## **MASK-ALIGN: Self-Supervised Neural Word Alignment**

**Chi Chen** [1] _[,]_ [3] _[,]_ [4] **, Maosong Sun** [1] _[,]_ [3] _[,]_ [4] _[,]_ [5] **, Yang Liu** _[∗]_ [1] _[,]_ [2] _[,]_ [3] _[,]_ [4] _[,]_ [5]

1Department of Computer Science and Technology, Tsinghua University, Beijing, China
2Institute for AI Industry Research, Tsinghua University, Beijing, China
3Institute for Artificial Intelligence, Tsinghua University, Beijing, China
4Beijing National Research Center for Information Science and Technology
5Beijing Academy of Artificial Intelligence



**Abstract**


Word alignment, which aims to align translationally equivalent words between source and
target sentences, plays an important role in
many natural language processing tasks. Current unsupervised neural alignment methods
focus on inducing alignments from neural machine translation models, which does not leverage the full context in the target sequence. In
this paper, we propose MASK-ALIGN, a selfsupervised word alignment model that takes
advantage of the full context on the target side.
Our model parallelly masks out each target token and predicts it conditioned on both source
and the remaining target tokens. This two-step
process is based on the assumption that the
source token contributing most to recovering
the masked target token should be aligned.
We also introduce an attention variant called
_leaky attention_, which alleviates the problem
of high cross-attention weights on specific tokens such as periods. Experiments on four language pairs show that our model outperforms
previous unsupervised neural aligners and obtains new state-of-the-art results.


**1** **Introduction**


Word alignment is an important task of finding
the correspondence between words in a sentence
pair (Brown et al., 1993) and used to be a key
component of statistical machine translation (SMT)
(Koehn et al., 2003; Dyer et al., 2013). Although
word alignment is no longer explicitly modeled in
neural machine translation (NMT) (Bahdanau et al.,
2015; Vaswani et al., 2017), it is often leveraged to
analyze NMT models (Tu et al., 2016; Ding et al.,
2017). Word alignment is also used in many other
scenarios such as imposing lexical constraints on
the decoding process (Arthur et al., 2016; Hasler
et al., 2018), improving automatic post-editing (Pal


_∗_ Corresponding author



**Tokyo**


Induced alignment link: **Tokio - Tokyo**


Figure 1: An example of inducing an alignment link for
target token “Tokyo” in MASK-ALIGN. First, we mask
out “Tokyo” and predict it with source and other target
tokens. Then, the source token “Tokio” that contributes
most to recovering the masked word (highlighted in
red) is chosen to be aligned to “Tokyo”.


et al., 2017), and providing guidance for translators
in computer-aided translation (Dagan et al., 1993).

Compared with statistical methods, neural methods can learn representations end-to-end from raw
data and have been successfully applied to supervised word alignment (Yang et al., 2013; Tamura
et al., 2014). For unsupervised word alignment,
however, previous neural methods fail to significantly exceed their statistical counterparts such
as FAST-ALIGN (Dyer et al., 2013) and GIZA++
(Och and Ney, 2003). Recently, there is a surge of
interest in NMT-based alignment methods which
take alignments as a by-product of NMT systems
(Li et al., 2019; Garg et al., 2019; Zenkel et al.,
2019, 2020; Chen et al., 2020). Using attention
weights or feature importance measures to induce
alignments for to-be-predicted target tokens, these
methods outperform unsupervised statistical aligners like GIZA++ on a variety of language pairs.

Although NMT-based unsupervised aligners
have proven to be effective, they suffer from two
major limitations. First, due to the autoregressive
property of NMT systems (Sutskever et al., 2014),


Alignment Attention Weights - ��� ���� ����









Leaky Attention


|t1|Col2|Col3|Col4|Col5|
|---|---|---|---|---|
|t**2**|||||
|t**3**|||||
|t**4**|||||





h **1** h **2** h **3** h **4**




|Col1|Col2|Col3|Col4|
|---|---|---|---|
|t|t|t|t|
|t||||
|t||||





Feed Forward



✕ L



L ✕













��� ���� ������ ��� - ��� ���� ����


Figure 2: The architecture of MASK-ALIGN.



they only leverage part of the target context. This
inevitably brings noisy alignments when the prediction is ambiguous. Consider the target sentence
in Figure 1. When predicting “Tokyo”, an NMT
system may generate “1968” because future context is not observed, leading to a wrong alignment
link (“1968”, “Tokyo”). Second, they have to incorporate an additional guided alignment loss (Chen
et al., 2016) to outperform GIZA++. This loss requires pseudo alignments of the full training data
to guide the training of the model. Although these
pseudo alignments can be utilized to partially alleviate the problem of ignoring future context, they
are computationally expensive to obtain.


