A Comprehensive Review Of Automated Essay Scoring (AES) Research And Development

Pertanika J. Sci. & Technol. 29 (3): 1875 - 1899 (2021)
ISSN: 0128-7680
e-ISSN: 2231-8526
Journal homepage: http://www.pertanika.upm.edu.my/
© Universiti Putra Malaysia Press
SCIENCE & TECHNOLOGY
Article history:
Received: 20 January 2021
Accepted: 24 May 2021
Published: 31 July 2021
ARTICLE INFO
DOI: https://doi.org/10.47836/pjst.29.3.27
E-mail addresses:
limchunthen95@gmail.com (Chun Then Lim)
chbong@unimas.my (Chih How Bong)
weesian.wong@gmail.com (Wee Sian Wong)
nklee@unimas.my (Nung Kion Lee)
* Corresponding author
Review article
A Comprehensive Review of Automated Essay Scoring (AES)
Research and Development
Chun Then Lim1
*, Chih How Bong1
, Wee Sian Wong1
and Nung Kion Lee2
1
Faculty of Computer Science & Information Technology, Universiti Malaysia Sarawak,
94300 Kota Samarahan, Sarawak, Malaysia
2
Faculty of Cognitive Sciences & Human Development, Universiti Malaysia Sarawak,
94300 Kota Samarahan, Sarawak, Malaysia
ABSTRACT
Automated Essay Scoring (AES) is a service or software that can predictively grade
essay based on a pre-trained computational model. It has gained a lot of research interest
in educational institutions as it expedites the process and reduces the effort of human
raters in grading the essays as close to humans’ decisions. Despite the strong appeal, its
implementation varies widely according to researchers’ preferences. This critical review
examines variousAES development milestones specifically on different methodologies and
attributes used in deriving essay scores. To generalize existing AES systems according to
their constructs, we attempted to fit all of them into three frameworks which are content
similarity, machine learning and hybrid. In addition, we presented and compared various
common evaluation metrics in measuring the efficiency of AES and proposed Quadratic
Weighted Kappa (QWK) as standard evaluation metric since it corrects the agreement purely
by chance when estimate the degree of agreement between two raters. In conclusion, the
paper proposes hybrid framework standard as the potential upcomingAES framework as it
capable to aggregate both style and content to predict essay grades Thus, the main objective
of this study is to discuss various critical
issues pertaining to the current development
ofAES which yielded our recommendations
on the future AES development.
Keywords: Attributes, automatic essay scoring,
evaluation metrics, framework, human raters,
recommendation

1876 Pertanika J. Sci. & Technol. 29 (3): 1875 - 1899 (2021)
Chun Then Lim, Chih How Bong, Wee Sian Wong and Nung Kion Lee
INTRODUCTION
The introduction of automated grading essays is an innovative attempt to reduce the effort
of examining essays and eliminate assessment biases and its discrepancies. Automated
Essay Scoring (AES) appears as a standalone computer software or distributed services that
evaluates and scores a written prose (Shermis & Burstein, 2003). The objective ofAES is to
overcome the time, cost, and reliability issues in manual assessment of essays. It should be
made clear thatAES is not intended to fully replace the human assessors but to be employed
as part of low-stakes classroom assessments to assist teachers’ essay marking routine. On
the other hand, it can be adopted in large-scale high-stakes assessments for the purpose of
increasing reliability, where the AES serves as an additional rater for cross-examination.
AES aims at developing models that can grade essays automatically or with reduced
involvement of human raters. It is a Natural Language Processing (NLP) based method
and application of assessing educational works especially on writing tasks. AES systems
may rely not only on grammars, but also on more complex features such as semantics,
discourse, and pragmatics. Thus, a prominent approach to AES is to learn scoring models
from previously graded samples, by modelling the scoring process of human raters. When
given the same set of essays to evaluate and enough graded samples, AES systems tend
to achieve high agreement levels with trained human raters. There are three common
AES frameworks: Content Similarity Framework (CSF) which assigns grades or scores to
new essays based on closer similarity of the reference essays’ scores, Machine Learning
Framework (MLF) which treats AES as classification or regression task and classifies the
new essays into correspond grade category by using machine learning algorithms, and
Hybrid Framework which combines the characteristics of both frameworks.
In a generalAES process, the collected essays usually stored as text or Microsoft Word
format. Hence, first step is to convert the group of collected essays into Microsoft Excel
or json format which contains the content and grade for each essay. The essay content will
then undergo pre-processing step such as tokenization, stop word removal, stemming and
lemmatization to remove noise. After that, there will be a major difference in different
frameworks. In cosine similarity framework, the pre-processed essay content will undergo
word representation step which convert the essay into vector form and compared their
similarity with gold standard’s grade. For machine learning framework, pre-processed
essay content will undergo feature selection, useful features will be extracted and act as
input data for predict model. Then, suitable machine learning algorithms will be used to
train the predict model to classify new essays into corresponding grades. The performance
of AES system is then evaluated with measurement metric. The most common evaluation
metric is using accuracy of to show the proportion of true results against the total number
of predicted grades examined.

1877
Recent Automated Essay Scoring (AES) Research and Development
Project Essay Grade®
(PEG) proposed by Ellis in 1966, is one of the earliest AES
systems. The project determines the quality of the essay by focusing on essay writing style
(Page, 2003; Rudner & Gagne, 2001). Subsequently, Intelligent Essay AssessorTM
(IEA)
by Pearson (2010) introduced an AES which could consider the essay’s semantic context.
In addition, e-rater®
, from Educational Testing Service (ETS) is another revolutionary
grading tool that uses computational methods to make sense of human natural language by
the means of tagging, chunking, and other labels, based on a collection of learner’s actual
language uses. To date, My Assess, from IntelliMetric®
model, is probably the pioneer
essay scoring tools solely based on Artificial Intelligence (AI), which with is modelled by
45 computers on human intelligence (Shermis & Burstein, 2003).
Over the course of 40 years, we have started noticing that the trend ofAES development
and many commercial applications are pretty much emerged in the Western continents,
especially in the United States. However, in recent years, there were many literatures on
AES reported in Asia: Malaysia, Thailand, Philippine, and Indonesian covering mainly
the English language as well as other languages.
To understand how AES has grown to its current state to meet the need in different
regions and to anticipate its future development, a detailed survey on AES is essential.
Thus, the purpose of this paper is to determine the recent progress of AES and generalize
frameworks accordingly. Moreover, this paper lays a spectrum of the development
frameworks for the reader by discussing the findings presented in recent research papers.
BACKGROUND
In this section, we provide an overview of AES systems that have achieved great success
in commercialization and attracted greater publicity. These systems are mainly proprietary
software developed in Western countries. For each product, we will present its vendor/
developer, primary focus, essay feature, scoring mechanism, and number of training
samples required.
Project Essay Grader®
(PEG)
Ellis Page’s Project Essay Grader (PEG) is considered as the first AES (Page, 1966). It
focuses on evaluating essays based on its writing style by using trins and proxes. PEG
assumes that there exist intrinsic qualities in a person’s writing style known as trins, which
can be measured or correlated with observable components denoted as proxes (Rudner
& Gagne, 2001). For example, the fluency of an essay (trin) can be correlated with the
amount of vocabulary (proxe). With training set of 100 to 400, PEG facilitates statistical
regression analysis to estimate essay scores. To date, PEG has developed more than 500
trins to be used to score essays (Measurement Incorporated, 2020).

