By Han Xiao, Tencent AI.In the last post of this series, I have introduced the task of machine reading comprehension (MRC) and presented a simple neural architecture for tackling such task..If you are a professional researcher who already knows well of the problem and the technique, please read my research paper “Dual Ask-Answer Network for Machine Reading Comprehension” on arXiv for a more comprehensive and formal analysis.As we already know, there are three modalities in the reading comprehension setting: question, answer and context..It is like an object stands side-on to a mirror, whereas the context is the mirror itself, as illustrated in the next figure.In the real world, we know that a good reading comprehension ability means not only giving perfect answer but also asking good question..Specifically, the model should be able to infer answers or questions when given the counterpart based on context.Formally, let’s denote a context paragraph as C:={c1,…,cn}, a question sentence as Q:={q1,…,qm} and an answer sentence as A:={a1,…,ak}..This neural sequence transduction model receives string sequences as input and processes them through an embedding layer, an encoding layer, an attention layer, and finally to an output layer to generate sequences.The rectangle super-block on the side can be viewed as a decoder for QG and QA, respectively..During testing, the shifted input is replaced by the model’s own generated words from the previous steps.The embedding layer maps each word to a high-dimensional vector space..The parameters of this layer are shared by context, question and answer..The final output of the embedding layer is the concatenation of the word and the character embeddings.The encoding layer contains three encoders for context, question and answer, respectively..This aids accurate reproduction of information especially in QA, while retaining the ability to generate novel words.During training, the model is fed with a question-context-answer triplet (Q,C,A), and the decoded Q and A from the output layer are trained to be similar to Q and A, respectively..To demonstrate the effectiveness of dual learning, I now present some generated questions and answers from our model and the mono-learning models..Question or answer is generated given the gold counterpart based on context..In the third sample, the generated question from our model is more readable comparing to the groundtruth..You may also find out some awesome attention matrices there.As both question and answer are generated in our model, we adopt BLEU-1,2,3,4 and Meteor scores from machine translation, and ROUGE-L from text summarization to evaluate the quality of the generation.. More details