关键词: 多项选择问答, 数据增强, 图卷积网络, 打分器, 课程学习

Abstract: Abstract: A masked data augmentation method based on graph convolutional networks (GCN) is proposed to address issues of inaccurate information, low sample quality, and poor model generalization in the domain of multiple-choice question answering (MCQA). In this method, the words in the articles are abstracted as graph nodes and connected with question candidate answer pair (QA) nodes to establish relationships with relevant article nodes. By calculating the similarity between the nodes and applying masking techniques to the graph nodes, augmented samples are generated. The augmented samples are then subjected to feature expansion using GCN to enhance the model's information representation capability. Furthermore, a scorer mechanism is incorporated to evaluate the original and augmented samples. The evaluation is accompanied by curriculum learning strategies to improve the accuracy of answer prediction. The proposed method was comprehensively evaluated on three datasets: RACE-M, RACE-H, and DREAM. Experimental results show that compared with the typical baseline model EAM on the RACE data set, the proposed method improves the accuracy by an average of 0.8 and 0.4 percentage points respectively, and compared with the typical baseline model STM on the DREAM data set, the proposed method The average accuracy is improved by 1.4 percentage points. Comparative experiments prove the effectiveness of the proposed method in MCQA tasks, and provide new inspiration for further research and application of data augmentation technology in this field.

Key words: Multiple-Choice Question Answering, Data Augmentation, Graph Convolutional Network, Scorer, Curriculum Learning