Research Article
An Improved Transformer-Based Neural Machine Translation Strategy: Interacting-Head Attention
Table 4
Overall evaluation scores using four attention-based models on the WMT17 EN-DE evaluation set and test set.
| Model | WMT17 subset | Number of heads/head size | 2/256 | 4/128 | 8/64 | 16/32 | 32/16 | 64/8 |
| (a) |
| Multihead attention | dev | 8.38 | 12.56 | 12.48 | 15.15 | 12.20 | 12.45 | newstest2014 | 8.77 | 11.31 | 12.85 | 15.73 | 11.61 | 12.28 | newstest2015 | 9.36 | 12.27 | 13.80 | 16.79 | 11.90 | 13.23 | newstest2016 | 11.40 | 14.83 | 16.79 | 20.97 | 15.17 | 16.16 | newstest2017 | 8.83 | 11.83 | 13.15 | 16.41 | 11.63 | 12.99 |
| Multihead attention (head size = n) | dev | 11.66 | 12.51 | 13.70 | 14.57 | 14.61 | 15.29 | newstest2014 | 11.64 | 13.26 | 14.93 | 15.42 | 14.96 | 15.00 | newstest2015 | 12.04 | 14.07 | 15.78 | 16.57 | 16.67 | 17.57 | newstest2016 | 14.71 | 17.57 | 19.77 | 20.33 | 19.80 | 20.08 | newstest2017 | 11.72 | 13.54 | 14.92 | 16.03 | 16.07 | 16.33 |
| Talking-head attention | dev | 10.90 | 13.89 | 15.31 | 15.37 | 15.23 | 15.50 | newstest2014 | 11.21 | 12.64 | 14.33 | 15.28 | 15.29 | 15.39 | newstest2015 | 11.85 | 12.81 | 14.61 | 15.95 | 15.55 | 15.36 | newstest2016 | 14.22 | 16.33 | 18.84 | 19.53 | 19.34 | 20.43 | newstest2017 | 11.15 | 12.81 | 14.35 | 15.49 | 14.44 | 15.36 |
| Interacting-head attention | dev | 13.07 | 14.84 | 14.53 | 17.46(+2.31) | — | — | newstest2014 | 10.91 | 14.72 | 15.75 | 17.35(+1.62) | — | — | newstest2015 | 11.99 | 15.43 | 16.90 | 18.40(+1.21) | — | — | newstest2016 | 14.95 | 19.61 | 21.12 | 22.36(+1.39) | — | — | newstest2017 | 11.09 | 15.00 | 17.30 | 17.67(+1.26) | — | — |
| (b) |
| Multihead attention | dev | 39.23 | 37.80 | 28.23 | 15.57 | 32.93 | 22.47 | newstest2014 | 46.85 | 45.59 | 32.10 | 19.68 | 39.59 | 28.80 | newstest2015 | 45.22 | 44.31 | 32.64 | 22.68 | 42.78 | 28.58 | newstest2016 | 40.14 | 39.35 | 28.51 | 15.91 | 32.04 | 23.97 | newstest2017 | 43.85 | 42.74 | 32.12 | 19.07 | 38.38 | 26.50 |
| Multihead attention (head size = n) | dev | 39.37 | 36.73 | 25.11 | 22.97 | 16.67 | 13.42 | newstest2014 | 46.40 | 35.77 | 32.56 | 28.65 | 18.56 | 18.77 | newstest2015 | 46.93 | 38.45 | 33.16 | 28.03 | 23.13 | 20.07 | newstest2016 | 42.91 | 30.24 | 27.66 | 22.77 | 17.15 | 13.26 | newstest2017 | 44.65 | 34.60 | 29.82 | 26.78 | 19.75 | 16.52 |
| Talking-head attention | dev | 48.17 | 40.63 | 39.03 | 24.80 | 20.13 | 16.25 | newstest2014 | 48.19 | 48.02 | 52.98 | 33.93 | 27.58 | 20.32 | newstest2015 | 61.09 | 53.80 | 52.33 | 33.43 | 31.13 | 22.05 | newstest2016 | 47.02 | 47.28 | 47.42 | 31.08 | 23.16 | 14.82 | newstest2017 | 51.13 | 47.34 | 50.63 | 34.52 | 28.01 | 20.05 |