Intelligent Essay AssessorTM
(IEA)
Intelligent EssayAssessor (IEA) is introduced by Pearson Knowledge Technologies (PKT)
to assess the quality of essay contents (Foltz et al., 1999; Pearson, 2010)). IEA scores
essay by using the Latent Semantic Analysis (LSA), which is a computational distribution
model to assess the semantics similarity of texts (Landauer et al., 1998). LSA is operated
on domain-specific corpus and the essays are represented through the multidimensional
semantic space of the meaning of their contained words and the similarity is derived by
comparing with other essay semantic representation (Foltz et al., 1999). IEAdiffered from
other AESs on the aspects that the scores derived from LSA are aligned closely to human
graders (Landauer et al., 2020), compared to the scores which are derived by correlation
of essay features. In addition, IEA uses NLP techniques to extract essay attributes such
as sophistication of lexical uses, grammatical, mechanical, stylistic, and organizational
aspects of essays (Zupanc & Bosnic, 2015). Comparing with other ASE, IEA requires
only a relatively small number of 100 pre-scored training essays sample for scoring a
prompt-specific essay (Dikli, 2006).
IntelliMetric®
Vantage Learning proposed IntelliMetric as a proprietaryAES to score essays operationally
since 1998 (Vantage Learning, 2020). IntelliMetric is regarded as the very first AES
system leveraging on Artificial Intelligence (AI) and Machine Learning (ML) to simulate
the scoring process (Dikli, 2006; Hussein et al., 2019). In producing an essay score,
IntellMetric uses more than 400 features (including semantics, syntactics, and discourses),
which can be categorized into five groups of IntelliMetric Feature Model: focus and unity
(coherence), organization, development and elaboration, sentence structure, mechanics,
and conventions in its scoring process. IntelliMetric claims itself of having multiple
automated scoring systems at work, each using a different mathematical model (e.g.,
Linear Analysis, Bayesian and LSA) for essay scoring (Vantage Learning, 2005; Vantage
Learning, 2020). Such multiple scoring engine within IntelliMetric emulates the equivalent
of a panel of multiple judges for achieving a more accurate final score as compared with
a single scoring engine in others. Another distinctive feature of IntelliMetric is its ability
in scoring essays in other languages besides English (Elliot, 2003). However, one of the
downsides of IntelliMetric is it requires a minimum of at least 300 scored essays to be
operated (Zupanc & Bosnic, 2015).
E-rater®
E-rater is developed and used by the Educational Testing Service (ETS) since 1999 (Attali
& Burstein, 2006). It relies on patented NLP techniques to extract linguistic features
for evaluating the style and content of an essay. E-rater 2.0 makes used of of syntactic,

1879
discourse and topical-analysis module to analyze essay features (Dikli, 2006). The features
are grammatical errors, word usage errors, mechanics error, style, organization segments
and vocabulary content (Shermis et al, 2010). To date, E-rater version extends the essay
scoring features into two areas:
(i) writing quality: grammar, usage, mechanics, style, organization, development,
word choice, average word length, proper prepositions, and collocation usage
(ii) content or use of prompt-specific vocabulary (Ramineni & Williamson, 2018).
E-rater uses regressing modelling to assign a final score to an essay. In addition, a
collection of approximately 250 training essay samples is required for the regression model
(Zupanc & Bosnic, 2015).
The comparison of these well-known AES is shown in Table 1.
Table 1
Summary of well-known AES
AES System
Vendor/
Developer
Main
Focus
Essay Scoring
Mechanism
Essay-Scoring Features
Training
Samples
Required
PEG Measurement
Incorporated
Style Statistical Trins & Proxes 100 - 400
IEA Pearson
Knowledge
Technologies
Content LSA • Content
Style
• Mechanics
100
IntelliMetric Vantage
Learning
Style &
Content
AI - cognitive
processing,
computational
linguistics, and
classification
• Focus & Unity
(Coherence)
• Organization
• Development &
Elaboration
• Sentence Structure
• Mechanics &
Conventions
300
e-rater Style &
Content
Regression
Analysis
• Grammatical Errors
• Word Usage Errors
• Mechanics Errors
• Style
• Organizational Segment
• Vocabulary Contents
250
PAST LITERATURE
In this section, we summarize the development ofAES in previous studies and categorized
their findings based on the type of attribute, methodology, prediction model and findings
(Table 2).
The attribute refers to the aspects evaluated by the proposed models which include style,
content, and hybrid. Style attributes focus on linguistic features such as spelling mistakes,
essay length and stop word count. Content attributes works to verify the correctness of

Table 2
Summary of AES
Type of
Attribute
Methodology Prediction Model
Measure &
Finding
Reference
Style Natural language
processing (Linguistic
features)
Nonparametric
Weighted Feature
Extraction, Stepwise
Regression and
Discriminant Analysis
Accuracy (51.3%) (Pai et al., 2017)
Linear Regression Close to human
rater
(Ramalingam et
al., 2018)
Random Forest Quadratic Weighted
Kappa
(0.8014)
(Chen & He,
2013)
Bayesian Linear
Ridge Regression
(BLRR)
Quadratic Weighted
Kappa (0.784)
(Phandi et al.,
2015)
Hybrid Natural language
processing (Rhetoric,
Organisation, Content)
Rule-based Expert
System
Correlation with
rater (0.57)
(Ishioka &
Kameda, 2006)
Vector Space Models
(VSM)
Support Vector
Regression (SVR)
Average correlation
(0.6107)
(Peng et al., 2010)
Natural language
processing (Rhetoric,
Organisation, Content
and Length)
Rule-based Expert
System
Pearson’s
correlation
coefficient (0.562)
(Imaki & Ishihara,
2013)
Latent Semantic Analysis,
number of words, number
of spelling mistakes, and
word distance.
Linear Regression Correlation with
rater (0.78)
Accuracy (96.72%)
(Alghamdi, et al.,
2014)
Latent semantic features Support Vector
Machine for Ranking
Pearson Corelation
(0.7248)
(Jin & He, 2015)
Rhetorical Structure
Theory (RST) and hand-
crafted features
Rule-based Expert
System
Accuracy (78.33%) (Al-Jouie & Azmi,
2017)
Latent Semantic Analysis
and Feature extraction
Linear Regression Accuracy (47.16%) (Contreras et al.,
2018)
Content Latent Semantic Analysis Artificial Neural
Network (ANN)
Mean of error
(0.44)
(Loraksa &
Peachavanish,
2007)
Cosine similarity Accuracy for small
class (69.80 % –
94.64 %)
Accuracy for
medium class
(77.18 % - 98.42 %)
(Ratna et al.,
2007)

