We gratefully acknowledge support from
the Simons Foundation and member institutions.
Full-text links:


Current browse context:


Change to browse by:

References & Citations

DBLP - CS Bibliography


(what is this?)
CiteULike logo BibSonomy logo Mendeley logo del.icio.us logo Digg logo Reddit logo ScienceWISE logo

Computer Science > Computation and Language

Title: DeBERTaV3: Improving DeBERTa using ELECTRA-Style Pre-Training with Gradient-Disentangled Embedding Sharing

Abstract: This paper presents a new pre-trained language model, DeBERTaV3, which improves the original DeBERTa model by replacing mask language modeling (MLM) with replaced token detection (RTD), a more sample-efficient pre-training task. Our analysis shows that vanilla embedding sharing in ELECTRA hurts training efficiency and model performance. This is because the training losses of the discriminator and the generator pull token embeddings in different directions, creating the "tug-of-war" dynamics. We thus propose a new gradient-disentangled embedding sharing method that avoids the tug-of-war dynamics, improving both training efficiency and the quality of the pre-trained model. We have pre-trained DeBERTaV3 using the same settings as DeBERTa to demonstrate its exceptional performance on a wide range of downstream natural language understanding (NLU) tasks. Taking the GLUE benchmark with eight tasks as an example, the DeBERTaV3 Large model achieves a 91.37% average score, which is 1.37% over DeBERTa and 1.91% over ELECTRA, setting a new state-of-the-art (SOTA) among the models with a similar structure. Furthermore, we have pre-trained a multi-lingual model mDeBERTa and observed a larger improvement over strong baselines compared to English models. For example, the mDeBERTa Base achieves a 79.8% zero-shot cross-lingual accuracy on XNLI and a 3.6% improvement over XLM-R Base, creating a new SOTA on this benchmark. We have made our pre-trained models and inference code publicly available at this https URL
Comments: 16 pages, 10 tables, 2 Figures. The DeBERTaV3 model significantly improves performance of the downstream NLU tasks over models with a similar structure, e.g. DeBERTaV3 large achieves 91.37% average GLUE score which is 1.37% over DeBERTa large. XSmall has only 22M backbone parameters, but significantly outperforms RoBERTa/XLNet-base
Subjects: Computation and Language (cs.CL); Machine Learning (cs.LG)
MSC classes: cs.CL, cs.GL
ACM classes: I.2; I.7
Cite as: arXiv:2111.09543 [cs.CL]
  (or arXiv:2111.09543v2 [cs.CL] for this version)

Submission history

From: Pengcheng He [view email]
[v1] Thu, 18 Nov 2021 06:48:00 GMT (432kb,D)
[v2] Wed, 8 Dec 2021 22:07:23 GMT (413kb,D)

Link back to: arXiv, form interface, contact.