BIDWESH: Bangla Dialect Hate Speech Dataset
- BIDWESH is a curated dataset of 9,183 instances capturing hate speech in Barishal, Noakhali, and Chittagong dialects.
- It preserves parallel annotations for hate presence, hate type, and target group to address detection gaps in regional dialects.
- The resource supports research on dialectal bias, cross-dialect transfer, and fair moderation in low-resource Bangla NLP.
Searching arXiv for BIDWESH and closely related Bangla hate-speech/dialect resources to ground the article in current papers. BIDWESH is a manually translated and annotated multi-dialectal Bangla hate speech dataset introduced to support automatic hate-speech detection beyond standard Bangla, specifically in the Barishal, Noakhali, and Chittagong regional varieties. It is presented as the first Bangla regional hate speech dataset and is constructed by adapting a subset of the BD-SHS benchmark into three dialectal versions, yielding a parallel corpus of 9,183 dialectal instances with labels for hate presence, hate type, and target group (Fayaz et al., 22 Jul 2025). The dataset addresses a specific failure mode in Bangla content moderation: existing resources centered on standard written Bangla or transliterated Bangla do not adequately capture dialect-specific vocabulary, orthography, slang, and informal morphology, which can produce under-detection and biased moderation in linguistically diverse settings (Fayaz et al., 22 Jul 2025).
1. Definition and scope
BIDWESH is a curated research dataset rather than a predictive model. The name is glossed by its authors as “Hatred,” while the title expands it informally as “Bangla Regional Based Hate Speech Detection dataset” (Fayaz et al., 22 Jul 2025). Its core design objective is to make automatic hate-speech detection work for Bangla regional dialects rather than only for standard Bangla.
The dataset covers three dialects: Barishal, Noakhali, and Chittagong. Its total size is 9,183 dialectal instances, with 3,061 instances per dialect, all derived from a selected subset of BD-SHS. Each instance is labeled along three dimensions: binary hate versus non-hate, hate type, and target group. Hate type includes slander, gender, religion, call to violence, and combinations of these categories; target group includes individual, male, female, group, and several combination labels (Fayaz et al., 22 Jul 2025).
In relation to the broader Bangla hate-speech literature, BIDWESH is positioned as a dataset-level intervention rather than a modeling contribution. The paper explicitly states that it does not provide baseline experiments, benchmark scores, or train/dev/test splits, and therefore its primary significance lies in resource construction and annotation design rather than algorithmic performance claims (Fayaz et al., 22 Jul 2025). This distinguishes it from earlier Bangla hate-speech work built around standard-language classification benchmarks, including BD-SHS (Fayaz et al., 22 Jul 2025).
2. Motivation in Bangla hate-speech research
The paper locates BIDWESH in the context of pervasive hate speech on digital platforms in Bangladesh, including social media, comment sections, and streaming platforms (Fayaz et al., 22 Jul 2025). The central argument is that a substantial portion of online communication in Bangladesh is dialectal rather than standard Bangla, and that dialectal writing often includes regional slurs, local idioms, non-standard spelling, and informal grammar (Fayaz et al., 22 Jul 2025).
From a technical standpoint, the dataset is motivated by a mismatch between the linguistic distribution of deployed moderation targets and the training distribution of available corpora. Existing Bangla hate-speech datasets, including BD-SHS, BanglaHateBERT data, BANTH, and G-BERT data, are described as focusing on standard written Bangla or transliterated Bangla rather than regional dialectal text (Fayaz et al., 22 Jul 2025). BIDWESH argues that this leaves dialect-specific hate speech systematically underrepresented, especially when users exploit non-standard forms to evade detection (Fayaz et al., 22 Jul 2025).
This framing is consistent with a broader tendency in low-resource NLP: benchmark coverage often lags behind sociolinguistic variation. BIDWESH makes that issue explicit for Bangla hate-speech detection by treating dialectal variation not as noise to normalize away, but as a first-class representational requirement (Fayaz et al., 22 Jul 2025). A plausible implication is that the dataset is useful not only for supervised classification but also for transfer-learning, robustness, and fairness studies where the gap between standard and regional language varieties is itself the object of analysis.
