Various automated essay scoring (AES) methods have been proposed over the past five decades, but the application of AES in the educational field has gained popularity since the COVID-19 pandemic, as most educational institutions have shifted to online teaching modes. Consequently, the assessment of student knowledge has become a major challenge. Researchers are focusing on new state-of-the-art techniques to devise a more performant AES to facilitate online grading tasks. However, few studies have analyzed the common features of existing AES. There are no general guiding principles for the implementation and improvement of AES. This work aims to address the research gap by proposing a unified conceptual hybrid framework for AES, adapted for short answers and inspired by an in-depth analysis of existing AES based on short answers. The unified framework consists mainly of the most frequently used components in existing AES except that a new important module namely the data augmentation module has been identified and added to the framework and also two existing modules have been modified. Following the proposed unified framework, the different essential components in an AES can be easily understood, and the implementation and improvement of AES can be achieved effortlessly using a hybrid approach. Furthermore, experiments have been carried out to validate the framework’s performance. During the experimentation, single, ensemble, and hybrid models were compared with and without the data augmentation technique respectively. Results confirmed that the data augmentation module did help in improving results for all the models. Notably, in the hybrid models, results demonstrated an average increase of 40.75

Automated Essay Scoring (AES) systems are computer-based tools to evaluate students’ responses. While ensemble and hybrid models have demonstrated effectiveness in various Natural Language Processing (NLP) tasks, their specific applications within AES remain underexplored. This study addresses this gap through a systematic literature review. Guided by the Kitchenham methodology, we formulated five research questions and conducted a structured search and evaluation process. Twenty-five (25) articles published between 2004 and 2024 were included based on predefined inclusion and quality criteria. The review identifies two types of ensemble models and outlines diverse hybrid model techniques used in AES. It highlights the models’ effectiveness, limitations, dataset types, evaluation metrics, and the best-performing approaches. This review provides researchers with insights into trends, challenges, and opportunities in applying ensemble and hybrid techniques for AES, thereby supporting the development of performance-oriented AES systems.

PurposeThe performance of Wireless Sensor Networks (WSNs) applications is bounded by the limited resources of battery-enabled Sensor Nodes (SNs), which include energy and computational power; the combination of which existing research seldom focuses on. Although bio-inspired algorithms provide a way to control energy usage by finding optimal routing paths, those which converge slower require even more computational power, which altogether degrades the overall lifetime of SNs.Design/methodology/approachHence, two novel routing protocols are proposed using the Red-Deer Algorithm (RDA) in a WSN scenario, namely Horizontal PEG-RDA Equal Clustering and Horizontal PEG-RDA Unequal Clustering, to address the limited computational power of SNs. Clustering, data aggregation and multi-hop transmission are also integrated to improve energy usage. Unequal clustering is applied in the second protocol to mitigate the hotspot problem in Horizontal PEG-RDA Equal Clustering.FindingsComparisons with the well-founded Ant Colony Optimisation (ACO) algorithm reveal that RDA converges faster by 85 and 80% on average when the network size and node density are varied, respectively. Furthermore, 33% fewer packets are lost using the unequal clustering approach which also makes the network resilient to node failures. Improvements in terms of residual energy and overall network lifetime are also observed.Originality/valueProposal of a bio-inspired algorithm, namely the RDA to find optimal routing paths in WSN and to enhance convergence rate and execution time against the well-established ACO algorithm. Creation of a novel chain cluster-based routing protocol using RDA, named Horizontal PEG-RDA Equal Clustering. Design of an unequal clustering equivalent of the proposed Horizontal PEG-RDA Equal Clustering protocol to tackle the hotspot problem, which enhances residual energy and overall network lifetime, as well as minimises packet loss.