1881
Table 2 (continue)
Type of
Attribute
Measure &
Finding
Reference
Accuracy (83.3%) (Amalia et al.,
2019)
Learning Vector
Quantization
Accuracy (96.3%) (Ratna et al.,
2018)
Topic classified by SVM
and assessed by LSA
Frobenius norm Average accuracy
for Japanese
(89.175%);
(Ratna et al.,
2019a)
Accuracy for
Bahasa Indonesia
(72.01%)
(Ratna et al.,
2019b)
GLSA Cosine similarity Precision, Recall
and F1 scores (0.98)
(Islam & Hoque,
2013)
Disco2, Damera-
levenshtein and N-gram
Similarity degree Correlation with
rater (0.82)
(Shehab et al.,
2018)
Latent Semantic Analysis
and Winnowing algorithm
Cosine similarity Accuracy for LSA
(87.78%); Accuracy
for Winnowing
(86.72%)
(Ratna et al.,
2019c)
Latent Semantic
Analysis with multi-level
keywords
Compared document
vector
Human raters’
agreement
(86%)
(Ratna et al.,
2015)
Modified LSA and
syntactic features
Cosine similarity RMSE (0.268) (Omar & Mezher,
2016)
Enhanced Latent
Semantic Analysis
Cosine similarity Gap with rater
(0.242)
(Sendra et al.,
2016)
Probabilistic Latent
Semantic Analysis
Cosine similarity Spearman
correlation (0.78)
(Kakkonen et al.,
2005)
Arabic WordNet (AWN) Cosine similarity Pearson Corelation
(98%)
(Awaida et al.,
2019)
Concept Indexing Cosine similarity Exact Agreement
Accuracy (0.452)
(Ong et al., 2011)
Contextualized Latent
Semantic Indexing
Support Vector
Machine (SVM)
Rating Agreement
(89.67)
(Xu et al., 2017)
Statistic (One-hot
encoding)
Convolutional Neural
Network (CNN),
Long Short-Term
Memory (LSTM),
Bi-directional Long
Short-Term Memory
(BiLSTM)
Quadratic Weighted
Kappa for LSTM +
CNN (0.761)
(Taghipour & Ng,
2016)
Statistic (Word
embedding)
Convolutional Neural
Network (CNN)
Average kappa
value (0.734)
(Dong & Zhang,
2016)

CNN and Ordinal
Regression (OR)
Accuracy (82.6%) (Chen & Zhou,
2019)
Siamese Bidirectional
Long Short-Term
Memory Architecture
(SBLSTMA)
Average Quadratic
Weighted Kappa
(0.801)
(Liang et al.,
2018)
Statistic (Training word
vector)
2-Layer Neural
Networks
Quadratic Weighted
Kappa (0.9448)
(Nguyen & Dery,
2016)
Hierarchical Recurrent
Neural Network
CNN, LSTM,
BiLSTM
Average Quadratic
Weighted Kappa
(0.773)
(Chen & Li, 2018)
Natural language
processing (Unigram
Language Model)
Machine Learning
Classifier
Mean accuracy
(51.5%)
(Wong & Bong,
2019)
Natural language
processing (Word
features, syntactic
features,
and dependency relation
features)
Logistic Regression
and k-
Nearest Neighbors
Correlation with the
rater (0.92)
(Cheon et al.,
2015)
Table 2 (continue)
Type of
Attribute
Measure &
Finding
Reference
content meaning and similarity between an essay with the graded essays. Hybrid attribute
facilitating both style and content.
Methodology refers to the methods used to identify the features from the essays. The
prediction model records the techniques or algorithms used to predict the score or grade
of the essay.
From Table 2, we can see that different researchers have developed their own
methodology and used different prediction models. This has proliferated the development
of AES because each methodology and prediction model does not seem to be universally
accessible to other researchers. Hence, a standard framework ofAES needs to be proposed
so that all researchers can use, modify, and enhance it in the future. Moreover, the evaluation
metric of AES also needs to be standardized so that the performance of AES system can
be compared with one and another. Lastly, most of the articles did not described in details
of their dataset used in research and therefore other researchers cannot reproduce the same
result as they stated in their article.
THE GENERAL FRAMEWORKS OF AES
Despite the increasing number of literature reporting novel approaches for AES
implementation, we can summarize them into three major general frameworks: content
similarity, machine learning and hybrid.

1883
Content Similarity Framework
The idea of content similarity framework (CSF) is to assign grades or scores to new
essays based on closer similarity of the reference essays’ scores. The framework requires
a gold standard: a collection of human graded reference essays, covering all spectrum of
grades or scores on the respective topics. The workflow of content similarity framework
is illustrated in Figure 1.
Preprocessing
New Essays
Graded Essays
Vector
Representation
Content
Similarity
Measures
Scores/Grades
Figure 1. The workflow of a content similarity framework
In this framework, selected essays firstly undergo pre-processing step, which includes
tokenization, stop words removal, stemming and lemmatization to reduce the noise in
essays. The subject of similarity can be based on (a) syntactic or, (b) semantics indicators,
or a combination of both.
(a) Syntactic Indicators. Syntactic indicators refer to the essays’surface features such as
part-of-speech, stemmed words, word connectors, and word count. Whereas the semantics
indicators refer to the meaning of word, phrase, sentence, and text. It is commonly
regarded that the semantics indicators are used to justify the whole or partial essays’
semantics similarity (Islam & Hoque, 2013; Omar & Mezher, 2016; Sendra et al., 2016;
Landauer et al., 2000; Ghosh & Fatima, 2008). The common syntactic indicators used
in AES are spelling checking, stemming, lemmatization, word segmentation (Loraksa
& Peachavanish, 2007), n-gram (Islam & Hoque, 2013; Chen & Zhou, 2019; Xu et al.,
2017), and normalization (Taghipour & Ng, 2016; Ratna et al., 2019a). These were the
essays’ surface features, and they are found to be useful in grading essays (Ong et al.,
2011). In addition, there were works reported to facilitate external knowledge bases such
as WordNet (Omar & Mezher, 2016; Shehab et al., 2018) and ontology (Contreras et al.,
2018) to improve grading efficacy.
In addition, the Japanese Scoring System (JESS) demonstrates an example of
facilitating syntactic features on essay grades based on three syntactic categories: rhetoric,
organization, and contents (Ishioka & Kameda, 2006). These categories are quantified by
readability, percentage of long, different words, passive sentences, orderly presentation

idea, and topical vocabularies. The essay score is then derived based on the deduction
mechanism of the essay’s perfect score. However, the uses of syntax and style alone are
not enough to determine the merits of the essay. Thus, JESS used syntactic indicators and
semantic indicators in its content analysis.
(b) Semantics Indicators. In recent years, AES solely based on syntactic indicators are
getting scarce, as many developments discovered that semantics indicators render more
accurate grades or scores. In natural language modelling, a semantic space aims to create
representations of the natural language that can represent the context. The most basic
semantic space can be tracked back to Vector Space Model (VSM), which was used to derive
content similarity based on the co-occurrence word in essays (Al-Jouie & Azmi, 2017).
One of the most popular syntactic-blind semantics indicators is Latent Semantic
Analysis (LSA), which excels at deriving content analysis (Landauer et al., 1998). LSA is
a distributional model used to derive meaning from a text. LSA was deemed “… a theory
and method for extracting and representing the contextual-usage meaning of words by
statistical computations applied to a large corpus of text”. With LSA, essays are represented
as a term-document matrix, which in turn is approximately reduced using singular value
decomposition (SVD). The dimension reduction process in LSA is to induce the probable
similarity of every word to every other if they are ever occurred in in other essays at a
common context. Experiment results showed that the addition of LSA over syntactic
features improves the scoring performance ofAES (Omar & Mezher, 2016). Many contents
similarity-basedAES used LSAor any of its variations in deriving grades or scores (Awaida
et al., 2019; Amalia et al., 2019; Alghamdi, et al., 2014; Contreras et al., 2018; Ong et al.,
2011; Shehab et al., 2018).
On the other hand, Generalized Latent Semantic Analysis (GLSA) is a variant of
LSA which considers word sequence and has been reported to be capable of improving
the efficacy of AESs (Islam & Hoque, 2013; Sendra et al., 2016). Almost all modern
development of AESs reported are composed of both syntactic and semantics features.
The study of essay grading using CSF took two inputs: key answers and student answers
(Amalia et al., 2019). Both essay inputs are pre-processed through noise removal, case
conversion, tokenization, stopwords removal negation, conversion, stemming, synonym
conversion, which are then represented using a term-document matrix where each row
corresponded to the term and each column corresponded to the document. Each cell in the
matrix represents the occurrence of the term to the documents. A zero value indicate the
absence of the term in the documents.
The most common similarity computation is derived through LSA, which is a 2-step
process: Singular Vector Decomposition (SVD) and cosine similarity measure. SVD is
responsible to decompose the term-document matrix as D=UΣVT. The k largest singular

