DeBERTa: Decoding-enhanced BERT with Disentangled Attention

By Pengcheng He, Xiaodong Liu, Jianfeng Gao, Weizhu Chen. 2020

3,946 citations recorded by the source corpus.

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 DeBERTa (Decoding-enhanced BERT with disentangled attention) 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 replace the output softmax layer to predict the masked tokens for model pretraining. 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 at this https URL.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition, Large Language Models, Model Architectures

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Computer Vision and Pattern Recognition

In Large Language Models

In Model Architectures

Citations in the corpus

Llama 2: Open Foundation and Fine-Tuned Chat Models (2023, methodology): “Recent literature has extensively explored the risks and challenges linked with Large Language Models.”
Reasoning Like Program Executors (2022, methodology): “Pre-training on self-supervised tasks, such as masked language modeling (Devlin et al., 2019; He et al., 2021) using large amounts of NL sentences, boosts the language understanding of models by a large margin (Wang et al., 2018a).”
Evaluating Large Language Models Trained on Code (2021, background): “A common approach to adjusting the behavior of Trans-formers is to ﬁne-tune large pre-trained models with curated or human-generated datasets of the desired behavior (e.g., Raffel et al. (2020); He et al. (2020)).”
Measuring Mathematical Problem Solving With the MATH Dataset (2021, methodology): “Even difﬁcult logical understanding tasks such as LogiQA (Liu et al., 2020) will soon be straightforwardly solved by enormous Transformers should trends continue, which we also show in the Supplementary Materials.”
Unnatural Instructions: Tuning Language Models with (Almost) No Human Labor (2022, positive): “6Using the deberta-xlarge-mnli model (He et al., 2021).”
Adversarial Training for High-Stakes Reliability (2022, background): “The candidate tokens were plausible “fill-in-the-blank” suggestions from deberta-v2-xlarge [52].”
Generating Training Data with Language Models: Towards Zero-Shot Language Understanding (2022, background): “Pretrained language models (PLMs) [5, 8, 11, 19, 34, 40, 41] have achieved human-level performance on natural language understanding (NLU) tasks [66, 67] when fine-tuned on a large amount of task-specific training data.”
WARP: Word-level Adversarial ReProgramming (2021, methodology): “Besides our best WARP models, we also include the human baselines, current state-of-the-art model (He et al., 2020), the regular ﬁne-tuned pretrained model we use, and also include relatively small language models, including (Jiao et al., 2020), (Clark et al., 2020), (Houlsby et al., 2019).”
M6: A Chinese Multimodal Pretrainer (2021, background): “Pretraining has become a focus in the research in natural language processing (NLP) [1, 2, 7, 16, 18, 19, 27, 31, 37, 44, 49].”
P-Tuning v2: Prompt Tuning Can Be Comparable to Fine-tuning Universally Across Scales and Tasks (2021, positive): “We include BERT-large (Devlin et al., 2018), RoBERTa-large (Liu et al., 2019), DeBERTa-xlarge (He et al., 2020), GLMxlarge/xxlarge (Du et al., 2021) for evaluation.”
OFA: Unifying Architectures, Tasks, and Modalities Through a Simple Sequence-to-Sequence Learning Framework (2022, methodology): “In comparison with the state-of-the-art natural language pretrained models, including RoBERTa [28], XLNET [25], ELECTRA [82], and DeBERTa [83], OFA reaches a comparable performance.”
What Will it Take to Fix Benchmarking in Natural Language Understanding? (2021, background): “…test set in order to reveal test set results for any other target task, a large majority of papers that report test set numbers omit this metric and decline to report potentially unﬂattering bias numbers (Raffel et al., 2020; Pruksachatkun et al., 2020; Schick and Schütze, 2020; He et al., 2020).”
QuALITY: Question Answering with Long Input Texts, Yes! (2021, methodology): “We test Longformer, RoBERTa, DeBERTaV3, and T5 models, using as much of the full source text as possible.”
GLM: General Language Model Pretraining with Autoregressive Blank Infilling (2021, methodology): “T5 (Raffel et al., 2020) uniﬁes NLU and conditional generation via encoder-decoder models but requires more parameters to match the performance of BRET-based models such as RoBERTa (Liu et al., 2019) and DeBERTa (He et al., 2021).”