摘要
There are several ethical issues that have arisen in recent years due to the ubiquity of the Internet and the popularity of social media and community platforms.Among them is cyberbullying,which is defined as any violent intentional action that is repeatedly conducted by individuals or groups using online channels against victims who are not able to react effectively.An alarmingly high percentage of people,especially teenagers,have reported being cyberbullied in recent years.A variety of approaches have been developed to detect cyberbullying,but they require time-consuming feature extraction and selection processes.Moreover,no approach to date has examined the meanings of words and the semantics involved in cyberbullying.In past work,we proposed an algorithm called Cyberbullying Detection Based on Deep Learning(CDDL)to bridge this gap.It eliminates the need for feature engineering and generates better predictions than traditional approaches for detecting cyberbullying.This was accomplished by incorporating deep learning—specifically,a convolutional neural network(CNN)—into the detection process.Although this algorithm shows remarkable improvement in performance over traditional detection mechanisms,one problem with it persists:CDDL requires that many parameters(filters,kernels,pool size,and number of neurons)be set prior to classification.These parameters play a major role in the quality of predictions,but a method for finding a suitable combination of their values remains elusive.To address this issue,we propose an algorithm called firefly-CDDL that incorporates a firefly optimisation algorithm into CDDL to automate the hitherto-manual trial-and-error hyperparameter setting.The proposed method does not require features for its predictions and its detection of cyberbullying is fully automated.The firefly-CDDL outperformed prevalent methods for detecting cyberbullying in experiments and recorded an accuracy of 98%within acceptable polynomial time.
