BanHateMe: Understanding Hate in Bangla Memes through Detection, Categorization, and Target Profiling

Md Ayon Mia and Md Fahim

🧩 Dataset Overview

• The dataset focuses on hierarchical hate detection in Bangla memes with multi-level annotations across binary labels, hate categories, and targeted groups.
• It contains 3,819 annotated meme samples collected from Facebook (2,517) and Instagram (1,302) between April 2022 – May 2025.
• Binary classification: Hate (1,769) vs. Non-Hate (2,050).
• Five hate categories: Abusive (27.3%), Political (20.1%), Gender (19.0%), Personal Offence (16.8%), and Religious (16.7%).
• Four targeted groups: Community (36.9%), Individual (30.8%), Organization (21.4%), and Society (10.8%).
• Each hateful meme is annotated with one hate category and one target group through a two-stage hierarchical process with majority voting.
• Manual text extraction by native Bangla typists to ensure accuracy.
• High inter-annotator agreement with average Cohen's κ = 0.78 (binary: 0.78, categories: 0.68, targets: 0.69).

📑 Data Format

Column	Description
image	Meme image filename
Text	Text on Bangla meme
Label	Hate / Non-Hate
Hate Categories	Abusive / Political / Gender / Personal Offence / Religious
Targeted Groups	Community / Individual / Organization / Society

---

🧩 Methodology Overview

Our approach leverages a hierarchical multimodal framework combining pretrained language encoders and vision encoders to extract modality-specific representations from Bangla memes. We systematically evaluate three fusion strategies to combine visual and textual features: summation-based fusion (element-wise addition of [CLS] tokens), concatenation-based fusion (self-attention over concatenated sequences with mean pooling), and co-attention fusion (cross-modal attention between modalities). The fused representations are passed through a linear classification module for hierarchical prediction. To address the multi-level annotation structure, we propose a composite loss function

L_total = L_binary + α · L_hate_cat + β · L_target_grp 

where L_binary handles hate/non-hate classification, L_hate_cat manages five-category classification, and L_target_grp identifies four target groups, with optimal weighting at α=0.5, β=0.5 ensuring balanced predictions across all hierarchical levels.

📊 Experimental Results

🔹Hate Detection Performance

Fusion Method	Non-Hate				Hate				Overall
	P	R	F1	Acc	P	R	F1	Acc	F1	Acc
BanglaBERT + ViT
Sum based	69.58	81.27	74.97	81.27	72.98	58.76	65.10	58.76	70.40	70.85
Concatenation	69.94	79.81	74.55	79.81	71.96	60.17	65.54	60.17	70.38	70.72
Co-Attention	69.27	66.91	68.07	66.91	63.04	65.54	64.27	65.54	66.31	66.27
BanglaBERT + Swin
Sum based	72.25	82.96	77.24	82.97	76.11	62.99	68.93	62.99	73.39	73.73
Concatenation	72.31	85.15	78.21	85.16	78.29	62.15	69.29	62.15	74.08	74.51
Co-Attention	69.48	74.21	71.76	74.21	67.48	62.15	64.71	62.15	68.50	68.63
XLM-RoBERTa + ViT
Sum based	62.03	51.68	56.39	51.69	63.62	72.75	67.88	72.75	62.56	63.01
Concatenation	65.04	64.71	64.88	64.72	59.27	59.60	59.44	59.60	62.36	62.35
Co-Attention	65.71	61.06	63.30	61.07	58.22	62.99	60.52	62.99	62.01	61.96
XLM-RoBERTa + Swin
Sum based	69.12	82.22	75.11	82.24	73.55	57.33	64.44	57.34	70.18	70.72
Concatenation	72.75	76.63	74.64	76.64	71.08	66.66	68.80	66.67	71.94	72.03
Co-Attention	65.05	81.50	72.35	81.51	69.60	49.14	57.62	49.15	65.53	66.54

---

🔹Performance Across Hate Categories and Target Groups (F1 Scores)

