A Deep Learning Model for Detecting Bullying Content in Swahili SMS Messages in Tanzania
TL;DR
A context-aware deep learning model for binary classification of bullying and non-bullying Swahili SMS messages, an empirically evaluated FastText-BiLSTM architecture, and deployment guidance that positions automated detection as a triage mechanism for human review rather than an autonomous enforcement tool are contributed.
Abstract
Automated bullying-content detection has advanced rapidly for English and other high-resource languages, yet comparable evidence for Swahili remains limited, particularly for short message service (SMS) communication in Tanzania. This study developed and evaluated a context-aware deep learning model for binary classification of bullying and non-bullying Swahili SMS messages. A multi-source corpus of 7,228 messages was initially assembled from prior Swahili datasets, voluntary student contributions, and Google Forms; duplicate records were removed during data cleaning before the train-validation partition was created. Four Kiswahili graduates applied a common annotation framework that considered the target, communicative intent, and surrounding linguistic context rather than treating offensive vocabulary as a sufficient label criterion. The messages were normalised, tokenised with a 10,000-token vocabulary, padded to 100 tokens, and represented using trainable 300-dimensional FastText embeddings. A Bidirectional Long Short-Term Memory network used 160 units in each direction, followed by dropout, a 96-unit rectified linear dense layer with L2 regularisation, and a two-class Softmax output. Candidate configurations were assessed through Keras Tuner and validation-based model selection. On the 1,470-message internal validation set created after duplicate removal, the selected model achieved 90.07% accuracy, 90.09% macro precision, 90.05% macro recall, and 90.06% macro F1-score. The confusion matrix contained 641 true negatives, 683 true positives, 80 false positives, and 66 false negatives. Bullying recall reached 91.19%, indicating that the model identified most harmful messages, although the 8.01-percentage-point training-validation gap and divergent loss curves showed moderate overfitting. The study contributes a Tanzania-focused Swahili SMS resource, an empirically evaluated FastText-BiLSTM architecture, and deployment guidance that positions automated detection as a triage mechanism for human review rather than an autonomous enforcement tool.