What Do You Get When You Cross Beam Search with Nucleus Sampling?

Uri Shaham, Omer Levy


Abstract
We combine beam search with the probabilistic pruning technique of nucleus sampling to create two deterministic nucleus search algorithms for natural language generation. The first algorithm, p-exact search, locally prunes the next-token distribution and performs an exact search over the remaining space. The second algorithm, dynamic beam search, shrinks and expands the beam size according to the entropy of the candidate’s probability distribution. Despite the probabilistic intuition behind nucleus search, experiments on machine translation and summarization benchmarks show that both algorithms reach the same performance levels as standard beam search.
Anthology ID:
2022.insights-1.5
Volume:
Proceedings of the Third Workshop on Insights from Negative Results in NLP
Month:
May
Year:
2022
Address:
Dublin, Ireland
Editors:
Shabnam Tafreshi, João Sedoc, Anna Rogers, Aleksandr Drozd, Anna Rumshisky, Arjun Akula
Venue:
insights
SIG:
Publisher:
Association for Computational Linguistics
Note:
Pages:
38–45
Language:
URL:
https://aclanthology.org/2022.insights-1.5
DOI:
10.18653/v1/2022.insights-1.5
Bibkey:
Cite (ACL):
Uri Shaham and Omer Levy. 2022. What Do You Get When You Cross Beam Search with Nucleus Sampling?. In Proceedings of the Third Workshop on Insights from Negative Results in NLP, pages 38–45, Dublin, Ireland. Association for Computational Linguistics.
Cite (Informal):
What Do You Get When You Cross Beam Search with Nucleus Sampling? (Shaham & Levy, insights 2022)
Copy Citation:
PDF:
https://preview.aclanthology.org/nschneid-patch-4/2022.insights-1.5.pdf
Video:
 https://preview.aclanthology.org/nschneid-patch-4/2022.insights-1.5.mp4