Mnemonic Reader for Machine Comprehension

Recently, several end-to-end neural models have been proposed for machine comprehension tasks. Typically, these models use attention mechanisms to capture the complicated interaction between the context and the query and then point the boundary of answer. To better point the correct answer, we introduce the Mnemonic Reader for machine comprehension tasks, which enhance the attention reader in two aspects. Firstly, we use a self-alignment attention to model the long-distance dependency among context words, and obtain query-aware and self-aware contextual representation for each word in the context. Second, we use a memory-based query-dependent pointer to predict the answer, which integrates both explicit and implicit query information, such as query category. Our experimental evaluations show that our model obtains the state-of-the-art result on the large-scale machine comprehension benchmarks SQuAD.