There was a mistake in the article as published. Tables 1–7 and Figures 1–8 were published as supplementary material when they should have been added to the main article. The corrected figures and tables appear below.
Distribution of answers by question type.
Question type
Question type (In Arabic)
Total questions
Total answers
Define the scientific term
6
291
Explain
21
830
What are the consequences of
6
282
Justify or give reasons for
10
465
What is the difference between
5
217
Total
5 types
48
2,085
Detailed distribution of randomly sampled responses across selected questions.
Q–No.
Question type
Total answers
Training answers
Test answers
1
Define the scientific term
46
36
10
26
Explain
47
37
10
28
What are the consequences of
48
38
10
35
Justify or give reasons for
51
40
11
45
What is the difference between
36
28
8
Performance evaluation of AraBERTv2 with MLP model using different feature sets: training vs. testing results.
Model
Stage
No. of feature
MAE
RMSE
Pearson correlation
Spearman's correlation
Epoch 1–5
AraBERTv2 with MLP
Training
2-feature
1.14
1.51
0.847
0.85
898 → 533 → 347 → 250 → 156
3-feature
1.2
1.58
0.818
0.816
1,026 → 614 → 263 → 185
4-feature
0.18
0.2
0.999
0.998
713 → 34 → 13 → 9 → 7
Testing
2-feature
1.31
1.76
0.803
0.808
3-feature
1.48
1.9
0.744
0.746
4-feature
1.77
2.22
0.691
0.689
Performance evaluation of AraBERTv2 with CNN model using different feature sets: training vs. testing results.
Model
Stage
No. of features
MAE
RMSE
Pearson correlation
Spearman's correlation
Epoch 1–5
AraBERTv2 with CNN
Training
2-feature
1.22
1.59
0.849
0.843
1,092 → 610 → 427 → 306 → 227
3-feature
1.17
1.53
0.833
0.832
1,057 → 567 → 379 → 280 → 205
4-feature
0.24
0.27
0.999
0.998
773 → 28 → 12 → 8 → 6
Testing
2-feature
1.45
1.93
0.784
0.788
3-feature
1.6
2.02
0.746
0.75
4-feature
2.63
3.07
0.607
0.613
Performance evaluation of AraBERTv2 with LSTM model using different feature sets: training vs. testing results.
Model
Stage
No. of features
MAE
RMSE
Pearson correlation
Spearman's correlation
Epoch 1–5
AraBERTv2 with LSTM
Training
2-feature
1.26
1.62
0.821
0.825
1,147 → 718 → 524 → 356 → 262
3-feature
1.27
1.66
0.811
0.818
1,141 → 675 → 456 → 349 → 267
4-feature
0.14
0.19
0.998
0.998
728 → 62 → 31 → 22 → 19
Testing
2-feature
1.48
1.86
0.757
0.759
3-feature
1.6
2.03
0.757
0.77
4-feature
3.62
4.19
0.388
0.419
Performance comparison of AraBERTv2 fine-tuned models with MLP, CNN, and LSTM architectures using different feature sets.
Fine-tuned models
MAE
RMSE
Pearson correlation
Spearman's correlation
2-features-AraBERTv2 with MLP
1.31
1.76
0.803
0.808
2-features-AraBERTv2 with CNN
1.45
1.93
0.784
0.788
2-features-AraBERTv2 with LSTM
1.48
1.86
0.757
0.759
3-features-AraBERTv2 with MLP
1.48
1.9
0.744
0.746
3-features-AraBERTv2 with CNN
1.6
2.02
0.746
0.75
3-features-AraBERTv2 with LSTM
1.6
2.03
0.757
0.77
4-features-AraBERTv2 with MLP
1.77
2.22
0.691
0.689
4-features-AraBERTv2 with CNN
2.63
3.07
0.607
0.613
4-features-AraBERTv2 with LSTM
3.62
4.19
0.388
0.419
The bold values represent the optimal results obtained from our experimental analysis.
Comparative performance evaluation of Arabic Automated Short Answer Grading (ASAG) systems.
