16.2.1 AI in Auto-Grading

Manual grading is a time-consuming task. With the help of computers this work can be made faster. Computer aided with AI techniques can evaluate faster and better analyze the test results and show the statistics in terms of easily understandable forms like graphs and charts. AI techniques can provide personalized feedback to students. Computers can evaluate multiple choice questions accurately and descriptive answer evaluation requires text processing with natural language processing capability.

The research work in [2] highlights the need and value of AIED’s Educational Cobots (Colleague Robots) and Smart classrooms. This work conveys the usefulness of building a humanoid robot as there are easier for humans to interact with; Humanoid robots are helpful in the classroom both from the teacher and the student’s perspective. This work opinions that humanoid robots should possess speech recognition, text recognition along with the capability to understand the human emotions to be more user friendly.

The work [3] explains how AI transformed the educational field from machines to cobots that work alongside or independently of teachers or educators to perform teacher-like tasks. Some students indulge in disgraceful act of plagiarism and other kinds of malpractice during assessments which needs to be taken care of. Computer vision-based AI systems [4] interpret and identify images of handwritten papers. These systems can be of great help in evaluating exam papers using pre-set benchmark and criteria. These systems can help reduce, such systems and put a stop to student from conducting unethical acts.

Automated essay grading is a difficult domain because computers struggle with several of the tasks an expert human performs during grading, such as implicitly correcting syntax mistakes and evaluating the logical structure of an argument.

An automated grading system for essay using machine learning proposed in [5] considers several attributes like proper grammar usage, domain knowledge, fluency in language and length of the essay which contribute significantly to the evaluation process. For a given text Latent Semantic Analysis (LSA) checks the similarity in words, catch phrases and paragraphs. The Latent Semantic Analyzer analyzes the content based on the context that was given. The evaluation method focuses on mistakes committed in grammar usage, spelling errors and improper usage of punctuation symbols. Bayes’ theorem was used in automated essay scoring focusing on specific words and phrases.

The keyword-based auto grading system evaluates the students answer script by considering the expected keywords in it. Different students present the same answer in different ways hence keyword based grading system is not efficient. To promote knowledge-based learning and to avoid rote learning an auto grading system with the combination of keyword, ontology and domain knowledge for theory subject is proposed in [6]. Student answer scripts which is unstructured data is transformed into a context map which is subjected to tokenization and ontology. Tokenization splits the data into words and eliminates those words that are not relevant or convey little or no information. Ontology maps the keywords based on facts to evaluate the answers.

Machine learning grading system for essay discussed in [7] considers the features such as total word count per essay, sentence count, number of long words and part of speech counts for training the model. This work used linear regression to generate parameters and to learn from the selected features. The proposed work uses a combination of features helping to predict better scores and employs 5-cross validation to train and test the model. Quadratic Weighted Kappa was used as the error metric.

The work presented in [8] efforts to produce essay evaluation model that best approximate human graders. This work discusses various features and techniques that are useful for accurate essay evaluation. Mean Quadratic Weighted Kappa is used as error metric. Two types of support vector machines: Rank SVM and Regression SVM were used in evaluation. Cross validation has found to be useful in the work.

In [9] proposes a computer assisted student answer assessment based on Natural language processing for evaluating the descriptive answers. The teacher is made to enter questions and correct answers. The grade assignment module compares every word in student response with the correct answer. Scores were assigned only when an exact match is observed in word as well as parts of speech tag and position of the word in the sentence.

Ontology is a concept map, makes the evaluation process holistic considering the presence of keywords, synonyms, proper and meaningful word combinations and coverage of concepts can be verified. A machine learning techniques for evaluating answer scripts proposed in [10] uses ontology to make the evaluation process more effective and accurate. This work efforts to provide a comparison of performance of answer evaluation with and without ontology and concludes that ontology improves the accuracy to a greater level.

16.2.2 AI in Smart Content

Digital teaching has several benefits over conventional teaching particularly in subjects that require demonstration. Animations and graphical applications have proved to convey the course content in an interesting and understandable manner. AI techniques can further contribute to generating smart course content. AI can recommend educational related suitable content like eBooks, related websites and videos on YouTube.

The author in [11] addresses various AI applications such as an educational method or instructive platform or model based guidance including the use of several technologies like virtual reality to explain concepts to learners or to demonstrate resources.

Authors in [12] predict student course selection taking into consideration the course and instructor characteristics, workload, mode of delivery and examination time using machine learning methods and ANN. This forecast allows the course teacher to schedule and prepare the content of the course as the specifications in an effective way.