3. Source corpus and construction pipeline
BIDWESH is derived from BD-SHS, described in the paper as a Bangla hate speech benchmark dataset introduced by Romim et al. in 2022. BD-SHS contains 50,281 Bangla comments from online platforms, of which 24,156 are labeled as hate speech, and provides binary as well as more detailed labels that BIDWESH preserves in adapted form (Fayaz et al., 22 Jul 2025).
From BD-SHS, the authors selected 3,061 instances: 1,513 hate and 1,548 non-hate. This subset functions as the shared source set for dialectal expansion. Each of these 3,061 standard Bangla sentences was translated into Barishal, Noakhali, and Chittagong, producing 3 dialects × 3,061 sentences = 9,183 dialectal instances (Fayaz et al., 22 Jul 2025).
The paper describes a multi-stage pipeline for dataset development. Data selection from BD-SHS was followed by translation or adaptation into the three dialects. The paper stresses that the translation was not literal: translators adapted vocabulary, orthography, and expression patterns to match authentic dialectal usage while preserving hate semantics (Fayaz et al., 22 Jul 2025). Translation was performed entirely by human translators who were native speakers; no automatic tools were used. Translators were instructed to preserve semantic content, preserve offensive strength, use natural dialectal forms rather than standard Bangla, and remain consistent with the original BD-SHS annotations (Fayaz et al., 22 Jul 2025).
A five-stage validation process was then applied: re-evaluation by the same translators; removal of unrelated text such as extraneous English fragments, unusual textual elements, missing or partial translations, and missing hate-bearing words or phrases; sentence verification by translators and other research team members; emoji removal; and balancing and stability checks (Fayaz et al., 22 Jul 2025). Because the dialects lack standardized orthography, spelling verification was performed manually according to common usage and phonetic fit (Fayaz et al., 22 Jul 2025).
The construction logic is therefore not merely translational but annotation-preserving and dialect-normalizing within each regional variety. This suggests that BIDWESH can be read as a parallel corpus with controlled semantic invariance and deliberately introduced surface variation.
4. Dialectal representation and linguistic properties
The paper does not provide a formal linguistic grammar of Barishal, Noakhali, or Chittagong. Instead, it identifies phenomena that matter for hate-speech detection across these dialects: lexical variation, orthographic variation, informal grammar, slang and regional swear words, and code-switching (Fayaz et al., 22 Jul 2025). It further notes that a single standard spelling for these dialects does not exist, so even within a dialect the same word may appear in multiple written forms (Fayaz et al., 22 Jul 2025).
BIDWESH addresses this by relying on native speakers from each region, who selected culturally correct swear words, slurs, and everyday phrasing rather than literal standard-Bangla renderings (Fayaz et al., 22 Jul 2025). The authors emphasize the preservation of target identity and hate intensity during translation, so that a sentence targeting a woman, a religious group, or a specific individual in the source corpus remains aligned with the same target and type labels in the dialectal versions (Fayaz et al., 22 Jul 2025).
The sample table described in the paper illustrates this parallelism. In a hate example, all four versions—standard Bangla and the three dialectal renderings—express the same insulting proposition, while exhibiting phonologically and orthographically distinct forms. In a non-hate example, all versions express a neutral proposition about a result making new history, again with dialect-specific realizations but preserved semantics (Fayaz et al., 22 Jul 2025). The significance of these examples lies not in lexical novelty alone, but in the corpus design principle: dialectal variation is encoded while label semantics are held constant.
This suggests BIDWESH is especially relevant for studying invariance under dialectal transformation. It can support analyses of whether a model’s decision boundary is anchored in target semantics and abusive intent, or instead overfits to standard-language lexical proxies.
5. Annotation schema
BIDWESH uses a multi-level annotation schema. The primary label is binary: Hate Speech = 1 for hate and Hate Speech = 0 for non-hate (Fayaz et al., 22 Jul 2025). The paper defines hate as content containing offensive, abusive, discriminatory, or toxic language directed toward individuals or groups, and non-hate as neutral, informative, or otherwise non-offensive content with no harmful intent (Fayaz et al., 22 Jul 2025).
