A Bert based Multi-step Attention Network for Chinese Reading Comprehension

• 1、State Key Laboratory of Networking and Switching Technology, Beijing University of Posts and Telecommunications

Abstract：Reading comprehension is an important task in current natural language understanding. It can well measure the ability of a natural language processing model. To facilitate the development of reading comprehension tasks, many related datasets and reading comprehension models have been proposed. However, most of these data sets and models are aimed at English reading comprehension, and the related work of Chinese reading comprehension is still relatively small. In this article, we study the task of Chinese reading comprehension and propose a Chinese reading comprehension model based on bert and multi-step inference mechanism. The model is mainly composed of a BERT layer and a multi-step attention layer.We used BERT to obtain good results on multiple Chinese reading comprehension datasets, and based on this, we introduced a multi-step inference mechanism to further improve the performance of the model.

Keywords： Chinese Reading Comprehension BERT Muli-step reasoning