Exploring neural architecture search for language tasks

Neural architecture search (NAS), the task of finding neural architectures automatically, has recently emerged as a promising approach for unveiling better models over human-designed ones. However, most success stories are for vision tasks and have been quite limited for text, except for a small language modeling setup. In this paper, we explore NAS for text sequences at scale, by first focusing on the task of language translation and later extending to reading comprehension. From a standard sequence-to-sequence models for translation, we conduct extensive searches over the recurrent cells and attention similarity functions across two translation tasks, IWSLT English-Vietnamese and WMT German-English. We report challenges in performing cell searches as well as demonstrate initial success on attention searches with translation improvements over strong baselines. In addition, we show that results on attention searches are transferable to reading comprehension on the SQuAD dataset.