Research On Machine Reading Comprehension Based On Supporting Information

The demand for explainability from users has grown along with the development of large-scale pre-trained models in natural language processing.Multiple-choice machine reading comprehension is one of the important tasks in natural language processing,which requires machines to select an answer from candidates based on the article and question.The supporting information is required to provide to the user when the machine reasoned out the answer to enhance the interpretability of the model.This paper focuses on the supporting information,which can be divided as single and multiple supporting information based on the number of the supporting information.The main contents are as follows:Firstly,to address the problem that the current models for finding single supporting information only takes into account word matching information and lacks deep semantics,this paper proposes a syntactic knowledge-guided single supporting information extraction model.The model is designed with a syntactic knowledge-guided sentence embedding network,which can obtain a dual knowledge sentence embedding containing syntactic knowledge and word meaning.Second,in order to unify the answer selection task and the supporting information extraction task,this model designs a weighted hierarchical fusion network.When the performance of answering questions using the fused document representation vector with weights improves,the performance of extracting supporting information based on weights also improves.Finally,this paper conducts experimental comparison and analysis on RACE~+and C3 datasets for proposed syntax knowledge-guided single supporting information extraction model,verifying its effectiveness and superiority.Secondly,current models based on multiple supporting information usually use semi-supervised or remotely supervised approaches,which can introduce noisy data and lead to inaccurate features.In addition,when only comparing supporting information with question-option pairs,context is often ignored,resulting in information loss.To address these problems,this paper proposes a document representation adjustment method that prefers multiple supporting information.Firstly,this paper designs an adaptive information gap mechanism to capture supporting information,avoiding expensive manual annotation.Secondly,this paper constructs an information gap narrowing mechanism that makes the document representation closer to the supporting sentences representation,thus making the document representation more preferential to the supporting sentence representation.When using document to select the answer,both supporting and context information can be considered,alleviating loss problem and improving reading comprehension performance.This paper evaluates proposed method on RACE and DREAM datasets for multiple-choice machine reading comprehension tasks and verify its effectiveness and superiority.