How shivaat/mtlb-struct works
MTLB-STRUCT trains and runs a BERT-based system for verbal multiword-expression tagging, optionally learning dependency parses jointly. The main workflow prepares PARSEME data, encodes examples, trains or loads a model, evaluates predictions, and converts token tags back to .cupt output. The README also documents configuration-driven training and saved-model testing. Some connections are based only on README statements because the testing entry point and several implementation details were not sampled.
Main entry points: main.py Training entry point · model.py Joint tagger · load_test.py Saved-model testing · berteval.py Evaluation loop
Components 14
Corpus preparation
- PARSEME data
data/1.2/GA/train.cupt - Corpus structures
code/corpus.py - Corpus reader
code/corpus_reader.py - Example encoding
code/preprocessing.py
Training and inference
- Run configurations
code/configs/config.json - Training entry point
code/main.py - Saved-model testing
code/load_test.py
Model and evaluation
- Pretrained BERT
- Joint tagger
code/model.py - Evaluation loop
code/berteval.py - Tagging metrics
code/berteval.py
Prediction output
- CUpt conversion
code/evaluation.py - Prediction files
Other components
- Researcher
Connections 16
- Researcher → Training entry point: starts training
README.md - Run configurations → Training entry point: configures run
README.md - PARSEME data → Corpus structures: reads corpus
code/corpus.py - Corpus structures → Corpus reader: provides sentences
code/corpus_reader.py - Corpus reader → Example encoding: provides tagged sequences
code/preprocessing.py - Example encoding → Training entry point: prepares examples
code/main.py - Training entry point → Joint tagger: trains model
code/main.py - Pretrained BERT → Joint tagger: provides encoder
code/model.py - Training entry point → Evaluation loop: runs evaluation
code/main.py - Evaluation loop → Joint tagger: scores batches
code/berteval.py - Evaluation loop → Tagging metrics: computes metrics
code/berteval.py - Training entry point → CUpt conversion: converts predictions
code/main.py - CUpt conversion → Prediction files: writes CUpt
code/evaluation.py - Researcher → Saved-model testing: starts testing
README.md - Saved-model testing → Joint tagger: loads saved model
- Saved-model testing → Prediction files: saves predictions
README.md