Two-stage data preparation pipeline: rule-based synthetic generation plus LLM paraphrase augmentation without data leakage

Design · All designs
Evidence · 2 claims (1 for, 1 mixed) · 1 study (1 design), q2 · 0 of 1 report an effect size · 2 claims rest on one study

Description

The pipeline first generates synthetic score-level-1 samples in the training partition only, using "template-based generation (40%), quality degradation (35%), and a combined approach (25%)", including truncation to 20–40% of original length and removal of domain-critical keywords. Second, paraphrase-based augmentation via the Gemini API balances minority classes at scores 2–4 to a uniform 25% per level. Splitting was done once at the student level so each student's responses appear in a single partition, "thereby preventing cross -dimensional data leakage," and the test set stayed locked and unmodified.

Design Implications

Context

Requirements

• Augmentation and generation confined strictly to the training partition; fixed random seed; target counts set to the majority class size per dimension

Constraints

• Alternative techniques (back-translation, synonym substitution, EDA) were excluded after preliminary experiments showed lower output quality and inconsistent meaning preservation in the Indonesian physics essay domain; in Overview, augmentation marginally decreased QWK

Target Learners

• Indonesian high school physics students whose essays form the scored dataset

Learning Goals

• Reliable ordinal automated scoring of critical thinking dimensions despite severe class imbalance

Claims

• Augmentation Critical Ordinal Qwk Gains [+M]
• Overview Augmentation Qwk Tradeoff [~M]

Related Designs

Examples

-

Key Sources

• Firdausi, H.; Wasis; Sholikah, R. W.; Lemantara, J.; Rosyadi, F. D.; Ginardi, R. V. H.; Pradityo, M. H. A.; Hakim, A. R. (2026). Educational Innovation through Automated Essay Scoring: A Multidimensional Framework for Evaluating Critical Thinking in High School Physics Essays. International Journal of Intelligent Engineering and Systems, Vol.19, No.8. https://doi.org/10.22266/ijies2026.0831.09

Page metadata

Field	Value
type	design
id	two-stage-synthetic-paraphrase-augmentation-pipeline
title	Two-stage data preparation pipeline: rule-based synthetic generation plus LLM paraphrase augmentation without data leakage
description	The pipeline first generates synthetic score-level-1 samples in the training partition only, using "template-based generation (40%), quality degradation (35%), and a combined approach (25%)", including truncation to 2...
status	review
generated	by process:settle-candidates on 2026-10-10
trust	unverified — no one has confirmed this page's content
sources	1 — firdausi-2026

Frontmatter is the page's machine-readable record — see the Schema & Guide.