CS5720 - Week 7
Slide 128 of 140

Bidirectional RNNs

What are Bidirectional RNNs?

Bidirectional RNNs process sequences in both forward and backward directions, combining information from past and future context to make better predictions.
Key Insight:

• Standard RNNs only see past context
• Humans use both past and future context
• Bidirectional RNNs model complete context
• Two separate RNNs process opposite directions
🔄 Two-Way Processing
Like reading a sentence twice - once from left to right, once from right to left - to fully understand its meaning.

Why Bidirectional?

  • 🔮 Future Context Matters
    Many tasks benefit from knowing what comes next, not just what came before
  • 📝 Complete Sentence Understanding
    Word meaning often depends on the entire sentence context
  • 🎯 Ambiguity Resolution
    Future words can clarify the meaning of previous ambiguous words
  • 📈 Significant Performance Gains
    10-30% improvement on many NLP tasks compared to unidirectional RNNs

Unidirectional vs Bidirectional Processing

Standard (Unidirectional) RNN
The
cat
sat
?
Forward Processing Only
➜ ➜ ➜
Output: Based on past context only
Bidirectional RNN
The
cat
sat
down
Forward: ➜ ➜ ➜
Backward: ⬅ ⬅ ⬅
Two-Way Processing
Output: Based on complete context
🧠
Richer Representations
Captures both past and future dependencies in a single model
🎯
Higher Accuracy
Consistently outperforms unidirectional models on sequence labeling
🔍
Context Awareness
Better understanding of word relationships and sentence structure
🔧
Versatile Applications
Excellent for NER, POS tagging, sentiment analysis, and translation
Prepared by Dr. Gorkem Kar