nkazi/MohlerASAG

Dataset Card for "Mohler ASAG"

The Mohler ASAG dataset is recognized as one of the first publicly available and widely used benchmark datasets for Automatic Short Answer Grading (ASAG). It was first introduced by Michael Mohler and Rada Mihalcea in 2009. An extended version of the dataset with additional questions and corresponding student answers was released in 2011. This repository presents the 2011 dataset along with a code snippet to extract the 2009 subset.

The dataset was collected from an introductory data structures course at the University of North Texas. It covers 87 assessment questions in total, including 81 open-ended and 6 closed-ended selection or ordering questions. These questions are distributed across 10 assignments and 2 examinations. Altogether, the dataset contains 2,442 student responses, with 2,273 corresponding to open-ended questions and 169 to closed-ended questions.

• Authors: Michael Mohler, Razvan Bunescu, and Rada Mihalcea.
• Paper: Learning to Grade Short Answer Questions using Semantic Similarity Measures and Dependency Graph Alignments

A curated version of the dataset is available on Hugging Face at nkazi/MohlerASAG-Curated , created to improve its quality and usability for NLP research, particularly for LLM-based approaches.

Known Errata

1. The 2009 paper reports 30 student answers per question for each assignment. In reality, assignment 1 contains 29 answers per question, assignment 2 contains 30 answers per question, and assignment 3 contains 31 answers per question.
2. The 2011 paper states that the dataset contains student answers for 80 questions. According to the README file included with the data, it actually includes answers for 81 open-ended questions.

Dataset Conversion Notebook

The Python notebook I developed to convert the Mohler ASAG dataset from its source files into a Hugging Face Dataset is available on my GitHub profile. It exhibits the process of parsing questions, instructor answers, student answers, scores, and annotations from their respective source files for each stage, correcting mojibakes in the raw data, structuring and organizing the information, dividing and transforming the data into subsets and splits, and exporting the final dataset in Parquet format for the Hugging Face repository. This demonstration ensures transparency, reproducibility, and traceability of the conversion process.

GitHub Link: https://github.com/nazmulkazi/ML-DL-NLP/blob/main/HF%20Dataset%20-%20Mohler%20ASAG.ipynb

Dataset Structure and Details

The dataset underwent several processing stages, each represented as a separate subset. The raw subset contains the original and unaltered student answers exactly as written. In the cleaned subset, the authors preprocessed the data by cleaning the text and tokenizing it into sentences using the LingPipe toolkit, with sentence boundaries marked by <STOP> tags. The parsed subset includes outputs from the Stanford Dependency Parser with additional postprocessing performed by the authors. The annotations subset contains manually annotated data. However, only 32 student answers were randomly selected for annotation.

The authors ignored responses to the closed-ended questions in all of their work. Therefore, the raw, cleaned, and parsed subsets are divided into open-ended and closed-ended splits.

Each sample in the raw, cleaned, and parsed subsets includes a unique identifier, the question, the instructor's answer, the student's answer, scores from two graders, and the average score. Samples in the annotations subset contain a unique identifier and the corresponding annotations. The unique identifiers are consistent across all subsets and follow the format EXX.QXX.AXX, where each component corresponds to its exercise (i.e., assignment), question, and answer, respectively, and XX are zero-padded numbers. For consistency, reproducibility, and traceability, the identifiers are constructed following the same indexing scheme used by the authors, with 1-based numbering for exercises and questions and 0-based numbering for student answers.

Exercises E01 through E10 were graded on a 0-5 scale, while E11 and E12 were graded on a 0-10 scale. The scores for E11 and E12 were converted to a 0-5 scale before computing the average by the authors, so all values in the score_avg column are in the 0-5 range. Grader 1 was the course teaching assistant, and Grader 2 was Michael Mohler.

For further details, please refer to the README (a formatted and styled version of the README provided by the authors) and the associated publications.

