Automated Essay Scoring — NLP Feature Grading Model

Scoring Model Weights

Vocabulary diversity 0.34

Syntactic complexity 0.33

Semantic coherence 0.33

Rater Behaviour

Human-rater noise (σ, pts) 12

Passing threshold 60

Live Readouts

Avg AI score

0.0

Passing (AI)

0 / 0

Agreement r

0.00

RMSE vs human

0.0

How it works

Automated essay-scoring systems don't read prose the way a person does — they reduce each essay to a small set of numeric features and combine them with a trained regression model:

score = 100 · (w1·vocab + w2·syntax + w3·coherence) / (w1+w2+w3)

vocab      = type–token ratio (unique words ÷ total words)
syntax     = normalised average sentence length / clause depth
coherence  = mean cosine similarity between consecutive
             paragraph embeddings

Each point in the cube is one synthetic essay positioned by its three feature values. Point colour encodes the current AI-predicted score (red → low, green → high). The translucent plane is the model's decision boundary: everything on the green side of it scores above your passing threshold.

Every essay also has a simulated "true quality" that a human rater would score, with added Gaussian noise σ to represent real inter-rater disagreement. Agreement r is the Pearson correlation between the AI score and that simulated human score — in real deployments this is exactly the metric (quadratic weighted kappa's simpler cousin) used to validate an automated grader before it's trusted in production. Raise the noise slider and even a well-tuned model's agreement with any single human rater degrades — a real limitation discussed in every AES validation study.

Weight sliders — change how much each feature counts toward the AI score; watch the point cloud recolour and the plane retilt live.
Rater noise — simulates how consistent (or not) a human grader is essay to essay.
Passing threshold — sets the pass/fail cutoff line the plane visualises.
New Essay Batch — resamples 70 synthetic essays from the same underlying feature-correlation model.

Low predicted score High predicted score

x: vocabulary
y: syntax
z: coherence

⚠ Couldn't load the 3D engineThree.js failed to load from the CDN. Check your connection and reload.