Multilingual Toxicity Detection with Enhanced Balancing and Contextual Learning

Saju, Jacob (2025) Multilingual Toxicity Detection with Enhanced Balancing and Contextual Learning. Masters thesis, Dublin, National College of Ireland.

Preview	PDF (Master of Science) Download (860kB) | Preview
Preview	PDF (Configuration Manual) Download (1MB) | Preview

Abstract

Online toxicity detection systems struggle immensely in scaling across multiple, diverse linguistic and cultural environments, frequently privileging high-resource languages and offering poor protection to low-resource languages speakers. This work presents MultiToxiGuard, a multilingual toxicity detection system that solves these problems using three new components: a Smart Balancing Module using hierarchical sampling and dynamic weighting, a Contextual Enhancement Layer leveraging cultural embeddings for enhanced semantic awareness, and a Confidence Estimation System that includes robust uncertainty estimation. Utilizing a dataset of 15 languages from 9 language families, rigorous data augmentation processes are implemented that greatly enhanced representation of low-resource languages (Japanese +1518%, Vietnamese +1208%). Results of the validation indicate high overall performance (F1=0.7944, accuracy=0.8278) with impressive uniformity spanning linguistic boundaries, and having a cultural fairness score of 0.96. Specifically, a few low-resource languages (Estonian, Swahili) performed better than medium-resource languages, highlighting the efficacy of these balancing techniques. Whereas performance objectives of F1 (≥0.88) and the rate of false positives (≤0.03) are still daunting, MultiToxiGuard is a major step forward in fair content moderation that closes the high to low-resource languages' performance gap, a sore problem of past techniques. This system presents a single, integrated framework for detection of toxicity which performs at a consistent rate without the need for distinct models per language, markedly improving the best available multilingual content moderation technologies.

Item Type:	Thesis (Masters)
Supervisors:	Name Email Niculescu, Hamilton UNSPECIFIED
Subjects:	Q Science > QA Mathematics > Electronic computers. Computer science T Technology > T Technology (General) > Information Technology > Electronic computers. Computer science P Language and Literature > P Philology. Linguistics > Computational linguistics. Natural language processing Z Bibliography. Library Science. Information Resources > ZA Information resources > ZA4150 Computer Network Resources > The Internet > World Wide Web > Websites > Online social networks T Technology > TK Electrical engineering. Electronics. Nuclear engineering > Telecommunications > The Internet > World Wide Web > Websites > Online social networks
Divisions:	School of Computing > Master of Science in Data Analytics
Depositing User:	Ciara O'Brien
Date Deposited:	18 Nov 2025 17:47
Last Modified:	18 Nov 2025 17:47
URI:	https://norma.ncirl.ie/id/eprint/8945

Actions (login required)

View Item