-v2PythonScikit-learnNLP

Overview
LexiGrade predicts the reading level of a passage by combining semantic and linguistic signals. Rather than depending only on formulas like Flesch-Kincaid, it fuses sentence-transformer embeddings with classical readability features.
The project was designed for skewed real-world data, where expert-level text is underrepresented and naive classifiers tend to underperform on the hardest class.
Problem
Classical readability indices are useful but shallow: they cannot fully capture semantic complexity. Transformer embeddings add semantic signal, but on small expert-class samples they can overfit. The challenge was to combine both without letting label imbalance dominate the model.
Solution
I trained a LightGBM regressor on a combined feature vector: all-MiniLM-L6-v2 embeddings plus handcrafted linguistic features. A stratified split and 2.5x expert-class weighting helped the model pay attention to the rare upper-grade examples.
A calibrated fusion layer blends the ML score with FRE, FKG, FOG, SMOG, CLI, and ARI. An ML trust override catches short high-FRE sentences that classical formulas incorrectly mark as easy.
Highlights
- LightGBM over transformer embeddings and handcrafted linguistic features
- Stratified split with 2.5x expert-class sample weight boosting
- Fusion layer across FRE, FKG, FOG, SMOG, CLI, and ARI readability indices
- ML trust override for high-confidence expert-level edge cases
Outcome
The calibrated fusion approach improved expert-class recall compared with either the ML model or classical indices alone, showing that semantic embeddings and readability formulas capture complementary signals.