This repository implements a Neural Conditional Random Field (CRF) tagger for Named Entity Recognition (NER), developed as part of a CSE project submission for 291E — Advanced Statistical NLP at UC San Diego.
It reproduces the Viterbi decoding algorithm used for structured sequence inference in CRF-based models, integrated with BiLSTM encoders for contextual representation learning.
The core focus is the implementation and comparison of a baseline BiLSTM tagger and a BiLSTM+CRF model for sequence labeling on standard NER datasets.
Example: Viterbi decoding visualized across state transitions in CRF inference.
| Component | Description |
|---|---|
| Goal | Compare BiLSTM and BiLSTM+CRF models for Named Entity Recognition |
| Core Algorithm | Viterbi decoding for globally optimal sequence prediction |
| Model | BiLSTM encoder + CRF layer |
| Dataset | CoNLL-2003 Named Entity Recognition corpus |
| Framework | PyTorch |
| Performance | BiLSTM F1 ≈ 72.46 • BiLSTM+CRF F1 ≈ 73.50 |
Neural-CRF_NER-Tagger/
│
├── cse291_assignment2_starter_code.ipynb # Baseline BiLSTM implementation
├── v_2.ipynb # BiLSTM + CRF with Viterbi decoding
├── Project2.pdf # Full project report (CSE 291E submission)
├── metadata.yml # Submission metadata
├── resources/ # Algorithmic references and visuals
│ ├── Viterbi algorithm - Wikipedia.pdf
│ ├── viterbi-algorithm.png
│ ├── viterbi-pseudocode.png
│ └── viterbi-visualizer.jpg
└── Project2_assessment.zip.gitignore
# Clone the repository
git clone https://github.com/keerthanap8898/Neural-CRF_NER-Tagger.git
cd Neural-CRF_NER-Tagger
# Create environment (Python 3.9+ recommended)
python -m venv venv
source venv/bin/activate # or venv\Scripts\activate on Windows
# Install dependencies
pip install -r requirements.txtTrain and evaluate directly in Google Colab or locally: jupyter notebook v_2.ipynb
The notebook contains:
- Implementation of BiLSTM Tagger
- Implementation of BiLSTM + CRF model with Viterbi decoding
- Comparative analysis, training curves, and F1 evaluation
The Viterbi Algorithm computes the most probable sequence of labels (tags) given emission and transition scores from the CRF layer.
At each time step, it maintains dynamic programming states that recursively compute the maximum path probability and backtrack pointers to recover the optimal label sequence.
Dynamic programming flow for the Viterbi decoding process.
Canonical explanation of the dynamic programming approach for maximum-likelihood sequence estimation.
- vocal.com/echo-cancellation/viterbi-algorithm-...-hmm
- frazierbutler.medium.com/intro-to-the-viterbi-algorithm...cf3
Initialization, recursion, and termination stages of the Viterbi decoding process.
Model F1 Score Observations
BiLSTM Tagger 72.46 Accuracy drops after epoch 10; mild overfitting after epoch 3
BiLSTM + CRF 73.50 Stable learning after epoch 7; improved consistency via transition modeling
These results demonstrate the added stability and grammatical awareness of CRF-based decoding compared to token-independent LSTM tagging.
- Viterbi Algorithm — Wikipedia
- Huang, Zhiheng et al. Bidirectional LSTM-CRF Models for Sequence Tagging. arXiv:1508.01991 (2015)
- Lafferty, McCallum, & Pereira (2001) — Conditional Random Fields: Probabilistic Models for Segmenting and Labeling Sequence Data.
- Lample, Ballesteros, Kawakami, & Dyer (2016) — Neural Architectures for Named Entity Recognition.
- UC San Diego — CSE 291E: Advanced Statistical NLP.
Keerthana Purushotham Graduate Coursework — Advanced Statistical NLP (CSE 291E) LinkedIn : linkedin.com/in/keerthanapurushotham
This repository is released under the MIT License. All figures and documents are provided for educational and academic use.
This project demonstrates:
- Integration of neural sequence encoders with structured probabilistic decoders
- Application of dynamic programming via Viterbi for sequence tagging
- Comparative evaluation between neural-only and neural-structured approaches
It serves as a clear, reproducible reference for understanding how BiLSTM-CRF architectures achieve global sequence consistency and correctness in NLP tasks.
“Structured learning meets sequence modeling — decoding meaning through structure.”
Copyright ⓒ 2025 Keerthana Purushotham <keerthanap0808@gmail.com>, <kpurusho@ucsd.edu>, <keep.consult@proton.me>.
Licensed under MIT. See LICENSE for details.