Shortformer: Better Language Modeling using Shorter Inputs

TL;DR

This paper introduces methods to improve language modeling by using shorter inputs, including training on short subsequences and enhancing recurrence efficiency, leading to faster training and better perplexity.

Contribution

It proposes novel techniques for reducing input length and improving recurrence efficiency in transformers, resulting in faster training and improved language modeling performance.

Findings

Training on short subsequences reduces training time and perplexity.

Replacing relative with absolute position embeddings improves recurrence efficiency.

Combined methods speed up training by 1.65 times and enhance perplexity on WikiText-103.

Abstract

Increasing the input length has been a driver of progress in language modeling with transformers. We identify conditions where shorter inputs are not harmful, and achieve perplexity and efficiency improvements through two new methods that decrease input length. First, we show that initially training a model on short subsequences before moving on to longer ones both reduces overall training time and, surprisingly, substantially improves perplexity. Second, we show how to improve the efficiency of recurrence methods in transformers, which let models condition on previously processed tokens when generating sequences that exceed the maximal length the transformer can handle at once. Existing methods require computationally expensive relative position embeddings; we introduce a simple alternative of adding absolute position embeddings to queries and keys instead of to word embeddings, which efficiently produces superior results. We show that these recurrent models also benefit from short input lengths. Combining these techniques speeds up training by a factor of 1.65, reduces memory usage, and substantially improves perplexity on WikiText-103, without adding any parameters.