Multi-task Learning Based Densely Connected Attention Network On Machine Comprehension

Machine comprehension is one of the core tasks in the field of natural language processing.It aims to teach machines to understand a given article to answer rele-vant questions.Machine comprehension with unanswerable questions,in addition to extracting the answer from the article,needs to determine whether the question is repliable,which requires the model to have a deeper semantic understanding of the text.The current popular models are mostly pipelined models.They divide the answer prediction into two stages:answer extraction and answer verification.In this thesis we propose a Multi-task Learning based Densely connected Attention Network to achieve one-stop implementation of answer extraction and judgment.The densely connected encoding layer makes the model deeper and extracts differ-ent levels of semantic information.The multi-layer attention mechanism allows the model to fuse problems and article information at different granularity,making the model answer positioning and judgment better and more accurate.We combine the answer extraction module and answer verification module closely when designing the model architecture and use multi-task learning to train so that the model has stronger generalization ability.In addition,we propose a new data augmentation method for machine comprehension with unanswerable questions.The experiment proves that the new data augmentation method is simple and effective.The model presented in this paper achieved F1 76.8%on the SQuAD 2.0 validation dataset.