DESIGN AND DEVELOPMENT OF FRAMEWORK FOR DETECTION OF HATE CONTENT

Abstract

Due to the widespread usage of social media platforms, an alarming problem has emerged in an era characterised by the rapid spread of hateful contents in any of the multimedia formats. The combination of the simplicity and complexity of these innovations presents a substantial risk to the clean and reliable conversations. This thesis emphasises the necessity to create novel systems for detecting hate content by using the potential of machine learning and deep learning techniques. The susceptibility of multimodal content towards hatefulness has significantly increased, reaching unprecedented levels. This results in implementing modern technologies that facilitate the production of counterfeits with a high degree of authenticity. The objective of this study is to leverage the capabilities of machine and deep learning to identify and mitigate hateful content effectively. Given that social media platforms are the main channels for sharing information, the suggested detection systems utilising machine learning and deep learning aim to ensure the mitigation of hate content detection. As a result, this will enhance the establishment of a digital ecosystem characterised by increased reliability and credibility. This thesis tackles this detection challenge by proposing four novel architectures. The first two techniques are dedicated to the problem of tacking the textual hate content in an efficient manner. In the first approach, it is seen that reducing features using Truncated SVD along with hyper parameter tuning helped in increasing balanced accuracy and F1 score for algorithms like Logistic Regression, SVM and XGBoost when compared to the baseline results. Still, the proposed approach is lacking in handling uncertain or imprecise data. The second approach focuses on handling the uncertainty and vagueness in the data by implementing the fuzzy classifiers. An empirical evaluation of seven classifiers is presented for hate speech detection on two commonly used benchmarks of different data characteristics providing essential insights into their detection in terms of accuracy for their deployment in real-world applications. Fuzzy classifiers outperformed the other two classifiers out of the three. Next two models are dedicated to the multimodal hate content detection. The first framework presents a dual-branch network which is composed of knowledge distillation attention for extracting the essential information from the caption modality and multi-kernel attention for collecting pertinent information from the images. Extensive testing on three publicly accessible datasets showed that the suggested architecture outperformed baseline models, claiming better results in terms of accuracy and AUC scores. Numerous ablation trials on the v available multimodal datasets are conducted to conclude that the proposed architecture is contributing to the performance of hate content identification in memes. The second proposed model “MHS-STMA” explored the problem of learning complementary information between multimodal data. The architecture utilizes transformers for capturing the dependencies and relationships between the elements in a sequence. The proposed architecture also utilizes attention mechanisms at multiple levels and focuses on crucial regions in the images based on the attended textual features. Self-attention mechanism is implied at the end to remove any redundancy from the multimodal data. The experimental results conducted on three popular datasets show that our method performs efficiently. Lastly, A novel robust approach MHM-HGraph is proposed to effectively capture the contextual dependencies within two modalities (Visual and Text). To better capture the underlying patterns within the data, this model makes use of hypergraph convolution layers to investigate the application of non-local information, identifying high-order correlations on hypergraph, and exploit the “enhancement connection” to perform non-linear mapping on the features. In conclusion, this thesis presents substantial discoveries and identifies potential areas for future research on the subject of identifying hate content detection.