pretraining_morphologizer_oscar

🪐 Weasel Project: Enhancing Morphological Analysis with spaCy Pretraining

This project explores the effectiveness of pretraining techniques on morphological analysis (morphologizer) by conducting experiments on multiple languages. The objective of this project is to demonstrate the benefits of pretraining word vectors using domain-specific data on the performance of the morphological analysis. We leverage the OSCAR dataset to pretrain our vectors for tok2vec and utilize the UD_Treebanks dataset to train a morphologizer component. We evaluate and compare the performance of different pretraining techniques and the performance of models without any pretraining.

📋 project.yml

The project.yml defines the data assets required by the project, as well as the available commands and workflows. For details, see the Weasel documentation.

⏯ Commands

The following commands are defined by the project. They can be executed using weasel run [name]. Commands are only re-run if their inputs have changed.

Command	Description
install_requirements	Download and install all requirements
download_oscar	Download a subset of the oscar dataset
download_model	Download the specified spaCy model for vector-objective pretraining
extract_ud	Extract the ud-treebanks data
convert_ud	Convert the ud-treebanks data to spaCy's format
train	Train a morphologizer component without pretrained weights and static vectors
evaluate	Evaluate the trained morphologizer component without pretrained weights and static vectors
train_static	Train a morphologizer component with static vectors from a pretrained model
evaluate_static	Evaluate the trained morphologizer component with static weights
pretrain_char	Pretrain a tok2vec component with the character objective
train_char	Train a morphologizer component with pretrained weights (character_objective)
evaluate_char	Evaluate the trained morphologizer component with pretrained weights (character-objective)
pretrain_vector	Pretrain a tok2vec component with the vector objective
train_vector	Train a morphologizer component with pretrained weights (vector_objective)
evaluate_vector	Evaluate the trained morphologizer component with pretrained weights (vector-objective)
train_trf	Train a morphologizer component without transformer embeddings
evaluate_trf	Evaluate the trained morphologizer component with transformer embeddings
evaluate_metrics	Evaluate all experiments and create a summary json file
reset_project	Reset the project to its original state and delete all training process
reset_training	Reset the training progress
reset_metrics	Delete the metrics folder

⏭ Workflows

The following workflows are defined by the project. They can be executed using weasel run [name] and will run the specified commands in order. Commands are only re-run if their inputs have changed.

Workflow	Steps
data	download_oscar → download_model → extract_ud → convert_ud
training	train → evaluate
training_static	train_static → evaluate_static
training_char	pretrain_char → train_char → evaluate_char
training_vector	pretrain_vector → train_vector → evaluate_vector
training_trf	train_trf → evaluate_trf

🗂 Assets

The following assets are defined by the project. They can be fetched by running weasel assets in the project directory.

File	Source	Description
assets/ud-treebanks-v2.5.tgz	URL