Skip to main content
. 2025 Jul 24;11:e3017. doi: 10.7717/peerj-cs.3017

Table 6. Description of the dataset for low-resource languages.

Article Language Source Number of categories Categories of data Size Way of labeling Balanced or no Use a technique to balance the data or not After balanced Splitting the data
Cruz & Cheng (2020) Filipino X Platform 2 -hate
-non-hate		 10,000 N/A Relatively N/A – N/A
Rahman et al. (2020) Bengali Open source
  • 1-BARD,
  • 2-OSBC
  • 3-ProthomAlo

  • 1-5

  • 2-11

  • 3-6

 Include crime

  • 1-50,560

  • 2-78,796

  • 3-128,761

 Annotated No No – Train: 80%
Test: 20%
Djandji et al. (2020) Arabic X Platform 2 Offensive
Not offensive		 10,000 Annotated No Yes, over/undersampling – Train: 70%
Test: 20%
Development: 10%
Kittask, Milintsevich & Sirts (2020) Estonian Postimees Estonian newspaper 4 Include crime Topic:
Negative Ambiguous Positive Neutral		 4,088 Annotated with sentiment &
with rubric labels.		 No No – Train: 70%
Test: 20%
Development: 10%
Mubarak et al. (2020) Arabic X Platform 4 - Offensive
- Vulgar
- Hate speech
- Clean		 10,000 Experienced annotator No No – N/A
Aluru et al. (2021) 1. Indonesian
2. Polish
3. Arabic		 Public Dataset
From X platform		 2 -Hate speech
-Normal		 13,169 & 713
9,788
4,120 & 1,670		 Annotated Yes
No
No & YES		 No – Train: 70%
Test: 20%
Validation: 10%
Arcila-Calderón et al. (2022) 1. Greek
2. Italian		 PHARM datasets 2 -Racist/xenophobic hate tweets
-Non-racist/xenophobic hate tweets		 1-10,399
2-10,752		 Manual Yes – – N/A
Modha, Majumder & Mandl (2022) Hindi TRAC 3 Non-aggressive (NAG), Overtly Aggressive (OAG), & Covertly Aggressive (CAG) 15,001 Annotated Yes – – N/A
Sharif & Hoque (2022) Bengali BAD
Facebook & YouTube

  • 1-2

  • 2-4

 1-AG & NoAG
2-ReAG & PoAG & VeAG & GeAG.		 14,443 Manual

  • 1-Yes

  • 2-No

  • 1-  -

  • 2-No

 – Train: 80%
Test: 10%
Validation: 10%
Kapil & Ekbal (2022) Hindi Eight datasets from social media

  • 1-2

  • 2-3

 Hate, normal
Hostile, non-hostile
CAG, NAG, OAG
Abusive, hate, natural		 Different sizes Annotated Some are balanced and some not No – Train: 80%
Test: 20%
Shapiro, Khalafallah & Torki (2022) Arabic X Platform 2 Offensive
Not offensive		 12,700 Annotated No No – Train: 70%
Test: 20%
Development: 10%
El-Alami, Alaoui & Nahnahi (2022) Arabic SemEval’2020 competition
-Arabic dataset		 2 1 = Offensive
0 = Not offensive		 7,800 Annotated No No – Train: 80%
Test: 20%
Akram, Shahzad & Bashir (2023) Urdu (Pakistan) X Platform 2 Hatful
Neutral		 21,759 Manual No No – Train: 90%
Test: 10%
Hossain et al. (2023) Bengali Kaggle & Bangla Newspaper

  • 1-2

  • 2-4

 1-Crime & Others
2-Murder, Drug, Rape & Others		 Approximately 5.3 million entries Annotated & Manual N/A N/A – N/A
Ullah et al. (2024) Urdu (Pakistan) X Platform 5 Cyber Terrorism
Hate Speech
Cyber Harassment
Normal
Offensive		 7,372 Three Manual Annotator No No – Train: 80%
Test: 20%
Mozafari et al. (2024) Persian X Platform 2 Offensive
Not offensive		 6,000 Manual No No – N/A
Adam, Zandam & Inuwa-Dutse (2024) Hausa (Africa) X & factbook Platforms 2 Offensive
Not offensive		 N/A Manual N/A N/A – N/A