Introduction
Detecting hate speech on social media is a critical challenge for platform safety, but it becomes significantly more complex in multilingual environments. This research focuses on "Hinglish"—a code-switched blend of Hindi and English. Hinglish is considered a low-resource language in NLP because it lacks large-scale labeled datasets and often utilizes Roman script for Hindi words, leading to inconsistent spelling and grammar that standard mono-lingual models cannot easily parse.
Beyond linguistic complexity, models often suffer from "stereotypical bias," where they incorrectly flag neutral discussions or general opinions as hate speech simply because they contain specific religious terms or profane words used in a non-hateful context. This paper proposes a multi-tiered approach to address these linguistic and social nuances.
Proposed Methodology
The authors introduce a three-tier pipeline designed to capture both the textual content and the social context of users.
1. Linguistic Modeling
The core text analysis uses a hybrid architecture of Convolutional Neural Networks (CNN), Long Short-Term Memory (LSTM) networks, and Attention layers. This combination allows the model to capture both local spatial features and long-term temporal dependencies in the text. To further assist the model, the authors incorporate a Profanity Vector (PV)—a 210-dimensional flag for specific offensive terms.
2. Author Profiling via Graph Embeddings
Recognizing that hate speech often originates from specific social clusters, the researchers utilized social network analysis. By building graphs based on Twitter followers and retweets, they generated user embeddings using DeepWalk and node2vec. This captures "linguistic homophily," the principle that individuals within the same social circles tend to share similar speech patterns.
3. Semi-Supervised Bias Elimination (BE)
To mitigate algorithmic bias, an "expert-in-the-loop" system was implemented. Human experts identify bias-inducing word pairs (e.g., unfairly linking a specific religion to a negative trait). The system then mathematically neutralizes and equalizes these word embeddings, ensuring the model treats neutral mentions of protected groups fairly.
Experimental Results
The pipeline was evaluated on two datasets: the HS Dataset (3,475 tweets with social data) and the HEOT Dataset (3,189 tweets for linguistic evaluation).
- Performance: The full pipeline achieved F1 scores of 0.73 on HS and 0.77 on HEOT, representing a significant improvement (0.04–0.07) over previous state-of-the-art models.
- Social Signal: Integrating social network features (DeepWalk) alone improved accuracy on the HS dataset from 0.64 to 0.71, proving that social context is a powerful predictor of behavior.
- Qualitative Impact: t-SNE visualizations demonstrated that the debiasing step successfully separated clusters based on actual intent rather than religious or social identity markers.
Conclusion and Future Work
By combining deep learning with social graph analysis and expert-guided debiasing, this research provides a robust framework for hate speech detection in low-resource, code-switched languages. While the current model requires some manual intervention for bias identification, future work aims to automate this process and adapt the methodology to other language pairs like Spanish-English or Chinese-English.
Summary updated