《中国教育现代化2035》指出，我国发展特色优质教育需构建教育质量评估监测机制，建立更科学更公正的考试评估制度。考试是我国衡量教育质量、筛选人才的重要途径。试题难度是评估考试质量的关键因素。然而，传统试题难度评估方法既会花费许多人力又会消耗大量物力。因此，如何自动化评估试题难度成为了教育领域乃至计算机领域关注的热点研究问题。
目前，自动化预测试题难度主要通过深度学习技术实现。大部分研究者将自然语言处理领域相关方法如卷积神经网络和预训练模型等应用到试题难度预测领域。虽然这种方式取得了较为丰硕的成果，但忽略了试题文本内部关联特征对试题难度预测的影响，且较少关注到外部知识对试题内容表征的影响。本文基于英语阅读理解试题，利用深度神经网络、预训练模型及注意力机制等技术提出了两种试题难度预测模型。本研究主要有以下两方面工作：
（1）基于多视角注意力机制的试题难度预测模型
首先，我们提出了一种具有多视角注意力机制的卷积神经网络模型（MACNN），用于从英语阅读理解试题的多部分文本中提取不同的关系特征，从而自动预测试题的难度。阅读理解试题一般由问题、对应的四个选项和阅读文档构成。首先利用CNN编码英语阅读理解试题文本。然后我们根据影响试题难度的主要因素设计了三个模块，分别是阅读模块、回忆模块和混淆模块。它们利用注意力机制获得问题、选项和文档之间不同关系的表征。最后，结合不同模块所获特征预测试题难度。此外，我们在三种数据集上验证了MACNN模型的有效性，并通过注意力权重可视化实验分析了多视角注意力机制获取文本间关系特征的效果。
（2）基于外部知识的试题难度预测模型
其次，本文提出了基于外部知识的试题难度预测模型（RKF+）。我们首次将现有外部知识引入到试题难度预测模型中。其核心是引入了带有哨兵向量的注意力机制，可以动态获取试题文本表征与相关外部知识表征。为了进一步融合所获外部知识，本研究增添了双向交互层。最后，在三种不同的数据集上验证了此模型的有效性，并通过消融实验和可视化实验进一步分析了带有哨兵向量的注意力机制对动态获取试题文本和外部知识表征的重要性。另外，本文基于真实英语阅读理解试题数据集，探索了两种不同外部知识对其试题难度预测模型的影响。

China's Education Modernization 2035 points out that the development of high-quality education needs to build a mechanism of educational quality evaluation and monitoring, and establish a scientific and impartial system of examination evaluation. In China, examination is an important approach to evaluate education quality and select talents. The question difficulty is the key factor to evaluate the examination quality. Traditional methods for question difficulty not only cost a lot of manpower, but also consume a lot of material resources. Therefore, how to automatically evaluate the question difficulty has become a hot research topic in education and computer.
Recently, automatic question difficulty prediction is mainly realized through deep learning technology. Most researchers applied methods in natural language processing, such as convolutional neural network and pre-train model, to prediction of question difficulty. Although these methods have obtained fruitful results, they ignore the influence of the internal correlation feature among several parts of text on the question difficulty prediction. Based on English reading comprehension question, this paper proposes two difficulty prediction models by using deep neural network, pre-train model and attention mechanism. The main contributes are as follows.
(1) Difficulty prediction model with multi-view attention mechanism
Firstly, we propose a convolutional neural network with multi-view attention mechanism (MACNN), to extract different relation from multi-part text in reading comprehension questions, and then automatically predict the difficulty of the questions. English reading comprehension exercises consist of a question, four options and corresponding reading documents. CNN is first used to encode them. Then we design three modules according to the main factors affecting the question difficulty, namely, the reading module, the recall module and the confusion module. They use attention mechanism to obtain the representation of different relationship among the question, documents and options. Finally, the question difficulty is predicted combined with feature obtained by different modules. In addition, we verify the effectiveness of MACNN we proposed on three datasets, and analyze the effect of multi-view attention mechanism through attention weight visualization experiments.
(2) Difficulty prediction model with external knowledge
Secondly, we propose the difficulty prediction model with external knowledge. We first introduce the external knowledge into the difficulty prediction model. Its core is the attention mechanism with sentinel vector in the knowledge fusion layer to dynamically obtain context and external knowledge related to questions. To further integrate external knowledge acquired before, a bidirectional interaction layer is added. Finally, the effectiveness of this model is verified on three different datasets, and we analyze the significance of attention mechanism with sentinel vector through ablation experiments and visualization experiments. In addition, based on the real English reading comprehension dataset we collected, this paper explores the influence of two different external knowledge on question difficulty prediction model.