Conventional learning strategies facilitate learning but may not be greatly engaging as students find game-based learning environment interesting. Students actively take part in game-based learning environment but do not uphold learning. Authors of [13] find strong positive relationship between learning outcomes, in-game problem solving and increased engagement. Narrative-centered learning environments offer a promising vehicle for delivering experiences that are both effective and engaging [13]. AI tools and techniques can make learning fun and engaging and thus ensure better learning experience through virtual education environment.

Educational content is easily accessible to students by various means, but relevant content desirable to the student is of greater help. Education content may be of different levels such as beginner or foundation, moderate and advanced. A context aware recommendation system discussed in [14] tries to cater to the needs of the student based on the student’s assessment score and the time taken to complete the assessment. This work applies fuzzy logic data mining method to suggest relevant learning content.

16.2.3 AI in Auto Analysis on Student’s Grade

The traditional education system treats every student alike. Learning pace of every student may not be the same and traditional system fails to address the issue of slow learners. AI as described in [15] can detect the learning shortcomings of students and take suitable measures to handle them in the earlier stages of education. AI systems can help to design the instructional material with required skills and knowledge that caters to the needs of the individual student.

The work in [16] depicts a scenario where teacher interacts with the AI tool to get an idea of students understanding of concept with respect to a particular subject. Digital information systems with storage and processing capability can be further benefitted with specialized AI algorithm which can identify and make sense of student commitment and behavioral patterns. The inferences from the analysis are presented along with suitable suggestions which can help the teacher to plan the upcoming classes according to the understanding level of the students.

A student grade prediction system for higher studies using machine learning techniques is proposed in [17]. This work uses Collaborative Filtering (CF), Matrix Factorization (MF), and Restricted Boltzmann Machines (RBM) techniques to predict the student grades for the enrolled courses. The prediction system aimed at identifying the students facing difficulty in enrolled courses so that the course instructor can take appropriate measures in this regard.

A student performance prediction system in [18] uses supervised machine learning techniques to identify poor performance in a distant learning system. Students demographic characteristics, marks obtained in previous written assessments are fed as input to train the model. The trained model is given new data to check with the accuracy. Naïve Bayes algorithm is found to have satisfactory performance.

Authors in [19] propose a machine learning system that predicts student performance. Total number of activities, average number of keystrokes and idle time for each session were considered for prediction. Artificial neural networks (ANNs), Naïve bayes classifiers, support vector machines (SVMs), logistic regression, and decision tree. Algorithms were employed to predict performance. The machine learning algorithms were used to analyze the data collected through digital electronics education and design suite (DEEDS).

Authors in [20] proposed a project on collaborative working model. This work enables peer review. A student can assess the work of another student on a web-based learning platform. The students review their peers by useful comments, evaluating assignments and giving valuable suggestions. This approach found to have positive impact and the students achieved better grades. Peer review enforces effective learning and improves motivation.

16.2.4 AI Extends Free Intelligent Tutoring

Among the most popular AI applications in education are Intelligent Tutoring Systems (ITS) discussed in [21]. With topics in well-defined standardized subjects such as mathematics or physics, ITS provides easy to understand bit by bit tutorials, individualized for each student. The ITS has experience in the topic and pedagogy and can recommend the best possible steps for the achievement of students. ITS guarantees successful learning by periodically reviewing the student, approaching the material with a growing level of learning.

Two kinds of digital tutoring systems namely Computer Aided-Instruction (CAI) and Intelligent Tutoring System (ITS) was discussed in [22]. CAI is helping the learner by providing immediate feedback and clue on their answers. In this type of system, the CAI poses a problem to the learner and the learner works out the problem and enters the answer. The CAI on receiving the answer from the learner congratulates the learner for the right answer. If the answer is found to be wrong, then the CAI helps the learner by giving suitable clues or feedback to make the learner solve the problem. ITS provides students an electronic type, natural language dialogue, virtual instrument panel, or other user interfaces that allow them to enter the steps necessary for solving the problem. In ITS the student is given a question and can enter his answer in a sequence of steps. The ITS is cable of interpreting Intermediate steps and results. Unlike the CAI this system provides helpful feedback and hints at every step instead of giving them at the end.

Self-Regulated Learning (SRL) skills allow a learner to self-assess and guide their own learning, demonstrates the ability of an educational robotic tutor in assisting the children to build up SRL skills with an Open Learning Model (OLM) more effectively. The robotic tutor [23] uses an OLM helps the learner to monitor their developing skills, set goals, and use appropriate tools. The robotic tutor is adaptable to the required circumstance and is capable of motivating the students to use SRL processes with it well-placed suggestions.

