POS Tagging Case Study

This repository is organized for submission in a notebook-first format.

Main notebooks for Moodle submission

1. notebooks/01_data_and_benchmark.ipynb Data-focused notebook for loading the UD benchmark, checking split statistics, inspecting labels, and verifying the dataset before training.

2. notebooks/02_averaged_perceptron.ipynb Experiment notebook for the averaged perceptron baseline.

3. notebooks/03_hmm_viterbi.ipynb Experiment notebook for the HMM + Viterbi baseline.

4. notebooks/04_crf.ipynb Experiment notebook for the CRF model.

5. notebooks/05_compact_bilstm.ipynb Experiment notebook for the compact BiLSTM model.

Repository structure

• notebooks/ Primary experiment narrative and implementation used for the submitted case study.
• data/ Raw benchmark data used by the notebooks.

Dataset setup

The notebooks expect the raw UD English EWT files in data/raw/:

• data/raw/en_ewt-ud-train.conllu
• data/raw/en_ewt-ud-dev.conllu
• data/raw/en_ewt-ud-test.conllu

The repository does not bundle these raw files. Download them from the Universal Dependencies repository and place them into data/raw/ before running the notebooks.

Dataset repository:

• https://github.com/UniversalDependencies/UD_English-EWT

Example:

mkdir -p data/raw
curl -L https://raw.githubusercontent.com/UniversalDependencies/UD_English-EWT/master/en_ewt-ud-train.conllu -o data/raw/en_ewt-ud-train.conllu
curl -L https://raw.githubusercontent.com/UniversalDependencies/UD_English-EWT/master/en_ewt-ud-dev.conllu -o data/raw/en_ewt-ud-dev.conllu
curl -L https://raw.githubusercontent.com/UniversalDependencies/UD_English-EWT/master/en_ewt-ud-test.conllu -o data/raw/en_ewt-ud-test.conllu

Recommended review order

1. Open notebooks/01_data_and_benchmark.ipynb
2. Open the four model notebooks in order from 02 to 05
3. Open FINAL_REPORT.md for the final written conclusions

Notes

• The notebooks are intentionally structured, commented, and separated by purpose.
• The main implementation is intentionally embedded directly into the notebooks so the repository matches the expected notebook-centric submission format.
• Generated experiment artifacts are intended to appear in notebook cell outputs rather than be committed as separate result files.