When doing beam search in sequence to sequence models, one explores next words in order of their likelihood. However, during decoding, there may be other constraints we have or objectives we wish to maximize. For example, sequence length, BLEU score, or mutual information between the target and source sentences. In order to accommodate these additional desiderata, the authors add an additional term Q onto the likelihood capturing the appropriate criterion and then choose words based on this combined objective.
After an inspiring day of speakers, panels, and conversations, Lilt Ascend 2020 has come and gone. First and foremost, thank you to everyone who attended our first-ever all digital conference - we couldn’t have done it without your support! We also want to send a thank you to all of our incredible speakers: Loïc Dufresne de Virel, Tom Davenport, Alessandra Binazzi, Kyunghyun Cho, Paul Buckley, Shaun Johnson, Paula Shannon, John DeNero, and Spence Green.