Research On Several Key Algorithms Of Machine Reading Comprehension

Reading comprehension is a cross-cutting issue throughout production and life,and can only be accomplished by correctly understanding the meaning expressed by others.In the same way,machine reading comprehension has a pivotal role in natural language problems.There are a large number of researchers who have proposed many algorithms with unique insights,and this paper provides some insights on their ideas and validates them in multiple choice and extractive question-and-answer tasks for machine reading comprehension.In multiple choice tasks for machine reading comprehension,to enhance the model’s ability to extract information,the features of the data are usually fused using matching algorithms.Most fusion schemes,use one-way or two-way matching.However,both schemes can only match two of the triad of articles,questions and options in a multiple choice task,and do not fully fuse the three features.Therefore,this paper proposes a triple matching and comparison regularity algorithm.The triple matching module can fully fuse the elements in the triad through the attention mechanism,and the contrast regularity module can pull apart the gap between correct and incorrect answers.In the extractive question-and-answer task for machine reading comprehension,two aspects are investigated in this paper.First,for the migration learning algorithm of extractive quizzing task,a large number of studies focus on single-source domain to single-target domain,multi-source domain to multi-target domain and multi-source domain to single-target domain schemes,and few articles propose a single-source domain to multi-target domain scheme.For this reason,a multi-target adaptive migration algorithm is proposed in this paper.This algorithm removes the private features of the dataset from the model by the domain identification module,so that the features learned by the model can be compatible with different target datasets.The semantic alignment module is used to close the distance between the source and target domain features and exclude the influence of semantic style on the model prediction results as much as possible.Second,in order to improve the performance of the model many algorithms introduce external knowledge,and although this introduction can improve the performance of the model,it invariably expands the dataset.Therefore,this paper proposes a self-supervised learning algorithm based on [Mask],which does not use external knowledge to generate copies of data,but uses [Mask] masking to generate copies.The masked copies enable the model to learn more useful information and reduce the risk of model overfitting as much as possible.The algorithm generates copies of the training data by different masking strategies and trains them together with the training data.The semantic alignment module then finds a balance between the training data and the masked data so that it does not lose the original information and is more robust at the same time.