Sarcasm Detection in Tweets using BERT

Sarcasm flips meaning — "Oh great, another Monday" means the opposite of what its words say — and it breaks every sentiment tool that reads text literally. Because sarcasm lives in context, contrast and exaggeration rather than in keywords, it is a classic hard problem for NLP and a genuinely interesting final-year build. This project tackles it with transfer learning: bert-base-uncased, pre-trained on massive English text, is fine-tuned on iSarcasmEval's author-annotated tweets, where the labels come from the tweet authors themselves rather than third-party guessers. The demo scores any pasted tweet and visualizes token attention so the verdict is explainable — the report can point at exactly which words the model treated as sarcasm cues, and the error analysis can discuss the cases sarcasm detection still gets wrong, such as world-knowledge jokes and deadpan delivery.

1. iSarcasmEval tweets are WordPiece-tokenized (max 128 tokens); author-provided sarcastic/non-sarcastic labels are kept as the binary target.
2. bert-base-uncased feeds its CLS embedding into a dropout + linear head; class weights in the cross-entropy loss compensate for the sarcastic-class minority.
3. The model is fine-tuned for about 4 epochs with AdamW; every epoch logs accuracy, precision, recall, macro-F1 and the confusion matrix on the validation split.
4. At inference, the demo app tokenizes the pasted tweet, runs it through the fine-tuned model and renders the verdict with a confidence bar.
5. Attention-based token influence scores color each word by how strongly it pushed the decision toward sarcastic or non-sarcastic.
6. Batch mode reads a CSV of tweets and writes per-tweet verdicts and confidences to a report file.

Tech stack:

• Python 3.10, PyTorch, Hugging Face transformers (bert-base-uncased)
• scikit-learn (metrics, confusion matrix), NumPy, pandas
• Matplotlib (training curves, label distribution)
• Single-file HTML/CSS/JS demo app (analyzer, dataset explorer, training views)
• iSarcasmEval benchmark (SemEval-2022 Task 6, author-annotated tweets)
• Trained weights exported from the included fine-tuning run

• [Tweet analyzer] Paste any tweet and get a sarcastic/non-sarcastic verdict with a confidence bar; token-attention highlighting colors the words that most pushed the decision.
• [BERT fine-tuning pipeline] bert-base-uncased with a dropout + linear classification head, fine-tuned with AdamW and a linear warmup schedule; class-weighted loss counters the label imbalance; all hyperparameters exposed in the notebook.
• [Author-annotated data] Trains on iSarcasmEval (SemEval-2022 Task 6), where sarcasm labels come from the tweet authors themselves — the gold standard for this task.
• [Dataset explorer view] Browse the iSarcasmEval label distribution, tweet-length stats and train/test splits inside the demo app, with real example tweets from each class.
• [Training-curve view] Per-epoch macro-F1 and train/validation loss plots that the notebook reproduces during the build.
• [Batch scoring mode] Score a CSV of tweets into a verdict report for larger-scale analysis.
• [Full evaluation logging] Accuracy, precision, recall, macro-F1 and the confusion matrix computed on the official iSarcasmEval test split — never pre-claimed.

• Complete source code (preprocessing, fine-tuning, attention extraction, inference, demo app)
• Jupyter training and evaluation notebook (buyer-run procedure: fine-tune, evaluate, inspect errors)
• Project report PDF (background, iSarcasmEval analysis, methodology, evaluation, error analysis)
• PPT presentation for final review
• Viva Q&A preparation document (transformers, attention, fine-tuning, class imbalance, evaluation metrics)
• Setup guide (environment, dataset download, training, running the demo)