Edge machine learning brings intelligence to low-power devices at the periphery of a network. By running machine learning algorithms on the Edge, classification can be performed faster without the need to transmit large data volumes across a network. However, on-device training is often not feasible since Edge devices have limited computing and storage resources. Improved, Scalable, Efficient, and Fast classifieR (iSEFR) is a classifier that performs both training and testing on low-power devices using linearly separable balanced datasets. The novelty of this work is the improvement of the iSEFR accuracy by fine-tuning the algorithm with datasets having an uneven class distribution. Three adaptive linear function transformation techniques were proposed to improve the decision threshold which is in the form of a linear function. Experiments using stratified sampling with 5-fold cross-validation demonstrate that one of the proposed techniques significantly improved F1-score, Recall and Matthews Correlation Coefficient (MCC) by an average of 23 %, 35 % and 21 % compared to iSEFR. Further evaluation of this technique in a Fog environment using highly imbalanced datasets such as credit card fraud, network intrusion and diabetic retinopathy also showed a significant increase of 38 %, 44 % and 30 % in F1-score, Recall and MCC with a Precision of 97 %. The adaptive binary classifier maintained the time complexity of iSEFR without altering the class imbalance.

In machine learning, multiclass classification is the process of classifying data that contains three or more classes. One-vs-Rest and One-vs-One are two popular binary transformation approaches that reduce the multiclass problem into several binary problems. One-vs-Rest is usually faster and less complex than One-vs-One and is therefore the most commonly used strategy. The main contribution of this paper is the formulation of a novel One-vs-Next method for multiclass classification. The proposed approach consists in fitting one classifier per next class pair. This technique reduces the number of binary combinations and is expected to be faster since it requires to fit less number of classifiers compared to the One-vs-Rest approach. As a proof of concept, One-vs-Next was evaluated with the improved Scalable, Efficient, and Fast classifieR (iSEFR) designed for embedded machine learning on the Edge. In addition to its computational efficiency, experiments demonstrate that the proposed approach combined with the iSEFR decision threshold gives an average increase of similar to 20% in accuracy for a multiclass dataset with equal number of class instances.

The extraction of information from unstructured documents is performed in a wide range of business processes including Know Your Customer (KYC). Such business processes require significant human resources and the main blocking factor to automate such processes is that these documents are unstructured. In other words, these documents do not follow a pre-defined layout, and the layout varies from client to client. It is beyond the control of the business process owners to predict the layout of the documents to be received for the business process and the sought data from such documents are not located in the same area. Commercially available document processing engines such as AI Builder require collections of all possible layouts during training. Otherwise, the performance parameters (recall, precision and F1 score) do not meet the benchmark for automation. In this study, we propose a Named Entity Recognition (NER) based solution to seamlessly extract key information from unstructured documents irrespective of their layouts. The unstructured documents used are annual account documents of UK-based companies and the extracted key information are Company Name, Company Number and Accounting year. The documents for the training and testing dataset were downloaded from the Companies House web portal leveraging Power Automate Desktop. The developed NER-based architecture was successfully trained leveraging the created dataset and the model achieved an impressive precision of 91 %, a recall of 97% and F1 score of 94%. The output from the trained model was compared with the ones from a trained AI Builder model. In contrast to the trained AI Builder model, the trained NER model performed better, and it must be noted that the developed architecture utilized solely open-source frameworks and libraries.

This paper investigates the location estimation accuracy of an indoor localization system using both the Received Signal Strength Indicator (RSSI) and Angle of Arrival (AoA) values from the User Equipment (UE). The Quadriga channel model was used to simulate the 5G indoor scenario, operating in the mm-Wave bands and simulations were conducted at two specific frequencies, 2 GHz and 28 GHz. Results showed that the proposed weighted hybrid RSSI/AoA localization system gave a horizontal localization accuracy of 1.76 m. The proposed system showed an improvement of 56 % compared to a TDoA-based approach. However, at 28 GHz, a horizontal localization error of 11.5 m was obtained. Hence, at this frequency, additional localization measurements and more accurate channel modelling are required.

Due to its boundless potential applications, Wireless Sensor Networks have been subject to much research in the last two decades. WSNs are often deployed in remote environments making replacement of batteries not feasible. Low energy consumption being of prime requisite led to the development of energy-efficient routing protocols. The proposed routing algorithms seek to prolong the lifetime of sensor nodes in the relatively unexplored area of 3D WSNs. The schemes use chain-based routing technique PEGASIS as basis and employ genetic algorithm to build the chain instead of the greedy algorithm. Proposed schemes will incorporate an energy and distance aware CH selection technique to improve load balancing. Clustering of the network is also implemented to reduce number of nodes in a chain and hence reduce delay. Simulation of our proposed protocols is carried out for homogeneous networks considering separately cases for a static base-station inside and outside the network. Results indicate considerable improvement in lifetime over PEGASIS of 817% and 420% for base station inside and outside the network respectively. Residual energy and delay performance are also considered.