Target-group labels are applied only to hate instances. The target categories are: Individual (ind), Male, Female, Group, Male_female, Male_group, and Female_group (Fayaz et al., 22 Jul 2025). The paper characterizes ind as a specific person and describes the other categories as demographic or collective targets, including combination labels such as hate toward both men and women or toward men or women within a group (Fayaz et al., 22 Jul 2025).
Hate-type labels are also applied only to hate instances. The base categories are Slander, Gender, Religion, and callToViolence (Fayaz et al., 22 Jul 2025). In addition, BIDWESH encodes combinations as distinct string labels. The combinations listed in the paper are: callToViolence_slander, Gender_slander, callToViolence_gender, Religion_slander, callToViolence_religion, Gender_religion, callToViolence_religion_slander, callToViolence_gender_slander, and Gender_religion_slander (Fayaz et al., 22 Jul 2025).
The annotation logic is categorical rather than formally axiomatized. The paper does not provide equations or schema diagrams, but it makes several operational rules clear: target and type labels are assigned only when hate is present, and non-hate content receives - for those fields in the sample representation (Fayaz et al., 22 Jul 2025). It also implies that severe personal slander can still be labeled as hate even when it is not directed at a protected class (Fayaz et al., 22 Jul 2025).
6. Annotation process, statistics, and structure
The translators and annotators were the same regional experts. For each dialect there were two native speakers from the relevant region, giving six translators in total: T01 and T02 for Noakhali, T03 and T04 for Chittagong, and T05 and T06 for Barishal (Fayaz et al., 22 Jul 2025). The paper reports annotator demographics: T01 was 23 and a student; T02 was 25 and a researcher with translation experience; T03 was 22 and a student; T04 was 23 and a student; T05 was 23 and a student; and T06 was 24 and a researcher. Across the six annotators, two were female and four were male, and all were at least 18 years old (Fayaz et al., 22 Jul 2025).
The annotation protocol was manual. Each dialect team annotated its own set of 3,061 sentences for hate versus non-hate and, where hate was present, for target and type (Fayaz et al., 22 Jul 2025). The paper does not specify the exact number of annotators per instance or a formal adjudication procedure for disagreements beyond the iterative verification stages in the translation workflow (Fayaz et al., 22 Jul 2025). It also does not report numerical inter-annotator agreement measures such as Cohen’s kappa or Krippendorff’s alpha (Fayaz et al., 22 Jul 2025). Reliability is instead justified qualitatively through native-speaker expertise and repeated review (Fayaz et al., 22 Jul 2025).
The class distribution is balanced within each dialect because each dialect mirrors the same 3,061-source-instance subset. The binary distribution is 1,513 hate instances, corresponding to 49.24%, and 1,548 non-hate instances, corresponding to 50.76% (Fayaz et al., 22 Jul 2025).
The hate-type distribution for the 1,513 hate instances per dialect is as follows:
| Hate type | Count | Percentage |
|---|---|---|
| Slander | 822 | 53.65% |
| callToViolence_slander | 214 | 13.96% |
| Gender_slander | 210 | 13.71% |
| Gender | 186 | 12.14% |
| callToViolence | 119 | 7.77% |
| Religion | 47 | 3.07% |
| callToViolence_gender | 36 | 2.35% |
| Religion_slander | 23 | 1.50% |
| callToViolence_religion | 22 | 1.44% |
| Gender_religion | 8 | 0.52% |
| callToViolence_religion_slander | 8 | 0.52% |
| callToViolence_gender_slander | 4 | 0.26% |
| Gender_religion_slander | 3 | 0.20% |
The target distribution for the 1,513 hate instances per dialect is:
| Target category | Count | Percentage |
|---|---|---|
| Individual (ind) | 520 | 34.37% |
| Male | 443 | 29.28% |
| Female | 284 | 18.77% |
| Group | 240 | 15.87% |
| Male_female | 200 | 13.22% |
| Male_group | 3 | 0.20% |
| Female_group | 3 | 0.20% |
These statistics indicate that slander dominates the hate-type distribution and that individual-targeted hate is the largest target category (Fayaz et al., 22 Jul 2025). The paper itself highlights that pure religion and pure call-to-violence categories are comparatively smaller, while multi-type combinations remain present but less frequent (Fayaz et al., 22 Jul 2025). A plausible implication is that models trained on BIDWESH may face pronounced class imbalance within the hate subspace even though the top-level hate/non-hate distinction is nearly balanced.
