Dataset Card for "SciEntsBank"

SciEntsBank is one of the two distinct subsets within the Student Response Analysis (SRA) corpus, the other subset being the Beetle dataset. Derived from student answers gathered by Nielsen et al. [1], this dataset comprises nearly 11K responses to 197 assessment questions spanning 15 diverse science domains. The dataset features three labeling schemes: (a) 5-way, (b) 3-way, and (c) 2-way. The dataset includes a training set and three distinct test sets: (a) Unseen Answers (test_ua), (b) Unseen Questions (test_uq), and (c) Unseen Domains (test_ud).

• Authors: Myroslava Dzikovska, Rodney Nielsen, Chris Brew, Claudia Leacock, Danilo Giampiccolo, Luisa Bentivogli, Peter Clark, Ido Dagan, Hoa Trang Dang
• Paper: SemEval-2013 Task 7: The Joint Student Response Analysis and 8th Recognizing Textual Entailment Challenge

Loading Dataset

from datasets import load_dataset
dataset = load_dataset('nkazi/SciEntsBank')

Labeling Schemes

The authors released the dataset with annotations using five labels (i.e., 5-way labeling scheme) for Automated Short-Answer Grading (ASAG). Additionally, the authors have introduced two alternative labeling schemes, namely the 3-way and 2-way schemes, both derived from the 5-way labeling scheme designed for Recognizing Textual Entailment (RTE). In the 3-way labeling scheme, the categories "partially correct but incomplete", "irrelevant", and "non-domain" are consolidated into a unified category labeled as "incorrect". On the other hand, the 2-way labeling scheme simplifies the classification into a binary system where all labels except "correct" are merged under the "incorrect" category.

The label column in this dataset presents the 5-way labels. For 3-way and 2-way labels, use the code provided below to derive it from the 5-way labels. After converting the labels, please verify the label distribution. A code to print the label distribution is also given below.

5-way to 3-way

from datasets import ClassLabel

dataset = dataset.align_labels_with_mapping({'correct': 0, 'contradictory': 1, 'partially_correct_incomplete': 2, 'irrelevant': 2, 'non_domain': 2}, 'label')
dataset = dataset.cast_column('label', ClassLabel(names=['correct', 'contradictory', 'incorrect']))

Using align_labels_with_mapping(), we are mapping "partially correct but incomplete", "irrelevant", and "non-domain" to the same id. Subsequently, we are using cast_column() to redefine the class labels (i.e., the label feature) where the id 2 corresponds to the "incorrect" label.

5-way to 2-way

from datasets import ClassLabel

dataset = dataset.align_labels_with_mapping({'correct': 0, 'contradictory': 1, 'partially_correct_incomplete': 1, 'irrelevant': 1, 'non_domain': 1}, 'label')
dataset = dataset.cast_column('label', ClassLabel(names=['correct', 'incorrect']))

In the above code, the label "correct" is mapped to 0 to maintain consistency with both the 5-way and 3-way labeling schemes. If the preference is to represent "correct" with id 1 and "incorrect" with id 0, either adjust the label map accordingly or run the following to switch the ids:

dataset = dataset.align_labels_with_mapping({'incorrect': 0, 'correct': 1}, 'label')

Saving and loading 3-way and 2-way datasets

Use the following code to store the dataset with the 3-way (or 2-way) labeling scheme locally to eliminate the need to convert labels each time the dataset is loaded:

dataset.save_to_disk('SciEntsBank_3way')

Here, SciEntsBank_3way depicts the path/directory where the dataset will be stored. Use the following code to load the dataset from the same local directory/path:

from datasets import DatasetDict
dataset = DatasetDict.load_from_disk('SciEntsBank_3way')

Printing Label Distribution

Use the following code to print the label distribution:

def print_label_dist(dataset):
    for split_name in dataset:
        print(split_name, ':')
        num_examples = 0
        for label in dataset[split_name].features['label'].names:
            count = dataset[split_name]['label'].count(dataset[split_name].features['label'].str2int(label))
            print(' ', label, ':', count)
            num_examples += count
        print('  total :', num_examples)

print_label_dist(dataset)

Label Distribution

5-way

Label	Train	Test UA	Test UQ	Test UD
Correct	2,008	233	301	1,917
Contradictory	499	58	64	417
Partially correct but incomplete	1,324	113	175	986
Irrelevant	1,115	133	193	1,222
Non-domain	23	3	-	20
Total	4,969	540	733	4,562

3-way

Label	Train	Test UA	Test UQ	Test UD
Correct	2,008	233	301	1,917
Contradictory	499	58	64	417
Incorrect	2,462	249	368	2,228
Total	4,969	540	733	4,562

2-way

Label	Train	Test UA	Test UQ	Test UD
Correct	2,008	233	301	1,917
Incorrect	2,961	307	432	2,645
Total	4,969	540	733	4,562

Citation

@inproceedings{dzikovska2013semeval,
  title = {{S}em{E}val-2013 Task 7: The Joint Student Response Analysis and 8th Recognizing Textual Entailment Challenge},
  author = {Dzikovska, Myroslava  and Nielsen, Rodney  and Brew, Chris  and Leacock, Claudia  and Giampiccolo, Danilo  and Bentivogli, Luisa  and Clark, Peter  and Dagan, Ido  and Dang, Hoa Trang},
  year = 2013,
  month = jun,
  booktitle = {Second Joint Conference on Lexical and Computational Semantics ({SEM}), Volume 2: Proceedings of the Seventh International Workshop on Semantic Evaluation ({S}em{E}val 2013)},
  editor = {Manandhar, Suresh  and Yuret, Deniz}
  publisher = {Association for Computational Linguistics},
  address = {Atlanta, Georgia, USA},
  pages = {263--274},
  url = {https://aclanthology.org/S13-2045},
}

References

1. Rodney D. Nielsen, Wayne Ward, James H. Martin, and Martha Palmer. 2008. Annotating students' understanding of science concepts. In Proceedings of the Sixth International Language Resources and Evaluation Conference, Marrakech, Morocco.