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.
|Title of host publication||Insights 2022 - 3rd Workshop on Insights from Negative Results in NLP, Proceedings of the Workshop|
|Editors||Shabnam Tafreshi, Joao Sedoc, Anna Rogers, Aleksandr Drozd, Anna Rumshisky, Arjun Reddy Akula|
|Publisher||Association for Computational Linguistics (ACL)|
|Number of pages||8|
|State||Published - 2022|
|Event||3rd Workshop on Insights from Negative Results in NLP, Insights 2022 - Dublin, Ireland|
Duration: 26 May 2022 → …
|Name||Insights 2022 - 3rd Workshop on Insights from Negative Results in NLP, Proceedings of the Workshop|
|Conference||3rd Workshop on Insights from Negative Results in NLP, Insights 2022|
|Period||26/05/22 → …|
Bibliographical noteFunding Information:
This work was supported by the Tel Aviv University Data Science Center, the Blavatnik Fund, the Alon Scholarship, and Intel Corporation. We would like to thank Ari Holtzman, Jonathan Berant, Ori Yoran, Lior Vassertail, and Yuval Kirstain for their valuable feedback.
© 2022 Association for Computational Linguistics.