thejosango/nuha-dataset

Jordanian NUHA Dataset

Dataset Summary

The Jordanian NUHA dataset is an Arabic hate speech corpus collected from Jordanian social media platforms (Facebook and X/Twitter) and annotated for online gender-based violence (OGBV). It was developed as a pilot proof-of-concept for the NUHA project — an initiative by the Jordan Open Source Association (JOSA), in collaboration with the TAMAM coalition, to detect and analyse technology-facilitated gender-based violence in Arabic-speaking online spaces.

The dataset contains 134,293 comments and is provided in three configurations suited to different research needs:

Configuration	Description	Labels
raw	Full annotations with all three original levels	Level 1 label, pipe-separated Level 2 classifications, Level 3 basis
methodology	Three-class coarsened scheme	Not Online Violence · Offensive Language · Gender Based Violence
binary	Binary classification	non-hate-speech · hate-speech

As a pilot study, this dataset reflects the exploratory nature of an initial annotation effort. Annotation quality is variable — the same comments were sometimes annotated by multiple annotators with differing results, and the annotation guidelines evolved during the project. Users should interpret results in this context.

---

Dataset Details

Languages

Arabic (ar) — specifically Jordanian dialect social media text. Comments are in colloquial Jordanian Arabic, though some comments from adjacent dialects may be present given the nature of the source platforms.

Source Data

Comments were collected from public posts by 70 social media accounts — female Jordanian public figures and women's rights campaigns — on Facebook and X (Twitter). Monitoring covered September 2022 to December 2023, during which approximately 201,000 comments were collected across 555 posts. Facebook is the dominant source; X (Twitter) accounts for a small minority of comments (~0.4%). Annotation was performed using LabelStudio across nine annotation rounds (r1–r7), producing the nine raw export files that were consolidated into this dataset.

Annotation Process

Five women's rights experts from the TAMAM coalition served as annotators. Annotators completed a training workshop covering OGBV concepts, the multi-label classification scheme, and annotation guidelines before beginning work. Data validation was performed in two steps: a random manual check of approximately 10% of the first 40,000 annotated comments, followed by model-based validation in which a premature classifier was used to identify high-confidence disagreements with manual labels, triggering reannotation of approximately 10% of the overall dataset.

Annotations were structured in three levels:

Level 1 — Primary Classification (required): Each comment was assigned one of six labels:

• Hatespeech
• Neutral Opinion
• Positive Comment
• Disagreement
• Vague and not Related
• Not Applicable (excluded from this dataset)

Level 2 — Hate Speech Classification (only when Level 1 = Hatespeech): One or more of twelve labels:

• Misogynist or Sexist · Stereotype · Insult and Bullying · Negative Comments on Physical Appearance
• Accusation · Dehumanizing and Demonizing
• Sexual Harassment · Rape Threat · Death Threat · Threat of Harm · Information Threat
• Irony and Sarcasm (referred to as Degrading Puns in the NUHA methodology)

Level 3 — Hate Speech Basis (only when Level 1 = Hatespeech): One of six bases:

• Gender Identity or Expression · Socioeconomic class · Religion · National Origin · Disability · Race

Level 3 is frequently absent even for Hatespeech rows: 33,640 of 54,973 hate speech comments (61.2%) have no Level 3 annotation. A minor spelling variant (Gender Identity/expression) appearing in 144 annotations has been normalised to Gender Identity or Expression.

As a known limitation of the annotation process, Level 2 and Level 3 labels sometimes appear in the raw data even when Level 1 is not Hatespeech. These are treated as annotation noise and discarded in all three dataset configurations.

---

Dataset Structure

Data Instances

raw example:

{
  "comment_id": 575584684323946,
  "comment_date": "2022-10-09",
  "comment_time": "21:34:44",
  "platform": "Facebook",
  "comment": "يعطيكم العافية",
  "level1_label": "Positive Comment",
  "level2_labels": null,
  "level3_label": null
}

methodology example:

{
  "comment_id": 352303,
  "comment_date": "2022-04-28",
  "comment_time": "15:47:47",
  "platform": "Facebook",
  "comment": "خرفت",
  "label": "Gender Based Violence",
  "subcategory": "Negative Stereotypes"
}

binary example:

{
  "comment_id": 352303,
  "comment_date": "2022-04-28",
  "comment_time": "15:47:47",
  "platform": "Facebook",
  "comment": "خرفت",
  "label": "hate-speech"
}

Data Fields

All configurations share these base fields:

