💬 NLP

Enhancing Automated Essay Scoring With Three Techniques: Two-Stage Fine-Tuning, Score Alignment, and Self-Training

This paper proposes a novel approach to enhance Automated Essay Scoring performance in both limited-data and full-data settings by integrating three key techniques—Two-Stage fine-tuning, Score Alignment, and uncertainty-aware self-training—into the DualBERT model, achieving state-of-the-art results on multiple datasets including ASAP++.

Original authors: Hongseok Choi, Serynn Kim, Wencke Liermann, Jin Seong, Jin-Xia Huang

Published 2026-08-12

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Hongseok Choi, Serynn Kim, Wencke Liermann, Jin Seong, Jin-Xia Huang

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a world where a tired teacher has to grade hundreds of essays every week. It's a massive job, and doing it fairly takes time and energy. To help out, scientists have built "Automated Essay Scoring" (AES) systems—computer programs that read student writing and give it a grade, just like a human would. Think of these programs as digital teaching assistants. For a long time, these assistants were like smart but clumsy robots; they needed to read thousands of examples before they could get good at grading. If a teacher only had a few essays to grade, the robot would get confused and give bad scores. This is a big problem because, in the real world, teachers often don't have thousands of essays to train the computer on; they just have a pile of papers from their own class.

The challenge is that teaching a computer to understand writing is hard. It's not just about counting words or checking grammar; it's about understanding the flow of ideas, the tone, and the structure. While huge, powerful AI models (like the ones that write stories or chat with you) exist, they often struggle with this specific task unless they are given a massive amount of data. So, the big question for researchers is: How can we make a computer grader that is smart and fair, even when it hasn't seen many examples?

This paper introduces a clever new way to train these digital graders, especially when there are very few essays to learn from. The researchers, working with a model called DualBERT (which is like a smart reader that understands both individual sentences and the whole essay), developed three special techniques to boost performance.

First, they used a method called Two-Stage Fine-Tuning with LoRA. Imagine you are teaching a student who already knows the basics of English. Instead of making them relearn everything from scratch, you give them a special set of "highlighter pens" (LoRA) that let them focus on specific, tricky parts of the assignment without messing up what they already know. The researchers let the model practice with these pens in two rounds, trying different ways to weigh the importance of different essay traits (like content vs. organization) until it finds the perfect balance.

Second, they introduced Score Alignment. Sometimes, even a smart grader gets a little nervous and plays it safe. If a student writes a perfect essay, the computer might be afraid to give a perfect score and instead gives a 9.8 out of 10. If the essay is terrible, it might give a 0.2 instead of a 0. This technique acts like a "calibration tool." The researchers look at how the computer graded a small practice set of essays, see where it was being too shy or too bold, and then apply a simple math adjustment to fix the scores for the real test. It's like tuning a guitar so that every note hits the right pitch.

Third, they used Uncertainty-Aware Self-Training. This is where the computer gets to practice on essays it hasn't been graded yet. The computer reads these ungraded essays and gives them a "fake" score. But here's the catch: if the computer is unsure about its own score (it's "uncertain"), it throws that essay away. It only keeps the essays it feels confident about to use as extra practice material. This way, the computer learns from more examples without accidentally learning from its own mistakes.

When the researchers tested these three tricks together on a dataset called ASAP++, the results were impressive. In a scenario where the computer only had 32 essays to learn from (a very small amount), the new method improved the grading accuracy significantly. It reached 91.2% of the performance level of a model that had been trained on about 1,000 essays. In simpler terms, with just a tiny bit of data and these three smart tricks, the computer became almost as good as if it had read a whole library of essays.

Even more interestingly, when the computer had plenty of data (the "full-data" setting), the "Score Alignment" trick alone helped it beat all previous records, achieving a new best score. The researchers also tested this on other datasets with different types of essays and found the tricks worked there too, suggesting these methods are robust and not just a fluke for one specific test.

The paper suggests that these techniques are modular, meaning they can be mixed and matched with other systems. However, the authors are careful to note that while the results are strong, they still face limitations. For instance, the "Self-Training" method needs a large pile of ungraded essays to work, which might not always be available in every school. Also, the "Two-Stage" method takes a bit more time to train the computer initially, though it doesn't slow down the actual grading process later.

Ultimately, this study shows that we don't always need massive amounts of data to build great AI tools. By using smart strategies to fine-tune models, calibrate their scores, and carefully select practice material, we can create automated graders that are reliable even when data is scarce. This brings us one step closer to having helpful, fair, and efficient digital assistants for teachers everywhere.

