Optimizing the Performance of Text Classification Models by Improving the Isotropy of the Embeddings using a Joint Loss Function

Joseph Attieh*, Abraham Zewoudie, Vladimir Vlassov, Adrian Flanagan, Tom Bäckström

*Corresponding author for this work

Research output: Chapter in Book/Report/Conference proceedingConference article in proceedingsScientificpeer-review

2 Downloads (Pure)


Recent studies show that the spatial distribution of the sen-
tence representations generated from pre-trained language models is
highly anisotropic. This results in a degradation in the performance of
the models on the downstream task. Most methods improve the isotropy
of the sentence embeddings by refining the corresponding contextual
word representations, then deriving the sentence embeddings from these
refined representations. In this study, we propose to improve the quality
of the sentence embeddings extracted from the [CLS] token of the pre-
trained language models by improving the isotropy of the embeddings.
We add one feed-forward layer between the model and the downstream
task layers, and we train it using a novel joint loss function. The pro-
posed approach results in embeddings with better isotropy, that gener-
alize better on the downstream task. Experimental results on 3 GLUE
datasets with classification as the downstream task show that our pro-
posed method is on par with the state-of-the-art, as it achieves perfor-
mance gains of around 2–3% on the downstream tasks compared to the
Original languageEnglish
Title of host publicationDocument Analysis and Recognition – ICDAR 2023 - 17th International Conference, Proceedings
EditorsGernot A. Fink, Rajiv Jain, Koichi Kise, Richard Zanibbi
Number of pages16
ISBN (Electronic)978-3-031-41734-4
ISBN (Print)978-3-031-41733-7
Publication statusPublished - 19 Aug 2023
MoE publication typeA4 Conference publication
EventInternational Conference on Document Analysis and Recognition - San Jose, United States
Duration: 21 Aug 202326 Aug 2023
Conference number: 17

Publication series

NameLecture notes in computer science
ISSN (Print)0302-9743
ISSN (Electronic)1611-3349


ConferenceInternational Conference on Document Analysis and Recognition
Abbreviated titleICDAR
Country/TerritoryUnited States
CitySan Jose


Dive into the research topics of 'Optimizing the Performance of Text Classification Models by Improving the Isotropy of the Embeddings using a Joint Loss Function'. Together they form a unique fingerprint.

Cite this