This paper uses a hybrid bidirectional long short-term memory (BiLSTM) transformer model with syntactic, polarity and context feature extractions to classify expressions of agency from witness hearings held during the 1996-97 South Africa Truth and Reconciliation Commission (SA TRC). During hearings, witnesses recalled the details and impact of apartheid-era human rights violations that they experienced. Training on a dataset of 2,286 labeled witness expressions with an 80/20 test/validation split stratified by label class, the hybrid model achieves a 0.90 validation F1 score across all-label prediction (agency and non-agency), and 0.50-0.60 validation accuracy scores on agency-relevant (low, mixed, high) labels. The model slightly outperforms a baseline finetuned RoBERTa model, especially across agency-relevant labels. Word attention weights are further visualized to assess high-importance words during the hybrid BiLSTM model classification process. The hybrid BiLSTM model is then deployed on 125 unseen hearings transcripts, representing 5,346 witness expressions. These expressions are qualitatively assessed to identify key themes concerning education and livelihoods recovery. Overall, the paper demonstrates the broader utility of modeling agency expression in human rights research.