Undergraduate research · 2026
Explainable AI for Automated Grading of Khmer Language Short-Answer Questions in Education
An undergraduate research project combining a real Khmer examination benchmark, model comparison, and word-level explanations for teacher-assisted grading.
Overview
Khmer has limited resources for automated educational assessment. This undergraduate project investigates how multiple model families can support short-answer grading while exposing words associated with model predictions. Supervised by Dr. Khim Chamroeun at RUPP.
My Contribution
- Developed a benchmark of 1,184 teacher-graded Khmer answers from 203 students across two schools, covering 41 questions and four subjects.
- Fine-tuned Qwen-KhmerGrader-4B, Gemma-KhmerGrader-4B, and SEA-LION-KhmerGrader-E2B using QLoRA adaptation of existing Qwen, Gemma, and SEA-LION bases.
- Compared the grading approaches with SVR, BiLSTM, multilingual encoders, and report-listed GPT, Claude, Gemini, and DeepSeek baselines in bare and reasoning settings.
- Implemented SHAP word-level explanations and a Gradio teacher-assist prototype; released software and dataset documentation.
Key Features
- Real examination answers across Biology, History, Geography, and Earth Science
- Classical, recurrent, encoder, and language-model approaches
- QWK and prediction-error evaluation
- Word-level explanations and a teacher-assist interface
Technical Implementation
- Public dataset and curated subset are alternative configurations, not independent corpora to concatenate
- Label-stratified answer-level splitting in the historical report
- SHAP explanations are model attributions, not proof of pedagogical correctness
- Dataset CC BY-NC 4.0; software MIT; student records require human and ethical oversight
Results
Report-stated test QWK 0.850 for Qwen-KhmerGrader-4B (n = 137; agreement with one teacher's marks; 67% exact, 80% within one point), above classical 0.802, BiLSTM 0.789, encoder 0.777 (n = 178), and the best zero-shot frontier model, Claude Opus 4.8 (0.734).
Public dataset v1.0.1 contains 1,184 full-corpus answers and a 909-answer curated subset. Code and the final undergraduate report are publicly available.
Historical report results
Report-stated frontier evaluation date: . Frontier evaluation date stated in the undergraduate report; not independently verified against live APIs.
QWK is agreement with the recorded teacher grades. These tables describe the undergraduate report.
| Research model | Upstream base | Reported test n | Reported QWK |
|---|---|---|---|
| Qwen-KhmerGrader-4B | Qwen 3.5 4B | 137 | 0.850 |
| Gemma-KhmerGrader-4B | Gemma 4 E4B | Not established in reviewed summary | 0.844 |
| SEA-LION-KhmerGrader-E2B | Gemma-SEA-LION v4.5 E2B | Not established in reviewed summary | 0.713 |
Qwen's test count comes from report Table 10. Missing counts are not inferred for the other variants. SEA-LION includes AI Singapore's adaptation of Google's Gemma; my contribution is task adaptation, not creation of the upstream models.
| Recorded model version | Prompting | Parsed n | Unparsed n | Reported QWK |
|---|---|---|---|---|
| openai/gpt-5.5-20260423 | bare | 126 | 11 | 0.684 |
| openai/gpt-5.5-20260423 | reasoning | 136 | 1 | 0.708 |
| anthropic/claude-opus-4.8 | bare | 137 | 0 | 0.662 |
| anthropic/claude-opus-4.8 | reasoning | 137 | 0 | 0.734 |
| google/gemini-3.5-flash-20260519 | bare | 72 | 65 | 0.555 |
| google/gemini-3.5-flash-20260519 | reasoning | 117 | 20 | 0.585 |
| deepseek/deepseek-v4-flash-20260423 | bare | 134 | 3 | 0.553 |
| deepseek/deepseek-v4-flash-20260423 | reasoning | 136 | 1 | 0.579 |
How to interpret these results
- QWK (Quadratic Weighted Kappa) measures agreement with the recorded teacher grades, not independent correctness or classroom benefit.
- Fine-tuned-family QWK values are transcribed from report Table 12. Qwen's 137-answer test count is documented in Table 10; the reviewed summary does not establish separate test counts for the other two variants.
- Frontier rows are transcribed from report Tables 13 and 16 and the report-to-evidence guide. Unparsed responses are excluded from QWK, so the denominators differ. Bare and reasoning are the report's prompting settings.
- The evidence guide records unresolved champion-artifact, calibration, and training-provenance concerns. Matching report tables to a saved summary is not independent reproduction. Model version strings are recorded identifiers, not confirmation of live API availability.
- The corpus uses one teacher's marks and answer-level splits. These results do not establish generalization to unseen students, questions, or schools, or a general superiority over frontier models.
Limitations
- One teacher supplied the marks; reference answers are student-derived proxies. Answer-level splits do not establish generalization to unseen students, questions, or schools.
- This is a research prototype for human review. No classroom field study, autonomous-grading safety, or peer-reviewed publication is claimed.