1885
values (which is dimensionality) is used to approximate D as D≈UkΣkVT
k. Its purpose is to
the discover “latent” concepts in the matrix. SVD is first applied to the key answers, where
each of the student answer is then go through the same pre-processing processes and match
it to the most similar key answers using cosine similarity measure to determine the scores.
The workflow of essay grading using CSF is illustrated in Figure 2 (Amalia et al., 2019).
Figure 2. A case of essay grading using CSF (Amalia et al., 2019)
Figure 3. The workflow of a machine learning framework
Preprocessing
New Essays
Graded Essays
Scores
LSA
Cosine
Similarity
Measure
Machine Learning Framework
As shown in Figure 3, in Machine Learning Framework (MLF), essay grading is treated as
a multiclass classification problem in which each grade is represented as a class. Modelling
requires computational functions to generalize all essays into multi classes. SinceAES has
been seen as the document classification problem, the machine learning algorithms used
are mainly from the categories of regression and classification. The workflow of machine
learning framework is illustrated in Figure 3.
Pre-processing is the first process which prepares the data and removes noises.
Similar to CSF, all essays will undergo tokenization, stop word removal, stemming and
lemmatization processes. Next, the essays are processed to retain significant features in
the Feature Selection process. Typical features in MLF were words (Cheon et al., 2015),
syntactic and dependency features. Feature selection is an important step in many machine
Preprocessing
New Essays
Graded Essays
Feature
Selection
Machine
Learning
Algorithms
Scores/Grades

learning tasks with the purpose to identify a significant feature subspace in reducing
redundant features and reducing complex computational space, yielding the optimal essay
representation. In this context, feature selection reduced the number of words to prevent
the curse of dimensionality that can eventually degrade the accuracy of classification. In
general, there are two dimensionality reduction techniques: feature elimination or feature
selection. After dimensionality reduction, the selected features will act as inputs to train
the machine learning model such as Support Vector Machine (SVM), k-Nearest Neighbors
(kNN), Naive Bayes and Artificial Neural Network (ANN).
Like CSF, a machine learning framework requires a gold standard, however, the graded
essays are compulsorily to be processed and transformed into a computational model to
be used for prediction of grades and scores on new essays. This is one of the significant
differences between the CSF and MLF.
The study reported in Taghipour and Ng (2016) adopted MLF to derive essay scores.
Both key and student answers are pre-processed with tokenization, case conversion and
normalize the essay score in the range of [0,1]. Feature selection is performed through
Enhanced AI scoring engine (EASE) is used to derive length-based representation, POS,
word overlap with the key answers, and bag of n-gram. The features are then fed into
machine learning algorithms such as support vector regression (SVR), Bayesian linear
ridge regression (BLRR) and a variant of neural networks to derived student answer
scores, resulted marginal increment against the baseline. The workflow of machine learning
framework reported in Taghipour and Ng (2016) is illustrated in Figure 4.
Figure 4. A MLF reported in (Taghipour & Ng, 2016)
Length-based
features/POS/
Word overlap/
n-grams
Key Answer
Essays Scores
Regression/
convolutional
& recurrent
neural network
Looking at the recent trends in AES, the machine learning framework is gaining
popularity in recent years due to the efficacy of SVM (Ratna et al., 2019b; Ratna, et al.,
2019a; Xu et al., 2017; Awaida et al., 2019; Chen & Li, 2018) and the ability to represent
text context with word embedding (Liang et al., 2018; Taghipour & Ng, 2016) propelled
by Artificial Neural Network (ANN) (Loraksa & Peachavanish, 2007; Taghipour & Ng,
2016; Dong & Zhang, 2016; Liang et al., 2018).
Hybrid Framework
The emerging hybrid framework (HF) combines both the goodness of content similarity
and machine learning framework in which it capable of aggregating both style and content

1887
to derive essay grades or scores. Different from the general MLF, where machine learning
algorithms are directly used to derive the grades or score, the machine learning algorithms in
the framework is used to generalize syntactic features (indices, topics, and domain specific
keywords) where CSF is used to retrieve the closest key answer score in the semantic space.
The process pipeline of hybrid framework is similar with the ML framework except both
machine learning algorithms and content similarity measure are used to derive essay scores
or grades. Figure 5 illustrates a general workflow of the hybrid framework.
The emerging of recent studies incorporating the hybrid framework have been seen the
combining the use of linear regression on selected features, derived from an ontology and
using LSAto measure content similarity (Contreras et al., 2018), generalizing essay scores
through vectorization with artificial neural network and LSA (Loraksa & Peachavanish,
2007) and twostep process classification with SVM and LSA Framework (Ratna et al.,
2019a; Ratna, et al., 2019b)
The studies reported in Ratna et al. (2019c) used two-step grading process as the HF:
used Support Vector Machine (SVM) to classify essay’s topic and LSA to compute the
similarity between student and key answers. A pre-trained SVM model on key answers’
topics is meant to rule out unrelated topic essays and to route it to the related key answers
Figure 5. The workflow of a hybrid framework
Figure 6. A Hybrid framework reported in (Ratna et al., 2019c)
Preprocessing
New Essays
Graded Essays
Content
Similarity
Measures
Scores/Grades
Feature
selection/
Machine
Learning
Algorithms
Preprocessing,
TFIDF
New Essays
SVM trained
with key
answer topics
Scores
LSA Content
Similarity with
key answer
scores
Any match to
the existing
topics?
Find the most
similar content
score

essay. The LSA is intended to build a semantic space for key answer essays, where each
of the student essay’s score in the range of [0,100] is obtained by finding the most similar
key answer score, using Frobenius norm similarity measure. The studies report substantial
accuracy (>95%) as compared to human raters. The workflow of hybrid framework reported
in Ratna et al. (2019c) is illustrated in Figure 6.
AES EVALUATION METHOD
In this section, we categorized the automated essay scoring evaluation methods from past
literature into five categories. All evaluation methods require the human rater to provide
a score or a grade for each essay to act as a reference. Then, the evaluation method will
compare the human rater’s score with the AES model predicted score to examine the
accuracy of the proposed model.
Overall Accuracy
This evaluation method compares the average score manually marked by the human rater
and scores automatically generated by the proposed model to examine the accuracy. The
predicted result is classified into three classes, which are ExactAgreementAccuracy (EAA),
Adjacent Agreement Accuracy (AAA) and Overall Accuracy (OA). Exact Agreement
Accuracy is the number of essays with human score equal to AES score over the total
numbers of essays, as denoted in Equation 1.
[1]
The Adjacent Agreement Accuracy is defined as the ratio of the number of test essays
with a human score equal to machine score ± 1 over the total number of test essays and is
given by the Equation 2.
[2]
OverallAccuracy is the sum of EAAandAAAand is defined as the ratio of the number
of test essays with a human score equal to AES score ± 1 over the total number of test
essays (Equation 3).
[3]
This evaluation method is simple and able to evaluate the accuracy of the scoring
model intuitively. However, the result can be inaccurate when the dataset contains an
imbalance ratio of data.