Personalization is easier in a one-to-one teaching environment where a student receives instructions from a proficient tutor. Personalization cannot be ensured when the teacher is dealing more large number of students. Teacher can utilize the power of AI to personalize the experience of students. AI helps to promote personalized learning through the formative assessment in the system [24].

AI systems can suggest the teacher regarding lessons that the students found difficult to understand and the teacher can conduct doubt clearing classes, share the recorded videos, related videos and relevant study materials to students pertaining to the subject.

Adaptive Intelligent Educational Systems utilize intelligent techniques to adjust educational content as per the requirement of the students. Authors in [25] propose a methodology for computerized dynamic analysis of learning objects using ontology models.

16.2.5 AI in Predicting Student Admission and Drop-Out Rate

Machine learning techniques, Support vector machine (SVM) and Artificial Neural Network (ANN) are used in several research studies to predict the academic performance and admission services. To predict dropouts using ANN, decision trees (DT) and logistic regression classification techniques, a [26] study examined students with different ethnic, academic, and financial features. With a score of 81.19%, ANN outperformed the other two, thus coming out with the conclusion that the financial, previous, and current academic status influence the most in predicting the student drop-out rate.

The work [27] used the SVM technique to identify students and predict admission decisions at a 97.17% accuracy level. Research clearly shows that AI-assisted admission decisions can predict accuracy with higher levels of precision, thus simplifying the administrative staff’s role.

A machine learning approach to predict dropouts proposed in [28] brings out the various factors influencing the drop-out rate. This work considers several useful factors that contribute to student dropout in universities. Personal data, academic data and institutional data were considered. Along with basic details personal data like amount of time spent with friends, social media, political engagements, current job, and time spent in tuition service is also considered for decision making. SVM, Random forest and neural network are used in designing the prediction model.

Authors in [29] propose a hybrid framework to determine first-year student retention. Feed Forward back propagation network architecture with 3-hidden layer neurons was used with activation function being hyperbolic tangent activation function. The output layer for predicting student retention has the hyperbolic tangent activation function. This work made use of Levenberg–Marquardt function and 10-fold cross-validations for training and validation, respectively.

16.3.1 Data Collection Module

The data collection module gathers variety of student related data that influences the students learning process. We gather the below mentioned data about the student.

1. a) Learner’s style of learning
2. b) Behavioral characteristics
3. c) Subject that the learn like
4. d) Knowledge about pre-requisite subject
5. e) Grades obtained in previous assessment
6. f) Attendance status.

Learner’s Style of Learning

Teachers come across different types of learners. Identifying the learner’s learning style plays crucial role in preparing and delivering the course content. Different learning aids are essential to effectively impart the course content to the students.

Walter Burke Barbe and his team proposed three learning modalities and are as follows:

1. Visualizing modality: Visual learners love to learn by visualizing concepts; they find diagrams, maps, charts and images with colours and shapes as interesting and useful. Suitable images, videos, demonstration, and animation can help in captivating their attention.
2. Auditory modality: Auditory learners learn by listening to sounds and like to learn from discussions. Learning environment free from noise, lectures with appropriate examples, music and group discussions can help them in drawing attention to the class.
3. Kinesthetic modality: Kinesthetic learners learn by performing activities. Hands on sessions, experimentation and team activities help this kind of learners to learn better.

Students are made to answer a questionnaire with the intent to know their learning styles. The questionnaire can be like the one given below.

• When you get a new word, how would you like to know the meaning (dictionary, Wikipedia, textbook)?
• What will you do or think of when you are waiting in a long queue?
• According to you, what is the best way to prepare for a test? Which one would you prefer when you want to learn something new (recorded lecture, friend, video and games)
• What distracts you the most when you are ready with all the required things for your class?
• Of these three classes (programming, theory, and mathematics) which is your favourite?
• When you hear a song on the radio, what are you most likely to do?
• If you were a winner in a game, what do you desire as a prize?

With the quiz the teacher can get to know the learner’s style of learning and the information about proportions of auditory, visual, and kinaesthetic learners in the class.

Behavioral Characteristics

Student behavioral aspects affect the way the student interacts with the teacher and his/her peers. Teachers in the classroom experience variety of student responses. Some students express their views openly, some may be shy to express their views. Knowing the student behaviour is necessary while forming groups. Student personality traits like sociality, helpfulness, empathy, extraversion, gregariousness, will achieve, assignment, ambition, aggressiveness, trouble tendency, hostility, openness to experiences, creativity is important to choose the proper partner. Students are made to answer a questionnaire with the objective to determine their characteristics.