| Interacting-head attention | dev | 58.10 | 35.20 | 20.83 | 9.63(5.94) | — | — | newstest2014 | 68.39 | 47.29 | 29.01 | 13.64(6.04) | — | — | newstest2015 | 79.39 | 48.55 | 29.64 | 16.05(6.63) | — | — | newstest2016 | 67.12 | 44.21 | 25.94 | 11.27(4.64) | — | — | newstest2017 | 81.56 | 49.07 | 30.89 | 15.23(3.84) | — | — |
| (c) |
| Multihead attention | dev | 19.36 | 19.54 | 20.31 | 23.16 | 20.56 | 20.53 | newstest2014 | 19.39 | 19.64 | 20.83 | 23.71 | 20.72 | 20.97 | newstest2015 | 19.36 | 19.60 | 21.15 | 24.11 | 20.70 | 21.06 | newstest2016 | 20.43 | 21.27 | 22.73 | 26.29 | 22.44 | 22.63 | newstest2017 | 19.52 | 19.85 | 21.05 | 23.96 | 20.70 | 21.11 |
| Multihead attention (head size = n) | dev | 19.47 | 20.96 | 21.04 | 23.01 | 23.32 | 23.79 | newstest2014 | 19.69 | 21.45 | 22.30 | 23.52 | 23.91 | 24.45 | newstest2015 | 19.64 | 21.54 | 22.61 | 23.70 | 24.02 | 24.46 | newstest2016 | 20.95 | 23.26 | 24.47 | 25.65 | 25.44 | 26.99 | newstest2017 | 19.68 | 21.50 | 22.25 | 23.73 | 24.05 | 24.49 |
| Talking-head attention | dev | 17.71 | 19.55 | 21.38 | 23.22 | 23.51 | 23.19 | newstest2014 | 17.73 | 19.59 | 21.87 | 23.65 | 23.73 | 23.75 | newstest2015 | 17.33 | 19.39 | 22.03 | 23.81 | 24.14 | 24.26 | newstest2016 | 18.60 | 20.52 | 23.19 | 25.40 | 26.02 | 25.88 | newstest2017 | 17.74 | 19.39 | 21.90 | 23.71 | 23.66 | 24.15 |
| Interacting-head attention | dev | 22.43 | 22.61 | 23.87 | 24.62(+1.46) | — | — | newstest2014 | 22.67 | 22.66 | 24.35 | 25.10(+1.39) | — | — | newstest2015 | 22.81 | 23.09 | 24.52 | 25.53(+1.42) | — | — | newstest2016 | 24.69 | 25.00 | 26.68 | 27.27(+0.98) | — | — | newstest2017 | 22.12 | 22.76 | 24.38 | 25.66(+1.70) | — | — |
| (d) |
| Multihead attention | dev | 40.08 | 40.19 | 41.54 | 45.77 | 42.47 | 42.79 | newstest2014 | 36.65 | 37.81 | 40.22 | 49.04 | 40.40 | 41.18 | newstest2015 | 37.33 | 39.63 | 42.56 | 46.84 | 41.69 | 42.71 | newstest2016 | 39.88 | 41.98 | 44.40 | 44.88 | 44.52 | 45.13 | newstest2017 | 36.94 | 39.01 | 41.43 | 46.09 | 41.02 | 42.02 |
| Multihead attention (head size = n) | dev | 41.39 | 43.67 | 44.89 | 45.79 | 45.83 | 46.01 | newstest2014 | 39.75 | 42.35 | 43.62 | 44.69 | 45.07 | 47.25 | newstest2015 | 40.76 | 43.37 | 45.54 | 46.18 | 46.22 | 48.15 | newstest2016 | 43.18 | 46.51 | 48.21 | 49.34 | 48.96 | 50.08 | newstest2017 | 40.41 | 43.05 | 44.87 | 45.52 | 45.55 | 46.08 |
| Talking-head attention | dev | 38.42 | 40.20 | 42.03 | 43.99 | 44.67 | 45.33 | newstest2014 | 35.67 | 39.23 | 42.51 | 42.74 | 45.86 | 46.09 | newstest2015 | 37.94 | 38.61 | 41.00 | 44.15 | 44.92 | 46.67 | newstest2016 | 38.83 | 41.21 | 42.88 | 46.28 | 48.05 | 46.99 | newstest2017 | 37.29 | 39.58 | 40.93 | 43.30 | 45.23 | 45.65 |
| Interacting-head attention | dev | 45.47 | 46.03 | 46.54 | 47.12(+1.35) | — | — | newstest2014 | 43.95 | 43.77 | 45.15 | 48.20 | — | — | newstest2015 | 45.33 | 45.83 | 46.61 | 47.31(+0.51) | — | — | newstest2016 | 47.97 | 48.83 | 49.94 | 50.47(+5.59) | — | — | newstest2017 | 43.61 | 44.58 | 45.91 | 45.99 | — | — |
