DEBERTA: DECODING-ENHANCED BERT WITH DISENTANGLED ATTENTION

TLDR

Recent advances in pre‑trained neural language models have markedly boosted performance across many NLP tasks. This work introduces DeBERTa, a new architecture that enhances BERT and RoBERTa by employing two novel techniques. The first technique is a disentangled attention mechanism that represents each token with separate content and position vectors and computes attention using disentangled matrices, while the second adds an enhanced mask decoder that incorporates absolute positions during pre‑training. The combined techniques yield faster pre‑training and superior downstream results, with a DeBERTa model trained on half the data outperforming RoBERTa‑Large on MNLI (+0.9 %), SQuAD v2.0 (+2.3 %), and RACE (+3.6 %), and the code and models will be released publicly.

Abstract

Recent progress in pre-trained neural language models has significantly improved the performance of many natural language processing (NLP) tasks. In this paper we propose a new model architecture \textbf{DeBERTa} (\textbf{D}ecoding-\textbf{e}nhanced \textbf{BERT} with disentangled \textbf{a}ttention) that improves the BERT and RoBERTa models using two novel techniques. The first is the disentangled attention mechanism, where each word is represented using two vectors that encode its content and position, respectively, and the attention weights among words are computed using disentangled matrices on their contents and relative positions. Second, an enhanced mask decoder is used to incorporate absolute positions in the decoding layer to predict the masked tokens in model pre-training. We show that these two techniques significantly improve the efficiency of model pre-training and performance of downstream tasks. Compared to RoBERTa-Large, a DeBERTa model trained on half of the training data performs consistently better on a wide range of NLP tasks, achieving improvements on MNLI by +0.9% (90.2% vs. 91.1%), on SQuAD v2.0 by +2.3% (88.4% vs. 90.7%) and RACE by +3.6% (83.2% vs. 86.8%). The DeBERTa code and pre-trained models will be made publicly available.