Research Article
Multitask Learning for Aspect-Based Sentiment Classification
Table 4
Performance comparison with previous models, with average accuracy as the evaluation metric; the best result for each dataset is boldfaced.
| | Method | 14Lap | 14Res | 15Res |
| Conventional network | TD-LSTM (2016) | 68.13 | 75.63 | 76.39 | ATAE-LSTM (2016) | 68.70 | 77.20 | 78.48 | MemNet (2016) | 70.33 | 78.16 | 77.31 | TNET (2016) | 74.65 | 80.05 | 78.47 | IAN (2017) | 72.10 | 78.60 | 78.58 | RAM (2017) | 74.49 | 80.23 | 79.98 | TG-SAN (2020) | 75.27 | 81.66 | — |
| Multitask learning | PRET + MULT (2018) | 71.15 | 79.11 | 81.30 |
| BERT | IMN (2019) | 75.36 | 83.89 | 85.64 | BERT-FC (2018) | 76.54 | 81.28 | 81.52 | BERT-pair-QA-M (2019) | 77.93 | 85.12 | 81.89 | AEN-BERT (2019) | 78.35 | 81.46 | — | BERT-PT (2019) | 78.07 | 84.95 | — | TD-BERT (2019) | 78.87 | 85.10 | — | SK-GCN-BERT (2020) | 79.00 | 83.48 | 83.20 | SPRN-BERT (2021) | 79.31 | 85.03 | 85.30 |
| Our model | SMLN | 80.09 | 85.67 | 86.31 |
|
|