Table 5 Generalizability and robustness of the proposed TRABSA model on both the extended and external datasets.
Dataset type | Dataset name | DL models | Evaluation metrics | Training time (s) | Inference time (s) | |||
|---|---|---|---|---|---|---|---|---|
Macro average precision | Macro average Recall | Macro average F1-score | Accuracy | |||||
Extended | Global COVID-19 Dataset | Single Hidden Layer NN | 97% ± 0% | 97% ± 0% | 97% ± 0% | 97% ± 1% | 7365 | 491 |
3 Hidden Layers NN | 97% ± 0% | 97% ± 0% | 97% ± 0% | 97% ± 1% | 7755 | 517 | ||
BiLSTM+3 Hidden Layers NN | 97% ± 1% | 97% ± 1% | 97% ± 1% | 97% ± 1% | 5709 | 518 | ||
BiLSTM+CNN | 11% ± 5% | 33% ± 4% | 17% ± 5% | 33% ± 4% | 8789 | 536 | ||
Proposed TRABSA Model | 98% ± 0% | 98% ± 0% | 98% ± 0% | 98% ± 1% | 8288 | 518 | ||
Extended | USA COVID-19 Dataset | Single Hidden Layer NN | 81% ± 3% | 81% ± 3% | 81% ± 3% | 83% ± 3% | 870 | 58 |
3 Hidden Layers NN | 85% ± 3% | 83% ± 1% | 84% ± 2% | 85% ± 3% | 696 | 58 | ||
BiLSTM+3 Hidden Layers NN | 85% ± 1% | 85% ± 1% | 85% ± 1% | 86% ± 0% | 1218 | 59 | ||
BiLSTM+CNN | 17% ± 5% | 33% ± 5% | 22% ± 4% | 51% ± 1% | 413 | 59 | ||
Proposed TRABSA Model | 87% ± 1% | 86% ± 1% | 86% ± 1% | 87% ± 1% | 1081 | 47 | ||
External | Twitter dataset | Single Hidden Layer NN | 93% ± 3% | 93% ± 3% | 93% ± 3% | 93% ± 3% | 9891 | 495 |
3 Hidden Layers NN | 92% ± 1% | 92% ± 1% | 92% ± 1% | 92% ± 1% | 4608 | 288 | ||
BiLSTM+3 Hidden Layers NN | 92% ± 3% | 92% ± 3% | 92% ± 3% | 92% ± 3% | 8700 | 291 | ||
BiLSTM+CNN | 53% ± 2% | 49% ± 4% | 46% ± 3% | 49% ± 4% | 1752 | 292 | ||
Proposed TRABSA model | 97% ± 1% | 97% ± 1% | 97% ± 1% | 97% ± 1% | 6602 | 287 | ||
External | Reddit dataset | Single Hidden Layer NN | 94% ± 3% | 93% ± 3% | 94% ± 3% | 94% ± 3% | 1494 | 90 |
3 Hidden Layers NN | 94% ± 1% | 94% ± 1% | 94% ± 1% | 94% ± 2% | 2415 | 119 | ||
BiLSTM+3 Hidden Layers NN | 94% ± 1% | 94% ± 2% | 94% ± 2% | 94% ± 1% | 2464 | 101 | ||
BiLSTM+CNN | 94% ± 1% | 94% ± 0% | 94% ± 0% | 94% ± 0% | 1944 | 119 | ||
Proposed TRABSA Model | 94% ± 1% | 93% ± 0% | 94% ± 0% | 95% ± 1% | 2200 | 94 | ||
External | Apple dataset | Single Hidden Layer NN | 81% ± 1% | 82% ± 2% | 81% ± 1% | 84% ± 3% | 96 | 11 |
3 Hidden Layers NN | 83% ± 2% | 81% ± 3% | 82% ± 3% | 85% ± 1% | 55 | 10 | ||
BiLSTM+3 Hidden Layers NN | 87% ± 2% | 85% ± 4% | 86% ± 3% | 89% ± 0% | 140 | 12 | ||
BiLSTM+CNN | 85% ± 1% | 83% ± 3% | 84% ± 2% | 87% ± 0% | 130 | 11 | ||
Proposed TRABSA Model | 88% ± 1% | 86% ± 2% | 86% ± 2% | 90% ± 0% | 210 | 12 | ||
External | US Airline dataset | Single Hidden Layer NN | 93% ± 3% | 93% ± 3% | 93% ± 3% | 94% ± 2% | 1201 | 48 |
3 Hidden Layers NN | 94% ± 2% | 93% ± 3% | 93% ± 3% | 94% ± 2% | 1166 | 53 | ||
BiLSTM+3 Hidden Layers NN | 94% ± 1% | 93% ± 2% | 94% ± 1% | 94% ± 1% | 1012 | 46 | ||
BiLSTM+CNN | 94% ± 3% | 93% ± 4% | 93% ± 4% | 94% ± 3% | 842 | 40 | ||
Proposed TRABSA Model | 95% ± 0% | 95% ± 0% | 95% ± 0% | 96% ± 1% | 897 | 39 | ||