1889
Root Mean Square Error
Root Mean Square Error (RMSE) is used to examine the similarity between the human
rater’s score set and system predicted score set. RMSE is calculated by Euclidean distance
hence it can use to evaluate real values such as essay score. Therefore, the accuracy of
predict model depends on the similarity between the human rater’s score and system
predicted score based on Euclidean distance (Omar & Mezher, 2016). RMSE can be
calculated as Equation 4.
[4]
where HSi is human rater’s score, SSi is the system predicted score and n is number of
essays. The smaller value of RMSE indicates that the predicted score is more similar to
human rater’s score and therefore achieve a better scoring.
Mean of Error and Standard Deviation of Errors
The mean error and standard deviation of errors are used to examine the difference between
the human rater’s score, and the system predicted score. This evaluation method requires a
state-of-art system as a baseline to examine the performance of proposed system. The mean
of error between human rater’s score and baseline system predicted score is calculated and
compare with the mean of error between human rater’s score and proposed system predicted
score. The smaller value of mean of error and standard deviation of error represent the
system has better performance in essay scoring. The mean of error and standard deviation
of error is calculated by using Equation 5 and 6:
[5]
[6]
is the arithmetic mean from all errors,
is an absolute value of an error between human score and machine score computed
by
n is the number of data set
SD is a standard deviation of error

Pearson’s correlation coefficient, Cohen’s Kappa Coefficient and Quadratic
Weighted Kappa (QWK)
Pearson’s correlation coefficient, Cohen’s Kappa coefficient and QWK scores are used to
examine the degree of agreement between the human rater’s score, and the system predicted
score. The output of these evaluate methods will between 0 and 1 when 1 means that the
system predicted score is complete agreement towards human rater’s score and 0 means that
is random agreement between system predict score and human rater’s score. In interpreting
any kappa value, K can be considered as poor when lower than 0.4, fair to good when K
between 0.4 and 0.75, and excellent when K is greater than 0.75. To interpret Pearson’s
correlation, r can be considered as very small when r lower than 0.2, small when r between
0.2 and 0.4, medium when r between 0.4 and 0.6, large when r between 0.6 and 0.8, and
very large when r greater than 0.8 (Cheon et al., 2015). Theses evaluation methods are
suitable for evaluating the multiclass classification model since it calculates a confusion
matrix between the predicted and actual values. Usually, these methods require a state-of-
art scoring model as a baseline to examine the performance of the proposed model.
Mean of Accuracy
The scoring model is evaluated by obtaining the mean of accuracy of each test date. In
achieving this, the difference between human rater’s score and system predicted score will
be derived. The accuracy for each test data will be calculated by using Equation 7.
[7]
It should be noted that |Human score – System score| is an absolute value. By using
this formula, the smaller the difference between human rater’s score and system predicted
score, the higher the accuracy of the scoring model. This evaluation method is suitable for
evaluate the actual score of essays but not essay grade due to the gap between essay grade
is usually small and result in overrated scoring model accuracy.
ISSUES
There Is No One-Size-Fits-All Solution
Most of the existing essay scoring systems reported thus far performed well in grading
pure English essays or essays written in pure European language. However, the system
graded a 10-15% lower score on essays containing Asian local content (Ghosh & Fatima,
2008). This is due to the influence of local languages and English written by non-native
English speakers. For example,Asian students tend to obtain lower scores in TOEFLexams.
Besides, Wong and Bong (2019) claim that the direct adoption of contemporary automated

1891
essay scoring system in the Asian context may not be practical as it may lead to the issue
of assessment reliability and validity. Reliability means how well an assessment tool can
produce stable and consistent results even in different time and place, whereas validity
means how good an assessment tool measures what it is supposed to measure.
Existing Automated Essay Scoring Systems Predict Essay Grade Ineffectively
under a Prompt-Independent Setting
Essay prompt refers to an essay title or topic which the writer requires to treat as the
main content of the essay. From the literature, majority of the essay scoring mechanism
rely on rated essays as the gold standard and the performance of the system is highly
dependent on these training data (Jin et al., 2018). However, this approach of training is
hard to perform especially when the rated essays for a target prompt are difficult to obtain
or even inaccessible due to legal, copyright or privacy issues which lead to inefficacy. In
addition, essays for different prompts may differ a lot in the uses of vocabulary, structure,
and grammatical characteristics. Hence, these models can hardly be generalized and fail to
grade them accurately for non-target prompts. This situation is because prompt-dependent
models are designed to learn the features from prompt-specific essays.
Effective AES Systems Requires Expert Tuning
A study by Alikaniotis et al. (2016) stated that the predictive features of the automated
essay scoring system need to be manually crafted by human experts to achieve satisfactory
performance and the process will consume a lot of time and work. The lack of human experts
results in a decrease in AES performance and an increase in the AES development time.
The Relationship between Essay’s Features and Grade is not Linear
Most of the existing automated essay scoring systems assume linear relationship between
the features of the essay and the essay grade. However, the study by Fazal et al. (2011)
stated that there exists a non-linear relationship between the feature vector and essay
grade. For example, most of the existing automated essay scoring systems treat the length
of essay as an important feature and use it to indicate the quality of essay which means
the longer the essay, the higher the essay grade. However, in some specific cases, an essay
will be assigned with a lower grade even though it is very long because of the irrelevant
content in the essay. Moreover, automated essay scoring system will delete unnecessary
and non-meaningful words to filter out keywords to evaluate during processing the essay.
This makes automated essay scoring system easy to be fooled by students because it is
unable to distinguish between good writing and baloney. In order words, the act of using
and repeating some key words from the prompt, fill up lots of space may result in a higher

grade from the automated essay scoring system (Greene, 2018). It is also reported in Davis
(2014) that essays can obtain high scores even with gibberish which makes no sense to
human readers.
AES are Sensitive to Noise
Fazal et al. (2011) also reported that the noise from the essays will have a negative effect
on the performance and efficiency of automated essay scoring system. These noises include
punctuation errors, syntax-based errors, morphological, context-based spelling errors,
and misspellings. Hence, it is important to rule out the noise before the modeling process
because it will lower the performance.
Essay Feedback Given by Automated Essay Scoring System is Unable to Increase
Student’s Grade Significantly
Darus et al. (2003) had conducted a study in Malaysia to investigate the improvement
of students after revising their essays based on the feedbacks given by automated essay
scoring system. Based on the results, the revisions made did not significantly increase the
score of revised essays for most of the students. Furthermore, students who participate in
the relevant research found that the feedback given by automated essay scoring system
is useful and informative to a certain extent although their score remained the same. This
suggests that feedbacks given in AES are not sufficient in pointing out areas that could
potentially increase the students’ scores.
AES do not Consider Context and Rating Criteria
To develop a context-aware AES, the essay context information should be reinforced to
build anAES which can distinguish poor, ordinary, and excellent essays. Besides, semantic
features should be added to assist the AES in context grading process. On the other side,
some existing AES only focus on the essay content and did not consider rating criteria
behind the essay during grading process (Liang et al., 2018). It is because a few rating
criteria are difficult to integrate in AES and therefore lower the accuracy of the system.
DISCUSSION
Standardizing AES Framework
The review of previous studies onAES has shown that various frameworks have been used
and this has proliferated the development of AES because each framework does not seem
to be universally accessible to other researchers. After reviewed all frameworks used in
AES system, we suggest hybrid framework as standard framework for AES as shown in
Figure 5 because it is a hybrid framework combines both the goodness of content similarity