| (e) |
| Multihead attention | dev | 1.26 | 1.30 | 1.37 | 1.64 | 1.37 | 1.38 | newstest2014 | 1.13 | 1.19 | 1.26 | 1.52 | 1.26 | 1.26 | newstest2015 | 1.24 | 1.31 | 1.42 | 1.63 | 1.36 | 1.36 | newstest2016 | 1.39 | 1.50 | 1.62 | 1.93 | 1.61 | 1.60 | newstest2017 | 1.17 | 1.22 | 1.35 | 1.61 | 1.29 | 1.33 |
| Multihead attention (head size = n) | dev | 1.23 | 1.38 | 1.45 | 1.63 | 1.63 | 1.66 | newstest2014 | 1.13 | 1.30 | 1.39 | 1.54 | 1.55 | 1.58 | newstest2015 | 1.19 | 1.37 | 1.45 | 1.67 | 1.68 | 1.65 | newstest2016 | 1.38 | 1.65 | 1.80 | 1.94 | 1.85 | 1.99 | newstest2017 | 1.16 | 1.35 | 1.43 | 1.60 | 1.71 | 1.70 |
| Talking-head attention | dev | 1.06 | 1.20 | 1.38 | 1.42 | 1.49 | 1.58 | newstest2014 | 0.96 | 1.11 | 1.35 | 1.38 | 1.46 | 1.51 | newstest2015 | 1.04 | 1.20 | 1.37 | 1.39 | 1.36 | 1.39 | newstest2016 | 1.16 | 1.40 | 1.67 | 1.45 | 1.48 | 1.62 | newstest2017 | 0.98 | 1.15 | 1.43 | 1.39 | 1.48 | 1.47 |
| Interacting-head attention | dev | 1.63 | 1.64 | 1.67 | 1.71(+0.07) | — | — | newstest2014 | 1.43 | 1.51 | 1.57 | 1.63(+0.11) | — | — | newstest2015 | 1.63 | 1.61 | 1.74 | 1.81(+0.18) | — | — | newstest2016 | 1.89 | 1.88 | 2.02 | 2.17(+0.24) | — | — | newstest2017 | 1.52 | 1.50 | 1.65 | 1.74(+0.13) | — | — |
| (f) |
| Multihead attention | dev | 43.74 | 44.71 | 45.35 | 48.82 | 46.28 | 45.37 | newstest2014 | 41.50 | 42.57 | 43.80 | 47.84 | 44.44 | 43.77 | newstest2015 | 43.17 | 44.10 | 45.99 | 49.68 | 45.81 | 45.49 | newstest2016 | 44.90 | 45.96 | 47.37 | 52.29 | 47.86 | 47.13 | newstest2017 | 42.11 | 43.36 | 44.99 | 48.71 | 44.91 | 44.71 |
| Multihead attention (head size = n) | dev | 44.58 | 46.00 | 47.06 | 48.51 | 48.99 | 49.01 | newstest2014 | 43.09 | 44.63 | 45.63 | 47.40 | 48.04 | 47.99 | newstest2015 | 44.37 | 46.09 | 47.54 | 48.83 | 49.15 | 49.56 | newstest2016 | 46.24 | 48.30 | 49.65 | 51.20 | 51.30 | 51.06 | newstest2017 | 43.82 | 45.47 | 46.38 | 48.34 | 48.55 | 48.77 |
| Talking-head attention | dev | 38.11 | 39.63 | 40.73 | 44.07 | 46.14 | 48.60 | newstest2014 | 35.46 | 38.32 | 39.02 | 42.77 | 44.52 | 47.33 | newstest2015 | 37.48 | 38.71 | 40.35 | 44.35 | 46.84 | 49.52 | newstest2016 | 37.09 | 40.25 | 41.43 | 45.96 | 48.36 | 50.70 | newstest2017 | 36.20 | 38.93 | 39.47 | 43.75 | 45.85 | 48.05 |
| Interacting-head attention | dev | 48.03 | 48.23 | 48.78 | 49.15(+0.33) | — | — | newstest2014 | 46.72 | 46.63 | 47.78 | 48.71(+0.87) | — | — | newstest2015 | 48.11 | 48.52 | 49.09 | 50.20(+0.52) | — | — | newstest2016 | 50.88 | 50.77 | 51.47 | 52.71(+0.42) | — | — | newstest2017 | 47.73 | 48.48 | 49.83 | 50.01(+1.30) | — | — |
|
|
Note. The units of the performances are (a) BLEU, (b) WER, (c) METEOR, (d) ROUGE_L, (e) CIDEr, and (f) YiSi.
|