Beam Search and BLEU score

<Beam Search>

- Greedy decoding has no way to undo decisions!

- Exhaustive search : This O(𝑉𝑡)complexity is far too expensive!

 

<Beam search>

- on each time step of the decoder, we keep track of the 𝑘 most probable partial translations (which we call hypothese)

- k : beam size (in practice around 5 to 10)

 

- longer hypotheses have lower scores

- 그러므로 Normalize가 필요함

 

<BLEU score>

- model 2는 100%가 나왔지만 문법적으로 전혀 말이 안 됨. 이 부분을 보완하자!

 

- BLUE score : precision를 계산, N-gram overlap(연속된 단위로 비교하기)