A Unified Recombination and Adversarial Framework for Machine Reading Comprehension

The ability of the machine to conduct reading comprehension plays a vital role in artificial intelligence. Multi-choice machine reading comprehension requires the machine to understand the semantics better, since most of its corresponding candidates are paraphrases of the references. State-of-the-art methods concentrate on the single type question and design ad-hoc models. Nevertheless, in practical reading comprehension scenarios, given a passage, there are usually various angles to examine students. This leads to the challenges of multi-type questions and uncertain number of associated candidates. It is evident that existing methods cannot cope with these real-life scenarios since they merely focus on the single type question. To address the aforementioned challenges, we propose a unified machine reading comprehension framework via recombination and adversarial learning, which can handle these issues in a single model. First, we introduce the recombination layer to recombine candidates, which converts all types of questions into the same form. Next, to enrich the information contained by the representations of candidates, we design the encoding and the fusion layers to achieve the interaction between the passage and combined candidates. Finally, the adversarial learning mechanism is employed to add stochastic disturbance to the training procedure, which can help the model avoid the convergence problem and produce discriminative representations. Experimental results on the real-world reading comprehension datasets from the high school science textbooks validate the superiority of our method, particularly in tackling the challenges of multi-type question and uncertain number of candidates.