Safer virtual environments

MetaHate

How can a metaverse platform detect harmful language in real time without carrying a model too heavy to deploy?

Published in the IEEE ICTC 2023 proceedings

Why I built it

Can an explainable, compressed text classifier support real-time hate-speech moderation inside a metaverse application?

Immersive social platforms reproduce the safety challenges of the wider internet, including harassment, hate speech, and offensive language. Detection models must be accurate enough to moderate interactions, explainable enough to inspect, and compact enough to run within an interactive platform.

MetaHate evaluates deep-learning text classifiers, develops a lightweight CNN with GloVe embeddings, applies LIME explanations, quantizes the model, and connects it to a live Roblox environment through a web service.

MetaHate system model from metaverse interaction through hate-speech detection and deployment

The complete system model connects live metaverse interaction to a labeled speech dataset, tokenization, GloVe embeddings, deep-model training, quantization, server deployment, and a moderation response returned to the virtual environment.

How the system works

Multi-class language detection

Deep-learning models classify text as hate, offensive, or normal using pretrained word representations and a benchmark social-media dataset.

Explainability and compression

LIME exposes influential words while quantization reduces the CNN footprint for practical serving.

Roblox deployment

A hosted web service receives messages from Roblox, performs real-time analysis, and returns moderation results to the virtual environment.

93.59%CNN model-size reduction

3speech classes

LIMElocal explanation method

Robloxinteractive deployment platform

What this work contributes

From a research question to an inspectable system.

A metaverse-focused system for hate, offensive, and normal language classification.
Comparative evaluation of deep-learning approaches for real-time moderation.
LIME explanations that expose the words influencing individual classifications.
A quantized CNN whose model size is reduced by 93.59 percent.
A working connection between the hosted detector and a Roblox environment.

MLP CNN and deep CNN architectures evaluated in MetaHate — The model comparison covers multilayer perceptron, convolutional neural network, and deep convolutional neural network architectures built for three-class text moderation.

How I read the results

What the evidence means

The deployment makes the research concrete: moderation is not evaluated only as an offline classification score. Model size, explanation quality, web serving, and the behavior of a live Roblox connection all become part of the system.

Current scope

What this study validates

MetaHate is a research prototype using text-based moderation and a three-class dataset. Broader deployment would require multilingual evaluation, continuously evolving language data, fairness analysis, and platform-specific moderation policies.

Inside the system

More evidence from the research.

LIME explanations comparing influential words across MetaHate models — LIME exposes the words that move each model toward hate, offensive, or normal predictions, making individual moderation decisions inspectable.

MetaHate CNN confusion matrices on CPU GPU and TPU — Processor-specific confusion matrices show whether the deployed CNN preserves class behavior across CPU, GPU, and TPU execution environments.

MetaHate web server and Roblox moderation demonstration — The end-to-end demonstration connects the hosted inference interface to live Roblox chat, showing a submitted message and the moderation result returned inside the virtual environment.

CNN confusion matrix for hate, offensive, and normal language — The confusion matrix reveals class-specific behavior, including the difficulty of separating hate speech from the larger offensive-language class.

MetaHate model and processor performance comparison — The performance analysis compares model behavior and deployment cost to support selection of a lightweight real-time classifier.

MetaHate web interface for submitting text to the detection model — The web service provides the inference endpoint that connects the trained language model to interactive virtual platforms.

Roblox environment connected to the MetaHate moderation service — The Roblox deployment demonstrates how live virtual messages can be sent to the hosted model for real-time analysis.