Fusion Method	Hate Category						Target Group
	Ab	Po	Ge	Per	Re	Avg	Co	Ind	Org	So	Avg
BanglaBERT + ViT
Sum based	60.87	64.29	31.03	25.40	32.88	42.89	44.86	58.23	59.20	09.23	42.88
Concatenation	65.74	71.53	52.57	15.79	61.54	53.43	61.32	57.76	63.24	08.13	47.61
Co-Attention	67.23	64.75	40.30	24.24	33.85	46.07	66.67	62.01	67.63	48.78	61.27
BanglaBERT + Swin
Sum based	54.02	77.61	32.99	31.84	76.47	54.59	59.11	58.25	78.57	7.14	50.77
Concatenation	69.53	75.56	53.89	19.18	74.00	58.43	69.63	56.45	74.45	12.32	53.21
Co-Attention	66.67	71.76	56.65	25.87	66.67	57.52	71.27	65.71	71.11	38.64	61.68
XLM-RoBERTa + ViT
Sum based	54.27	45.40	43.98	37.62	10.34	38.32	64.67	60.50	07.23	6.22	34.66
Concatenation	65.38	50.00	45.45	38.71	25.35	44.98	63.26	63.76	25.00	29.03	45.26
Co-Attention	64.73	45.83	38.34	18.18	34.48	40.31	61.62	57.00	08.92	32.84	40.10
XLM-RoBERTa + Swin
Sum based	62.46	70.83	50.37	08.21	68.09	51.99	64.29	60.39	70.27	31.33	56.57
Concatenation	68.80	75.36	50.00	22.50	72.00	57.73	66.94	62.01	77.03	45.78	62.94
Co-Attention	63.26	70.34	57.83	26.51	64.65	56.52	70.80	65.09	71.14	38.36	61.35

---

(Ab = Abusive, Po = Political, Ge = Gender, Per = Personal Offence, Re = Religious; Co = Community, Ind = Individual, Org = Organization, So = Society.)

---

🔹 Impact of Hierarchical Loss Parameters (α, β)

Value of α & β	α=0.2, β=0.8			α=0.5, β=0.5			α=0.8, β=0.2			α=1.0, β=1.0
	H/NH	Cat	Tar	H/NH	Cat	Tar	H/NH	Cat	Tar	H/NH	Cat	Tar
BanglaBERT + ViT
Sum based	72.05	51.89	58.75	70.40	42.89	42.88	69.96	56.39	58.75	68.25	49.33	48.06
Concatenation	70.69	53.25	47.84	70.38	53.43	47.61	70.38	53.43	45.58	69.57	59.93	63.42
Co-Attention	67.02	56.39	58.75	66.31	46.07	61.27	67.44	58.68	63.21	65.71	57.01	60.55
BanglaBERT + Swin
Sum based	73.39	54.59	50.77	73.39	54.59	50.77	70.63	54.26	48.34	66.44	60.57	53.42
Concatenation	70.91	57.10	49.64	74.51	58.43	53.21	72.03	60.40	59.24	66.89	36.77	42.49
Co-Attention	69.78	55.78	61.89	68.50	57.50	61.68	63.97	50.87	49.09	66.99	55.69	63.72
XLM-RoBERTa + ViT
Sum based	62.73	36.72	31.44	62.56	51.99	56.57	63.53	40.35	31.20	65.65	61.63	65.64
Concatenation	63.84	40.13	37.78	62.36	44.98	45.26	61.29	46.45	36.06	64.16	57.33	54.42
Co-Attention	62.20	36.06	40.48	62.01	40.31	40.10	62.68	36.55	37.42	66.06	49.88	58.11
XLM-RoBERTa + Swin
Sum based	61.99	52.16	51.59	70.18	51.99	56.57	62.87	53.87	49.87	62.78	23.02	33.23
Concatenation	70.37	57.13	51.54	71.94	57.73	62.94	72.77	63.58	56.04	66.24	57.08	51.74
Co-Attention	63.79	51.02	58.97	65.53	56.52	61.35	64.05	50.23	56.10	64.63	52.93	58.47

---

🔑 Key Findings

• Hierarchical loss function balances predictions across binary, category, and target levels, with optimal weighting (α=0.5, β=0.5) achieving the most consistent performance and preventing bias toward binary detection at the expense of fine-grained classification.
• BanglaBERT + Swin Transformer achieves best overall performance, with concatenation fusion most effective for binary and category detection, while co-attention fusion excels at target group identification.
• Monolingual models (BanglaBERT) consistently outperform multilingual models (XLM-RoBERTa) by 3-5%, demonstrating the importance of language-specific pretraining for capturing cultural and linguistic nuances in Bangla hateful content.
• Fine-grained classification remains challenging, with ~40% of hateful memes misclassified as non-hate due to implicit expressions, and notable confusion between overlapping categories (Political/Abusive, Personal Offence/Gender), highlighting the difficulty of detecting subtle cultural cues and sarcasm.

---