Neural Machine Translation

写在前面

机器翻译主要用的是seq2seq的模型，配合上Attentional Model
项目参见: Seq2seq-Translation

---

Introduction of Sequence Model

Language Model

Language Modeling is the task of predicting what word comes next

language model.png

Conventional LM

1. Probability Presentation Problem
2. Sparsity Problem
3. Model Size huge

conventional_lm.png

Neural Language Model

1. fixed-window size too small
2. Can't process any input length cases

nlm.png

Recurrent Neural Networks Language Model

Advantages:

1. process any input length
2. share Weights

Disadvantages:

1. gradient vanishes or explodes
2. training costs time

RNN.png

Pointer Sentinel Mixture Models

1. integrate RNN with pointer-sentinel networks
2. RNN networks predicts word on given Vocabulary while pointer-sentinel networks predicts word on previous window
3. Pointer Sentinel Network makes effort to QA, Passage Summarization et.
4. 论文参见: Pointer networks | Point Sentinel Mixture Models
   

pointer-sentinel.png mixture networks.png

---

Machine Translation

Neural Machine Translation

Advantages:

1. More fluent
2. Better use of context
3. Better use of phrase similarities
4. A single neural network to be optimized end-to-end
5. Requires much less human engineering effort

BottleNecks:
Encoding of the source sentence. This needs to capture all information about the source sentence

Neural Machine Translation.png

Attentional Model

1. In contrast of Neural MT Model above, we don't feed the decoder with all encoder outputs
2. Integrate decoder hidden state with encoder outputs to get attention scores
3. apply attention scores to encoder outputs, generating context
4. feed context to decoder input

Attentional model.png

Intuition on Attention

Attention.png

---

Fundamental Attentional Model

Bahandanau Attentional Model

1. 使用的是 global attention
2. encoder使用了Bi-LSTM或者Bi-GRU
3. 使用concat方式生成attention
4. 解析顺序: ht−1 → at → ct → ht
5. 参考论文 Neural Machine Translation by Jointly Learning to Align and Translate

Luong Attentional Model

1. 这一篇论文主要是改进了 Bahdanau 中的attentional model
2. 提供了 global attention 以及 local attention，并提供了两种确定local position的方法(线性对应或是预测对应)
3. encoder使用了stack-LSTM或者stack-GRU
4. 提供了dot general concat三种方式计算attention
5. decoder解析顺序与 Bahdanau有所不同 ht →at →ct →h ̃t
6. 提出了input-feed，将上一步生成的attention作为当前decoder的输入
7. 参考论文 Effective Approaches to Attention-based Neural Machine Translation

Global Attention

global.png

Local Attention

1. the model first generates an aligned position pt for each target word at time t
2. The context vec-tor ct is then derived as a weighted average over the set of source hidden states within the window [pt−D,pt+D]
3. Monotonic alignment (local-m) pt = t
4. Predictive alignment (local-p) pt = S · sigmoid(v⊤p tanh(Wpht))

local.png

Input Feed

input feed.png

---

Advanced Attentional Model

Intra-Decoder attention for Summarization

1. Reinforcement learning
2. 论文参见: A Deep Reinforced Model for Abstractive Summarization

More advanced attention

1. More advanced similarity function than simple inner product
2. Temporal attention function, penalizing input tokens that have obtained high attention scores in past decoding steps
3. Improves coverage and prevent repeated attention to same inputs
4. Combine softmax’ed weighted hidden states from encoder

Self-attention on decoder

1. each hidden state attends to the previous hidden states of the same RNN
2. Apply softmax to get attention distribution over previous hidden hd(t) states for t’ = 1,...,t-1

Hybrid NMT

1. When applying to more languages, <UNK> tag may occur more often
2. Char-LSTM used to translate <UNK> tag char-by-char

hybrid_nmt.png

---

Results

DataSet

English - Chinese
Small Sample: 10k sentences, 3k eng words, 2.9k cn words

Mary came in.   瑪麗進來了。
Mary is tall.   瑪麗很高。
May I go now?   我现在能去了吗？
Move quietly.   轻轻地移动。
My eyes hurt.   我的眼睛痛。
No one knows.   沒有人知道。
Nobody asked.   没人问过。

