Introduction
Related work
BERT and attention
Syntactic knowledge and GNN
Pipeline and end-to-end
Proposed method
Syntactic features representation
LCFS processing
Context feature dynamic weighted (CDW)
Interactive learning
Loss function
Experiments
Datasets and experimental settings
Datasets | Positive | Negative | Neural | |||
---|---|---|---|---|---|---|
Train | Test | Train | Test | Train | Test | |
Laptop | 994 | 341 | 870 | 128 | 646 | 169 |
Restaurant | 2164 | 728 | 807 | 196 | 637 | 196 |
Twitter | 1561 | 173 | 1560 | 173 | 3127 | 346 |
Hyper-parameter | Setting |
---|---|
BERT dim | 768 |
BiLSTM | 128 |
MaxPool1d kernel | 3 |
Average Pooling kernel | 128 |
Multi-Head | 12 |
SRD thresholds | 4 |
epochs | 50 |
batch_size | 64 |
learning_rate | 2e-5 |
dropout | 0.3 |
optimizer | Adam |
L2 regularization | 0.01 |
Baseline model
Model | Laptop | Restaurant | Twitter | |||
---|---|---|---|---|---|---|
Accuracy (%) | F1 (%) | Accuracy (%) | F1 (%) | Accuracy (%) | F1 (%) | |
DTBCSNN | 78.45 | 74.66 | 84.50 | 82.97 | 74.52 | 73.20 |
BiDTreeCRF | 78.74 | 74.46 | 84.10 | 82.73 | 74.46 | 72.98 |
Seq2Seq4ATE | 82.61 | 75.57 | 84.24 | 82.41 | 76.47 | 73.37 |
RINANTE | 82.67 | 76.34 | 81.58 | 79.92 | 72.67 | 70.51 |
SPAN-BERT | 82.66 | 75.36 | 83.79 | 81.52 | 76.24 | 72.82 |
JET-BERT | 83.25 | 76.40 | 84.57 | 82.75 | – | – |
Peng-to-stage | 82.13 | 75.02 | 82.60 | 80.30 | 74.41 | 72.04 |
OTE-MTL | 83.34 | 75.68 | 84.71 | 85.71 | 77.57 | 75.08 |
RACL-BERT | 82.79 | 76.59 | 85.38 | 85.27 | 77.63 | 75.31 |
DREGCN | 82.54 | 75.26 | 83.64 | 81.45 | – | – |
BART-ABSA | 83.50 | 75.92 | 85.20 | 83.56 | 77.42 | 74.12 |
PD-GAT | 83.64 | 75.82 | 84.87 | 82.64 | 77.52 | 74.36 |
SSEGCN | 83.54 | 74.62 | 84.50 | 82.57 | 76.75 | 73.50 |
Sentic GCN | 84.06 | 76.50 | 86.35 | 85.45 | 76.45 | 75.25 |
KGAN | 83.98 | 76.12 | 86.25 | 85.59 | 77.23 | 74.85 |
WSIN | 84.15 | 76.54 | 86.47 | 85.68 | 78.20 | 75.64 |
KDGN | 84.24 | 76.85 | 85.94 | 85.71 | 77.69 | 75.45 |
SIASC | 84.11 | 77.14 | 86.65 | 85.90 | 78.42 | 75.90 |
Comparison study
Model | Laptop | Restaurant | Twitter | |||
---|---|---|---|---|---|---|
Accuracy (%) | F1 (%) | Accuracy (%) | F1 (%) | Accuracy (%) | F1 (%) | |
ATAE-LSTM | 68.70 | – | 77.20 | – | – | – |
AEN-BERT | 79.93 | 76.31 | 83.12 | 74.76 | 74.71 | 73.13 |
LCFS-BERT | 79.68 | 75.55 | 85.21 | 78.16 | 75.83 | 74.40 |
R-GAT+BERT | 79.73 | 75.50 | 85.50 | 79.33 | 76.15 | 74.88 |
RINANTE | 79.93 | 74.31 | 83.12 | 74.76 | 75.18 | 74.01 |
SPAN-BERT | 80.50 | 76.24 | 86.35 | 79.72 | 75.94 | 74.52 |
JET-BERT | 80.39 | 75.47 | 83.25 | 74.94 | 72.15 | 70.40 |
Peng-to-stage | 80.21 | 76.20 | 85.91 | 79.12 | 75.57 | 73.82 |
OTE-MTL | 80.46 | 75.68 | 84.57 | 77.40 | – | – |
RACL-BERT | 80.94 | 76.72 | 86.54 | 80.16 | 76.19 | 74.68 |
DREGCN | 80.32 | 75.54 | 84.25 | 77.08 | – | – |
BART-ABSA | 80.57 | 76.08 | 85.95 | 76.96 | 75.66 | 73.66 |
PD-GAT | 80.79 | 76.85 | 86.24 | 78.95 | 76.09 | 74.85 |
SSEGCN | 80.95 | 76.98 | 85.75 | 79.68 | 75.85 | 74.63 |
