An Empirical Study on Multi-class Classification of Hostile Posts in Hindi Language
Loading...
Date
2025-07-30
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Department of Computer Science & Engineering, NITS
Abstract
The exponential growth of social media usage has intensified the dissemination of hostile
content online, posing significant risks to individuals and communities. While consider
able research has focused on hostility detection in English, other linguistically rich but
under-resourced languages such as Hindi remain severely neglected due to the lack of
high-quality annotated datasets and specialized classification models. This thesis ad
dresses these challenges by presenting a holistic framework for the detection of hostile
content in Hindi, with contributions spanning dataset creation, model development,
and architectural innovation.
The first major contribution is the development of a comprehensive and fine-grained
hostility detection dataset comprising approximately 8300 Hindi tweets manually anno
tated into five distinct classes: Hate, Abusive, Defamation, Offensive, and Non-Hostile.
To exploit this dataset effectively, we developed a suite of baseline and hybrid models,
including Convolutional Neural Networks (CNN), Long Short-Term Memory networks
(LSTM), Multi-Layer Perceptrons (MLP), and Decision Trees. These were trained
using multilingual embeddings such as GloVe, MuRIL, and XLM-RoBERTa, and fur
ther enhanced with emoji-based semantic features to improve contextual interpretation.
The hybrid models were fine-tuned to address challenges inherent in Hindi text, such
as morphological complexity, and ambiguous phrasing.
We leveraged MuRIL’s pooled output to improve Hindi text representation, benefiting
from its multilingual understanding and proposed novel hybrid models by combining
RNN and CNN architectures with traditional machine learning classifiers.
As a methodological advancement, this thesis proposes the Scaled Multi-Head Atten
tion–CNN Fusion Model (SMA-CFM), a novel architecture that combines Transformer
based attention with convolutional layers to capture both long-range dependencies and
local linguistic features from textual as well as emoji features. A trainable attention
scaling mechanism is introduced within the multi-head attention block to dynami
cally prioritize hostility-relevant tokens. To further enhance classification robustness,
a multi-model ensemble strategy integrates feature representations from neural back
bones using Random Forest and weighted voting mechanisms.
This research makes significant contributions to multilingual NLP by offering a scal
able, high-performance framework for hostile content detection in Hindi. It highlights
the importance of combining dataset granularity, linguistic contextualization, and ar
chitectural innovation to address challenges in low-resource language environments.
The dataset, models, and methodologies presented herein lay a robust foundation for
future advancements in socially responsible AI and content moderation technologies.