Technical Summary: Enhancing Automated Essay Scoring With Three Techniques

Problem Statement
Automated Essay Scoring (AES) is critical for scalable education but faces significant hurdles in real-world deployment, primarily due to the extreme scarcity of labeled data. Constructing AES datasets is labor-intensive, expensive, and prone to inter-rater inconsistency. While Large Language Models (LLMs) have succeeded in various NLP tasks, their performance in AES remains suboptimal (Quadratic Weighted Kappa [QWK] of ~0.5–0.6) compared to task-specific deep learning models, and their computational requirements pose barriers to adoption. Furthermore, existing deep learning approaches often struggle in low-resource settings where only a small number of labeled essays are available.

Methodology
The authors propose a novel approach to enhance AES performance in both limited-data and full-data settings by integrating three key techniques onto a baseline model, DualBERT, which captures inter-token and inter-sentence semantics.

Two-Stage Fine-Tuning with Low-Rank Adaptation (LoRA):
- Stage 1: The base DualBERT model is fine-tuned on the labeled dataset.
- Stage 2: LoRA layers are inserted into the frozen base model. The authors introduce a strategy to sweep various trait-specific loss-weight configurations. Instead of fully fine-tuning separate models for different weight combinations, the method trains only the adapter layers (LoRA) for each configuration. The configuration yielding the best performance on the development set is selected. This allows the model to capture fine-grained features and adapt to specific trait requirements efficiently.
Score Alignment (SA):
- This is a training-free, post-processing technique designed to correct distributional bias between predicted and true scores. Fine-tuned models often fail to predict extreme scores (e.g., predicting 0.95 instead of 1.0).
- SA applies a linear transformation to the test set predictions based on the discrepancy between true and predicted scores observed in the held-out development set. Specifically, it calculates the mean differences for the top and bottom $p\%$ of scores in the development set and uses these to adjust the minimum and maximum bounds of the test predictions, ensuring the predicted distribution aligns more closely with the true distribution.
Uncertainty-Aware Self-Training (UST):
- Designed to leverage unlabeled data (which is easier to collect than labeled data), UST estimates the uncertainty of predictions on unrated essays.
- Uncertainty is measured as the standard deviation of predictions obtained via Monte Carlo dropout (repeated dropout passes).
- Samples are partitioned into bins based on predicted score magnitude. From each bin, a fixed number of samples with the lowest uncertainty are selected to serve as pseudo-labeled data. This filtering mitigates the propagation of label noise. The augmented dataset (original labeled + selected pseudo-labeled) is then used to retrain the model.

Key Contributions

Novel Integration: The paper integrates Two-Stage fine-tuning, Score Alignment, and UST into a modular framework. This allows techniques to be combined or applied independently to address complementary bottlenecks: model adaptation, output calibration, and data augmentation.
Score Alignment Innovation: The authors demonstrate that Score Alignment consistently improves performance in both limited-data and full-data settings. When combined with LoRA, the DualBERT model achieves new state-of-the-art (SOTA) results in the full-data setting, outperforming existing AES methods.
Low-Resource Effectiveness: The study provides extensive evidence that the integrated approach significantly boosts performance when labeled data is scarce, reaching 91.2% of the full-data performance using only 32 labeled samples per prompt.

Experimental Results
Experiments were conducted on the ASAP++ dataset, with generalizability tested on TOEFL11 and ELLIPSE.

Full-Data Setting (ASAP++): The integration of LoRA and Score Alignment with DualBERT achieved a new SOTA result, outperforming previous methods like ArTS+RMTS and SaMRL. The QWK improved by 0.011 (per-trait) and 0.008 (per-prompt) over the baseline DualBERT, with the full integration (LoRA+SA) reaching 0.714 (per-trait) and 0.732 (per-prompt). While Score Alignment alone performed on par with the best supervised baselines, the combination with LoRA yielded the best overall results.
Limited-Data Setting (32-data): In the 32-data setting, the integrated method (LoRA+SA+UST) improved the QWK by 4.6 points (per-trait) and 5.4 points (per-prompt) over the base DualBERT. This resulted in a final score of 0.639 (per-trait), which is 91.2% of the full-data performance.
Comparison with LLMs: The proposed method outperformed few-shot LLM baselines (e.g., Llama-3.1-8B-Instruct) on ASAP++ in low-resource settings while requiring significantly fewer computational resources and parameters.
Generalizability: On TOEFL11 and ELLIPSE, Score Alignment and UST consistently improved performance over the baseline, though LoRA showed marginal or negative effects in multi-prompt settings due to overfitting on homogeneous trait sets.

Significance and Claims
The paper claims that its approach effectively addresses the data scarcity challenge in AES without relying on massive LLMs. The significance lies in the modularity of the proposed techniques, which allows for incremental performance improvements. The authors emphasize that Score Alignment is particularly robust, providing consistent gains regardless of data volume. The integrated method demonstrates that high-quality AES is achievable even with severe label scarcity (e.g., 32 samples), bridging the gap between limited-data performance and full-data capabilities. The study concludes that these techniques offer a practical, efficient pathway for deploying robust AES systems in real-world educational scenarios where labeled data is a bottleneck.

More like this