Readability Scorer using NLP

Teachers, editors and content teams share one practical problem: is this text too hard for its audience? A school textbook, a patient leaflet or a government form that overshoots its readers fails at its job, but "too hard" is subjective — until you measure it. Readability formulas turn text into numbers: sentence length, word length and syllable counts combine into grade-level scores that have been validated against comprehension tests for decades. Modern NLP goes one step further, training classifiers on corpora labeled by reading level. This project builds both: six classic formulas computed from real linguistic analysis, plus a readability-level classifier, all demonstrated in a browser app that analyzes any pasted text live and shows exactly how each score was derived.

1. Pasted text is tokenized into sentences and words; a rule-based counter measures syllables, characters and polysyllabic words.
2. The six formulas are computed from these counts (e.g. Flesch Reading Ease = 206.835 − 1.015·(words/sentences) − 84.6·(syllables/words)).
3. The same linguistic feature vector feeds a softmax classifier trained on reference texts labeled Elementary, Intermediate or Advanced.
4. The demo displays all scores with interpretation bands (e.g. Flesch 60–70 ≈ plain English) and the classifier's level prediction with confidences.
5. Sentences exceeding complexity thresholds are highlighted in the source text.
6. The notebook documents the formulas, the classifier training and the evaluation on reference texts.

Tech stack:

• Python (text processing, formula implementations)
• NLTK / regex tokenization (sentence and word splitting)
• NumPy, scikit-learn (softmax classifier training and evaluation)
• Matplotlib (score distributions, classifier evaluation plots)
• Jupyter notebook (formula walkthrough and classifier training)
• Single-file HTML/CSS/JS demo app (live analysis, metric dashboard)

• [Six readability formulas] Flesch Reading Ease, Flesch-Kincaid Grade Level, Gunning Fog, SMOG, Coleman-Liau Index and ARI — each computed from measured linguistic counts, not approximations.
• [Real syllable counting] A rule-based syllable counter handles silent-e, diphthongs and edge cases so formula inputs are genuine measurements.
• [Level classifier] A softmax classifier trained on reference texts assigns Elementary, Intermediate or Advanced reading level, with per-class confidence shown.
• [Live text analysis] Paste any text and the demo tokenizes it, counts sentences, words, syllables and polysyllabic words, then reports all six scores instantly.
• [Metric explanations] Every score ships with its formula, its interpretation scale and what drives it up or down — the report practically writes itself.
• [Reference text comparison] Built-in sample texts at each level let the viva compare a known-easy and known-hard passage side by side.
• [Sentence-level highlighting] Long or complex sentences are flagged in the input so the user sees exactly where difficulty concentrates.

• Complete source code (tokenization, syllable counter, six formulas, classifier, demo app)
• Jupyter notebook (formula derivations, classifier training and evaluation)
• Single-file browser demo app with live text analysis
• Project abstract PDF
• Project report (background, formula theory, classifier design, evaluation)
• PPT presentation and viva Q&A document (readability theory, NLP preprocessing, softmax classification)

• Formula scores measure surface features (length, syllables), not true comprehension difficulty — vocabulary rarity, coherence and background knowledge are outside their view, and the report says so.
• The syllable counter is rule-based and approximate on unusual words, abbreviations and proper nouns.
• Readability formulas were calibrated on English prose; they are not validated for other languages or for poetry, code or highly formatted documents.
• The classifier is trained on a small reference set; its level predictions are indicative, and its evaluation numbers come from your own build, not from a claim.