Exam paper generation based on performance prediction of student group
Introduction
The generation of exam questions is a challenging task in educational technology. The related research are roughly two categorizes. One is to use methods such as Natural Language Processing and Semantic Ontology to generate new questions from text or paragraphs [6], [28], and these methods focus on generating natural questions. Another is to extract questions from the question bank [13], [37], and the related methods consider the characteristics of the questions and their relevance to the student’s learning state. The task of exam paper generation (EPG) extracts several different questions from the question bank. Similarly, EPG depends on the characteristics of the questions, and needs to consider the students’ learning status.
EPG must consider various factors of an exam paper, such as its difficulty level, the coverage of assessed skills (synonyms of the knowledge points in this article), and the score of each question. Therefore, EPG is an optimization process with multiple objectives [8], [22]. However, with regard to translating the difficulty of each question into the difficulty of the entire exam paper, the existing methods do not propose a reasonable solution. Moreover, the exam is designed for the whole student group, rather than for a single student, and thus the difficulty is essentially a relative indicator. Therefore, different students may have different feelings about the difficulty level of the same question. Unfortunately, most existing EPG methods [36], [16], [24] rely on manually labeling the difficulty level of the question. This may be inefficient and produce label bias. Obviously, these methods cannot ensure that the difficulty of the generated exam paper is reasonable. Moreover, the existing EPG methods ignore an important issue that a good exam paper should be verified by the results of the exam [14], [10]. That is to say, the existing EPG methods cannot measure the quality of the exam papers generated by them in practical applications. Although some research has noticed the relationship between the EPG and the rationality of the exam results[12], there is still a long way to go before reaching the ideal goal.
In actual teaching activities, a common solution to EPG is to update an old version exam paper. The teacher can update the questions and adjust the difficulty of the exam paper based on his/her knowledge of students’ learning status. Constant adjustments make the distribution of skills and difficulty of the exam paper more and more reasonable. Inspired by this, we propose a novel exam paper generation approach based on performance prediction student groups. Where, the method of optimizing exam papers using dynamic programming named PDP-EG, and the other is using genetic algorithm named PGA-EG. In our study, skill is considered as the basic unit because it is the backbone embedded in various entities that appears during the learning process. Therefore, students’ skill mastery level determines whether they can correctly answer question related to the corresponding skill.
Our approach adopts Deep Knowledge Tracing [25] to achieve the students performance prediction task based on students’ exercise answered records. The exam paper generated by PDP-EG or PGA-EG could meet the difficulty level and skill weight requirements. The distribution of the achieved scores on our generated exam is close to that of requirement without manually setting the difficulty levels of the questions. The experimental results show that the exam papers generated by our approach are more advantageous than the baseline methods in terms of the main evaluation metrics of exam paper. In conclusion, the main contributions of our research are as follows:
- •
We propose a novel EPG approach that can generate the exam paper of a given difficulty without setting the difficult level of each question. In addition, the exam papers generated by our approach can ensure that the distribution of the achieved scores on exam are more reasonable.
- •
Our approach is able to well match the predicted student mastery levels of skills into the difficulty levels of the exam. This is achieved by using the multi-objective optimization algorithms.
- •
To the best of our knowledge, this is the first work on introducing the distribution of skill into EPG.
The organization of the paper is as follows: Section 2 starts with a focused review of some related work. Section 3 describes preliminaries and important notations. Sections 4 PDP-EG model, 5 PGA-EG model detail the proposed frameworks of PDP-EG and PGA-EG. Section 6 presents the setup of the experiments and results. And Section 7 draws conclusions for this paper.
Section snippets
Related work
In this section, we first review the relevant research on EPG, and then briefly introduce the recent research on learning performance prediction.
Preliminaries
In this section, we first describe how to represent question and exam paper by skill. Then, we explain the working process of Deep Knowledge Tracking in our research and how to predict the exam score based on it. Table 1 shows some important notations. In the following section, we will give a more detailed explanation of their roles.
PDP-EG model
An idealized exam paper should cover all the skills of the course, but when the number of skills is large, such exam papers are actually difficult to obtain. In practice, the closer the weight distribution of skills in the exam paper is to the weight distribution of skills in the course, the better the exam paper is. A group of students whose skill mastery level is influenced by a number of random factors, such as ability and intelligence. As we know, when the value of a variable is affected by
PGA-EG model
Generating a good exam paper needs to meet the three goals, namely skill weight, difficulty, and distribution of scores simultaneously, and thus it is a multi-objective optimization problem. In the solution of multi-objective problems, genetic algorithms have good performance. Therefore, in this section, we propose another performance prediction based EPG approach, which adopts improved genetic algorithm (abbr.PGA-EG). Like the PDP-EG method, this method also uses Dis, Dif and Div as
Experiments
In this section, we conduct experiments to test the performance of the proposed PDP-EG and PGA-EG, and compare them to two baselines.
