Fine-tuning BERTweet for two tweet classification tasks: irony detection and emotion classification. Both tasks are evaluated using macro F1-score.

Binary classification: predict whether a tweet is ironic (1) or non-ironic (0).

• Dataset: 2,862 training / 955 validation / 784 test tweets (test labels not released — challenge submission only)
• Model: vinai/bertweet-base + custom FC head (Dropout → Linear)
• Best checkpoint val F1: 0.7667 (model saved by best val loss)

Multi-class classification: predict the dominant emotion in a tweet.

• Dataset: 3,257 training / 374 validation / 1,421 test tweets (test labels not released — challenge submission only; class imbalance: anger >> joy, optimism, sadness)
• Model: vinai/bertweet-base + custom FC head (Dropout → Linear)
• Best checkpoint val F1: 0.8147 (model saved by best val F1)

Both tasks use domain-specific preprocessing tailored for BERTweet:

• URLs → HTTPURL, mentions → @USER (matching BERTweet's pretraining vocabulary)
• Emojis converted to textual form (e.g., 😊 → :smiling_face:); unknown emoji tokens added to the model vocabulary
• Emoticons converted via emot library
• Letter repetitions reduced (e.g., soooo → soo)
• Hashtags segmented using wordsegment

Custom classifier head on top of the fine-tuned BERTweet encoder:

BERTweet [CLS] representation → Dropout → Linear → logits

Training with early stopping on validation loss (irony) / F1 (emotion). Hyperparameter search over learning rate, dropout, and batch size.

The emotion dataset is imbalanced (anger is the majority class). To address this:

1. Cosine similarity computed between hand-crafted seed tweets and each class distribution using Twitter4SSE sentence embeddings — to verify that seeds are representative.
2. DeepSeek-V3 used to generate paraphrases of minority-class tweets (joy, optimism, sadness), preserving emotional tone.
3. Augmented samples merged with the original training set.

Reported on the validation set (test set labels were not released — test tweets were used for challenge submission only, see predictions/).

NLP-Tweet-CLS-Challenge/
├── irony_submission.ipynb           # Irony detection pipeline
├── emotion_submission.ipynb         # Emotion classification pipeline
├── data_augmentation_process.ipynb  # DeepSeek-V3 augmentation process
├── irony/                           # Irony task data files
│   ├── train_text.txt / train_labels.txt
│   ├── val_text.txt / val_labels.txt
│   ├── test_text.txt
│   ├── mapping.txt
│   └── abbreviations.json
├── emotion/                         # Emotion task data files
│   ├── train_text.txt / train_labels.txt
│   ├── val_text.txt / val_labels.txt
│   ├── test_text.txt
│   ├── mapping.txt
│   └── abbreviations.json
├── emotion_train_aug/               # Augmented training data for emotion task
└── predictions/
    ├── irony_predictions.csv
    └── emotion_predictions.csv

• PyTorch — model training and custom classifier
• HuggingFace Transformers — BERTweet tokenizer and model
• sentence-transformers — Twitter4SSE embeddings for augmentation validation
• DeepSeek-V3 — LLM-based tweet paraphrasing for data augmentation
• scikit-learn — metrics (macro F1), hyperparameter grid search
• emoji / emot / wordsegment — tweet-specific text preprocessing