Label Distribution Augmented Maximum Likelihood Estimation for Reading Comprehension.

Reading comprehension (RC) aims to locate a text span from a context passage to answer the given question. Despite the effectiveness of modern neural RC models, most existing work relies on maximum likelihood estimation (MLE) and ignores the structure of the output space. That is during training, one treats all the text spans do not match the ground truth as equally poor, leading to overconfident predictions on ground truth labels and reduced generalization ability in test. One way to bridge the gap between training and test is to take into account the task reward of alternative outputs using the reinforcement learning (RL) algorithms, which is often deficient in optimization as compared with MLE. In this paper, we propose a new learning criterion for the RC task which combines the merits of both MLE and RL-based methods. Specifically, we show that we are able to derive the distribution of the outputs, i.e., label distribution, using their corresponding task rewards based on the decomposition property of the RC problem. We then optimize the RC model by directly learning towards the auxiliary label distribution, instead of the ground truth label, using the MLE framework. In this way, we can make use of the structure of the output space for better generalization (as RL) via efficient optimization (as MLE). We name our approach as Label Distribution augmented MLE (LD-MLE), which is a general learning criterion that could be adopted by almost all the existing RC models. Experiments on three representative benchmark datasets demonstrate that RC models learned with the LD-MLE criterion can achieve consistently improved results over those based on the traditional MLE and RL-based criteria.