What Do You Get When You Cross Beam Search with Nucleus Sampling?
This work addresses the challenge of enhancing text generation efficiency for NLP applications, but it is incremental as it does not surpass existing methods.
The authors tackled the problem of improving natural language generation by combining beam search with nucleus sampling, resulting in two deterministic algorithms that achieved performance levels equivalent to standard beam search on machine translation and summarization benchmarks.
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.