Study On Machine Reading Comprehension Based On Multiway Bidirectional Attention And External Knowledge

Language understanding is a key component of the human cognitive system,and teaching machines to understand human language is one of the most elusive challenges in Artificial Intelligence.Machine reading comprehension,a crucial task introduced to evaluate how well computer systems understand natural language,is the only way to artificial general intelligence.It requires the machine to answer questions over a passage of text.In this thesis,we focus on multiple-choice reading comprehension,where every question comes with a set of candidate options and only one correct answer.Current methods on such tasks usually judge options independently and ignore their relations.Moreover,rare research studies how to integrate the reading comprehension models with external knowledge.Thus,this work analyzes these problems and proposes methods to improve the robustness and generalization ability of the model,which can be summarized as follows:(1)This thesis proposes a machine reading comprehension model MBAN,which is based on multiway bidirectional attention.In contrast to most existing models that ignore the option interaction,this model tries to simulate the comparing process of humans with an option comparison strategy,which can capture the option correlations.Specifically,the bidirectional attention mechanism is applied to model the interactions among the passage,question and candidate options.In particular,the option correlation is calculated to extract the relations among options.This enables the model to leverage the option comparison information for inferring the final answer accurately.(2)This study presents a machine reading comprehension model K-MBAN,which is based on external knowledge.There remains a problem in the MBAN model,i.e.,it fails to integrate external knowledge.On the contrary,K-MBAN introduces external knowledge based on the MBAN model,and its performance is further improved.More specifically,a framework named K-Adapter is used to infuse two kinds of knowledge with two knowledge-specific adapters.This enables the model to capture rich knowledge,which is more suitable for real application scenarios.The models are trained and evaluated on the Cosmos QA dataset.MBAN achieves81.2% and 82.4% accuracy on the development set and test set respectively,which outperforms the competitive baselines.Compared with the MBAN model,the performance of the K-MBAN model is further improved on both the development set and test set.All the results confirm the effectiveness of the models proposed in this thesis.