Named Entity Recognition

Overview

This project implements and compares BiLSTM-based models for Named Entity Recognition (NER), trained on the CoNLL-2003 dataset. The work demonstrates comprehensive preprocessing, model design, training strategies, and performance evaluation for identifying named entities in text.

Repository

• Technologies Used: PyTorch, Python, NumPy, Pandas, Scikit-learn, GloVe Embeddings
• Dataset: CoNLL-2003 Shared Task Dataset
• GitHub Repository: NagaHarshita/NamedEntityRecognition

Dataset & Preprocessing

The project utilizes the CoNLL-2003 dataset, a standard benchmark for Named Entity Recognition tasks. The dataset preprocessing includes:

• Data Split: train, dev, and test sets
• Cleaning: Removal of empty lines and normalization
• Tokenization: Converting words and tags to indexed sequences
• Entity Types: Person (PER), Organization (ORG), Location (LOC), and Miscellaneous (MISC)

Model Architecture

BiLSTM Base Model

BLSTM(
    embedding: Embedding(23700, 100)
    lstm: LSTM(100, 256, dropout=0.33, bidirectional=True)
    linear: Linear(512 → 128)
    elu: ELU(alpha=1.0)
    classifier: Linear(128 → 10)
)

BiLSTM with GloVe Embeddings

BLSTM_Glove(
    embedding: Embedding(400000, 100)
    lstm: LSTM(100, 256, dropout=0.33, bidirectional=True)
    linear: Linear(512 → 128)
    elu: ELU(alpha=1.0)
    classifier: Linear(128 → 9)
)

Training Configuration

The models were trained with carefully tuned hyperparameters:

• Optimizer: SGD with learning rate 0.8
• Learning Rate Scheduler: Dynamic adjustment during training
• Epochs: 100
• Batch Size: 32
• Loss Function: Weighted cross-entropy to handle class imbalance
• Padding Strategy: ignore_index = -1 for padded values

Results & Performance

Key Insights

Performance Analysis:

• Both models achieve high accuracy (>93%) on token-level classification
• Location entities show the best F1 scores (~80%), indicating clearer patterns
• Person and Organization entities present greater challenges due to contextual complexity
• Class imbalance significantly impacts performance, especially for PER and ORG entities

Technical Learnings:

• GloVe embeddings provide rich semantic representations but require careful tuning
• Proper padding and batch handling are crucial for consistent training
• Weighted loss functions effectively address class imbalance issues
• Bidirectional context significantly improves entity boundary detection