Search

Automated Short Answer Grading (ASAG) systems contribute immensely to providing prompt feedback to students, which eases the workload of instructors. This research focuses on the development of an optimized ASAG model using LSTM model and particle swarm optimization techniques to prevent model overfitting. The popular ASAG dataset by Mohler was utilized for the experiment. The dataset contains training samples from Computer Science department of the Federal University of Technology, Minna, Nigeria, with grades between 0 and 5. In order to effectively optimize the LSTM model parameters, which are learning rate and number of neurons in the LSTM layers, four experiments were performed, each with different particle population sizes (5, 10, 15 and 20). The results show that PS5 model produced the lowest RMSE and MAPE of 0.77697 and 44.5356%, respectively. The PS15 model, however, produced the highest RMSE and MAE of 0.80985 and 56.6192%, respectively. In order to validate the developed PSO-LSTM ASAG model, normal LSTM model for ASAG was implemented and tested. The PSO-LSTM has an RMSE value of 0.77687 and MAPE of 44.5356%, as compared with LSTM, which has an RMSE value of 0.9423 and MAPE of 85.73%. The results clearly show the superiority of the developed hybrid model in predicting the scores of short answer grading. The model's performance can be further improved by increasing the sample size and using other optimization algorithms, such as genetic algorithms or ant colony optimization. Further research can also investigate the effect of other variables, such as question complexity and student writing style, on the model's performance. [ABSTRACT FROM AUTHOR]

The increasing adoption of electronic emails as a means of communication, both at the commercial, government, and individual levels, serves as an impetus for attackers to compromise communication. Consequently, numerous machine learning techniques have been developed for identifying unwanted emails, commonly known as spam. Despite the significant progress reported in existing literature, most studies do not integrate the detection of both textual and image-based spam. In this paper, two deep learning techniques that detect both textual and image-based spam were evaluated. First, the Recurrent Neural Network (RNN) and Convolutional Neural Network (CNN) is studied, training them on various text-image features to explore their effectiveness on an improved dataset. Subsequently, in an effort to outsmart current spam detection techniques, a bi-modal architecture capable of detecting textual spam, image spam, and mixed spam is designed. The experimental results in conjunction with existing transfer learning for effective spam detection is provided. [ABSTRACT FROM AUTHOR]

Anomaly detection in motor bearings is a critical task for preventing downtime and ensuring efficient operation. This paper proposes a novel approach for anomaly detection using Fast Fourier Transform (FFT) and Long Short-Term Memory (LSTM)-Autoencoder (AE). A data processing approach based on FFT was developed to pre-process the raw sensor data. This helped to reduce noise and improve the Signal-to-Noise Ratio (SNR). Additionally, an anomaly detection model based on LSTM-Autoencoder was developed and trained on the pre-processed data. The proposed approach was able to detect anomalies at a low threshold and achieved a high accuracy score. [ABSTRACT FROM AUTHOR]

In recent times, big Internet companies have come under increased pressure from governments and NGOs to remove inappropriate materials from social media platforms (e.g., Twitter, Facebook, YouTube). A typical example of this problem is the posting of hateful, abusive, and violent tweets on Twitter which has been blamed for inciting hatred, violence and causing societal disturbances. Manual identification of such tweets and the people who post these tweets is very difficult because of the large number of active users and the frequency with which such tweets are posted. Existing approaches for identifying inappropriate tweets have focused on the detection of such tweets without identifying the users who post them. This paper proposes an approach that can automatically identify different types of inappropriate tweets together with the users who post them. The proposed approach is based on a user profiling algorithm that uses a deep Long Short-Term Memory (LSTM) based neural network trained to detect abusive language. With the support of word embedding features learned from the training set, the algorithm is able to classify the tweets of users into different abusive language categories. Thereafter, the user profiling algorithm uses the classes assigned to the tweets of each user to profile each user into different abusive language category. Experiments on the test set show that the deep LSTM-based abusive language detection model reached an accuracy of 89.14% on detecting whether a tweet is bigotry, offensive, racist, extremism-related and neutral. Also, the user profiling algorithm obtained an accuracy of 83.33% in predicting whether a user is a bigot, racist, extremist, uses offensive language and neutral. [ABSTRACT FROM AUTHOR]

Twitter is a social networking platform that has become popular in recent years. It has become a versatile information dissemination tool used by individuals, businesses, celebrities, and news organizations. It allows users to share messages called tweets with one another. These messages can contain different types of information from personal opinions of users, advertisement of products belonging to all kinds of businesses to the news. Tweets can also contain messages that are racist, bigotry, offensive, and of extremist views as shown by research. Manual identification of such tweets is impossible as hundreds of millions of tweets are posted every day and hence a solution to automate the identification of these types of tweets through classification is required for the Twitter administrators or an intelligence and security analyst. This paper presents a comparative study of traditional machine learning algorithms and deep learning algorithms for the task of tweet classification to detect different categories of abusive languages with the aim to determine which algorithm performs best in detecting abusive language that is prevalent on social media. Two approaches for building feature vectors were explored. Feature vectors based on the bag-of-words method and feature vectors based on word embeddings. These two methods of feature representation were evaluated in this paper using tweet messages representing five abusive language categories. The experiments show that the deep learning algorithms trained with word embeddings outperformed all the other machine learning algorithms that were trained with feature vectors based on the bag-of-words approach. [ABSTRACT FROM AUTHOR]

One of the main approaches for personalization of activity recognition is the generation of the classification model from user annotated data on mobile itself. However, giving the resource constraints on such devices there is a need to examine the effects of system parameters such as the feature extraction parameter that can affect the performance of the system. Thus, this paper examines the effects of window length of the sensor data and varying data set sizes on the classification accuracy of four selected supervised machine learning algorithms running on off the shelf smartphone. Our results show that out of the three window lengths of 32, 64 and 128 considered, the 128 window length yields the best accuracy for all the algorithms tested. Also, the time taken to train the algorithms with samples of this length is minimal compare to 64 and 32 window lengths. A smartphone based activity recognition is implemented to utilize the results in an online activity recognition scenario. [ABSTRACT FROM AUTHOR]