1893
and machine learning framework in which capable to aggregate both style and content
to predict essay grades or scores. Besides, this framework enables researchers to apply
different predictive models and NLP techniques on it and examine their performance. The
workflow of proposed standard AES framework is illustrated in Figure 7.
For Malaysia University English Test (MUET), the marking scheme of an essay is
the aggregation of task fulfilment, language, and organisation (Malaysian Examinations
Council, 2014). Task fulfilment refers to the ability to understand topic and developing
ideas which can be benchmarked with semantic indicators. Whereas the language and
organisation refer to the number of grammar error, number of spelling error, use of
appropriate vocabulary and coherence of content, are regarded as syntactic indicators.
Hence, the machine learning part of our proposed framework is aimed to evaluate language
and organisation using where the semantic indicators can be evaluated with content
similarity measures.
In this framework, the essays undergo pre-processing such as word tokenization,
predicting Parts of Speech, lemmatization and stop words removal to reduce the noises in
the essay. Next, the features of the essay such as essay length, word count and misspelled
word count may need to be extracted to ensure these features will not result in biased
decision after vectorization of essay content. Then, the pre-processed essay undergoes word
vectorization step to vectorize the essay content and feature selection to extract syntactic
features from essay. Finally, the features and vectors are both treated as the inputs and
used to train the predictive model using machine learning and content similarity measures.
Recommended Evaluation Metric
One of the issues with existing literatures is each researcher used different evaluation
metrics to examine the performance of AES. This causes discrepancy when comparing
the results with another works.
Accuracy has been seen as the most common evaluation metrics used in evaluating
the performance ofAES by calculating the percentage of matched pairs with human raters.
Figure 7. The workflow of proposed standard AES framework

However, the result may be skew and biased toward the majority class when dealing with
imbalanced datasets (Tanha et al., 2020). In real life, dataset normally contains imbalance
proportion of grades since the distribution of the essays grade are normal.
Hence, this paper suggests that using Quadratic Weighted Kappa (QWK) score as
a standard evaluation metric for AES. QWK is belong to kappa-like family and kappa
coefficient has been proven that it can provide valuable information on the reliability of
ordinal scale data (Sim & Wright, 2005). Moreover, Wong and Bong (2019) had stated
that Kappa value is a better measurement than simple percent agreement calculation in
AES because it corrects the percent-agreement for the case of agreement that would be
expected purely by chance when estimating the degree of agreement between two raters.
CONCLUSION
Automated Essay Scoring (AES) is a software or service which grade essays with high
human rater agreement. The objective of AES is to reduce the cost of time and effort on
grading essays. In the current stage, the presence of AES cannot replace human raters in
essay grading tasks, but it can assist human raters as a second rater. Most of the AES are
developed in Western countries and there is still no commercialAES developed in theAsian
region. According to past literature, most studies focused on evaluating essays based on
its content by using Latent Semantic Analysis (LSA) technique. Moreover, most of the
reported implementations treat AES as a supervised document classification task. Hence,
this paper proposed three types of supervised general frameworks (content similarity,
machine learning and hybrid) based on the past literature and a new framework which
evaluates essays based on content and linguistic features. This review has also shown that
different AES research used different evaluation methods to examine the proposed model
performance, and this causes difficulty in doing a proper comparative study. Hence, this
paper proposes Quadratic weighted kappa (QWK) as a standard method to evaluate AES
performance and this can help to standardize the development of AES.
ACKNOWLEDGEMENTS
This study is supported by Universiti Malaysia Sarawak (UNIMAS), Prototype Research
Grant Scheme #F04/PRGS/1801/2019 and Exploratory Research Grant Scheme #ERGS/
ICT07(01)/1018/2013(15) by the Malaysia Ministry of Higher Education.
REFERENCES
Alghamdi, M., Alkanhal, M., Al-Badrashiny, M., Al-Qabbany, A., Areshey, A., & Alharbi, A. (2014). A hybrid
automatic scoring system forArabic essays. AI Communications, 27(2), 103-111. https://doi.org/10.3233/
aic-130586

1895
Alikaniotis, D., Yannakoudakis, H., & Rei, M. (2016). Automatic text scoring using neural networks. In K.
Erk & N. A. Smith (Eds.), Proceedings of the 54th Annual Meeting of the Association for Computational
Linguistics (Volume 1: Long Papers) (pp. 715-725). Association for Computational Linguistics. https://
doi.org/10.18653/v1/p16-1068
Al-Jouie, M., &Azmi,A. (2017).Automated evaluation of school children essays inArabic. Procedia Computer
Science, 117, 19-22. https://doi.org/10.1016/j.procs.2017.10.089
Amalia, A., Gunawan, D., Fithri, Y., & Aulia, I. (2019). Automated Bahasa Indonesia essay evaluation with
latent semantic analysis. Journal of Physics: Conference Series, 1235, Article 012100. https://doi.
org/10.1088/1742-6596/1235/1/012100
Attali, Y., & Burstein, J. (2006).Automated essay scoring with e-rater® V.2. Journal of Technology, Learning,
and Assessment, 4(3), 1-29.
Awaida, S. A., Shargabi, B. A., & Rousan, T. A. (2019). Automated Arabic essays grading system based on
F-score andArabic wordnet. Jordanian Journal of Computers and Information Technology (JJCIT), 5(3),
170-180. https://doi.org/10.5455/jjcit.71-1559909066
Chen, H., & He, B. (2013).Automated essay scoring by maximizing human-machine agreement. In Proceedings
of the 2013 Conference on Empirical Methods in Natural Language Processing (pp. 1741-1752).
Association for Computational Linguistics.
Chen, M., & Li, X. (2018). Relevance-based automated essay scoring via hierarchical recurrent model. In
2018 International Conference on Asian Language Processing (IALP) (pp. 378-383). IEEE Conference
Publication. https://doi.org/10.1109/ialp.2018.8629256
Chen, Z., & Zhou, Y. (2019). Research on automatic essay scoring of composition based on CNN and OR. In
2019 2nd International Conference on Artificial Intelligence and Big Data (ICAIBD) (pp. 13-18). IEEE
Conference Publication. https://doi.org/10.1109/icaibd.2019.8837007
Cheon, M., Seo, H. W., Kim, J. H., Noh, E. H., Sung, K. H., & Lim, E. (2015). An automated scoring tool
for Korean supply-type items based on semi-supervised learning. In Proceedings of the 2nd Workshop
on Natural Language Processing Techniques for Educational Applications (pp. 59-63). Association for
Computational Linguistics andAsian Federation of Natural Language Processing. https://doi.org/10.18653/
v1/w15-4409
Contreras, J. O., Hilles, S., & Abubakar, Z. B. (2018). Automated essay scoring with ontology based on text
mining and nltk tools. In 2018 International Conference on Smart Computing and Electronic Enterprise
(ICSCEE) (pp. 1-6). IEEE Conference Publication. https://doi.org/10.1109/icscee.2018.8538399
Darus, S., Stapa, S. H., & Hussin, S. (2003). Experimenting a computer-based essay marking system at
Universiti Kebangsaan Malaysia. Jurnal Teknologi, 39(E), 1-18. https://doi.org/10.11113/jt.v39.472
Davis, B. (2014). Essay grading computer mistakes gibberish for genius. RetrievedAugust 28, 2020, from http://
www.realclear.com/tech/2014/04/29/essay_grading_computer_mistakes_gibberish_for_genius_6784.html
Dikli, S. (2006).An overview of automated scoring of essays. Journal of Technology, Learning, and Assessment
(JTLA), 5(1), 1-35.

