Page 290 - 《软件学报》2026年第2期
P. 290

张建新 等: 依存句法信息增强的完全非自回归翻译                                                         769





                                            表 2 不同模型在     WMT  任务上的结果展示

                                                                      WMT14           WMT16
                                     模型                       I dec                               Speedup
                                                                  EN→DE   DE→EN   EN→RO   RO→EN
                                        Transformer [1]       N    27.30    -       -       -       -
                   AT Models          Transformer-base [29]   N    27.48   31.27   33.70   34.05   1.0×
                                    Transformer-base w/KD     N    28.47   31.95   33.70   33.75   1.0×
                                         NAT-IR [30]          10   21.61   25.48   29.32   30.19   1.5×
                                         LaNMT [61]           4    26.30    -       -      29.10   5.7×
                                          LevT [10]           6+   27.27    -       -      33.26   4.0×
                                       Mask-Predict [9]       10   27.03   30.53   33.08   33.31   1.7×
                                      CMLM+HGLT [32]          -    27.46   31.03   33.08   33.12   2.6×
                                              [15]
                                        JM-NAT                10   27.69   32.24   33.52   33.72    -
                                             [62]
                                         DisCO               Adv   27.34   31.31   33.22   33.25   3.5×
                                                 [63]
                  Iterative NAT       Multi-Task NAT          10   27.98   31.27   33.80   33.60   1.7×
                                               [64]
                                       RewriteNAT            Adv   27.83   31.52   33.63   34.09    -
                                              [65]
                                        CMLMC                 10   27.63   31.17   33.88   34.09   1.75×
                                                     [7]
                                  Isotropy-Enhanced CMLMC     10   26.68   30.50   34.01   33.83    1.7
                                               [26]
                                       XlmF-NAT               10   27.63   31.17   33.88   34.09   1.75×
                                CMLM+N-gram 掩码+一致性学习   [27]   10   27.32   31.08   33.69   34.02    -
                                              [11]
                                         AMOM                 10   27.57   31.67   34.62   34.82    2.3
                                                 [66]
                                      CMLMC+EECR              10   28.04   31.65   34.33   34.32   3.77
                                         NAT-FT [6]           1    17.69   21.47   27.29   29.06   15.6×
                                       Mask-Predict [9]       1    18.05   21.83   27.32    28.2    -
                                         SynST [45]           1    20.74   25.50    -       -      4.96×
                                    BoN+源语言依存标签    [47]       1    21.42   25.15   29.41   30.01   9.25×
                                        imit-NAT [22]         1    22.44   25.67   28.61   28.90   18.6×
                                    BoN+源语言依存标签    [47]       1     -       -      29.17   30.24   9.23×
                                         Flowseq [19]         1    23.72   28.39   29.73   30.72   1.1×
                                       NAT-HINT [67]          1    25.80   28.40   32.30   31.70   18.6×
                                       NAT-DCRF [68]          1    25.80   28.40   32.30   31.70   18.6×
                                       ReorderNAT [69]        1    22.79   27.28   29.30   29.50   16.1×
                   Fully NAT             SNAT [46]            1    24.64   28.42   32.87   32.21   22.6×
                             NAT+ 门控增强输入, 课程学习 (NPD, k=3) [27]  1  25.68   30.45   32.00   32.71    -
                                        AlignNAT [70]         1    26.40   30.40   32.50   33.10   11.3×
                                    Fully-NAT+KD+CTC [57]     1    26.51   30.46   33.41   34.07   16.5×
                                       OAXE-NAT [71]          1    26.10   30.20   32.40   33.30   15.3×
                                       GLAT+DePA [72]         1    26.43   30.42   33.07   33.82   15.1×
                                     GLAT+RenewNAT [21]       1    25.75   29.67   32.53   32.93   12.5×
                                          NAT [26]            1    25.02   29.51   33.58    33.6   18.9×
                                   NAT 课程学习+增强网络    [28]      1    27.08    -      33.05    -      15.2×
                                       PCFG-NAT [73]          1    27.02   31.29   32.72   33.07   12.6×
                             DA-Transformer+BeamSearch+5-gram LM [74]  1  27.91  31.95  -   -      7.0×
                                        NAT-CTC [23]          1    16.56   18.64   19.54   24.67    -
                    w/CTC
                                         Imputer [75]         1    25.80   28.40   32.30   31.70   18.6×
   285   286   287   288   289   290   291   292   293   294   295