|
|
--- |
|
|
license: apache-2.0 |
|
|
task_categories: |
|
|
- text-classification |
|
|
language: |
|
|
- en |
|
|
pretty_name: TNCC |
|
|
size_categories: |
|
|
- 1K<n<10K |
|
|
--- |
|
|
Given the scarcity of datasets for understanding natural language in visual scenes, we introduce a novel textual entailment dataset, named Textual Natural Contextual Classification (TNCC). |
|
|
This dataset is formulated on the foundation of Crisscrossed Captions (https://github.com/google-research-datasets/Crisscrossed-Captions), an image captioning dataset supplied with human-rated semantic similarity ratings on a continuous scale from 0 to 5. |
|
|
We tailor the dataset to suit a binary classification task. Specifically, sentence pairs with annotation scores exceeding 4 are categorized as positive (entailment), whereas pairs with scores less than 1 are marked as negative (non-entailment). |
|
|
The TNCC dataset is partitioned into training, validation, and testing sets, containing 3,600, 1,200, and 1,560 instances, respectively. |
|
|
|
|
|
If you use this dataset for academic research, please cite the NeurIPS 2023 paper titled 'Back-Modality: Leveraging Modal Transformation for Data Augmentation'. |
|
|
|
|
|
|