Criterion/study
Methodology
Dataset
Best RMSE
Best Pearson/Spearman
Key strength
Primary limitation
Our study (AraBERTv2)
-Fine-tuned AraBERTv2 with MLP/CNN/LSTM -Tested 2/3/4 feature configurations
AS-ARSG (2,133 answers)
1.31
-Pearson: 0.803 -Spearman: 0.808
Optimal balance between generalizability and accuracy with limited data
Performance degradation in LSTM with added features
(4)
Latent Semantic Analysis (LSA) with local/hybrid weighting
AR-ASAG (2,133 answers)
N/A
N/A
Effective semantic weighting
Limited capacity for capturing complex contextual relationships
(19)
-BERT vs. Word2Vec/AWN comparison -Intensive text preprocessing
-AR-ASAG (2,133) -Jordanian History (550)
1.00308
Pearson: 0.841902
Demonstrated BERT's superiority over traditional approaches
Heavy dependency on text normalization and stemming
General workflow of the proposed automated Arabic short-answer grading model using AraBERTv2.
Flowchart illustrating the process of training and testing with the AR-ASAG dataset. The sequence includes dataset loading, preprocessing, and splitting into 80% training and 20% testing. The training subset undergoes feature selection and AraBERT training, leading to finetuned AraBERT models. These models are evaluated and compared, followed by visualization to determine the best AraBERT model. Arrows indicate the workflow and connections among the steps.
The AraBERT_MLP training methodology.
Bar charts compare AraBERTv2 with LSTM across training and testing phases. The top charts show MAE and RMSE, and Pearson and Spearman correlations for different features in training. The bottom charts depict the same metrics for testing, highlighting variations in error values and correlation coefficients across two, three, and four features.
The AraBERT_CNN training methodology.
Bar charts showing AraBERTv2 with CNN performance during training and testing. Training error bars compare MAE and RMSE for 2, 3, and 4 features, with RMSE generally higher. Correlation values for Pearson and Spearman increase with more features. Testing error values increase with more features, while correlation values decrease slightly as features increase.
The AraBERT_LSTM training methodology.
Flowchart depicting the AraBERT with MLP training stage. Feature selection leads to three models: 2-features (red), 3-features (green), and 4-features (purple). Each model connects to AraBERT with CNN, then fine-tuning AraBERT stages, ending with evaluation and comparison.
Performance evaluation of AraBERTv2 with MLP model using different feature sets: training vs. testing results.
Bar charts comparing AraBERTv2 with MLP performance in training and testing phases. In training, 4-feature shows minimal MAE and RMSE, with high Pearson and Spearman correlations. In testing, 2-feature has lower error values than 3-feature and 4-feature, though 4-feature performs slightly better in correlation values.
Performance evaluation of AraBERTv2 with CNN model using different feature sets: training vs. testing results.
Diagram illustrating a machine learning process, titled “AraBERT with MLP Training Stage.” It starts with “Feature Selection,” leads to three models: “2-features model,” “3-features model,” and “4-features model.” Each model goes to “AraBERT with CNN,” followed by “Fine tuning AraBERT,” and ends with “Evaluation and comparison.” Arrows indicate the flow direction.
Performance evaluation of AraBERTv2 with LSTM model using different feature sets: training vs. testing results.
Scatter plot titled “Model Performance: Error vs Spearman Correlation” showing different models' performance using colored markers: AraBERTv2 with MLP, CNN, and LSTM. The x-axis represents MAE (mean absolute error), where lower is better, and the y-axis represents Spearman’s rank correlation, where higher is better. The plot uses different shapes to indicate feature numbers. Most points cluster between 1.5 to 2.0 MAE and 0.65 to 0.80 correlation, with one outlier beyond 3.5 MAE and below 0.45 correlation.
Fine-tuned models performance: MAE vs. spearman correlation.
Diagram of a machine learning workflow featuring AraBERT with MLP. It starts with feature selection, separating into three models: a 2-feature model in red, a 3-feature model in green, and a 4-feature model in blue. These feed into the AraBERT with MLP stage, which then advances to fine-tuning AraBERT in individual boxes. An evaluation and comparison stage follows, indicated by arrows.
All in-text Supplementary Table and Supplementary Figure in-text citations have been changed to Table and Figure in-text citations.
The original version of this article has been updated.
Generative AI statement
Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.
Approved by: Frontiers Editorial Office, Frontiers Media SA, Switzerland