Basic Attentional Model

Train Loss
0m 30s (- 50m 7s) (1000 1%) 4.6231
1m 4s (- 52m 17s) (2000 2%) 4.3355
1m 39s (- 53m 35s) (3000 3%) 4.0968
2m 15s (- 54m 4s) (4000 4%) 3.8949
2m 51s (- 54m 9s) (5000 5%) 3.7295
3m 26s (- 53m 52s) (6000 6%) 3.6388
4m 1s (- 53m 34s) (7000 7%) 3.5656
4m 36s (- 53m 4s) (8000 8%) 3.4461
5m 12s (- 52m 39s) (9000 9%) 3.3745
5m 47s (- 52m 11s) (10000 10%) 3.2642
6m 23s (- 51m 45s) (11000 11%) 3.1839
6m 59s (- 51m 13s) (12000 12%) 3.0886
7m 35s (- 50m 46s) (13000 13%) 3.0016
8m 10s (- 50m 14s) (14000 14%) 2.9388
8m 46s (- 49m 44s) (15000 15%) 2.8103
9m 22s (- 49m 10s) (16000 16%) 2.7448
9m 56s (- 48m 30s) (17000 17%) 2.7605
10m 30s (- 47m 52s) (18000 18%) 2.6554
11m 7s (- 47m 24s) (19000 19%) 2.6391
11m 43s (- 46m 54s) (20000 20%) 2.5279
12m 19s (- 46m 22s) (21000 21%) 2.4900
12m 56s (- 45m 54s) (22000 22%) 2.4153
13m 33s (- 45m 22s) (23000 23%) 2.3262
14m 10s (- 44m 52s) (24000 24%) 2.3369
14m 46s (- 44m 19s) (25000 25%) 2.3140
15m 20s (- 43m 38s) (26000 26%) 2.2905
15m 55s (- 43m 4s) (27000 27%) 2.1696
16m 32s (- 42m 31s) (28000 28%) 2.1323
17m 7s (- 41m 56s) (29000 28%) 2.0508
17m 45s (- 41m 25s) (30000 30%) 2.0222
18m 21s (- 40m 51s) (31000 31%) 1.9533
18m 54s (- 40m 10s) (32000 32%) 1.9775
19m 29s (- 39m 34s) (33000 33%) 1.9728
20m 5s (- 38m 59s) (34000 34%) 1.8780
20m 41s (- 38m 25s) (35000 35%) 1.8047
21m 17s (- 37m 51s) (36000 36%) 1.8641
21m 54s (- 37m 18s) (37000 37%) 1.7819
22m 31s (- 36m 45s) (38000 38%) 1.7386
23m 8s (- 36m 11s) (39000 39%) 1.6892
23m 44s (- 35m 36s) (40000 40%) 1.6438
24m 21s (- 35m 3s) (41000 41%) 1.5385
24m 58s (- 34m 29s) (42000 42%) 1.6697
25m 36s (- 33m 56s) (43000 43%) 1.5461
26m 11s (- 33m 20s) (44000 44%) 1.5460
26m 46s (- 32m 43s) (45000 45%) 1.5115
27m 23s (- 32m 9s) (46000 46%) 1.5514
27m 59s (- 31m 34s) (47000 47%) 1.4065
28m 35s (- 30m 58s) (48000 48%) 1.4017
29m 12s (- 30m 23s) (49000 49%) 1.3494
29m 43s (- 29m 43s) (50000 50%) 1.3298
