CS5720 - Week 7
Slide 137 of 140

Sequence-to-Sequence Models

Seq2Seq Architecture

Seq2Seq models transform one sequence into another sequence of potentially different length
Key Components:

• Encoder - Processes input sequence
• Context Vector - Fixed-size representation
• Decoder - Generates output sequence
• End-to-end training - Learns mappings
⚠️ Information Bottleneck
Entire input compressed into single vector

Applications

🌍 Machine Translation
English → French, German, etc.
📝 Text Summarization
Long article → Short summary
❓ Question Answering
Context + Question → Answer
💬 Conversational AI
User input → Bot response

Encoder-Decoder Architecture

Encoder
Hello
LSTM
World
LSTM
→
Context
Vector
→
Decoder
LSTM
Bonjour
LSTM
Monde
Example: English "Hello World" → French "Bonjour Monde"
← Previous Next →
Prepared by Dr. Gorkem Kar