Google

Minimum Bayes Risk (MBR) decoding is an alternative to Maximum a Posteriori (MAP) decoding that modifies the search objective. Instead of selecting the single most probable output, MBR decoding aims to choose an output that minimizes the expected risk when evaluated against a distribution of possible correct outputs.

MBR Decoding as an Alternative to MAP Decoding

A company is developing a machine translation service for its public-facing website. They have two decoding algorithms to choose from for generating the final translation of a user's input.

Choosing a Decoding Strategy for a Translation Service

Imagine a text generation system has produced several potential output sentences. One decoding strategy selects the single sentence with the highest calculated probability. An alternative strategy selects the sentence that is, on average, most similar to all other high-probability sentences. Contrast these two strategies. Which one is more likely to avoid unusual or outlier generations, and why?

Contrasting Decoding Objectives

A language model is generating a summary of a document. It produces a set of candidate summaries. Candidate A has the highest individual probability (30%) but is an outlier, sharing very little phrasing with other candidates. Candidates B, C, and D have lower individual probabilities (20% each) but are all very similar to one another, forming a clear consensus. Which of the following statements most accurately predicts the behavior of two different decoding strategies?

Learn Before

Related