With the progressing complexity of process industries, alarm flooding has become a critical problem in alarm management. Granger causality (GC) analysis provides a pivotal role in investigating anomalous propagations and identifying alarm sources to reduce nuisance alarms. This paper proposes a novel attention-based long short-term memory (ALSTM) method for improved design of GC models to diagnose the alarm root-cause and identify the alarm propagation pathways. A novel training strategy is proposed for ALSTM-GC to enhance noise robustness and prevent overfitting. Also, a sensitivity-based method is presented to alleviate the spurious causalities caused by unobservable factors and slight indirect relationships. Compared with the traditional GC models, the main contribution of the proposed approach is 3-fold: (i) ALSTM-GC can effectively identify causality of process variables from varying and long-term delays without any prior data assumptions; (ii) with the incorporation of attention mechanism, ALSTM-GC can avoid the nested loops in causality analysis to reduce the computational complexity; (iii) ALSTM-GC can provide a more accurate and simplified network structure for late alarm traceability based on the sensitivity identification. The innovativeness of ALSTM-GC is discussed on a stochastic process based on comparative studies with different GC methods. In addition, the practicability of the proposed method is validated on a complicated case of the Tennessee Eastman process (TEP).