Capsule Networks for Chinese Opinion Questions Machine Reading Comprehension

摘要

In recent years, machine reading comprehension is becoming a more and more popular research topic. Promising results were obtained when the machine reading comprehension task had only two inputs, context and query. In this paper, we propose a capsule networks based model for Chinese opinion machine reading comprehension task which has three inputs: context, query and alternatives. First, we use a bi-directional LSTM to encode the three inputs. Second, model the complex interactions between context and query with a multiway attention layer. In addition to the attention mechanism used in BiDAF, the other two attention functions are designed to match the relationship between inputs. Finally, we present a capsule networks layer to route the right alternative. Specifically, we use two strategies to improve the dynamic routing process to filter noisy capsules, which may contain useless information such as stop words. Our single model achieves competitive results compared to the baseline methods on a Chinese dataset and obtains a significant improvement of 2.45% accuracy.