Dong, F., & Zhang, Y. (2016). Automatic features for essay scoring–an empirical study. In Proceedings of the
2016 conference on empirical methods in natural language processing (pp. 1072-1077). Association for
Computational Linguistics. https://doi.org/10.18653/v1/D16-1115
Elliot, S. M. (2003). IntelliMetric: From here to validity. In M. D., Shermis & J. C. Burstein (Eds.), Automated
essay scoring: A cross-disciplinary perspective (pp. 71-86). Routledge
Fazal, A., Dillon, T., & Chang, E. (2011). Noise reduction in essay datasets for automated essay grading. In
OTM Confederated International Conferences On the Move to Meaningful Internet Systems (pp. 484-493).
Springer. https://doi.org/10.1007/978-3-642-25126-9_60
Foltz, P. W., Laham, D., & Landauer, T. K. (1999). The intelligent essay assessor: Applications to educational
technology. Interactive Multimedia Electronic Journal of Computer-Enhanced Learning, 1(2), 939-944.
Ghosh, S., & Fatima, S. S. (2008). Design of an automated essay grading (AEG) system in Indian context. In
TENCON 2008-2008 IEEE Region 10 Conference (pp. 1-6). IEEE Conference Publication. https://doi.
org/10.1109/tencon.2008.4766677
Greene, P. (2018). Automated essay scoring remains an empty dream. Retrieved September 8, 2020, from
Forbes: https://www.forbes.com/sites/petergreene/2018/07/02/automated-essay-scoring-remains-an-
empty-dream/#4474e4f74b91
Hussein, M. A., Hassan, H., & Nassef, M. (2019). Automated language essay scoring systems: A literature
review. PeerJ Computer Science, 5, Article e208 https://doi.org/10.7287/peerj.preprints.27715v1
Imaki, J., & Ishihara, S. (2013). Experimenting with a Japanese automated essay scoring system in the L2
Japanese environment. Papers in Language Testing and Assessment, 2(2), 28-47.
Ishioka, T., & Kameda, M. (2006). Automated Japanese essay scoring system based on articles written by
experts. In Proceedings of the 21st International Conference on Computational Linguistics and 44th Annual
Meeting of the Association for Computational Linguistics (pp. 233-240). Association for Computational
Linguistics. https://doi.org/10.3115/1220175.1220205
Islam, M. M., & Hoque,A. L. (2013).Automated Bangla essay scoring system:ABESS. In 2013 International
Conference on Informatics, Electronics and Vision (ICIEV) (pp. 1-5). IEEE Conference Publication.
https://doi.org/10.1109/iciev.2013.6572694
Jin, C., & He, B. (2015). Utilizing latent semantic word representations for automated essay scoring. In
2015 IEEE 12th Intl Conf on Ubiquitous Intelligence and Computing and 2015 IEEE 12th Intl Conf
on Autonomic and Trusted Computing and 2015 IEEE 15th Intl Conf on Scalable Computing and
Communications and Its Associated Workshops (UIC-ATC-ScalCom) (pp. 1101-1108). IEEE Conference
Publication. https://doi.org/10.1109/UIC-ATC-ScalCom-CBDCom-IoP.2015.202
Jin, C., He, B., Hui, K., & Sun, L. (2018). TDNN: a two-stage deep neural network for prompt-independent
automated essay scoring. In Proceedings of the 56th Annual Meeting of the Association for Computational
Linguistics (Volume 1: Long Papers) (pp. 1088-1097).Association for Computational Linguistics. https://
doi.org/10.18653/v1/p18-1100
Kakkonen, T., Myller, N., Timonen, J., & Sutinen, E. (2005).Automatic essay grading with probabilistic latent
semantic analysis. In Proceedings of the second workshop on Building Educational Applications Using
NLP (pp. 29-36). Association for Computational Linguistics.

1897
Landauer, T. K., Foltz, P. W., & Laham, D. (1998). Introduction to latent semantic analysis. Discourse
Processes, 25, 259-284.
Landauer, T. K., Laham, D., & Foltz, P. W. (2000). The intelligent essay assessor. IEEE Intelligent Systems,
15, 27-31.
Liang, G., On, B. W., Jeong, D., Kim, H. C., & Choi, G. S. (2018). Automated essay scoring: A siamese
bidirectional LSTM neural network architecture. Symmetry, 10(12), Article 682. https://doi.org/10.3390/
sym10120682
Loraksa, C., & Peachavanish, R. (2007). Automatic Thai-language essay scoring using neural network and
latent semantic analysis. In First Asia International Conference on Modelling & Simulation (AMS’07)
(pp. 400-402). IEEE Conference Publication. https://doi.org/10.1109/ams.2007.19
Malaysian Examinations Council. (2014). Malaysian university English test (MUET): Regulations, test
specifications, test format and sample questions. Retrieved March 15, 2021, from https://www.mpm.edu.
my/images/dokumen/calon-peperiksaan/muet/regulation/Regulations_Test_Specifications_Test_Format_
and_Sample_Questions.pdf
Measurement Incorporated. (2020). Automated essay scoring. Retrieved June 4, 2020, from https://www.
measurementinc.com/products-services/automated-essay-scoring
Nguyen, H., & Dery, L. (2016). Neural networks for automated essay grading. CS224d Stanford Reports, 1-11.
Omar, N., & Mezher, R. (2016). A hybrid method of syntactic feature and latent semantic analysis for
automatic Arabic essay scoring. Journal of Applied Sciences, 16(5), 209-215. https://doi.org/10.3923/
jas.2016.209.215
Ong, D. A., Razon, A. R., Guevara, R. C., & Prospero C. Naval, J. (2011, November 24-25). Empirical
comparison of concept indexing and latent semantic indexing on the content analysis of Filipino essays.
In Proceedings of the 8th National Natural Language Processing Research Symposium (pp. 40-45). De
La Salle University, Manila.
Page, E. B. (1966). The imminence of... grading essays by computer. Phi Delta Kappan, 47(5), 238-243.
Page, E. (2003). Project essay grade: PEG. In M. Shermis, & J. Burstein (Eds.), Automated essay scoring: A
cross-disciplinary perspective (pp. 43-54). Lawrence Erlbaum Associates Publishers.
Pai, K. C., Lu, Y., & Kuo, B. C. (2017). Developing Chinese automated essay scoring model to assess college
students’essay quality. In Proceedings of the 10th International Conference on Educational Data Mining
(pp. 430-432).
Pearson. (2010). Intelligent essay assessor (IEA)™ fact sheet. Pearson Education. Retrieved June 4, 2020,
from https://images.pearsonassessments.com/images/assets/kt/download/IEA-FactSheet-20100401.pdf
Peng, X., Ke, D., Chen, Z., & Xu, B. (2010). Automated Chinese essay scoring using vector space models. In
2010 4th International Universal Communication Symposium (pp. 149-153). IEEE Conference Publication.
https://doi.org/10.1109/IUCS.2010.5666229
Phandi, P., Chai, K. M. A., & Ng, H. T. (2015). Flexible domain adaptation for automated essay scoring using
correlated linear regression. In Proceedings of the 2015 Conference on Empirical Methods in Natural

