Low-Rank Softmax Can Have Unargmaxable Classes in Theory but Rarely in Practice

Andreas Grivas, Nikolay Bogoychev, Adam Lopez

Research output: Chapter in Book/Report/Conference proceedingConference contribution

Abstract / Description of output

Classifiers in natural language processing (NLP) often have a large number of output classes. For example, neural language models (LMs) and machine translation (MT) models both predict tokens from a vocabulary of thousands. The Softmax output layer of these models typically receives as input a dense feature representation, which has much lower dimensionality than the output. In theory, the result is some words may be impossible to be predicted via argmax, irrespective of input features, and empirically, there is evidence this happens in small language models (Demeter et al., 2020). In this paper we ask whether it can happen in practical large language models and translation models. To do so, we develop algorithms to detect such unargmaxable tokens in public models. We find that 13 out of 150 models do indeed have such tokens; however, they are very infrequent and unlikely to impact model quality. We release our algorithms and code to the public.
Original languageEnglish
Title of host publicationProceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)
EditorsSmaranda Muresan, Preslav Nakov, Aline Villavicencio
Place of PublicationDublin, Ireland
PublisherAssociation for Computational Linguistics
Pages6738-6758
Number of pages21
DOIs
Publication statusPublished - 1 May 2022
Event60th Annual Meeting of the Association for Computational Linguistics - The Convention Centre Dublin, Dublin, Ireland
Duration: 22 May 202227 May 2022
https://www.2022.aclweb.org

Conference

Conference60th Annual Meeting of the Association for Computational Linguistics
Abbreviated titleACL 2022
Country/TerritoryIreland
CityDublin
Period22/05/2227/05/22
Internet address

Fingerprint

Dive into the research topics of 'Low-Rank Softmax Can Have Unargmaxable Classes in Theory but Rarely in Practice'. Together they form a unique fingerprint.

Cite this