Character-Level Language Modeling with Deeper Self-Attention

TL;DR

This paper demonstrates that a deep 64-layer transformer with fixed context surpasses RNNs in character-level language modeling, achieving state-of-the-art results on text8 and enwik8 benchmarks by using auxiliary losses.

Contribution

It introduces a deep transformer architecture with auxiliary losses for improved character-level language modeling performance.

Findings

Deep transformer outperforms RNNs on benchmarks

Auxiliary losses improve training at depth

Achieves state-of-the-art results on text8 and enwik8

Abstract

LSTMs and other RNN variants have shown strong performance on character-level language modeling. These models are typically trained using truncated backpropagation through time, and it is common to assume that their success stems from their ability to remember long-term contexts. In this paper, we show that a deep (64-layer) transformer model with fixed context outperforms RNN variants by a large margin, achieving state of the art on two popular benchmarks: 1.13 bits per character on text8 and 1.06 on enwik8. To get good results at this depth, we show that it is important to add auxiliary losses, both at intermediate network layers and intermediate sequence positions.