Figure 1
Character Model architecture.
Figure 2
Couplet Model architecture.
Figure 3
Poem-4 (left) and Poem-1 (right) Model architectures.
Table 1
Models and inference pipelines, with performance metrics (mean ± standard deviation) acquired from 100 trials.
| MODEL/INFERENCE | ACCURACY | PRECISION | RECALL | F1 |
|---|---|---|---|---|
| Character | 0.932 ± 0.034 | 0.930 ± 0.041 | 0.936 ± 0.037 | 0.932 ± 0.032 |
| Couplet | 0.947 ± 0.019 | 0.923 ± 0.040 | 0.979 ± 0.021 | 0.949 ± 0.017 |
| Char → Couplet | 0.881 ± 0.026 | 0.897 ± 0.038 | 0.864 ± 0.046 | 0.879 ± 0.026 |
| Poem-1 | 0.886 ± 0.029 | 0.860 ± 0.059 | 0.932 ± 0.042 | 0.892 ± 0.022 |
| Poem-1 (2 epochs) | 0.905 ± 0.022 | 0.881 ± 0.045 | 0.941 ± 0.037 | 0.908 ± 0.019 |
| Couplet → Poem-1 | 0.839 ± 0.051 | 0.777 ± 0.073 | 0.969 ± 0.031 | 0.860 ± 0.035 |
| Char → Poem-1 | 0.808 ± 0.027 | 0.847 ± 0.042 | 0.760 ± 0.076 | 0.797 ± 0.036 |
| Poem-4 | 0.696 ± 0.017 | 0.742 ± 0.055 | 0.601 ± 0.102 | 0.655 ± 0.048 |
| Poem-4 (2 epochs) | 0.735 ± 0.022 | 0.759 ± 0.052 | 0.686 ± 0.077 | 0.715 ± 0.031 |
| Poem-4 → Poem-1 | 0.659 ± 0.027 | 0.673 ± 0.066 | 0.674 ± 0.144 | 0.657 ± 0.061 |
Figure 4
F1 Score Distribution by Target. Metrics computed against silver labels generated by the teacher model.
Figure 5
Attention distribution in a regulated poem, top layer of the fine-tuned SikuBERT classifier (Poem-1). Each heatmap is a head, each row is a couplet, each cell is a Chinese character or punctuation. Darker color indicates higher attention score from the [CLS] token. The [CLS] and [SEP] tokens have been removed for better visibility. Notice the isomorphic attention distribution in the inner couplets: in Head 1 (top left), for example, the third (parallel) couplet elicits higher attention at positions 1, 2, and 5 in both lines. Punctuation marks often serve as “attention sinks,” providing a stable anchor for information flow across layers.