Tying Word Vectors and Word Classifiers: A Loss Framework for Language Modeling
arXiv:1611.01462
Abstract
Recurrent neural networks have been very successful at predicting sequences of words in tasks such as language modeling. However, all such models are based on the conventional classification framework, where the model is trained against one-hot targets, and each word is represented both as an input and as an output in isolation. This causes inefficiencies in learning both in terms of utilizing all of the information and in terms of the number of parameters needed to train. We introduce a novel theoretical framework that facilitates better learning in language modeling, and show that our framework leads to tying together the input embedding and the output projection matrices, greatly reducing the number of trainable variables. Our framework leads to state of the art performance on the Penn Treebank with a variety of network models.
References in corpus (4)
Cited by in corpus (16)
- Neural Architecture Search with Reinforcement Learning
- Regularizing and Optimizing LSTM Language Models
- Weighted Transformer Network for Machine Translation
- FastGRNN: A Fast, Accurate, Stable and Tiny Kilobyte Sized Gated Recurrent Neural Network
- Single Headed Attention RNN: Stop Thinking With Your Head
- Revisiting Activation Regularization for Language RNNs
- Compression of Recurrent Neural Networks for Efficient Language Modeling
- Tied & Reduced RNN-T Decoder
- Transformer on a Diet
- Early Improving Recurrent Elastic Highway Network
- Improving Human Text Comprehension through Semi-Markov CRF-based Neural Section Title Generation
- Learning Less-Overlapping Representations
- Kite: Automatic speech recognition for unmanned aerial vehicles
- Assessing the Memory Ability of Recurrent Neural Networks
- Improving Neural Language Models by Segmenting, Attending, and Predicting the Future
- An Embedded Deep Learning based Word Prediction