The dataset consists of 800 premise-hypothesis pairs that were handwritten and annotated by two people, excluding the author, for the task of natural language inference in Slovak.

The goal of creating this dataset is to provide a set of labeled and annotated human data that can be used to test abilities of fine-tuned language models. To our knowledge, other NLI datasets in Slovak are translated.

The dataset contains 204 entailment pairs, 215 neutral pairs, 178 contradiction pairs, 200 other pairs, and three undecided pairs. There are 629 pairs with 'Yes' in the Gold column and 171 pairs with 'No'.

Dataset Structure

The dataset consists of five columns: "Source," "Premise," "Hypothesis," "Label," and "Gold."

Source: Slovensko, Pixabay, or SlovakSum.
Premise: A string representing the truth sentence used to compare the hypothesis.
Hypothesis: A string that may or may not be true based on the premise.
Label: Entailment (hypothesis is true), Neutral (undecidable), Contradiction (hypothesis is not true), Other (premise and hypothesis are speaking about different things), Undecided (cannot reach conclusion)
Gold: Yes or No

Source Data

Premises are sourced from three distinct categories:

SlovakSum: the premises are drawn from the Slovak dataset SlovakSum (a summarisation of news articles), either from the title, a short paragraph, or a combination of both. The author of the dataset selected each premise and, where necessary, refined the grammar and syntax.
Slovakia: the author has written each premise by hand, based on info about Slovakia, like its traditions, holidays and geography.
Pixabay: the author has written each premise by hand, based on pictures from Pixabay, and has described what is in the picture. They come from different categories: forest, animals, winter, activities, buildings/places, people. This category was inspired by the SNLI dataset, where the premises come from the Flickr30k corpus.

Every sentence in the "Hypothesis" column is hand written.

Annotations

Annotation process

As the author has written hypotheses based on premises so labels are distributed evenly, the author counts as first annotator. Then we have annother 2 annotators who were given only pair premise-hypothesis and they needed to label the pair. Before labeling, they were given annotation manual with explanation and examples.

Therefore, we have three labels for each pair. If all three annotators agreed on a label, it was declared golden (Yes in the Gold column, otherwise, No). If only two annotators agreed, the label was decided based on their opinion. If each annotator labelled it differently, the label was given as 'Undecided'.

Who are the annotators?

Annotators are trained volunteers. They are native Slovak speakers.

Personal and Sensitive Information

This dataset does not contain personal or sensitive information.

Downloads last month: 12