Student Answer Distribution

Distribution of student answers in the raw, cleaned, and parsed subsets:

	Q01	Q02	Q03	Q04	Q05	Q06	Q07	Q08	Q09	Q10	Total
E01	29	29	29	29	29	29	29	-	-	-	203
E02	30	30	30	30	30	30	30	-	-	-	210
E03	31	31	31	31	31	31	31	-	-	-	217
E04	30	30	30	30	30	30	30	-	-	-	210
E05	28	28	28	28	-	-	-	-	-	-	112
E06	26	26	26	26	26	26	26	-	-	-	182
E07	26	26	26	26	26	26	26	-	-	-	182
E08	27	27	27	27	27	27	27	-	-	-	189
E09	27	27	27	27	27	27	27	-	-	-	189
E10	24	24	24	24	24	24	24	-	-	-	168
E11	30	30	30	30	30	30	30	30	30	30	300
E12	28	28	28	28	28	28	28	28	28	28	280

Distribution of student answers in the annotations subset/split:

	Q01	Q02	Q03	Q04	Q05	Q06	Q07	Total
E01	3	3	3	3	2	1	1	16
E02	1	1	1	2	1	1	1	8
E03	1	1	1	1	1	1	2	8

Code Snippets

Extracting 2009 Dataset

Exercises 1-3 are inherited from the 2009 dataset. The following code extracts the raw samples of the 2009 dataset from the raw subset:

from datasets import load_dataset

ds = load_dataset('nkazi/MohlerASAG', name='raw', split='open_ended')
ds_2009 = ds.filter(lambda row: row['id'].split('.')[0] in ['E01', 'E02', 'E03'])

Concatenating Splits

The following code creates a new dataset with rows from both open-ended and close-ended splits from the raw subset:

from datasets import load_dataset
from datasets import concatenate_datasets

ds = load_dataset('nkazi/MohlerASAG', name='raw')
ds_all = concatenate_datasets([ds['open_ended'], ds['close_ended']]).sort('id')

Joining Open-Ended Raw Data with Annotations

The following code joins the annotations with their corresponding samples from the raw subset.

from datasets import load_dataset

# Load the annotations split and create a mapping
# from IDs to their annotations.
ds_ann = load_dataset('nkazi/MohlerASAG', name='annotations', split='annotations')
ann_map = {row['id']: row['annotations'] for row in ds_ann}

# Load the raw open-ended subset and keep only rows
# with IDs present in the annotations set.
ds_raw = load_dataset('nkazi/MohlerASAG', name='raw', split='open_ended') \
  .filter(lambda row: row['id'] in ann_map)

# Collect annotations in the same order as the IDs in
# the filtered raw dataset.
ann_list = [ann_map.get(row_id, None) for row_id in ds_raw['id']]

# Add an annotations column to the filtered raw dataset,
# using the annotations list and feature specification
# from the annotations subset.
ds_joined = ds_raw.add_column(
  name = 'annotations',
  column = ann_list,
  feature = ds_ann.features['annotations']
)

Citation

In addition to citing Mohler et al. (2011), we kindly request that a footnote be included referencing the Hugging Face page of this dataset (https://huggingface.co/datasets/nkazi/MohlerASAG) in order to inform the community of this readily usable version.

@inproceedings{mohler2011learning,
  title = {Learning to Grade Short Answer Questions using Semantic
    Similarity Measures and Dependency Graph Alignments},
  author = {Mohler, Michael and Bunescu, Razvan and Mihalcea, Rada},
  year = 2011,
  month = jun,
  booktitle = {Proceedings of the 49th Annual Meeting of the Association
    for Computational Linguistics: Human Language Technologies},
  pages = {752--762},
  editor = {Lin, Dekang and Matsumoto, Yuji and Mihalcea, Rada},
  publisher = {Association for Computational Linguistics},
  address = {Portland, Oregon, USA},
  url = {https://aclanthology.org/P11-1076},
}