Length Normalization

What is length normalization in the context of beam search?

Intermédiaire

Apprentissage automatique


Length normalization is a technique used in beam search or other sequence generation algorithms to address biases towards shorter or longer sequences. It aims to ensure fair evaluation and ranking of sequences of different lengths, especially when using probability-based scoring methods.

In the context of beam search:

Problem Addressed

How Length Normalization Works

Example

Purpose and Impact

Importance in Sequence Generation

In essence, length normalization in beam search adjusts the scores of candidate sequences based on their lengths to ensure a fair comparison and ranking, promoting a more balanced exploration of sequences of different lengths.