ConvBERT: Improving BERT with Span-based Dynamic Convolution

Abstract

Pre-trained language models like BERT and its variants have recently achieved impressive performance in various natural language understanding tasks. However, BERT heavily relies on the global self-attention block and thus suffers large memory footprint and computation cost. Although all its attention heads query on the whole input sequence for generating the attention map from a global perspective, we observe some heads only need to learn local dependencies, which means the existence of computation redundancy. We therefore propose a novel span-based dynamic convolution to replace these self-attention heads to directly model local dependencies. The novel convolution heads, together with the rest self-attention heads, form a new mixed attention block that is more efficient at both global and local context learning. We equip BERT with this mixed attention design and build a ConvBERT model. Experiments have shown that ConvBERT significantly outperforms BERT and its variants in various downstream tasks, with lower training cost and fewer model parameters. Remarkably, ConvBERTbase model achieves 86.4 GLUE score, 0.7 higher than ELECTRAbase, while using less than 1/4 training cost. Code and pre-trained models will be released.

arXiv comment17 pages

References in corpus (16)

• Distributed Representations of Words and Phrases and their Compositionality
• Convolutional Sequence to Sequence Learning
• Pre-trained Models for Natural Language Processing: A Survey
• ALBERT: A Lite BERT for Self-supervised Learning of Language Representations
• A Convolutional Neural Network for Modelling Sentences
• Semi-supervised Sequence Learning
• Q8BERT: Quantized 8Bit BERT
• A Simple Method for Commonsense Reasoning
• Reducing Transformer Depth on Demand with Structured Dropout
• Depthwise Separable Convolutions for Neural Machine Translation
• The Evolved Transformer
• UNITER: UNiversal Image-TExt Representation Learning
• Lite Transformer with Long-Short Range Attention
• MobileBERT: a Compact Task-Agnostic BERT for Resource-Limited Devices
• Q-BERT: Hessian Based Ultra Low Precision Quantization of BERT
• SpeechBERT: An Audio-and-text Jointly Learned Language Model for End-to-end Spoken Question Answering

Cited by in corpus (15)

• A Survey on Visual Transformer
• DeepViT: Towards Deeper Vision Transformer
• CANINE: Pre-training an Efficient Tokenization-Free Encoder for Language Representation
• Explainable Identification of Dementia from Transcripts using Transformer Networks
• M6: A Chinese Multimodal Pretrainer
• Deciphering the Language of Nature: A transformer-based language model for deleterious mutations in proteins
• A Comparative Study of Transformer-Based Language Models on Extractive Question Answering
• ChatGPT v.s. Media Bias: A Comparative Study of GPT-3.5 and Fine-tuned Language Models
• DaCy: A Unified Framework for Danish NLP
• Analysis of the Evolution of Advanced Transformer-Based Language Models: Experiments on Opinion Mining
• Sampling Equivariant Self-attention Networks for Object Detection in Aerial Images
• Deep Insights into Cognitive Decline: A Survey of Leveraging Non-Intrusive Modalities with Deep Learning Techniques
• End-to-End Entity Detection with Proposer and Regressor
• Training ELECTRA Augmented with Multi-word Selection
• LV-BERT: Exploiting Layer Variety for BERT