Automated scoring of short answers is becoming increasingly important nowadays, as more classes are taking place online due to COVID-19 outbreaks. Over the past 50 years, different techniques have been applied to develop ways of assessing short answers automatically. The automated assessment of short answer questions has been used with varying kinds of questions in different domains. However, no interoperable tools have been developed to facilitate the automated assessment of short answers for common modules across educational institutions. In this paper, we present a novel system called OBMAAQ, which grades short answers for questions designed according to Bloom's revised taxonomy remember category. The proposed system uses publicly available domain ontologies to grade short answers in different subject areas. This method is unique as it does not use model answers since it relies solely on the domain ontologies by utilizing an ontology-based pruning technique to grade the students' responses. The purpose of the system is to allow teachers from different educational institutions to use a similar approach when assessing the short answers for common modules automatically. Two publicly available domain ontologies in the field of security and computer networks have been used to validate students' answers. Experimental results based on both datasets indicate some promising results even though it is less than the set thresholds. This paper discusses the lessons learned and future recommendations based on the OBMAAQ model.

A good transportation system is essential to modern society. Unfortunately, road networks deteriorate rapidly because of a range of reasons, namely wear and tear caused by the continuous and heavy flow of vehicles, weather elements such as heavy rain and high temperature and various excavation works that are carried out on the roads. Bad roads lead to accidents, damage to vehicles and increased fuel consumption. Therefore, the monitoring of road quality is crucial. This paper proposes a low-cost prototype that can be implemented in any vehicle. The device collects and transmits data on road quality to a cloud platform, where analytics is performed to assess and classify road condition. The pothole locations are viewable on a web page. While travelling, drivers receive real-time sound alerts on their mobile phones to warn them of imminent degradation in the road network, such as potholes. An Android application has been developed to synchronize with the database via the API and send real-time and up-to-date notifications to road users whenever they approach such areas. Authorities can also use the system to grasp the quality of the road network and automatically identify problems that may require immediate repair.

Two important criteria of wireless body area networks (WBANs) are low power consumption and delay. These criteria can be met by designing efficient medium access control (MAC) protocols. In this paper, two TDMA-based MAC protocols are proposed. The first protocol, TM–MAC makes use of only a main radio. The second proposed protocol, TWM–MAC makes use of a WUR alongside the main radio. The two proposed protocols are compared with different categories of standard MAC protocols and it is shown that they outperform the standard ones by improving the power consumption and delay. The TWM–MAC consumes 55% less power consumption than the Scheduled Channel Polling MAC (SCP–MAC) protocol for a high traffic scenario on the high-rate platform while the TM-MAC consumes 85% less power consumption than the SCP–MAC. For a low traffic scenario, the TWM–MAC performs 53.5% better than the SCP–MAC protocol and 77.5% better than the very low power MAC (VLPM) protocol on the high and low-rate platforms respectively. An improvement in delay was observed with the TWM–MAC protocol for high traffic situations. The TWM–MAC protocol surpasses the VLPM protocol by 81.1% in terms of latency for a high traffic scenario and 3.2% for a low traffic scenario.

A cell outage compensation algorithm based on Received Signal Strength Indicator (RSSI) for a three-tier hetrogeneous network (HetNet) is proposed in this paper. The algorithm is non-machine learning based to reduce the complexity of the compensation scheme, and to eliminate the need for training. Simulation results show that cell outage compensation is successfully achieved, provided that base stations (BSs) of sufficient capacity are deployed near users in outage. The RSSI values after compensation are also higher than those during the outage. Additionally, the proposed algorithm outperforms a k-means clustering scheme when allocating users in outage to neighbouring BSs.

