# Language Modeling with Deep Transformers

## TL;DR

This paper demonstrates that deep autoregressive Transformer models, configured specifically for language modeling, outperform LSTM baselines in speech recognition tasks and can operate effectively without positional encoding due to inherent positional information in autoregressive setups.

## Contribution

The study shows that well-configured deep Transformer language models outperform LSTM baselines and can function without positional encoding in autoregressive language modeling.

## Key findings

- Deep Transformers outperform LSTM baselines in speech recognition.
- Removing positional encoding can slightly improve deep autoregressive Transformer performance.
- Deep autoregressive models utilize inherent positional information without explicit encoding.

## Abstract

We explore deep autoregressive Transformer models in language modeling for speech recognition. We focus on two aspects. First, we revisit Transformer model configurations specifically for language modeling. We show that well configured Transformer models outperform our baseline models based on the shallow stack of LSTM recurrent neural network layers. We carry out experiments on the open-source LibriSpeech 960hr task, for both 200K vocabulary word-level and 10K byte-pair encoding subword-level language modeling. We apply our word-level models to conventional hybrid speech recognition by lattice rescoring, and the subword-level models to attention based encoder-decoder models by shallow fusion. Second, we show that deep Transformer language models do not require positional encoding. The positional encoding is an essential augmentation for the self-attention mechanism which is invariant to sequence ordering. However, in autoregressive setup, as is the case for language modeling, the amount of information increases along the position dimension, which is a positional signal by its own. The analysis of attention weights shows that deep autoregressive self-attention models can automatically make use of such positional information. We find that removing the positional encoding even slightly improves the performance of these models.