How nancy123a/ldsl-winter-school works
Project type and purpose
• This is an NLP/ML research project implemented as a training and evaluation pipeline for a convolutional neural network (CNN) that performs modal sense classification.
• Key indicators: a single Jupyter notebook (MSCtask_MASC_EN.ipynb), a README describing the CNN architecture and training details, and directories for logging.
File structure
Top-level items:
• MSCtask_MASC_EN.ipynb – the main notebook containing data loading, preprocessing, model definition, training loops, and evaluation.
• README.md – high-level description of the task, architecture, loss function, and dataset.
• log/ – directory for runtime logs (e.g., training loss, accuracy per epoch).
• log_file – possibly a flat file for serialized log outputs.
• yolo/ – a folder whose contents are unclear from the tree, but likely contains utility scripts or data submodules (treat it as an external or auxiliary component).
README
• Confirms the use of a CNN based on Germies (2020) for three-way classification: dynamic, epistemic, deontic.
• Indicates categorical cross-entropy loss and a custom dataset with uniform labels.
Components
a. Data Sources
• Raw dataset files (MASC_EN data)
b. Preprocessing Module
• Tokenization/embedding lookup
• Input formatting for CNN (e.g., sliding windows or sentence matrices)
c. Model Definition
• Convolutional layers, pooling layers, fully connected layers
• Loss function (categorical cross-entropy)
d. Training Pipeline
• Jupyter notebook orchestration
• Optimizer loop, checkpointing
e. Evaluation Module
• Test dataset evaluation
• Metrics calculation (accuracy, precision, recall)
f. Logging & Artifacts
• log/ directory for epoch logs
• log_file for consolidated run outputs
g. Auxiliary/External Components (yolo/)
• Any helper scripts, custom utility functions, or external data loaders
Relationships and interactions
• Model Definition ↔ Training Pipeline (forward pass, backpropagation)
• Training Pipeline → Logging & Artifacts (writes loss, accuracy per epoch)
• Trained Model → Evaluation Module → Logging & Artifacts (writes final metrics)
• Preprocessing Module ↔ Auxiliary Components (yolo/) if helpers are used
Architectural patterns & principles
• Linear pipeline (data→ preprocess→ train→ evaluate)
• Separation of concerns: keep preprocessing, model code, training orchestration, and logging in discrete modules
• Reproducibility: notebook plus logging directories for hyperparameter tracking and results
Detailed capture of architectural elements
• CNN block with layer types and dimensions (e.g., “Conv1D (kernel=3) → ReLU → MaxPool”)
• Checkpoints or model-save points in the Training Pipeline
• Categorical cross-entropy next to the loss node