Seq2Seq

1. Tokenization & Build Dictionary

• input_texts => [Eng_Tokenizer] => input_tokens

• target_texts => [Deu_Tokenizer] => target_tokens

  • Use 2 different tokenizers for the 2 languages.

  • Then build 2 different dictionaries.

• Tokenization can be char-level or word-level

Eng_Tokenizer: "I_am_okay." => ['i', '_', 'a', 'm', ..., 'a', 'y']
Deu_Tokenizer: "Es geht mir gut" => ['e', 's', '_', ..., 'u', 't']

Question: Why 2 different tokenizers and dictionaries?

Answer: In the char-level, languages have different alphabets/chars.

• English: A a, B b, C c …, Z z. (26 letters × 2).

• German: 26 letters, 3 umlauts (Ä,Ö,Ü), and one ligature (ß).

• Greek: Α α, Β β, Γ γ, Δ δ, …, Ω ω. (24 letters × 2).

• Chinese: 金 木 水 火 土 … 赵 钱 孙 李 (a few thousands characters).

Question: Why 2 different tokenizers and dictionaries?

Answer: In the word-level, languages have different vocabulary.

2. One-Hot Encoding

3. Training Seq2Seq Model

4. Inference

Summary

• Encoder’s final states (𝐡 < and 𝐜 < ) have all the information of the English sentence.

• If the sentence is long, the final states have forgotten early inputs.

• Bi-LSTM (left-to-right and right-to-left) has longer memory.

• Use Bi-LSTM in the encoder; use unidirectional LSTM in the decoder.

• Word-level tokenization instead of char-level.

  • The average length of English words is 4.5 letters.

  • The sequences will be 4.5x shorter.

  • Shorter sequence -> less likely to forget.

• But you will need a large dataset!

  • of (frequently used) chars is ~$10^2$ ) -> one-hot suffices.

  • of (frequently used) words is ~$10^4$ -> must use embedding.

  • Embedding Layer has many parameters -> overfitting!