30m 4s (- 28m 54s) (51000 51%) 1.3180
30m 40s (- 28m 19s) (52000 52%) 1.3411
31m 17s (- 27m 45s) (53000 53%) 1.3513
31m 53s (- 27m 10s) (54000 54%) 1.1807
32m 30s (- 26m 35s) (55000 55%) 1.2839
33m 6s (- 26m 0s) (56000 56%) 1.2610
33m 40s (- 25m 24s) (57000 56%) 1.1819
34m 13s (- 24m 46s) (58000 57%) 1.0888
34m 45s (- 24m 9s) (59000 59%) 1.1754
35m 17s (- 23m 31s) (60000 60%) 1.1238
35m 50s (- 22m 55s) (61000 61%) 1.1342
36m 23s (- 22m 18s) (62000 62%) 1.1035
36m 55s (- 21m 41s) (63000 63%) 1.1070
37m 23s (- 21m 1s) (64000 64%) 1.0512
37m 43s (- 20m 18s) (65000 65%) 1.0274
38m 5s (- 19m 37s) (66000 66%) 1.0114
38m 25s (- 18m 55s) (67000 67%) 0.9937
38m 46s (- 18m 14s) (68000 68%) 0.9291
39m 8s (- 17m 35s) (69000 69%) 0.9563
39m 33s (- 16m 57s) (70000 70%) 0.9308
40m 0s (- 16m 20s) (71000 71%) 0.9932
40m 28s (- 15m 44s) (72000 72%) 0.9481
40m 53s (- 15m 7s) (73000 73%) 0.9157
41m 15s (- 14m 29s) (74000 74%) 0.8805
41m 32s (- 13m 50s) (75000 75%) 0.9156
41m 54s (- 13m 14s) (76000 76%) 0.8637
42m 25s (- 12m 40s) (77000 77%) 0.8662
42m 57s (- 12m 6s) (78000 78%) 0.8124
43m 29s (- 11m 33s) (79000 79%) 0.8772
44m 0s (- 11m 0s) (80000 80%) 0.8268
44m 28s (- 10m 25s) (81000 81%) 0.8146
44m 56s (- 9m 51s) (82000 82%) 0.8086
45m 13s (- 9m 15s) (83000 83%) 0.7818
45m 41s (- 8m 42s) (84000 84%) 0.7374
46m 8s (- 8m 8s) (85000 85%) 0.7332
46m 38s (- 7m 35s) (86000 86%) 0.8005
47m 1s (- 7m 1s) (87000 87%) 0.7465
47m 30s (- 6m 28s) (88000 88%) 0.8085
47m 47s (- 5m 54s) (89000 89%) 0.7231
48m 13s (- 5m 21s) (90000 90%) 0.7111
48m 34s (- 4m 48s) (91000 91%) 0.7396
49m 3s (- 4m 15s) (92000 92%) 0.6561
49m 34s (- 3m 43s) (93000 93%) 0.7108
50m 4s (- 3m 11s) (94000 94%) 0.6574
50m 31s (- 2m 39s) (95000 95%) 0.6983
50m 56s (- 2m 7s) (96000 96%) 0.7017
51m 28s (- 1m 35s) (97000 97%) 0.6549
52m 0s (- 1m 3s) (98000 98%) 0.5979
52m 32s (- 0m 31s) (99000 99%) 0.6844
53m 2s (- 0m 0s) (100000 100%) 0.6341