Language Processing (pp. 431-439).Association for Computational Linguistics. https://doi.org/10.18653/
v1/D15-1049
Ramalingam, V. V., Pandian, A., Chetry, P., & Nigam, H. (2018). Automated essay grading using machine
learning algorithm. In Journal of Physics: Conference Series (Vol. 1000, No. 1, p. 012030). IOPPublishing.
https://doi.org/10.1088/1742-6596/1000/1/012030
Ramineni, C., & Williamson, D. (2018). Understanding mean score differences between the e‐rater® automated
scoring engine and humans for demographically based groups in the GRE® general test. ETS Research
Report Series, 2018(1), 1-31. https://doi.org/10.1002/ets2.12192
Ratna, A. A. P., Budiardjo, B., & Hartanto, D. (2007). SIMPLE: System automatic essay assessment for
Indonesian language subject examination. Makara Journal of Technology, 11(1), 5-11.
Ratna,A.A. P., Purnamasari, P. D., &Adhi, B.A. (2015). SIMPLE-O, the Essay grading system for Indonesian
Language using LSA method with multi-level keywords. In The Asian Conference on Society, Education
& Technology 2015 (pp. 155-164). The International Academic Forum.
Ratna, A. A. P., Arbani, A. A., Ibrahim, I., Ekadiyanto, F. A., Bangun, K. J., & Purnamasari, P. D. (2018).
Automatic essay grading system based on latent semantic analysis with learning vector quantization
and word similarity enhancement. In Proceedings of the 2018 International Conference on Artificial
Intelligence and Virtual Reality (pp. 120-126). Association for Computing Machinery. https://doi.
org/10.1145/3293663.3293684
Ratna, A. A. P., Kaltsum, A., Santiar, L., Khairunissa, H., Ibrahim, I., & Purnamasari, P. D. (2019a). Term
frequency-inverse document frequency answer categorization with support vector machine on automatic
short essay grading system with latent semantic analysis for japanese language. In 2019 International
Conference on Electrical Engineering and Computer Science (ICECOS) (pp. 293-298). IEEE Conference
Publication. https://doi.org/10.1109/ICECOS47637.2019.8984530
Ratna,A.A. P., Khairunissa, H., Kaltsum,A., Ibrahim, I., & Purnamasari, P. D. (2019b).Automatic essay grading
for Bahasa Indonesia with support vector machine and latent semantic analysis. In 2019 International
Conference on Electrical Engineering and Computer Science (ICECOS) (pp. 363-367). IEEE Conference
Publication. https://doi.org/10.1109/ICECOS47637.2019.8984528
Ratna, A. A. P., Santiar, L., Ibrahim, I., Purnamasari, P. D., Luhurkinanti, D. L., & Larasati, A. (2019c).
Latent semantic analysis and winnowing algorithm based automatic Japanese short essay answer
grading system comparative performance. In 2019 IEEE 10th International Conference on Awareness
Science and Technology (iCAST) (pp. 1-7). IEEE Conference Publication. https://doi.org/10.1109/
ICAwST.2019.8923226
Rudner, L., & Gagne, P. (2001). An overview of three approaches to scoring written essays by computer.
Practical Assessment, Research & Evaluation, 7, Article 26.
Sendra, M., Sutrisno, R., Harianata, J., Suhartono, D., & Asmani, A. B. (2016). Enhanced latent semantic
analysis by considering mistyped words in automated essay scoring. In 2016 International Conference on
Informatics and Computing (ICIC) (pp. 304-308). IEEE Conference Publication. https://doi.org/10.1109/
IAC.2016.7905734

1899
Shehab, A., Faroun, M., & Rashad, M. (2018). An automatic Arabic essay grading system based on text
similarity Algorithms. International Journal of Advanced Computer Science and Applications, 9(3),
263-268. https://doi.org/10.14569/IJACSA.2018.090337
Shermis, M. D., & Burstein, J. (2003). Automated essay scoring: A cross-disciplinary perspective. Lawrence
Erlbaum Associates Publishers.
Shermis, M. D., Burstein, J., Higgins, D., & Zechner, K. (2010).Automated essay scoring: Writing assessment
and instruction. International Encyclopedia of Education, 4(1), 20-26. https://doi.org/10.1016/B978-0-
08-044894-7.00233-5
Sim, J., & Wright, C. C. (2005). The Kappa statistic in reliability studies: Use, interpretation, and sample size
requirements. Physical Therapy, 85(3), 257-268.
Taghipour, K., & Ng, H. T. (2016). A neural approach to automated essay scoring. In Proceedings of the 2016
Conference on Empirical Methods in Natural Language Processing (pp. 1882-1891). Association for
Computational Linguistics. https://doi.org/10.18653/v1/d16-1193
Tanha, J., Abdi, Y., Samadi, N., Razzaghi, N., & Asadpour, M. (2020). Boosting methods for multi-class
imbalanced data classification: an experimental review. Journal of Big Data, 7(1), 1-47. https://doi.
org/10.1186/s40537-020-00349-y
Vantage Learning. (2005). How IntelliMetricTM
works. Retrieved June 4, 2020, from http://www.vantagelearning.
com/docs/intellimetric/IM_How_IntelliMetric_Works.pdf
Vantage Learning. (2020). IntelliMetric®: Frequently asked questions. Retrieved June 4, 2020, from http://
www.vantagelearning.com/products/intellimetric/faqs/#LongUsed
Wong, W. S., & Bong, C. H. (2019). A study for the development of automated essay scoring (AES) in
Malaysian English test environment. International Journal of Innovative Computing, 9(1), 69-78. https://
doi.org/10.11113/ijic.v9n1.220
Xu,Y., Ke, D., & Su, K. (2017). Contextualized latent semantic indexing:Anew approach to automated Chinese
essay scoring. Journal of Intelligent Systems, 26(2), 263-285. https://doi.org/10.1515/jisys-2015-0048
Zupanc, K., & Bosnic, Z. (2015). Advances in the field of automated essay evaluation. Informatica, 4(39),
383-396.

View publication stats
View publication stats

A Comprehensive Review Of Automated Essay Scoring (AES) Research And Development

More Related Content

Similar to A Comprehensive Review Of Automated Essay Scoring (AES) Research And Development (20)

More from Anna Landers (20)

Recently uploaded (20)

A Comprehensive Review Of Automated Essay Scoring (AES) Research And Development