The unprecedented advances in UAV (unmanned aerial vehicle) technology have allowed the deployment of drones for various communication related applications for the upcoming fifth generation of mobile networks. Similar to the actual cell towers, UAV base stations are expected to serve ground users continuously, which is not currently possible due to the limited flight time of the UAVs, a result of the finite battery capacity. This paper proposes a seamless handover mechanism for the swapping of UAVs acting as base stations based on an RSS decision algorithm. The proposed model ensures that no downtime is introduced during the swapping of UAVs, creating the impression that the UAVs fly forever.

In automated essay scoring (AES) systems, similarity techniques are used to compute the score for student answers. Several methods to compute similarity have emerged over the years. However, only a few of them have been widely used in the AES domain. This work shows the findings of a ten-year review on similarity techniques applied in AES systems and discusses the efficiency and limitations of current methods. In the final review, thirty-four (34) articles published between 2010 and 2020 were included. The metrics used to evaluate the performance of the AES systems are also elaborated. The review was conducted using the Kitchenham method, whereby three research questions were formulated and a search strategy was developed. Research papers were chosen based on pre-defined inclusion and quality assessment criteria. This review has identified two types of similarity techniques used in AES systems. In addition, several methods were used to compute the score for student answers in the AES systems. The similarity computation in AES systems is dependent on several factors, hence many studies have combined multiple methods in a single system yielding good results. In addition, the review found that the quadratic weighted kappa (QWK) was most frequently used to evaluate AES systems.

The field of natural language processing is being revolutionized with transformers. The latter is based on a novel type of neural network framework that is already pre-trained. Hence, large datasets to train models are no longer required. This framework is suitable for automated assessment systems (AAS), where a large number of labeled data is needed. The larger the dataset, the higher the accuracy of the AAS. In this work-in-progress paper, a prototype for an AAS has been built where two transformer models, namely the Sentence-Transformers from hugging face and the OpenAI GPT-3 models have been used. The transformer models generate the similarity index between students’ answers and reference answers from the Texas dataset. Then the similarity index is used to compute marks for students. The performance of the prototype is evaluated using the quadratic weighted kappa metric.

Internet of Things (IoT) is the future of communication technology and its applications range from smart cities, health, transport to agriculture among others. Embedded systems, networks, sensors, devices, communication protocols and software applications contribute to enable connectivity between devices located at the network boundary. Edge computing enables data processing on devices with limited hardware capabilities. Software engineering practices could be applied in the development of IoT systems to ensure applications are reliable and operate efficiently. However, there are no proper guidelines about which best software engineering practices might be applied to IoT systems. The main contribution of this study is to investigate the application of software design pattern in an IoT environment in terms of resource and power consumption of battery-powered Edge devices. Two variations of the Singleton design pattern have been implemented in an IoT computing scenario. Experiments demonstrate that the Singleton design pattern can reduce processor, memory, power and battery usage of the device by an average of 0.3%, 0.1%, 0.7% and 0.8% respectively.

Internet of Things , Edge Computing and 5G networks are rapidly becoming key enablers for a wide range of new services. This new infrastructure opens more possibilities for devices with embedded systems and microcontrollers to perform data processing and analysis. The need for real-time decisions now defeats the purpose of sending sensor data to the Cloud for further processing. Storing and analyzing data near its source introduce new constraints on resource-constrained and battery-powered devices. Scalable, Efficient, and Fast classifieR (SEFR) is one such algorithm that brings machine learning to low-power microcontrollers. The main contribution of this work is the definition of a new decision boundary, also known as the bias, to improve the accuracy of the SEFR binary classifier. Experiments performed on an 8-bit Arduino microcontroller using 5-fold cross-validation demonstrate that the improved algorithm (iSEFR) increases Precision, Recall and Accuracy by an average of 9%, 14% and 11% respectively . F1-score increases on average from 80% to 92%, which represents an increase of 12% in the model accuracy for even class distribution. MCC also increases from 64% to 85% by an average of 21%, and this represents a better binary classification quality. Training time and space complexity remain constant but the testing time increases marginally by an average of 0.0013 seconds for low-capacity processors. Moreover, iSEFR performs better on a 64-bit intel i5 processor with thousands of data points and features. Experiments also demonstrate that the accuracy of iSEFR is comparable to Support Vector Machine with a linear kernel.