DETOEX is a multilingual hate speech detection system that combines Large Language Models (LLMs) with a curated vocabulary of derogatory language and traditional Natural Language Processing (NLP) techniques. It also leverages LLMs to provide contextualized explanations for why a certain piece of text was labeled as toxic.

• Multilingual support for English, French, and Greek
• Combined detection approach using both term matching and direct LLM analysis
• Detailed explanations of detected toxic content

DETOEX uses a two-pipeline approach for toxicity detection:

1. Term-based pipeline: Matches specific terms from curated vocabularies
2. Non-term pipeline: Analyzes text directly using LLMs
3. Fusion: Combines results from both pipelines for comprehensive analysis

• Python 3.12+
• Docker (for containerized deployment)
• An LLM API endpoint that supports OpenAI-compatible API format

You can freely access the deployed tool at https://detoex.ails.ece.ntua.gr/

The easiest way to run DETOEX locally is using Docker with the provided Dockerfile.local:

# Build the Docker image
docker build -t detoex-local -f Dockerfile.local .
# Run the container
docker run -p 8000:8000 detoex-local

The API will be available at https://localhost:8000.

DETOEX supports configuring different LLM endpoints for each language. By default, the Dockerfile.local is configured to use the Docker host's IP (172.17.0.1:8080) for all languages, but you can modify these settings:

# In Dockerfile.local
ENV LLM_URL_EN=https://172.17.0.1:8080/v1
ENV LLM_URL_FR=https://172.17.0.1:8080/v1
ENV LLM_URL_EL=https://172.17.0.1:8080/v1

You can update these variables before building the image, or override them at runtime:

docker run -p 8000:8000 \
  -e LLM_URL_EN=https://your-custom-endpoint/v1 \
  detoex-local

For development with hot-reloading of code changes:

docker run -p 8000:8000 \
  -v $(pwd)/detoex:/app/detoex \
  detoex-local

This will mount your local detoex directory into the container, allowing changes to be reflected immediately.

POST /
{
  "language": "en",
  "texts": ["text to analyze for toxicity"]
}

{
  "results": ["Explanation of toxicity if detected"]
}

The Docker image includes:

• Prompt templates for all supported languages
• Vocabularies of potentially toxic terms
• Stanza NLP models (downloaded during image build)

This project is funded as part of an FSTP call from the EU project UTTER (Unified Transcription and Translation for Extended Reality), supported by the European Union's Horizon Europe programme under grant agreement No. 101070631.