The article presents Approximate Minimum Bayes-Risk (AMBR) decoding, a faster, hyperparameter-free method for MBR decoding. MBR decoding is a powerful alternative for text generation tasks, but its computational complexity makes it infeasible in many situations. AMBR is proposed as a solution, using the Correlated Sequential Halving (CSH) algorithm to compute the sample-based MBR objective. The method was evaluated on machine translation, text summarization, and image captioning tasks, where it performed on par with Confidence-based Pruning (CBP), another method to reduce inference time.
Publication date: 5 Jan 2024
Project Page: https://arxiv.org/abs/2401.02749
Paper: https://arxiv.org/pdf/2401.02749