Conclusions and future work
As an essential part of teaching, exam paper generation has to face the challenge of manual labeling. In this study, we propose a novel EPG approach, which applies the DKT model to predict learning performance, and use dynamic programming and genetic algorithm to optimize the quality of exam paper. The achieved AUC scores of the DKT model are not the same for different datasets, which measures the accuracy difference in the level of prediction of students’ skill mastery level in practice. With
CRediT authorship contribution statement
Zhengyang Wu: Conceptualization, Methodology, Writing - original draft. Tao He: Software, Investigation, Visualization. Chenjie Mao: Data curation, Validation. Changqin Huang: Supervision, Formal analysis, Writing - review & editing.
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgement
This work was supported by the National Natural Science Foundation of China (No. U1811263, 61877020), and the Foundation of China Scholarship Council (No. 201808440652), and the Key-Area Research and Development Program of Guangdong Province, China (No. 2018B010109002).
References (40)
- et al.
A multi-objective genetic algorithm for overlapping community detection based on edge encoding
Information Sciences
(2018) - et al.
An iterative adaptive dynamic programming algorithm for optimal control of unknown discrete-time nonlinear systems with constrained inputs
Information Sciences
(2013) - et al.
A three learning states bayesian knowledge tracing model
Knowledge-Based Systems
(2018) - et al.
Personalized learning full-path recommendation model based on lstm neural networks
Information Sciences
(2018) - et al.
An effective and efficient algorithm for high-dimensional outlier detection
The VLDB Journal
(2005) - Nabeela Altrabsheh, Mihaela Cocea, and Sanaz Fallahkhair. Predicting students’ emotions using machine learning...
- et al.
A tale of two moocs: How student motivation and participation predict learning outcomes in different moocs
Australasian Journal of Educational Technology
(2018) - et al.
An information-theoretic approach to automatic query expansion
ACM Transactions on Information Systems
(2001) - Dhawaleswar Rao Ch and Sujan Kumar Saha. Automatic multiple choice question generation from text: A survey. IEEE...
- et al.
Automated question paper generator system using apriori algorithm and fuzzy logic
International Journal for Innovative Research in Science & Technology
(2016)
A multi-objective optimization approach for question routing in community question answering services
IEEE Transactions on Knowledge and Data Engineering
Knowledge tracing: Modeling the acquisition of procedural knowledge
User modeling and user-adapted interaction
Algina. Introduction to classical and modern test theory
ERIC
The role of students’ motivation and participation in predicting performance in a mooc
Journal of Computer Assisted Learning
Automated test paper generation using utility based agent and shuffling algorithm
International Journal of Web-Based Learning and Teaching Technologies
Personalized question recommendation for english grammar learning
Expert Systems
Educational and psychological measurement and evaluation
ERIC
Ekt: Exercise-aware knowledge tracing for student performance prediction
IEEE Transactions on Knowledge and Data Engineering
Fuzzy logic based intelligent question paper generator
Introduction to educational statistics and evaluation
Cited by (24)
A methodological approach to enable natural language interaction in an Intelligent Tutoring System
2023, Computer Speech and LanguageReinforcement learning for automatic detection of effective strategies for self-regulated learning
2023, Computers and Education: Artificial IntelligenceAbility boosted knowledge tracing
2022, Information SciencesCitation Excerpt :The goal of KT is to estimate students’ degree of mastery of a specific knowledge concept based on students’ responses to items while tracing the development of learners’ knowledge concepts. KT has become a hot research topic in intelligent education and data mining communities [16,44]. In a general intelligent e-learning system, learners’ knowledge states are estimated by analyzing the history of the feedback on questions (exercises).
Improving paragraph-level question generation with extended answer network and uncertainty-aware beam search
2021, Information SciencesCitation Excerpt :In conversational systems and chatbots, such as Siri, Cortana, and Google Assistant, QG can serve as an important component to start a conversation or request feedback [17,34]. In education, QG can help with reading practices and assessment for educational purposes since manually generating questions is time-consuming [12,41]. Besides, QG can also be utilized to augment training data for other tasks, such as machine reading comprehension and QA, by generating large-scale question–answer corpora, which further assists in improving their models’ performance [18,36].
Exploring Multiple-Objective Optimization for Efficient and Effective Test Paper Design with Dynamic Programming Guided Genetic Algorithm
2024, Mathematical Biosciences and EngineeringStable Knowledge Tracing Using Causal Inference
2024, IEEE Transactions on Learning Technologies