Relation Classification Challenge - Solution

Overview

This solution implements a BERT-based relation classification model using the specified architecture: google/bert_uncased_L-4_H-256_A-4.

Dataset Information

• Training samples: 1,919
• Public test samples: 480
• Private test samples: 407
• Number of relation labels: 18

Relation Labels

1. Cause-Effect(e1,e2) / Cause-Effect(e2,e1)
2. Component-Whole(e1,e2) / Component-Whole(e2,e1)
3. Content-Container(e1,e2) / Content-Container(e2,e1)
4. Entity-Destination(e1,e2)
5. Entity-Origin(e1,e2) / Entity-Origin(e2,e1)
6. Instrument-Agency(e1,e2) / Instrument-Agency(e2,e1)
7. Member-Collection(e1,e2) / Member-Collection(e2,e1)
8. Message-Topic(e1,e2) / Message-Topic(e2,e1)
9. Product-Producer(e1,e2) / Product-Producer(e2,e1)
10. Other

Solution Architecture

Model

• Base Model: google/bert_uncased_L-4_H-256_A-4 (Small BERT)
• Classification Head: Linear layer on top of BERT's [CLS] token
• Special Tokens: Added [E1], [/E1], [E2], [/E2] to mark entity boundaries

Key Features

1. Entity Marking: Replaces <e1>, </e1>, <e2>, </e2> with special tokens
2. Data Split: 85% train, 15% validation (stratified by label)
3. Evaluation Metric: Macro F1 score across all 18 classes
4. Optimization: AdamW optimizer with learning rate 2e-5

Hyperparameters

• Max sequence length: 128
• Batch size: 16
• Epochs: 10
• Learning rate: 2e-5
• Dropout: 0.3
• Validation split: 15%

Usage

1. Install Dependencies

pip install -r requirements.txt

2. Train Model and Generate Predictions

python relation_classifier.py

This will:

• Train the model for 10 epochs
• Save the best model based on validation F1 score
• Generate predictions for both public and private test sets
• Create submission files:
  • public_test_submission.csv
  • private_test_submission.csv

Output Files

• best_model.pt: Best model checkpoint
• public_test_submission.csv: Public test predictions
• private_test_submission.csv: Private test predictions

Model Performance

The model is evaluated using Macro-averaged F1 score across all 18 relation classes, which ensures balanced performance across both frequent and rare relation types.

Implementation Details

Data Preprocessing

• Sentences are tokenized using BERT tokenizer
• Entity tags are replaced with special tokens for better entity representation
• Sequences are padded/truncated to max length of 128 tokens

Training Strategy

• Cross-entropy loss for multi-class classification
• Early stopping based on validation F1 score
• Progress bars for training visibility

Prediction

• Batch prediction for efficiency
• Argmax over logits for final label prediction
• Label mapping back to original relation names