Sample

> 我藏在桌子底下
= i hid under the table .
< i am the the the table<EOS>

> 我不想看起來傻
= i don t want to look stupid .
< i don t want to look . <EOS>

> 把鹽遞給我好嗎
= pass me the salt would you ?
< pass me the salt would ? <EOS>

> 新年快樂
= happy new year !
< happy new year ! ! <EOS>

> 她來這裡放鬆的嗎
= did she come here to relax ?
< did she come on the watch ? <EOS>

> 现在道歉也迟了
= it s too late to apologize .
< it s too late apologize. <EOS>

> 让我们回家吧
= let us go home .
< let us go home . <EOS>

> 它真的很便宜
= it is really cheap .
< it really is cheap . <EOS>

> 社區是安靜的
= the neighborhood was silent .
< the neighborhood was silent . <EOS>

> 你可以随便去哪儿
= you may go anywhere .
< you may go anywhere . <EOS>

Loung Attentional Model(2-Layer GRU, Global Attention, Dot, Teacher Forcing)

Train Loss
0m 34s (- 1427m 31s) (20 0%) 5.8900
1m 9s (- 1437m 13s) (40 0%) 4.4671
1m 43s (- 1438m 23s) (60 0%) 4.2377
2m 19s (- 1455m 7s) (80 0%) 4.0304
2m 55s (- 1461m 40s) (100 0%) 3.8420
3m 31s (- 1466m 16s) (120 0%) 3.7378
4m 7s (- 1469m 7s) (140 0%) 3.5625
4m 42s (- 1466m 36s) (160 0%) 3.3969
5m 18s (- 1468m 55s) (180 0%) 3.3081
5m 52s (- 1462m 29s) (200 0%) 3.1576
6m 27s (- 1462m 18s) (220 0%) 3.0534
7m 3s (- 1463m 35s) (240 0%) 2.9494
7m 39s (- 1465m 15s) (260 0%) 2.8771
8m 16s (- 1469m 28s) (280 0%) 2.7319
8m 53s (- 1473m 20s) (300 0%) 2.6896
9m 30s (- 1475m 46s) (320 0%) 2.6084
10m 6s (- 1477m 14s) (340 0%) 2.5069
10m 42s (- 1476m 17s) (360 0%) 2.4839
11m 17s (- 1474m 43s) (380 0%) 2.4190
11m 54s (- 1476m 36s) (400 0%) 2.3489
12m 30s (- 1476m 46s) (420 0%) 2.2978
13m 6s (- 1476m 9s) (440 0%) 2.1818
13m 42s (- 1476m 47s) (460 0%) 2.1761
14m 19s (- 1477m 54s) (480 0%) 2.0791
14m 55s (- 1478m 11s) (500 1%) 2.0150
15m 32s (- 1478m 16s) (520 1%) 1.9801
16m 9s (- 1479m 33s) (540 1%) 1.9012
16m 45s (- 1479m 33s) (560 1%) 1.8956
17m 22s (- 1479m 52s) (580 1%) 1.8302
17m 58s (- 1480m 34s) (600 1%) 1.7954
18m 35s (- 1480m 41s) (620 1%) 1.7853
19m 11s (- 1480m 26s) (640 1%) 1.6581
19m 48s (- 1480m 15s) (660 1%) 1.6432
20m 24s (- 1480m 16s) (680 1%) 1.6289
21m 0s (- 1479m 48s) (700 1%) 1.5508
21m 36s (- 1478m 49s) (720 1%) 1.5253
22m 12s (- 1478m 34s) (740 1%) 1.5154
22m 48s (- 1478m 9s) (760 1%) 1.4367
23m 25s (- 1478m 21s) (780 1%) 1.3914
24m 2s (- 1478m 6s) (800 1%) 1.4067
24m 38s (- 1477m 26s) (820 1%) 1.3517
25m 14s (- 1477m 25s) (840 1%) 1.2992
25m 51s (- 1477m 10s) (860 1%) 1.2663
26m 27s (- 1476m 42s) (880 1%) 1.2122
27m 3s (- 1475m 50s) (900 1%) 1.2101
27m 38s (- 1475m 0s) (920 1%) 1.1728
28m 15s (- 1474m 43s) (940 1%) 1.1274
28m 51s (- 1474m 27s) (960 1%) 1.1317
29m 28s (- 1474m 22s) (980 1%) 1.0582
