Abstract:

To address the issues of information overload， lack of content matching， and insufficient interpretability of the educational exam's models of English Multi-Choice Reading Comprehension （MCRC） in high school， the MCRC model integrating Knowledge augmentation and Contrastive Learning （KCL-MCRC） was proposed. First， the Large Language Model （LLM） was employed to generate a summary and extract clue sentences from the article for knowledge augmentation to reduce the input of irrelevant information. Second， the embedded representations of the article， questions and options were obtained by using the Bidirectional Encoder Representations from Transformers （BERT） model. Third， the information difference vectors of the article， questions and options were obtained through the contrastive learning strategy combining average pooling and absolute difference， and were subsequently concatenated with the original embeddings to jointly influence the answer selection. Finally， each question type was labeled by the LLM， and the performance of the KCL-MCRC model was evaluated to analyze the reasons for the model’s performance gains through accuracy in answering various types of questions. Experimental results on the RACE-H dataset show that KCL-MCRC achieves the optimal overall accuracies of 65.75% on validation set and 64.12% on test set， surpassing the models MDT （Multi-Decision-Transformer） and MMA （Multi-stage Maximization Attention） by 1.02， 0.35 and 0.88， 0.92 percentage points， respectively， achieving the best performance； both knowledge enhancement and contrastive learning are crucial for model's performance improvement， with significant the accuracy gains in three question types： logical reasoning， main idea summarization， and author's attitude. This validates effectiveness of designing modules for different question types.

Key words: Multi-Choice Reading Comprehension (MCRC), Large Language Model (LLM), Pre-trained Language Model (PLM), knowledge augmentation, contrastive learning

摘要：

为了解决高中英语多项选择阅读理解（MCRC）教育考试模型应用中信息过载、缺乏内容匹配和解释性不足的问题，提出一种融合知识增强和对比学习的多项选择阅读理解模型（KCL-MCRC）。首先，使用大语言模型（LLM）生成文章摘要和提取线索句子，通过知识增强减少无关信息的输入；其次，利用BERT （Bidirectional Encoder Representations from Transformers）模型获取文章、问题和选项的嵌入表示；再次，采用平均池化和绝对差值的对比学习策略获取文章、问题和选项的信息差异向量，并与原嵌入表示拼接，以共同作用于答案选择；最后，使用LLM进行问题类型标注，通过评估KCL-MCRC在各种问题类型上的答题准确率分析模型性能提升的原因。在RACE-H数据集上的实验结果表明， KCL-MCRC在验证集和测试集上的整体准确率分别达到了65.75%和64.12%，相比模型MDT （Multi-Decision-Transformer）和MMA （Multi-stage Maximization Attention）分别提高了1.02、0.35和0.88、0.92个百分点，取得了最佳性能；知识增强和对比学习对模型性能的提升都具有重要意义，在逻辑推理、主旨概括和作者态度这3种问题类型上的准确率提升明显，验证了针对问题类型进行模块设计的有效性。

关键词: 多项选择阅读理解, 大语言模型, 预训练语言模型, 知识增强, 对比学习