AI-Driven Media Analysis for Toxic and Xenophobic Language Detection

Reviewing more than explicit keywords

Harmful media language can appear through direct terms, implicit framing, questionable factual claims, sentiment, or recurring topic patterns. This NYU applied project explored how those signals could be brought into one multilingual review workflow instead of being assessed as isolated classification tasks.

Three modules, one review path

I built the preprocessing and evaluation pipeline, integrated the analysis modules, and developed the Streamlit application and API-based workflow. The system accepted heterogeneous source documents and returned structured signals for human review.

• Processed PDF, Word, and plain-text inputs through a unified document pipeline.
• Combined GPT-4-based analysis with a fine-tuned BERT classifier.
• Integrated toxicity and xenophobia detection, fact and terminology validation, and thematic analysis.

Signals designed for interpretation

The interface organized results into toxicity categories, explicit and implicit xenophobia indicators, sentiment, 1–10 severity scores, fact and terminology checks, and topic labels. The aim was to support a reviewer’s judgment with multiple evidence views rather than reduce a document to a single label.

Measured within the project setup

Internal project evaluation reported 87% xenophobic-language detection accuracy, 90% fact and language validation accuracy, and 85% thematic categorization accuracy, with sub-two-second response time in the tested environment. These figures apply only to the project dataset, task definitions, evaluation setup, and runtime environment; they are not general production-performance claims.

An applied academic collaboration

The work was completed as a UNICC-sponsored NYU applied project. This description does not imply employment by UNICC, deployment within UNICC systems, institutional endorsement, or validation beyond the project scope.