README.md

metadata

license: cc-by-sa-4.0
task_categories:
  - text-classification
language:
  - sk
size_categories:
  - 10K<n<100K

Slovak Hate Speech and Offensive Language Database

The dataset contains posts from a social network with human annotations.

Annotations

The posts are marked 1 if the post contain hateful or offensive language, 0 otherwise.

Dataset Creation

The source data were scraped from a social network from a selection of public pages for sport, politics or general discussion. The gathered data were cleaned from span with a text clustering. The posts were annotated by a group of students of the Technical University of Košice, Slovakia.

We removed annotations of users that had low level of agreement with others.

Data filtering

One item was annotated by multiple annotators, but some annotators are unreliable. We had to identify unreliable annotators.

1. We removed annotations from users that mostly (90%) clicked on the same option .
2. We calculated level of agreement for each annotator. Annotator gets a positive point for each annotation, if he annotated the same as other annotators and negative if he or she annotated differently. For each annotator we calculate ratio of positive and negative points.
3. We remove annotations from annotators with low ratio of agreement (less than 70%).
4. We calculate votes for positive, neutral and negative class for each annotation from the remaining annotators. We remove annotations where neutral class has majority.

Bias

Annotations are dependent on the personal opinions of the annotators. Class for most of the items was determined by voting of trustworthy annotators, but some items had only one vote available.

Credits

• NLP@ KEMT Technical University of Košice, Slovakia
• Vladimír Ferko: annotation application and preliminary experiments
• Daniel Hládek: data filtering and export