This is the 2D counterpart of the 3D Automated Essay Scoring simulator, built around the exact same real feature-based grading pipeline. A synthetic "essay" is genuinely generated word-by-word from your quality knobs, then a generic feature extractor — one that would work on any text, not just generated ones — scans the raw string and computes four real numbers:
wordCount = total tokens
avgSentLen = mean words per sentence (split on . ! ?)
ttr = unique(lowercased tokens) / wordCount
errorRate = tokens not found in the word dictionary / wordCount
Each raw feature is normalised to [0,1] and combined by a fixed-form weighted-regression model, identical in shape to the 3D version:
score = 100 · (w1·f1 + w2·f2 + w3·f3 + w4·f4) / (w1+w2+w3+w4)
The bar chart shows each term wi·fi as its own bar — the four bars sum exactly to the predicted score, so you can see precisely how much each feature contributed.
- Target length / Sentence length — control how many words are generated and how long each generated sentence runs, which directly changes the real measured word count and average sentence length.
- Vocabulary diversity — the probability that each word is drawn from a large 140-word pool instead of a small 30-word filler pool; higher values push more distinct words into the text, which raises the real computed type-token ratio.
- Spelling-error rate — the probability that a generated word is corrupted (a letter deleted, duplicated or swapped) before being written down. The extractor never sees this probability directly — it only flags a token as an error if it fails a real dictionary lookup, exactly like a real spell-checker.
- Weight sliders — change how much each of the four features counts toward the final score; watch the bars resize and the score update live.
Real-world relevance: this reduce-to-features-then-regress approach (word count, sentence complexity, lexical diversity, mechanical accuracy) is exactly how early, still widely deployed automated essay-scoring engines like e-rater and IntelliMetric work — interpretable and auditable, unlike a black-box neural scorer.