# Email Intent and Priority Detection with DistilBERT A Hugging Face-ready multi-task NLP project that classifies Enron emails into: - **Intent**: Meeting Request, Action Required, Info Sharing, Follow-up, Approval Request, Complaint - **Priority**: High, Medium, Low This project uses **DistilBERT** with two classification heads on top of a shared encoder. ## Model Card Summary - **Task**: Multi-label text classification (multi-task: intent + priority) - **Language**: English - **Dataset**: Enron Email Dataset - **Architecture**: DistilBERT backbone + separate intent and priority heads - **Framework**: PyTorch / Transformers ## Intended Use This model is designed for email triage and message understanding. It can help classify incoming emails by topic and urgency. ## Limitations - Labels are generated using **weak supervision**, not manual annotation. - Performance may drop on emails outside the Enron style or vocabulary. - The model is best suited for research and experimentation unless further fine-tuned with human-labeled data. ## Training Details The notebook performs the following steps: 1. Parse raw Enron emails into structured text fields. 2. Clean subjects and bodies. 3. Generate pseudo-labels using rule-based patterns. 4. Balance the dataset with sampling. 5. Tokenize text using `distilbert-base-uncased`. 6. Train a shared DistilBERT encoder with two heads. 7. Evaluate on a held-out test set. ## Performance Reported test results from the notebook: - **Intent F1**: 0.8047 - **Priority F1**: 0.8845 - **Intent Accuracy**: about 0.80 - **Priority Accuracy**: about 0.88 ## Files in This Repository - `enron_final.ipynb` — training and evaluation notebook - `README.md` — project overview - `output/` — saved model and plots ## Inference Example Input: - Subject: `URGENT Please sign the contract by EOD today` - Body: `This is time-sensitive. Please review and sign immediately.` Expected output: - Intent: `Action Required` - Priority: `High` ## Usage ### 1. Install dependencies ```bash pip install transformers safetensors torch scikit-learn pandas numpy matplotlib seaborn tqdm ``` ### 2. Run the notebook Open the notebook in Google Colab or Jupyter and execute all cells in order. ### 3. Load the saved model Use the saved `bestmodel.safetensors`, tokenizer files, and `labelmaps.json` for inference. ## Hugging Face Deployment Notes If you want to publish this project on Hugging Face: - Add a `README.md` model card like this one. - Include the trained weights and tokenizer files. - Provide a small inference script or pipeline example. - Mention that labels are pseudo-labeled and not manually annotated. ## Citation If you use this project in a report or demo, cite it as: - **Email Intent and Priority Detection using DistilBERT** ## Code Submission Instructions For submission, you may either: - Attach the code as a PDF export, or - Include the Google Colab link with edit access or view access enabled.