A Privacy-Compliant Hybrid Framework for Real-Time Offensive Text Moderation in Dual-Mode Chat Systems

View at publisher Summarise with VIRAT
Authors: Raghu Yashwanth T M, Parivel R, Priyan G S, Santhosh Kumar R
Cited by: 0
References: 14

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

An automatic text moderation system that provides secure, synchronized, real-time communication in chat-style venues for anonymous and non-anonymous modes is described. It is constructed to prevent inappropriate content by targeting and blocking messages at the edge, forming a protective layer that ensures inappropriate content is intercepted before delivery. Any flagged message is blocked and never transmitted to the recipient and never be transported to the end user. A hybrid filtering approach was developed to combine the rule-based approach with AIinduced detection to handle phonetically changed and structurally modified word forms, which are often employed to bypass the regular filters. This manually compiled dataset was used to mimic those linguistic variants and train moderation models that are tailored towards them. Furthermore, generative language models were integrated to increase the dataset and generalize to unseen inputs. The test results indicate that the detection ratio was 96.4% with a false positive ratio of 2.1%, and the average processing time was 117 milliseconds, regardless of the message size. The modular design allows for a lightweight deployment on various digital platforms, such as educational, corporate, and public-based systems. Continual adjustment is supported through the use of dynamically updated wordlists and scalable moderation logic. The system emphasizes privacy-preserving incognito modes alongside policy-enforced permanent chat, enabling flexible deployment while ensuring regulatory compliance and communication safety. Further work will add deeper semantic models informed by a context-aware language model and improve the quality of the moderation by refining the user feedback loop.

Fields of Study & Subjects

Hate Speech and Cyberbullying Detection Spam and Phishing Detection Text and Document Classification Technologies

Publication details

Journal / source
International Conference on Cognitive Systems and Computer Interaction (ICoSCI)
Pages
1–6
Publication date
15 January 2026
Language
EN