Sentic GCN | 81.07 | 77.25 | 86.54 | 80.25 | 76.20 | 75.27 |
KGAN | 80.95 | 76.92 | 86.27 | 79.82 | 76.15 | 75.29 |
WSIN | 81.14 | 77.24 | 86.45 | 80.35 | 75.85 | 74.50 |
KDGN | 81.29 | 77.48 | 86.67 | 80.54 | 76.19 | 74.85 |
SIASC (CDM) | 81.05 | 77.10 | 86.35 | 80.10 | 76.40 | 74.82 |
SIASC (CDW) | 81.35 | 77.45 | 86.71 | 80.31 | 76.56 | 75.40 |
Model | Laptop | Restaurant | Twitter | |||
---|---|---|---|---|---|---|
Accuracy (%) | F1 (%) | Accuracy (%) | F1 (%) | Accuracy (%) | F1 (%) | |
KGAN | 80.32 | 76.84 | 81.75 | 78.25 | 68.35 | 65.35 |
WSIN | 79.11 | 75.49 | 85.08 | 82.57 | 73.17 | 69.27 |
KDGN | 82.68 | 79.35 | 85.32 | 81.49 | 75.91 | 71.45 |
SIASC | 84.00 | 80.27 | 86.63 | 84.75 | 77.67 | 75.30 |
Laptop | Restaurant | Twitter | ||
---|---|---|---|---|
Task | Model | Accuracy (%) | Accuracy (%) | Accuracy (%) |
AE | SIASC | 84.11 | 86.65 | 78.42 |
w/o Part-of-speech | (83.34)\(\downarrow \)0.77 | (86.25)\(\downarrow \)0.40 | (77.90)\(\downarrow \)0.52 | |
w/o Dependency | (82.47)\(\downarrow \)1.64 | (85.56)\(\downarrow \)1.09 | (77.18)\(\downarrow \)1.24 | |
w/o Part-of-speech & Dependency | (81.68)\(\downarrow \)2.43 | (85.12)\(\downarrow \)1.53 | (76.67)\(\downarrow \)1.75 | |
ASC | SIASC(CDM) | 81.05 | 86.35 | 76.40 |
SIASC(CDW) | 81.35 | 86.71 | 76.56 | |
w/o LCFS(CDM) | (80.12)\(\downarrow \)0.93 | (85.37)\(\downarrow \)0.98 | (74.82)\(\downarrow \)1.58 | |
w/o LCFS(CDW) | (80.24)\(\downarrow \)1.11 | (85.39)\(\downarrow \)1.32 | (74.76)\(\downarrow \)1.80 |
Ablation study
Dataset category | LCFS method | Laptop | Restaurant | Twitter |
---|---|---|---|---|
Test set | CDW | 3 | 3 | 5 |
CDM | 4 | 3 | 4 | |
Validation set | CDW | 3 | 3 | 4 |
CDM | 3 | 3 | 3 |
The SRD threshold study
Attention visualization
WSIN | KDGN | SIASC | |||||
---|---|---|---|---|---|---|---|
# | Sentences | AE | ASC | AE | ASC | AE | ASC |
1 | I choose apple MacBook because of their design and the aluminum casing. | Design, casing | P✓ P✓ | Design, aluminum casing | P✓ P✓ | Design, aluminum casing | P✓ P✓ |
2 | I continued to take the computer in AGAIN and they replaced the hard drive and mother board. | Hard drive, mother board | P✗ N✓ | Hard drive, mother board | P✗ N✓ | Hard drive, mother board | N✓ N✓ |
3 | Similar to other Indian restaurants, they use the dinner special to attract customers. | Dinner | P✓ | Dinner | O✗ | Dinner special | O✓ |
4 | The blond wood decor is very soothing, the premium sake is excellent and the service is great. | Blond wood, sake, service | P✓ P✓P✓ | Blond wood, sake, service | P✓ P✓ P✓ | Blond wood decor, sake, service | P✓ P✓ P✓ |
5 | The duck confit is always amazing and the foie gras terrine with figs was out of this world. | Duck confit, foie gras | P✓ N✗ | Duck confit, foie gras terrine | P✓ N✗ | Duck confit, foie gras terrine with figs | P✓ P✓ |
6 | The food is nothing like its menu description. | Food, menu | N✓ N✓ | Food, menu description | N✓ N✓ | Food, menu description | N✓ N✓ |