Regarding file structure, the paper does not specify exact field names but shows a CSV-like organization with columns such as Standard Bangla, Chittagong, Noakhali, Barishal, Target, Type, and Hate Speech (Fayaz et al., 22 Jul 2025). It does not define train/dev/test splits, leaving downstream evaluation protocol design to later users (Fayaz et al., 22 Jul 2025).
7. Relation to existing resources, applications, and limitations
The paper compares BIDWESH with prior Bangla resources in two ways. First, relative to BD-SHS, BIDWESH is much smaller but linguistically richer: BD-SHS contains 50,281 standard Bangla comments and no explicit dialectal variants, whereas BIDWESH contains 9,183 instances derived from 3,061 source sentences and provides three regional dialect versions for each sentence while maintaining multi-label target and type annotations (Fayaz et al., 22 Jul 2025). Second, relative to other Bangla datasets such as BanglaHateBERT, G-BERT, and BANTH, BIDWESH is distinguished by its focus on regional dialects and by the granularity of its type and target combinations (Fayaz et al., 22 Jul 2025).
The paper also situates BIDWESH with respect to non-hate dialect resources. It names VASHANTOR as a dialect-to-standard translation resource, ANCHOLIK-NER as dialectal named-entity recognition, and ANUBHUTI as emotion detection in regional languages, while asserting that none of these targets hate speech in dialects (Fayaz et al., 22 Jul 2025). On that basis, the paper characterizes BIDWESH as the first dataset explicitly constructed for Bangla regional hate-speech detection (Fayaz et al., 22 Jul 2025).
The stated application areas include dialect-sensitive hate-speech classifiers, fair and inclusive moderation, sociolinguistic research, model robustness and domain adaptation, and low-resource NLP resource construction (Fayaz et al., 22 Jul 2025). Because the corpus is parallel across dialects and anchored to a standard Bangla source, it is especially suitable for cross-dialect transfer experiments and standard-to-dialect adaptation studies, although the paper itself does not report such experiments (Fayaz et al., 22 Jul 2025).
Several limitations are stated directly. BIDWESH covers only three dialects and excludes other major varieties such as Sylhet and Mymensingh, as well as India-Bangla dialects (Fayaz et al., 22 Jul 2025). Its domain is inherited from BD-SHS, so the data originates from social-media and streaming comments rather than from long-form blogs, news commentaries, or private chats (Fayaz et al., 22 Jul 2025). Its scale, 9,183 instances, is modest relative to larger multilingual hate corpora (Fayaz et al., 22 Jul 2025). Numerical inter-annotator agreement is absent, and no in-paper benchmarks are supplied (Fayaz et al., 22 Jul 2025).
The paper makes the dataset publicly available through Mendeley Data at https://data.mendeley.com/datasets/bpkrvf882k/1 and states that it is intended for research and academic purposes (Fayaz et al., 22 Jul 2025). No code release is mentioned, and no special restrictive license is specified in the paper itself (Fayaz et al., 22 Jul 2025).
Taken together, BIDWESH occupies a specific and methodologically important niche in Bangla NLP: it operationalizes dialect sensitivity at the dataset level by constructing a parallel, balanced, manually verified hate-speech corpus across Barishal, Noakhali, and Chittagong. Its principal contribution is not a new classifier but a substrate for evaluating whether Bangla hate-speech detection systems can generalize beyond the standard-language regime that has dominated previous resource construction (Fayaz et al., 22 Jul 2025).
