Language Models with Transformers

Abstract

The Transformer architecture is superior to RNN-based models in computational efficiency. Recently, GPT and BERT demonstrate the efficacy of Transformer models on various NLP tasks using pre-trained language models on large-scale corpora. Surprisingly, these Transformer architectures are suboptimal for language model itself. Neither self-attention nor the positional encoding in the Transformer is able to efficiently incorporate the word-level sequential context crucial to language modeling. In this paper, we explore effective Transformer architectures for language model, including adding additional LSTM layers to better capture the sequential context while still keeping the computation efficient. We propose Coordinate Architecture Search (CAS) to find an effective architecture through iterative refinement of the model. Experimental results on the PTB, WikiText-2, and WikiText-103 show that CAS achieves perplexities between 20.42 and 34.11 on all problems, i.e. on average an improvement of 12.0 perplexity units compared to state-of-the-art LSTMs. The source code is publicly available.

arXiv comment12 pages, 7 tables, 4 figures

References in corpus (15)

• Adam: A Method for Stochastic Optimization
• Google's Neural Machine Translation System: Bridging the Gap between Human and Machine Translation
• Neural Architecture Search with Reinforcement Learning
• DARTS: Differentiable Architecture Search
• Pointer Sentinel Mixture Models
• Regularizing and Optimizing LSTM Language Models
• SMASH: One-Shot Model Architecture Search through HyperNetworks
• Efficient Architecture Search by Network Transformation
• Hierarchical Representations for Efficient Architecture Search
• Simple And Efficient Architecture Search for Convolutional Neural Networks
• Tying Word Vectors and Word Classifiers: A Loss Framework for Language Modeling
• Progressive Neural Architecture Search
• Net2Net: Accelerating Learning via Knowledge Transfer
• Adaptive Input Representations for Neural Language Modeling
• Accelerating Neural Architecture Search using Performance Prediction

Cited by in corpus (22)

• Deep Neural Networks and Tabular Data: A Survey
• Controllable Unsupervised Text Attribute Transfer via Editing Entangled Latent Representation
• Encoder-Agnostic Adaptation for Conditional Language Generation
• Weakly-Supervised Video Moment Retrieval via Semantic Completion Network
• TraDE: Transformers for Density Estimation
• Neural Machine Translation with Byte-Level Subwords
• Towards Debiasing Sentence Representations
• Transformer on a Diet
• Transferring Monolingual Model to Low-Resource Language: The Case of Tigrinya
• GRUEN for Evaluating Linguistic Quality of Generated Text
• Beyond Dropout: Feature Map Distortion to Regularize Deep Neural Networks
• The NTNU Taiwanese ASR System for Formosa Speech Recognition Challenge 2020
• Detection of COVID-19 informative tweets using RoBERTa
• Neural Machine Translation: A Review and Survey
• PoD: Positional Dependency-Based Word Embedding for Aspect Term Extraction
• Meta-Path Constrained Random Walk Inference for Large-Scale Heterogeneous Information Networks
• Relational Weight Priors in Neural Networks for Abstract Pattern Learning and Language Modelling
• Neural Language Modeling With Implicit Cache Pointers
• Revisiting Simple Neural Probabilistic Language Models
• Actions Speak Louder than Listening: Evaluating Music Style Transfer based on Editing Experience
• Dispatcher: A Message-Passing Approach To Language Modelling
• Identifying and Exploiting Structures for Reliable Deep Learning