WaLDORf: Wasteless Language-model Distillation On Reading-comprehension (1912.06638v2)

Abstract: Transformer based Very LLMs (VLLMs) like BERT, XLNet and RoBERTa, have recently shown tremendous performance on a large variety of Natural Language Understanding (NLU) tasks. However, due to their size, these VLLMs are extremely resource intensive and cumbersome to deploy at production time. Several recent publications have looked into various ways to distil knowledge from a transformer based VLLM (most commonly BERT-Base) into a smaller model which can run much faster at inference time. Here, we propose a novel set of techniques which together produce a task-specific hybrid convolutional and transformer model, WaLDORf, that achieves state-of-the-art inference speed while still being more accurate than previous distilled models.