The questionnaire can be like the one given below.

• What do you like to do to relax?
• When you are angry, what are you most likely to do?
• When you are happy, what are you most likely to do?
• What are you most likely to do while you are waiting?
• What type of book interests you the most?
• In a new place, what approaches you take to find your destination?

Subject That the Learner Likes

In technical education the students study a variety of subjects. Considering subjects related to computer science, we broadly classify the subjects into three kinds namely mathematical, logical, and descriptive. It is commonly observed that all students do not perceive all the three kinds of subjects equally. Student may have interest in any of the subjects and may have trouble with other kinds of subjects. It is necessary to determine the kinds of subjects that the student is in passionate with or facing problems. Conducting a test comprising of questions from all the three kinds of subjects helps us in knowing the kind of subjects the student is comfortable and/or facing problems. Considering 100 questions, with 25 questions each in mathematical and logical subjects and 50 questions in theoretical or descriptive subjects, we set the threshold as shown in Table 16.1.

Knowledge About Pre-Requisite Subject

Connecting subjects require good knowledge of pre-requisite subjects as they serve as the foundation to understand the new subject. Teachers can get to know student’s knowledge in the pre-requisite subjects with the help of quizzes that contains questions targeting fundamental concepts. The quiz score can help to determine the knowledge level of students pertaining to the quiz subject. Students who lack proper understanding in the pre-requisite subjects face difficulty in understanding the new subjects and drawing conclusions from the learnt concepts. Teachers usually handle this problem by engaging the students with bridge course. Bridge courses provide supplementary knowledge that helps students to gain essential knowledge required for the advanced subjects that will be taught to them in the forthcoming course. After the bridge course the teacher can know the impact of bridge course by conducting a quiz again with the different questionnaire. AI tools can compare the student scores before and after the bridge course and provide an insight to the teacher whether the students are ready with the pre-requisite or they need additional help. Based on the scores of individual students the teacher can help the student by sharing necessary material for easy understanding of the content.

Grades Obtained in Previous Assessment

Grade obtained by the student in previous assessment gives an idea about the learning pace of the students. Knowing the learning pace of the students helps the teacher to plan the lessons in a way that is easy for students to digest. Teachers can get to know the fast, average, and slow learners considering their grades obtained in the previous assessments.

Attendance Status

Attendance status of the student is an important factor that affects the students understanding of the subject and grades. An irregular student may lack continuity in subject related linked topics and may face difficulty in interpreting the concept and its applications. The attendance of the student can help to know the days that the student was absent to the class. The dates of the lesson plan executed by the teacher and the student’s attendance report can be compared to know if the student had missed foundation classes. The respective student should be provided with supporting study material necessary to cope up with the taught subject. If majority of the strength in the class had missed the foundation classes, the teacher may be suggested to repeat the class or change the lesson plan.

16.3.2 Data Pre-Processing Module

Data collected from the students may not be suitable to be fed into the analysis phase directly. The data may contain missing values or out of range values, thus data should be subjected to cleaned.

The data is processed with the following steps as given below:

Data cleaning: Data collected from the user end devices may be incomplete or lacking in certain behaviors. Unlike the Ignore the tuple approach which ignore the entire tuple we concentrate on filling the missing values. In the data cleaning process, we focus on validating the data, filling in the missing values with the most probable value and correct inconsistent or out of range values.

Noise Reduction: Faulty data collection may result in noise and hardens the process of proper interpretation. Noise is handled with different methods.

1. Binning Method works by dividing the sorted data into segments of equal size. Segments are handled individually after which different methods are applied on the sorted data with the intent to smooth it.
2. Regression approach uses regression function to smoothen the noise. The regression can be with single or multiple independent variables.
3. Clustering technique works by grouping similar data in a cluster. This may result in the outliers to fall outside the cluster or may even go undetected.

We use binning method to reduce noise.

Data Transformation

Normalization helps the chosen attribute values to fall inside a little determined range and discretization replaces the raw values of numeric attribute via conceptual ranges.

16.3.3 Clustering Module

The clustering module forms two clusters. Based on the score obtained by the student in mathematical, logical, and theoretical subjects we form clusters. We form two clusters one with students finding mathematics and logical subjects easy and another cluster with students feeling theoretical subjects easy. We assume that students fall in one of the two clusters. Students are by a grouped using K-means algorithm.

K-Means Clustering Algorithm

K-means is one of the simplest unsupervised learning algorithms that solve the well-known clustering problem. K-means algorithm iteratively partitions the dataset into K pre-determined unique subgroups. Every data point is assigned to a group. The algorithm ensures maximum distance between the clusters as much as possible. Figure 16.2 shows the K-means algorithm on datasets. It assigns the dаtа роint to а cluster such that the sum of the squared distance between the dаtа роint and the cluster’s centroid.

