Partially Shuffling the Training Data to Improve Language Models
This addresses a bottleneck in language modeling for NLP researchers by improving training efficiency and model performance, though it is incremental as it builds on existing shuffling techniques.
The paper tackles the problem of training word-level language models without full data shuffling, which hinders learning inter-sentence dependencies, by introducing a partial shuffling method that randomizes batches while preserving most sentence order, achieving new state-of-the-art results on Penn Treebank and WikiText-2 datasets.
Although SGD requires shuffling the training data between epochs, currently none of the word-level language modeling systems do this. Naively shuffling all sentences in the training data would not permit the model to learn inter-sentence dependencies. Here we present a method that partially shuffles the training data between epochs. This method makes each batch random, while keeping most sentence ordering intact. It achieves new state of the art results on word-level language modeling on both the Penn Treebank and WikiText-2 datasets.