30m 4s (- 1473m 33s) (1000 2%) 1.0331
30m 41s (- 1473m 26s) (1020 2%) 1.0256
31m 18s (- 1473m 31s) (1040 2%) 0.9916
31m 54s (- 1473m 27s) (1060 2%) 0.9797
32m 31s (- 1473m 12s) (1080 2%) 0.9083
33m 8s (- 1473m 0s) (1100 2%) 0.8935
33m 44s (- 1472m 20s) (1120 2%) 0.9068
34m 21s (- 1472m 24s) (1140 2%) 0.8751
34m 56s (- 1471m 26s) (1160 2%) 0.8495
35m 33s (- 1471m 22s) (1180 2%) 0.8085
36m 10s (- 1471m 27s) (1200 2%) 0.8214
36m 45s (- 1469m 47s) (1220 2%) 0.8005
37m 21s (- 1469m 13s) (1240 2%) 0.7534
37m 59s (- 1469m 21s) (1260 2%) 0.7534
38m 34s (- 1468m 33s) (1280 2%) 0.7450
39m 11s (- 1468m 27s) (1300 2%) 0.7239
39m 48s (- 1468m 8s) (1320 2%) 0.6771
40m 25s (- 1467m 56s) (1340 2%) 0.6904
41m 1s (- 1467m 25s) (1360 2%) 0.6561
41m 38s (- 1466m 54s) (1380 2%) 0.6191
42m 14s (- 1466m 27s) (1400 2%) 0.5595
42m 50s (- 1465m 38s) (1420 2%) 0.5670
43m 26s (- 1465m 10s) (1440 2%) 0.5784
44m 2s (- 1464m 24s) (1460 2%) 0.5575
44m 39s (- 1464m 2s) (1480 2%) 0.5276
45m 12s (- 1461m 53s) (1500 3%) 0.5492
45m 48s (- 1461m 4s) (1520 3%) 0.5415
46m 25s (- 1460m 50s) (1540 3%) 0.5114
47m 2s (- 1460m 28s) (1560 3%) 0.4885
47m 38s (- 1459m 45s) (1580 3%) 0.4923
48m 14s (- 1459m 14s) (1600 3%) 0.4519
48m 51s (- 1458m 54s) (1620 3%) 0.4560
49m 27s (- 1458m 22s) (1640 3%) 0.4397
50m 3s (- 1457m 43s) (1660 3%) 0.4215
50m 38s (- 1456m 30s) (1680 3%) 0.4191
51m 13s (- 1455m 31s) (1700 3%) 0.4042
51m 49s (- 1454m 31s) (1720 3%) 0.4179
52m 25s (- 1453m 54s) (1740 3%) 0.3818
53m 1s (- 1453m 30s) (1760 3%) 0.3581
53m 38s (- 1453m 9s) (1780 3%) 0.3668
54m 15s (- 1453m 2s) (1800 3%) 0.3531
54m 53s (- 1452m 54s) (1820 3%) 0.3542
55m 30s (- 1452m 41s) (1840 3%) 0.3418
56m 6s (- 1452m 20s) (1860 3%) 0.3171
56m 43s (- 1451m 48s) (1880 3%) 0.3197
57m 19s (- 1451m 23s) (1900 3%) 0.3101
57m 55s (- 1450m 37s) (1920 3%) 0.3008
58m 32s (- 1450m 6s) (1940 3%) 0.2947
59m 8s (- 1449m 38s) (1960 3%) 0.2741
59m 45s (- 1449m 13s) (1980 3%) 0.2970
60m 21s (- 1448m 40s) (2000 4%) 0.2850
60m 57s (- 1447m 58s) (2020 4%) 0.2623

Sample

> 他騎腳踏車去
= he went by bicycle .
< he went by bicycle . <EOS>

> 雞肉還不夠熟
= the chicken is undercooked .
< the chicken is undercooked . <EOS>

> 她看起来很年轻
= she looks young .
< she looks very young . <EOS>

> 别让汤姆走了
= don t let tom leave .
< don t let tom leave . <EOS>

> 終於星期五了
= finally it s friday .
< finally it s friday . <EOS>

> 也许下一次吧
= maybe some other time .
< maybe it a piece s time . <EOS>

> 你真壞
= you re so bad .
< you re really terrible . <EOS>

> 坦率地说 他错了
= frankly speaking he is wrong .
< frankly speaking he s wrong . <EOS>

> 我是个高中生
= i m a high school student .
< i m a high school student . <EOS>