The k-means algorithm works is as follows:

1. Choose a value for k indicating the desired number of clusters.
2. The dataset is shuffled, and the centroids are initialized. After the initialization K data points are chosen for the centroids randomly
3. The process of assigning the data points to clusters is repeated until no further changes are observed.
4. Calculate the squared distance between data points and centroid. Determine their sum.
5. Each data point is assigned to the nearest cluster.
6. Calculate the centroids for each cluster by finding the average of all the data points that are a part of the cluster.

Euclidean distance between the two points is denoted by ||x_i – v_j||

Number of points in cluster i is denoted by c_i

Number of points in cluster j is denoted by c_j

16.3.4 Partner Selection Module

Students learn from their peers to a greater extent. Group activities encourage active participation and learning. Although group activities are effective means for engaging students but has several challenges. Formation of the groups is an action that needs to be done with intense care. Like minded individuals in the group provide happy learning environment and promotes active participation by the group members. Incompatible individuals in the group causes many problems like conflicts, dominant behaviour unequal contribution by group members and these things may discourage students.

Group activities and group studies provide many benefits and ease the job of teacher.

• Group activities cultivate the habit of learning from peers. Any group member if he/she had missed any class and is unable to follow the topics then the group members can help them with their knowledge.
• Develops team spirit in students.
• Students feel free to approach their friends and peers when compared to teachers.
• Improves communication and presentation skills.
• Group activities enable students to share their ideas and work on it.
• Develops qualities like adaptability, compassion, and patience.

For team activities involving more than two students, like minded students are identified by clustering them using K-means algorithm.

Various researchers opined that diverse groups let students in the same groups to be taught from one another, and consequently leads to more inventive and productive behaviors. In a small team of two students this idea can be extended by forming clusters followed by adopting a sample from each cluster to construct the varied groups. The learning style, behavioural characteristic, previous grades, and attendance are considered while forming a group. When the team size is small, close interaction exists among the students and choosing the right study partner is very important. Leaving the student to choose their study partner may pose challenges. Student’s choice of study partner may be biased. They may choose their desired ones without having into consideration their knowledge level. So, choosing the right study partner should be done without any bias.

In a normal classroom environment, the teacher can form groups impartially. Unlike the classroom, in the virtual/online learning environment the learners may be familiar with one-another or completely strangers. A virtual learning environment is not limited by small classroom strength. Learning can be made more joyful when students learn, share, and enthusiastically take part in activities. Irving’s algorithm is used to find the most suitable match between the learners.

Every student in a cluster ranks students in the other cluster, for example the preference list of students m₁ in cluster M is a vector [t₁ t₄ t₇ t₂ t₅ t₃ t₁₀ t₆ t₉ t₈] indicating student prefers t₁ over all others in the other cluster. After t₁, m₁ prefers t₄ and so on.

Irving’s algorithm works in three stages:

1. Making Initial proposals
2. Rule out any worst matches.
3. Finalize the best matching.

Pseudo Code

while there are unmatched people do

let i be the smallest value such that a_i is unmatched.

a_i proposes to his favourite roommate a_i who has not rejected him previously.

If a_i has not received a proposal before then

a_j accepts a_i

else

if a_j prefers ai over his current match ak then

a_j accepts a_i

a_k rejects a_k

else

a_j rejects a_i

end

end

end

for all accepted proposals (a_i,a_j) do

reject all (a_j,a_k) where a_j prefers a_i over a_k

for all cycles (p1,…….,p_n) and associated second preferences (q1,……………q_n)

such that :

q_i is the second preference of p_i

p_i+1 is the last preference of q_i

p_n ∈ {p1,…..pn-1} do

for I=1 …… n-1 do

q_i rejects p_i+1

end

end

Students are made to give their preferences based on their experience with the other student. To avoid biasing anonymity is maintained. Firstly, based on the learning style, prior knowledge, academic grades, and behavioural characteristics are considered and a set of suggestible students is obtained for every student. Student is made to take up some initial assessments with the other students in the set. Student after the assessment gives choices based on the experience during the assessment with other students. Assessments are designed in such a way that they force interaction and participation from both the learners. During the assessment, the students get an experience of behaviour characteristics of the other student with whom he/she interacted. The preferences made by the student depend on the experience he/ she gained during the assessment. Every student gives his opinion about the behavioural characteristic of all the students that he interacted with.