Social Media Toxicity Detection

A multi-stage NLP pipeline designed to detect and classify toxic language across social media platforms at scale. The system ingests raw user-generated content, applies platform-aware preprocessing, and combines classical machine learning with transformer-based deep learning to distinguish generic offensive speech from targeted hate speech.

The core challenge: toxic language looks very different on long-form forums versus short-form Twitter. Slang, abbreviations, and cultural context shift dramatically across platforms — a single model trained on one domain routinely fails on the other. This project explores whether domain-adaptive fine-tuning can close that gap.

Built as a university research project at CU Boulder, the pipeline covers the full NLP lifecycle from raw data ingestion through unsupervised exploration, supervised classification, and transformer fine-tuning — producing both quantitative benchmarks and qualitative insight into what makes language toxic.

Two publicly available datasets were combined to enable cross-platform analysis — one from long-form forum comments, one from short-form social posts:

The project was structured around three interconnected research questions:

The pipeline runs in three stages: data preparation, unsupervised exploration, and supervised modeling.

Data preparation applies different preprocessing for each modeling approach. For classical ML: contraction expansion → lowercase normalization → mention anonymization (@user → @USER) → URL and hashtag stripping → NLTK tokenization → stopword removal. For transformers: minimal normalization only — preserving subword context for the BERT tokenizer.

Unsupervised analysis uses TF-IDF (unigrams + bigrams, top 50K features) → TruncatedSVD (50K → 200 dimensions) → StandardScaler to project comments into a dense feature space. KMeans clustering (k=2 through k=6) and t-SNE visualization (5K sample) explore whether toxicity forms natural clusters in this space.

Supervised modeling runs four classical baselines — Logistic Regression, Multinomial Naive Bayes, LinearSVC, LightGBM — benchmarked on accuracy, F1, precision, and recall. The best classical model is then compared against two-stage DistilBERT fine-tuning: Stage 1 on 50K Jigsaw samples, Stage 2 on 9K TweetEval hate records.

Select a visualization below to explore the key findings interactively:

The pipeline is built entirely in Python with a clear separation between classical ML tooling and transformer-based deep learning.

Cross-platform toxicity detection is genuinely hard, and the 2.6% vocabulary overlap between Jigsaw and TweetEval is the clearest evidence of why. A model that achieves strong accuracy on forum comments will not transfer cleanly to tweets without domain-specific adaptation.

Classical bag-of-words features (TF-IDF + TruncatedSVD) provide useful signal and competitive baselines, but they miss contextual meaning — sarcasm, dog-whistles, and in-group slang all defeat surface-level features.

Two-stage DistilBERT fine-tuning is the most effective approach tested: Stage 1 on the large Jigsaw corpus provides a strong general toxicity prior; Stage 2 on TweetEval adapts the model to short-form, slang-heavy text. This sequential strategy is more data-efficient than training a single model on a merged dataset.

Several directions remain open for future exploration: