DeMRC: Dynamically Enhanced Multi-hop Reading Comprehension Model for Low Data

Abstract

AbstractMulti-hop reading comprehension requires the aggregation of multiple evidence facts to answer complex natural language questions, and the answer should be avoided when there is no answer. Training a model that can handle such difficult tasks requires a large number of data sets to support, but the labeling of data sets is very expensive and time-consuming, so it is very important to explore reading comprehension models suitable for low data, and external data related to large-scale tasks. It will also effectively improve the performance of the model. This paper proposes a two-stage model with dynamically context-enhanced method for multi-hop reading comprehension tasks under low data called DeMRC. The first stage sentence filtering model filters the top k sentences that are strongly related to the question, and the second stage answer prediction model dynamically constructs the training set every time during training to expand the data set, and uses sentences selected by sentence filtering model as input to reduce the interference of irrelevant sentences to the model during inference. In addition, the self-training method is used to pseudo-label the external data and use it as an auxiliary data set to improve the performance of the model. We conducted experiments on the multi-hop reading comprehension data set of the Chinese “CAIL 2020" Judicial Artificial Intelligence Challenge Reading Comprehension Track and English cross-document level data set called HotpotQA, which are 3.5% and 21.3% higher than the powerful baseline model, showing the effectiveness of the method.KeywordsData augmentationMachine reading comprehensionSelf trainingMulti-hot QA