Adaptive Input Representations for Neural Language Modeling

TL;DR

This paper proposes adaptive input representations for neural language models, significantly improving training speed and perplexity performance by extending adaptive softmax to input embeddings and systematically comparing different factorization choices.

Contribution

It introduces adaptive input embeddings for language models, extending adaptive softmax to variable capacity inputs, and provides a systematic comparison of modeling choices in self-attentional architectures.

Findings

Models with adaptive embeddings train over twice as fast as character CNNs.

Achieved 18.7 perplexity on WikiText-103, improving previous results by 10.5.

Attained 23.02 perplexity on the Billion Word benchmark.

Abstract

We introduce adaptive input representations for neural language modeling which extend the adaptive softmax of Grave et al. (2017) to input representations of variable capacity. There are several choices on how to factorize the input and output layers, and whether to model words, characters or sub-word units. We perform a systematic comparison of popular choices for a self-attentional architecture. Our experiments show that models equipped with adaptive embeddings are more than twice as fast to train than the popular character input CNN while having a lower number of parameters. On the WikiText-103 benchmark we achieve 18.7 perplexity, an improvement of 10.5 perplexity compared to the previously best published result and on the Billion Word benchmark, we achieve 23.02 perplexity.