When to Finish? Optimal Beam Search for Neural Text Generation (modulo beam size)

TL;DR

This paper introduces a provably optimal beam search algorithm for neural text generation that guarantees the best hypothesis is found efficiently, improving translation quality.

Contribution

It presents a new beam search method with proven optimality and a length reward mechanism to handle shorter hypotheses, advancing neural text generation techniques.

Findings

Improved BLEU scores in neural machine translation

Algorithm guarantees optimal hypothesis completion

Handles length bias with bounded length reward

Abstract

In neural text generation such as neural machine translation, summarization, and image captioning, beam search is widely used to improve the output text quality. However, in the neural generation setting, hypotheses can finish in different steps, which makes it difficult to decide when to end beam search to ensure optimality. We propose a provably optimal beam search algorithm that will always return the optimal-score complete hypothesis (modulo beam size), and finish as soon as the optimality is established (finishing no later than the baseline). To counter neural generation's tendency for shorter hypotheses, we also introduce a bounded length reward mechanism which allows a modified version of our beam search algorithm to remain optimal. Experiments on neural machine translation demonstrate that our principled beam search algorithm leads to improvement in BLEU score over previously proposed alternatives.