Neural Network-Based Detection of Gender-Biased Hate Speech in Social Media Using Optimized Text Classification

Abstract

This paper presents a robust neural network-based framework for detecting gender-biased hate speech in social media, leveraging optimized text classification techniques. The increasing prevalence of hate speech on platforms like Twitter poses a significant threat to digital discourse, necessitating automateddetectionsystemscapableofhandlingthecomplexities of natural language.Our proposed model achieves a state-of-the-art classification accuracy of 98.19%, surpassing conventional machine learning approaches. The methodology integrates advanced natural lan- guageprocessing(NLP)techniqueswithdeeplearning,employing adatasetof5,000tweetscollectedviatheSNSCRAPEAPI. The system features a meticulously designed text preprocessing pipeline utilizing the NLTK library, vocabulary optimization through frequency-based filtering, and a custom neural network architecture enhanced with dropout regularization for improved generalization. Akeycontributionofthisworkisitsabilitytoaddress three major challenges in gender-biased hate speech detection: the context-dependent nature of offensive language, extremeclass imbalance in real-world datasets, and linguistic variations inherent in short-text social media content. Through extensive experimentation, our model demonstrates a 12.4% improvement in classification performance over traditional approaches suchas Support Vector Machines (SVM) and Logistic Regression. Furthermore, the inclusion of an adaptive learning rate mechanism ensures stability in model convergence, while dropout regularization mitigates overfitting.This research contributes to ongoing efforts to develop au- tomated systems for hate speech detection by proposing an optimized deep learning approach capable of handling noisy, unbalanced,andcontext-sensitivesocialmediadata.Futurework will extend this framework to multilingual settings and real-time detection applications

Introduction

1. Introduction

Hate speech on social media, especially gender-biased content (e.g., misogyny, misandry), is a growing concern. Manual moderation is impractical at scale, and automated systems using NLP and deep learning offer a solution. However, challenges include:

• Context-dependency of language

• Data imbalance (hate vs. non-hate)

• Short text lengths in platforms like Twitter

2. Objective

To develop an optimized neural network-based framework that:

• Detects gender-biased hate speech

• Handles imbalanced and noisy text data

• Outperforms traditional models (e.g., SVM, Naive Bayes)

3. Key Contributions

1. A custom neural network achieving 98.19% accuracy in binary classification.

2. A 7-step text preprocessing pipeline achieving 95.6% noise reduction.

3. Effective class balancing using SMOTE.

4. Comparative analysis showing clear performance superiority over traditional methods.

4. Methodology

A. Data Collection & Preprocessing

• Collected 5,000 tweets using hashtags like #womenarestupid and #menaretrash via SNSCRAPE.

• Balanced to 3,973 entries using oversampling.

• Preprocessing included:

  • Removing URLs, punctuation, and stopwords

  • Lowercasing and token filtering

  • Minimum token length enforcement

B. Vocabulary Optimization

• Reduced vocabulary from 9,564 to 3,030 tokens

• Used frequency threshold θ=2\theta = 2θ=2

C. Model Architecture

• Input Layer: 3,030 nodes

• Hidden Layer: 50 nodes with ReLU

• Output Layer: 1 node with Sigmoid

• Dropout: 0.2 to reduce overfitting

• Optimizer: Adam (learning rate = 0.001)

• Loss: Binary cross-entropy

5. Dataset

• Comprised of diverse, annotated user-generated content.

• Focused on English-language posts, with care taken to avoid storing any personally identifiable information (PII).

6. Results

• Figure 1: Visuals of accuracy/loss over epochs and keyword frequencies like “genderpaygap”, “patriarchy”, “misandry”.

• Preprocessing and dropout helped prevent overfitting and improved generalizability.

7. Related Work Highlights

• Prior work spans traditional ML, LSTM, CNN, transformer-based models (BERT), and multimodal approaches.

• Limitations in earlier models included handling sarcasm, class imbalance, and lack of multimodal data support.

8. Discussion & Limitations

• Strengths:

  • High performance

  • Robust handling of noise and class imbalance

• Limitations:

  • English-only support

  • Binary classification may miss nuanced hate

  • May miss slang, coded, or implicit bias

  • Hashtag-based data collection may introduce sampling bias

• Future Work:

  • Multilingual capability

  • Multi-class classification

  • Incorporation of multimodal data (text + images)

  • Adaptive vocabulary updates

Conclusion

In this study, we have developed a neural network-based approach for detecting gender-biased hate speech in social media content, achieving a state-of-the-art classification accu- racy of 98.19%. The model’s architecture, combined with an optimized text preprocessing pipeline and vocabulary reduc- tionstrategy,hasproveneffectiveinaddressingthechallenges posed by noisy and unstructured social media data. To enhance the model’s applicability and robustness, future research should focus on several key areas: 1) Multilingual Support: Incorporate multilingual embed- dingsortranslationmodelstodetecthatespeechacrossvarious languages, thereby increasing the system’s global applicability. 2) Multi-Class Classification: Develop a more nuanced classification system that can distinguish between different types and severities of hate speech, enabling more targeted content moderation strategies. 3) Real-Time Deployment: Implement the model in a real-time environment, such as a Flask API, to facilitate immediate detection and response to hate speech on social media platforms. 4) Continuous Vocabulary Update: Establish mechanisms for the dynamic updating of the model’s vocabulary to capture emerging terms and slang used in hate speech, ensuring the system remains current with evolving language trends. 5) Comprehensive Data Collection: Expand data collection methods to include a wider array of hate speech examples, encompassing both explicit and implicit instances, to improve the model’s generalization capabilities. By addressing these areas, the proposed system can evolve intoamorerobustandversatiletoolformitigatingtheproliferationofgender-biasedhatespeechonsocialmedia, contributingtoasaferandmoreinclusiveonlineenvironment.

References

[1] Arango, J. Pe´rez, and B. Poblete, ”Hate Speech Detection is Not asEasy as You May Think,” Proc. 42nd Int. ACM SIGIR Conf., 2019, doi:10.1145/3331184.3331262. [2] J.H. Park and P. Fung, ”Abusive Language Detection Using Hierarchical LSTMs,” Proc. EMNLP, 2020, pp. 210-220. [3] H. Mubarak, ”A Literature Review of Textual Hate Speech Detection Methods,” Information, vol. 13, no. 6, 2022, doi: 10.3390/info13060273. [4] A. Das et al., ”Multilingual Hate Speech Detection Using Transformers,”arXiv:2401.11021, 2024. [5] A. Mandal et al., ”Multimodal Hate Speech Detection with Attentive Fusion,” arXiv:2401.10653, 2024. [6] S. Unnava and S.R. Parasana, ”Deep Learning for CyberbullyingDetec- tion with Focal Loss,” J. Comput. Eng. Sci., vol. 14, no. 4, 2024. [7] M.L. Ripoll et al., ”Transformer Ensembles for Multilingual Detection,” [8] Proc. HASOC, 2022, pp. 45-53. [9] K. Darwish et al., ”CBDC-Net: Advanced Cyberbullying Detection,” [10] IJCESEN, vol. 9, no. 2, 2024. [11] Z. Zhang et al., ”Hate Speech Detection Using CNNs,” Proc. WWW, 2018, pp. 1345-1354. [12] J. Devlin et al., ”BERT: Pre-training of Deep Bidirectional Transform- ers,” Proc. NAACL, 2019, pp. 4171-4186.

Copyright

Copyright © 2025 Shubham Hazra, Shushant Aggarwal, Shruti Kishore. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.