How ayanaussenbayeva/kazru-ner-representation-shift works
This repository is a research pipeline for training and analyzing multilingual Kazakh–Russian NER on public-transport complaints. The principal workflow loads tokenized train/test data, fine-tunes XLM-RoBERTa, evaluates entity spans, then examines representation geometry and train–test shift; analysis outputs include tables and figures. The README also documents an optional stratified re-split. The notebook is the main implementation, but its source was not sampled, so notebook-internal wiring is based on README descriptions; the statistical-control script’s input dependencies are explicit.
Main entry points: ner_representation_analysis.ipynb Training and analysis · make_stratified_split.py Stratified split · representation_shift_checks.py SWD controls · figures Analysis figures
Components 16
Data preparation
- Stratified split
scripts/make_stratified_split.py - Stratified split files
scripts/make_stratified_split.py
NER modeling
- Training and analysis
notebooks/ner_representation_analysis.ipynb - Entity evaluation
notebooks/ner_representation_analysis.ipynb
Representation analysis
- SWD controls
scripts/representation_shift_checks.py - Error geometry
notebooks/ner_representation_analysis.ipynb - Complaint-level shift
notebooks/ner_representation_analysis.ipynb - Entity-level shift
notebooks/ner_representation_analysis.ipynb - Error prediction
notebooks/ner_representation_analysis.ipynb - Entity string overlap
notebooks/ner_representation_analysis.ipynb - Exploratory projections
notebooks/ner_representation_analysis.ipynb
Research outputs
- Analysis figures
figures - SWD results table
scripts/representation_shift_checks.py
Other components
- Researcher
- Complaint data
- Pretrained XLM-R
Connections 17
- Researcher → Training and analysis: runs
- Complaint data → Training and analysis: loads splits
- Complaint data → Stratified split: provides splits
- Researcher → Stratified split: runs optionally
- Stratified split → Stratified split files: writes
- Stratified split files → Training and analysis: supplies splits
- Pretrained XLM-R → Training and analysis: initializes encoder
- Training and analysis → Entity evaluation: evaluates
- Training and analysis → Error geometry: computes features
- Training and analysis → Complaint-level shift: measures shift
- Training and analysis → Entity-level shift: measures shift
- Training and analysis → Error prediction: tests prediction
- Training and analysis → Entity string overlap: checks overlap
- Training and analysis → Exploratory projections: creates projections
- Training and analysis → SWD controls: provides analysis state
- SWD controls → SWD results table: writes results
- Training and analysis → Analysis figures: saves figures