Field	Type	Description
comment_id	int64	Unique identifier from the source platform
comment_date	string	Date of comment (YYYY-MM-DD)
comment_time	string	Time of comment (HH:MM:SS, Asia/Amman timezone)
platform	string	Source social media platform
comment	string	Cleaned Arabic comment text

Additional fields per configuration:

raw:

Field	Type	Description
level1_label	string	Primary annotation label
level2_labels	string	Pipe-separated Level 2 labels (null if Level 1 ≠ Hatespeech)
level3_label	string	Level 3 basis label (null if Level 1 ≠ Hatespeech)

methodology:

Field	Type	Description
label	ClassLabel	Not Online Violence · Offensive Language · Gender Based Violence
subcategory	string	OGBV subcategory (Negative Stereotypes · Dehumanization · Expression of Violence · Unknown), empty otherwise

binary:

Field	Type	Description
label	ClassLabel	non-hate-speech · hate-speech

Label Mapping (methodology → binary)

Methodology Label	Binary Label
Not Online Violence	non-hate-speech
Offensive Language	hate-speech
Gender Based Violence	hate-speech

Data Splits

The dataset is provided with predefined stratified splits across all three configurations:

Split	Examples	Share
train	107,434	80%
validation	13,429	10%
test	13,430	10%
Total	134,293

Splits are stratified by label to preserve class proportions across all three configurations. The random seed is fixed (42) for reproducibility.

Unsplit convenience files (nuha_binary.csv, nuha_methodology.csv, nuha_raw.csv) containing the full dataset are also included in data/.

Dataset Statistics

Label	raw	methodology	binary
Hatespeech / Gender Based Violence / hate-speech	54,973 (40.9%)	52,471 (39.1%)	54,973 (40.9%)
Not Online Violence / non-hate-speech	—	79,320 (59.1%)	79,320 (59.1%)
Offensive Language	—	2,502 (1.9%)	—
Other Level 1 labels (raw only)	79,320	—	—

---

Dataset Creation

Text Cleaning

Comment text was cleaned using the following procedure before inclusion in this dataset:

1. Comments with more than 50 words were discarded.
2. All characters except Arabic Unicode (\u0600–\u06FF), spaces, and emojis were stripped.
3. Comments consisting entirely of emojis after stripping were discarded.

Note that the NUHA models were trained on an earlier internal version of the dataset that had normalization applied directly to the text (diacritic removal, punctuation removal, URL/mention replacement) rather than this cleaned version.

Filtering

Stage	Rows removed
Empty after cleaning	22,169
Labelled Not Applicable or missing Level 1 annotation	27,643
Level 1 inter-annotator conflict (irresolvable)	11,543
Merged inter-annotator Level 2 conflicts	2,017 (merged, not dropped)
Final dataset	134,293

Inter-Annotator Disagreement

A subset of comments were annotated by multiple annotators. Duplicates were resolved as follows: where annotators disagreed on Level 1, all copies were discarded; where Level 1 agreed but Level 2 differed, labels were merged by union; where only Level 3 differed, the first annotation was kept; exact duplicates were deduplicated to one copy.

Personal and Sensitive Information

This dataset contains social media comments that may include offensive, misogynistic, or violent language by nature of the annotation task. Comment IDs are preserved from the source platforms but no usernames, profile information, or post content beyond the comment text is included.

---

Considerations for Using the Data

Social Impact

This dataset is intended to support research into automated detection of online gender-based violence in Jordanian Arabic, with the goal of enabling more effective content moderation and safety tooling. Misuse for targeting, surveillance, or profiling of individuals would be contrary to its intended purpose.

Bias and Limitations

• Pilot annotation quality: As an initial proof-of-concept study, the annotation process was not fully standardised across all rounds. Inter-annotator agreement was not formally measured, and the annotation guidelines evolved during the project.
• Dialect coverage: While the primary focus is Jordanian Arabic, the social media source may include adjacent dialects.
• Class imbalance: The dataset is imbalanced, with a majority of non-hate-speech examples. The NUHA model training used data augmentation to partially compensate, with weighted loss additionally applied in the binary classifier.
• Subjectivity: Hate speech annotation is inherently subjective. Labels reflect the judgements of the annotators involved in this pilot and may not generalise to other annotation schemes.

---

Citation

@misc{nuha-jo-dataset,
  title  = {Jordanian NUHA Dataset},
  author = {Jordan Open Source Association (JOSA)},
  year   = {2023},
  url    = {https://huggingface.co/datasets/thejosango/nuha-dataset}
}