In this paper, we propose a self-supervised
model specifically designed for the word alignment
task, namely MASK-ALIGN. Our model parallelly
masks out each target token and recovers it conditioned on the source and other target tokens. Figure 1 shows an example where the target token
“Tokyo” is masked out and re-predicted. Intuitively,
as all source tokens except “Tokio” can find their
counterparts on the target side, “Tokio” should be
aligned to the masked token. Based on this intuition, we assume that the source token contributing
most to recovering a masked target token should be
aligned to that target token. Compared with NMTbased methods, MASK-ALIGN is able to take full
advantage of bidirectional context on the target side
and hopefully achieves higher alignment quality.
We also introduce an attention variant called _leaky_
_attention_ to reduce the high attention weights on
specific tokens such as periods. By encouraging
agreement between two directional models both
for training and inference, our method consistently



outperforms the state-of-the-art on four language
pairs without using guided alignment loss.


**2** **Approach**


Figure 2 shows the architecture of our model. The
model predicts each target token conditioned on the
source and other target tokens and generates alignments from the attention weights between source
and target (Section 2.1). Specifically, our approach
introduces two attention variants, _static-KV atten-_
_tion_ and _leaky attention_, to efficiently obtain attention weights for word alignment. To better utilize
attention weights from two directions, we encourage agreement between two unidirectional models
during both training (Section 2.2) and inference
(Section 2.3).


**2.1** **Modeling**


Conventional unsupervised neural aligners are
based on NMT models (Peter et al., 2017; Garg
et al., 2019). Given a source sentence **x** =
_x_ 1 _, . . ., xJ_ and a target sentence **y** = _y_ 1 _, . . ., yI_,
NMT models the probability of the target sentence
conditioned on the source sentence:



where **y** _<i_ is a partial translation. One problem of
this type of approaches is that they fail to exploit the
future context on the target side, which is probably
helpful for word alignment.
To address this problem, we model the same
conditional probability but predict each target token
_yi_ conditioned on the source sentence **x** and the



_P_ ( **y** _|_ **x** ; _θ_ ) =



_I_


_P_ ( _yi|_ **y** _<i,_ **x** ; _θ_ ) (1)

_i_ =1


remaining target tokens **y** _\yi_ :



Alignment





_P_ ( **y** _|_ **x** ; _θ_ ) =



_I_


_P_ ( _yi|_ **y** _\yi,_ **x** ; _θ_ ) (2)

_i_ =1



This equals to masking out each _yi_ and then recovering it. We build our model on top of Transformer
(Vaswani et al., 2017) which is the state-of-the-art
sequence-to-sequence architecture. Next, we will
discuss in detail the implementation of our model.


**Static-KV Attention**


As self-attention is fully-connected, directly computing [�] _i_ _[I]_ =1 _[P]_ [(] _[y][i][|]_ **[y]** _[\][y][i][,]_ **[ x]** [;] _[ θ]_ [)][ with a vanilla Trans-]
former requires _I_ separate forward passes, in each
of which only one target token is masked out
and predicted. This is costly and time-consuming.
Therefore, how to parallelly mask out and predict
all target tokens in a single pass is important.
To do so, a major challenge is to avoid the representation of a masked token getting involved in
the prediction process of itself. Inspired by Kasai
et al. (2020), we modify the self-attention in the
Transformer decoder to perform the forward passes
concurrently. Given the word embedding **w** _i_ and
position embedding **p** _i_ for target token _yi_, we first
separate the query inputs **q** _i_ from key **k** _i_ and value
inputs **v** _i_ to prevent the to-be-predicted token itself
from participating in the prediction:


**q** _i_ = **p** _i_ **W** _[Q]_ (3)

**k** _i_ = ( **w** _i_ + **p** _i_ ) **W** _[K]_ (4)

**v** _i_ = ( **w** _i_ + **p** _i_ ) **W** _[V]_ (5)


where **W** _[Q]_, **W** _[K]_ and **W** _[V]_ are parameter matrices.
The hidden representation **h** _i_ for _yi_ is computed by
attending to keys and values, **K** = _i_ and **V** = _i_, that
correspond to the remaining tokens **y** _\yi_ :


**h** _i_ = Attention( **q** _i,_ **K** = _i,_ **V** = _i_ ) (6)

**K** = _i_ = Concat( _{_ **k** _m|m ̸_ = _i}_ ) (7)

**V** = _i_ = Concat( _{_ **v** _m|m ̸_ = _i}_ ) (8)


In this way, we ensure that **h** _i_ is isolated from the
word embedding **w** _i_ in a single decoder layer. However, there exists a problem of information leakage
if we update the key and value inputs for each position across decoder layers since they will contain
the representation of each position from previous
layers. Therefore, we keep the key and value inputs unchanged and only update the query inputs



Attention Weights


Figure 3: An example of inducing alignments from attention weights where the source token “.” has high attention weights. The two “in”s in the target sentence
are wrongly aligned to “.” because of the high attention
weights on it.


to avoid information leakage:


**h** _[l]_ _i_ [= Attention(] **[q]** _i_ _[l][,]_ **[ K]** [=] _[i][,]_ **[ V]** [=] _[i]_ [)] (9)

**q** _[l]_ _i_ [=] **[ h]** _i_ _[l][−]_ [1] **W** _[Q]_ (10)


where **q** _[l]_ _i_ [and] **[ h]** _i_ _[l]_ [denote the query inputs and hidden]
states for _yi_ in the _l_ -th layer, respectively. **h** [0] _i_ [is ini-]
tialized with **p** _i_ . We name this variant of attention
the **static-KV attention** . By static-KV, we mean
the keys and values are unchanged across different
layers in our approach. Our model replaces all selfattention in the decoder with static-KV attention.


**Leaky Attention**


Extracting alignments from vanilla cross-attention
often suffers from the high attention weights on
some specific source tokens such as periods, [EOS],
or other high frequency tokens (see Figure 3). This
is similar to the “garbage collectors” effect (Moore,
2004) in statistical aligners, where a source token
is aligned to too many target tokens. Hereinafter,
we will refer to these tokens as _collectors_ . As a
result of such effect, many target tokens (e.g., the


not 1.0



**vanilla**
**attention**


**leaky**
**attention**



true



0.5 0.5


not true



true



1.0 falsch


falsch



|0.4|0.6|
|---|---|
|0.2|0.8|


falsch



falsch 0.2 0.4 0.4




[NULL]




[NULL]



not true



Figure 4: An illustrative example of the attention
weights from two directional models using vanilla and
leaky attention. Leaky attention provides a leak position “[NULL]” to collect extra attention weights.


two “in”s in Figure 3) will be incorrectly aligned
to the collectors according to the attention weights.
This phenomenon has been studied in previous
works (Clark et al., 2019; Kobayashi et al., 2020).
Kobayashi et al. (2020) show that the norms of the
value vectors for the collectors are usually small,
making their influence on attention outputs actually
limited. We conjecture that this phenomenon is due
to the incapability of NMT-based aligners to deal
with tokens that have no counterparts on the other
side because there is no empty (NULL) token that
is widely used in statistical aligners (Brown et al.,
1993; Och and Ney, 2003).
We propose to explicitly model the NULL token with an attention variant, namely **leaky atten-**
**tion** . As shown in Figure 4, when calculating crossattention weights, leaky attention provides an extra
“leak” position in addition to the encoder outputs.
Acting as the NULL token, this leak position is expected to address the biased attention weight problem. To be specific, we parameterize the key and
value vectors as **k** NULL and **v** NULL for the leak position in the cross-attention, and concatenate them
with the transformed vectors of the encoder outputs.
The attention output **z** _i_ is computed as follows:


**z** _i_ = Attention( **h** _[L]_ _i_ **[W]** _[Q][,]_ **[ K]** _[,]_ **[ V]** [)] (11)

**K** = Concat( **k** NULL _,_ **H** enc **W** _[K]_ ) (12)

**V** = Concat( **v** NULL _,_ **H** enc **W** _[V]_ ) (13)


where **H** enc denotes encoder outputs. [1] We use a
normal distribution with a mean of 0 and a small


1A similar attention implementation can be found
in [https://github.com/pytorch/fairseq/blob/master/fairseq/](https://github.com/pytorch/fairseq/blob/master/fairseq/modules/multihead_attention.py)
[modules/multihead](https://github.com/pytorch/fairseq/blob/master/fairseq/modules/multihead_attention.py) ~~a~~ ttention.py.



deviation to initialize **k** NULL and **v** NULL to ensure
that their initial norms are rather small. When extracting alignments, we only consider the attention
matrix without the leak position.

Note that leaky attention is different from adding
a special token in the source sequence, which will
share the same high attention weights with the existing collector instead of calibrating it (Vig and Belinkov, 2019). Our parameterized method is more
flexible than Leaky-Softmax (Sabour et al., 2017)
which adds an extra dimension with the value of
zero to the routing logits. In Section 2.2, we will
show that leaky attention is also helpful for applying agreement-based training on two directional
models.

We remove the cross-attention in all but the last
decoder layer. This makes the interaction between
the source and target restricted in the last layer.
Our experiments demonstrate that this modification improves alignment results with fewer model
parameters.


**2.2** **Training**


To better utilize the attention weights from two directions, we apply an agreement loss in the training
process to improve the symmetry of our model,
which has proven effective in statistical alignment
models (Liang et al., 2006; Liu et al., 2015). Given
a parallel sentence pair _⟨_ **x** _,_ **y** _⟩_, we can obtain the
attention weights from two different directions, denoted as _**W**_ **x** _→_ **y** and _**W**_ **y** _→_ **x** . As alignment is bijective, _**W**_ **x** _→_ **y** is supposed to be equal to the transpose of _**W**_ **y** _→_ **x** . We encourage this kind of symmetry through an agreement loss:


_La_ = MSE    - _**W**_ **x** _→_ **y** _,_ _**W**_ _[⊤]_ **y** _→_ **x**    - (14)


where MSE represents the mean squared error.
For vanilla attention, _La_ is hardly small because
of the normalization constraint. As shown in Figure
4, due to the use of softmax activation, the minimal
value of _La_ is 0 _._ 25 for vanilla attention. Using
leaky attention, our approach can achieve a lower
agreement loss ( _La_ = 0.1) by adjusting the weights
on the leak position.

However, our model may converge to a degenerate case of zero agreement loss where attention
weights are all zero except for the leak position.
We circumvent this case by introducing an entropy


loss on the attention weights:



_J_

- _W_ ˜ **x** _[ij]_ _→_ **y** [log ˜] _[W][ij]_ (15)

_j_ =1



(Zenkel et al., 2019, 2020) and used the preprocessing scripts from Zenkel et al. (2019) [3] . Following
Ding et al. (2019), we take the last 1000 sentences
of the training data for these three datasets as validation sets. We used a joint source and target Byte
Pair Encoding (BPE) (Sennrich et al., 2016) with
40k merge operations. During training, we filtered
out sentences with the length of 1 to ensure the
validity of the masking process.


**3.2** **Settings**


We implemented our model based on the Transforemr architecture (Vaswani et al., 2017). The encoder consists of 6 standard Transformer encoder
layers. The decoder is composed of 6 layers, each
of which contains static-KV attention while only
the last layer is equipped with leaky attention. We
set the embedding size to 512, the hidden size to
1024, and attention heads to 4. The input and output
embeddings are shared for the decoder.

We trained the models with a batch size of 36K
tokens. We used early stopping based on the prediction accuracy on the validation sets. We tuned
the hyperparameters via grid search on the ChineseEnglish validation set as it contains gold word alignments. In all of our experiments, we set _λ_ = 0 _._ 05
(Eq. (16)), _α_ = 5, _β_ = 1 (Eq. (17)) and _τ_ = 0 _._ 2
(Eq. (19)). The evaluation metric is Alignment Error Rate (AER) (Och and Ney, 2000).


**3.3** **Baselines**


We introduce the following unsupervised neural
baselines besides two statistical baselines FASTALIGN and GIZA++:


  - NAIVE-ATT (Garg et al., 2019): a method
that induces alignments from cross-attention
weights of the best (usually penultimate) decoder layer in a vanilla Tranformer.

  - NAIVE-ATT-LAST: same as NAIVE-ATT except that only the last decoder layer performs
cross-attention.

  - ADDSGD (Zenkel et al., 2019): a method that
adds an extra alignment layer to repredict the
to-be-aligned target token.

  - MTL-FULLC (Garg et al., 2019): a method
that supervises an attention head with symmetrized NAIVE-ATT alignments in a multitask learning framework.


[3https://github.com/lilt/alignment-scripts](https://github.com/lilt/alignment-scripts)



_Le,_ **x** _→_ **y** = _−_ [1]

_I_



_I_



_i_ =1



_W_ **x** _[ij]_ _→_ **y** [+] _[ λ]_
_W_ ˜ **x** _[ij]_ _→_ **y** [=] ~~�~~ (16)
_j_ [(] _[W][ ij]_ **x** _→_ **y** [+] _[ λ]_ [)]


where _W_ [˜] **x** _[ij]_ _→_ **y** [is the renormalized attention weights]
and _λ_ is a smoothing hyperparamter. Similarly, we
have _Le,_ **y** _→_ **x** for the inverse direction.
We jointly train two directional models using the
following loss:


_L_ = _L_ **x** _→_ **y** + _L_ **y** _→_ **x** + _αLa_ +

(17)
_β_ ( _Le,_ **x** _→_ **y** + _Le,_ **y** _→_ **x** )


where _L_ **x** _→_ **y** and _L_ **y** _→_ **x** are NLL losses, _α_ and _β_
are hyperparameters.


**2.3** **Inference**


When extracting alignments, we compute an alignment score _Sij_ for _yi_ and _xj_ as the harmonic mean
of attention weights _W_ **x** _[ij]_ _→_ **y** [and] _[ W][ ji]_ **y** _→_ **x** [from two]
directional models:


**x** _→_ **y** _[W][ ji]_ **y** _→_ **x**
_Sij_ = [2] _[ W][ ij]_ (18)
_W_ **x** _[ij]_ _→_ **y** [+] _[ W][ ji]_ **y** _→_ **x**


We use the harmonic mean because we assume a
large _Sij_ requires both _W_ **x** _[ij]_ _→_ **y** [and] _[ W][ ji]_ **y** _→_ **x** [to be]
large. Word alignments can be induced from the
alignment score matrix as follows:


     - 1 if _Sij ≥_ _τ_
_Aij_ = (19)
0 otherwise


where _τ_ is a threshold.


**3** **Experiments**


**3.1** **Datasets**


We conducted our experiments on four public
datasets: German-English (De-En), English-French
(En-Fr), Romanian-English (Ro-En) and ChineseEnglish (Zh-En). The Chinese-English training set
is from the LDC corpus that consists of 1.2M sentence pairs. For validation and testing, we used the
Chinese-English alignment dataset from Liu et al.
(2005) [2], which contains 450 sentence pairs for validation and 450 for testing. For other three language pairs, we followed the experimental setup in


[2http://nlp.csai.tsinghua.edu.cn/](http://nlp.csai.tsinghua.edu.cn/~ly/systems/TsinghuaAligner/TsinghuaAligner.html) _∼_ ly/systems/
[TsinghuaAligner/TsinghuaAligner.html](http://nlp.csai.tsinghua.edu.cn/~ly/systems/TsinghuaAligner/TsinghuaAligner.html)


**Method** **Guided** **De-En** **En-Fr** **Ro-En** **Zh-En**


FAST-ALIGN (Dyer et al., 2013) N 25.7 12.1 31.8     GIZA++ (Och and Ney, 2003) N 17.8 6.1 26.0 18.5


NAIVE-ATT (Garg et al., 2019) N 31.9 18.5 32.9 28.9
NAIVE-ATT-LAST N 28.4 17.7 32.4 26.4
ADDSGD (Zenkel et al., 2019) N 21.2 10.0 27.6     MTL-FULLC (Garg et al., 2019) N 20.2 7.7 26.0     BAO (Zenkel et al., 2020) N 17.9 8.4 24.1     SHIFT-ATT (Chen et al., 2020) N 17.9 6.6 23.9 20.2


MTL-FULLC-GZ (Garg et al., 2019) Y 16.0 4.6 23.1     BAO-GUIDED (Zenkel et al., 2020) Y 16.3 5.0 23.4     SHIFT-AET (Chen et al., 2020) Y 15.4 4.7 21.2 17.2


MASK-ALIGN N **14.4** **4.4** **19.5** **13.8**


Table 1: Alignment Error Rate (AER) scores on four datasets for different alignment methods. The lower AER, the
better. “Guided” denotes whether the guided alignment loss is used during training. All results are symmetrized.
We highlight the best results for each language pair in bold.




  - BAO (Zenkel et al., 2020): an improved version of ADDSGD that extracts alignments
with Bidirectional Attention Optimization.

  - SHIFT-ATT (Chen et al., 2020): a method that
induces alignments when the to-be-aligned
tatget token is the decoder input instead of the
output.


We also included three additional baselines with
guided training: (1) MTL-FULLC-GZ (Garg et al.,
2019) which replaces the alignment labels in MTLFULLC with GIZA++ results, (2) BAO-GUIDED
(Zenkel et al., 2020) which uses alignments from
BAO for guided alignment training, (3) SHIFTAET (Chen et al., 2020) which trains an additional alignment module with supervision from
symmetrized SHIFT-ATT alignments.


**3.4** **Main Results**


Table 1 shows the results on four datasets. Our
approach significantly outperforms all statistical
and neural baselines. Specifically, it improves over
GIZA++ by 1.7-6.5 AER points across different
language pairs without using any guided alignment
loss, making it a good substitute to this commonly
used statistical alignment tool. Compared to SHIFTATT, the best neural methods without guided training, our approach achieves a gain of 2.2-6.4 AER
points with fewer parameters (as we remove some
cross-attention sublayers in the decoder).
When compared with baselines using guided
training, we find MASK-ALIGN still achieves sub


Masked Leaky Agree AER


_×_ _×_ _×_ 28.4
✓ _×_ _×_ 27.2
_×_ ✓ _×_ 28.3
_×_ _×_ ✓ 26.6
_×_ ✓ ✓ 23.4
✓ _×_ ✓ 17.6
✓ ✓ _×_ 17.2


✓ ✓ ✓ **14.4**


Table 2: Ablation study on the German-English dataset.
We use “Masked” to denote the masked modeling with
static-KV attention in Section 2.1, “Leaky” to denote
the leaky attention in Section 2.1 and “Agree” to denote
the agreement-based training and inference in Sections
2.2 and 2.3.


stantial improvements over all methods. For example, on the Romanian-English dataset, it improves over SHIFT-AET by 1.7 AER points. Recall
that our method is fully end-to-end, which does
not require a time-consuming process of obtaining
pseudo alignments for full training data.


**3.5** **Ablation Study**


Table 2 shows the ablation results on the GermanEnglish dataset. As we can see, masked modeling
seems to play a critical role since removing it will
deteriorate the performance by at least 9.0 AER.
We also find that leaky attention and agreementbased training and inference are both important.
Removing any of them will significantly diminish


human


rights


throughout


the


world


in


1995


       

1996



human


rights


throughout


the


world


in


1995


       

1996



(a) Vanilla Attention



(b) Leaky Attention



Figure 5: Attention weights from vanilla and leaky attention. “MR” is short for “menschenrechte”, which means
“human rights” in English. We use “[NULL]” to denote the leak position.


source sentence [NULL] MR in der welt 1995 _\_ 1996


vanilla attention    - 21.1 11.7 **5.2** 15.0 21.2 17.7 21.8
leaky attention **1.9** 28.5 17.2 18.1 20.2 24.2 21.4 23.8


Table 3: Norms of the transformed value vectors of different source tokens in Figure 5. We mark the minimum
norm for each variant of attention with boldface.



the performance.


**3.6** **Effect of Leaky Attention**


Figure 5 shows the attention weights from vanilla
and leaky attention and Table 3 presents the norms
of the transformed value vectors of each source token for two types of attention. For vanilla attention,
we can see large weights on the high frequency
token “der” and the small norm of its transformed
value vector. As a result, the target token “in” will
be wrongly aligned to “der”. While for leaky attention, we observe a similar phenomenon on the leak
position “[NULL]”, and “in” will not be aligned to
any source tokens since the weights on all source tokens are small. This example shows leaky attention
can effectively prevent the collector phenomenon.


**3.7** **Analysis**


**Removing End Punctuation** To further investigate the performance of leaky attention, we tested
an extraction method that excludes the attention
weights on the end punctuation of a source sentence.
The reason behind this is that when the source sentence contains the end punctuation, it will act as the
collector in most cases. Therefore removing it will



Method w/ punc. w/o punc.


vanilla attention 27.2 17.7
leaky attention 17.2 17.4


Table 4: Comparison of AER with and without considering the attention weights on end punctuation.


alleviate the effect of collectors to a certain extent.
Table 4 shows the comparison results. For vanilla attention, removing end punctuation obtains a gain of
7.7 AER points. For leaky attention, however, such
extraction method brings no improvement on alignment quality. This suggests that leaky attention can
effectively alleviate the problem of collectors.


**Case Study** Figure 6 shows the attention weights
from four different models for the example in Figure 1. As we have discussed in Section 1, in this
example, NMT-based methods might fail to resolve
ambiguity when predicting the target token “tokyo”.
From the attention weight matrices, we can see that
NMT-based methods (Figures 6(b) and 6(c)) indeed put high weights wrongly on “1968” in the
source sentence. As for MASK-ALIGN, we can see


i

was


born


in


tokyo


in


1968

.



(d) MASK-ALIGN



(a) Reference



(b) NAIVE-ATT-LAST



(c) SHIFT-ATT



Figure 6: Attention weights from different models for the example in Figure 1. Gold alignment is shown in (a). For
target token “tokyo”, NMT-based methods NAIVE-ATT-LAST (b) and SHIFT-ATT (c) assign high weights to the
wrongly aligned source token “1968”, while MASK-ALIGN (d) focuses on the correct source token “tokio”.






|7 cPcA<br>wPcA<br>6 cPwA<br>wPwA<br>5<br>4<br>3<br>2<br>1<br>0<br>Naive-Att Naive-Att-Last Shift-Att Mask-Align|cPcA<br>wPcA<br>cPwA<br>wPwA|Col3|Col4|
|---|---|---|---|
|Naive~~-~~Att<br>Naive~~-~~Att~~-~~Last<br>Shift~~-~~Att<br>Mask~~-~~Align<br>0<br>1<br>2<br>3<br>4<br>5<br>6<br>7<br>cPcA<br>wPcA<br>cPwA<br>wPwA|cPcA<br>wPcA<br>cPwA<br>wPwA|ift~~-~~Att<br>Ma|k~~-~~Align|



Figure 7: Relations between prediction and alignment
for different methods.


that the attention weights are highly consistent with
the gold alignment, showing that our method can
generate sparse and accurate attention weights.


**Prediction and Alignment** We analyzed the relevance between the correctness of word-level prediction and alignment. We regard a word as correctly
predicted if any of its subwords are correct and as
correctly aligned if one of its possible alignment
is matched. Figure 7 shows the results. We divide
target tokens into four categories:


1. cPcA: correct prediction & correct alignment;


2. wPcA: wrong prediction & correct alignment;


3. cPwA: correct prediction & wrong alignment;


4. wPwA: wrong prediction & wrong alignment.


Compared with other methods, MASK-ALIGN
significantly reduces the alignment errors caused by
wrong predictions (wPwA). In addition, the number of the tokens with correct prediction but wrong



alignment (cPwA) maintains at a low level, indicating that our model does not degenerate into a
target masked language model despite the use of
bidirectional target context.


**4** **Related Work**


Our work is closely related to unsupervised neural
word alignment. While early unsupervised neural
aligners (Tamura et al., 2014; Alkhouli et al., 2016;
Peter et al., 2017) failed to outperform their statistical counterparts such as FAST-ALIGN (Dyer et al.,
2013) and GIZA++ (Och and Ney, 2003), recent
studies have made significant progress by inducing
alignments from NMT models (Garg et al., 2019;
Zenkel et al., 2019, 2020; Chen et al., 2020). Our
work differs from prior studies in that we design a
novel self-supervised model that is capable of utilizing more target context than NMT-based models
to generate high quality alignments without using
guided training.

Our work is also inspired by the success of
conditional masked language models (CMLMs)
(Ghazvininejad et al., 2019), which have been applied to non-autoregressive machine translation.
The CMLM can leverage both previous and future
context on the target side for sequence-to-sequence
tasks with the masking mechanism. Kasai et al.
(2020) extend it with a disentangled context Transformer that predicts every target token conditioned
on arbitrary context. By taking the characteristics
of word alignment into consideration, we propose
to use static-KV attention to achieve masking and
aligning in parallel. To the best of our knowledge,
this is the first work that incorporates a CMLM into
alignment models.


**5** **Conclusion**


We have presented a self-supervised neural alignment model MASK-ALIGN. Our model parallelly
masks out and predicts each target token. We
propose static-KV attention and leaky attention
to achieve parallel computation and address the
“garbage collectors” problem, respectively. Experiments show that MASK-ALIGN achieves new stateof-the-art results without using the guided alignment loss. In the future, we plan to extend our
method to directly generate symmetrized alignments without leveraging the agreement between
two unidirectional models.


**Acknowledgments**


This work was supported by the National Key
R&D Program of China (No. 2017YFB0202204),
National Natural Science Foundation of China
(No.61925601, No. 61772302) and Huawei Noah’s
Ark Lab. We thank all anonymous reviewers for
their valuable comments and suggestions on this
work.


**References**


Tamer Alkhouli, Gabriel Bretschner, Jan-Thorsten Peter, Mohammed Hethnawi, Andreas Guta, and Her[mann Ney. 2016. Alignment-based neural machine](https://doi.org/10.18653/v1/W16-2206)
[translation. In](https://doi.org/10.18653/v1/W16-2206) _Proceedings of the First Conference_
_on Machine Translation: Volume 1, Research Papers_,
pages 54–65, Berlin, Germany. Association for Computational Linguistics.


Philip Arthur, Graham Neubig, and Satoshi Nakamura.
2016. [Incorporating discrete translation lexicons](https://doi.org/10.18653/v1/D16-1162)
[into neural machine translation. In](https://doi.org/10.18653/v1/D16-1162) _Proceedings of_
_the 2016 Conference on Empirical Methods in Natu-_
_ral Language Processing_, pages 1557–1567, Austin,
Texas. Association for Computational Linguistics.


Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. 2015. [Neural machine translation by jointly](http://arxiv.org/abs/1409.0473)
[learning to align and translate.](http://arxiv.org/abs/1409.0473) In _3rd Inter-_
_national Conference on Learning Representations,_
_ICLR 2015, San Diego, CA, USA, May 7-9, 2015,_
_Conference Track Proceedings_ .


Peter F. Brown, Stephen A. Della Pietra, Vincent J.
[Della Pietra, and Robert L. Mercer. 1993. The math-](https://www.aclweb.org/anthology/J93-2003)
[ematics of statistical machine translation: Parameter](https://www.aclweb.org/anthology/J93-2003)
[estimation.](https://www.aclweb.org/anthology/J93-2003) _Computational Linguistics_, 19(2):263–
311.


Wenhu Chen, Evgeny Matusov, Shahram Khadivi, and
Jan-Thorsten Peter. 2016. Guided alignment training for topic-aware neural machine translation. _As-_
_sociation for Machine Translation in the Americas_,
page 121.



Yun Chen, Yang Liu, Guanhua Chen, Xin Jiang, and
[Qun Liu. 2020. Accurate word alignment induction](https://doi.org/10.18653/v1/2020.emnlp-main.42)
[from neural machine translation. In](https://doi.org/10.18653/v1/2020.emnlp-main.42) _Proceedings of_
_the 2020 Conference on Empirical Methods in Natu-_
_ral Language Processing (EMNLP)_, pages 566–576,
Online. Association for Computational Linguistics.


Kevin Clark, Urvashi Khandelwal, Omer Levy, and
[Christopher D. Manning. 2019. What does BERT](https://doi.org/10.18653/v1/W19-4828)
[look at? an analysis of BERT’s attention. In](https://doi.org/10.18653/v1/W19-4828) _Pro-_
_ceedings of the 2019 ACL Workshop BlackboxNLP:_
_Analyzing and Interpreting Neural Networks for_
_NLP_, pages 276–286, Florence, Italy. Association
for Computational Linguistics.


Ido Dagan, Kenneth Church, and Willian Gale. 1993.

[Robust bilingual word alignment for machine aided](https://www.aclweb.org/anthology/W93-0301)
[translation. In](https://www.aclweb.org/anthology/W93-0301) _Very Large Corpora: Academic and_
_Industrial Perspectives_ .


Shuoyang Ding, Hainan Xu, and Philipp Koehn. 2019.

[Saliency-driven word alignment interpretation for](https://doi.org/10.18653/v1/W19-5201)
[neural machine translation. In](https://doi.org/10.18653/v1/W19-5201) _Proceedings of the_
_Fourth Conference on Machine Translation (Volume_
_1: Research Papers)_, pages 1–12, Florence, Italy. Association for Computational Linguistics.


Yanzhuo Ding, Yang Liu, Huanbo Luan, and Maosong
Sun. 2017. [Visualizing and understanding neural](https://doi.org/10.18653/v1/P17-1106)
[machine translation. In](https://doi.org/10.18653/v1/P17-1106) _Proceedings of the 55th An-_
_nual Meeting of the Association for Computational_
_Linguistics (Volume 1: Long Papers)_, pages 1150–
1159, Vancouver, Canada. Association for Computational Linguistics.


Chris Dyer, Victor Chahuneau, and Noah A. Smith.
2013. [A simple, fast, and effective reparameter-](https://www.aclweb.org/anthology/N13-1073)
[ization of IBM model 2.](https://www.aclweb.org/anthology/N13-1073) In _Proceedings of the_
_2013 Conference of the North American Chapter of_
_the Association for Computational Linguistics: Hu-_
_man Language Technologies_, pages 644–648, Atlanta, Georgia. Association for Computational Linguistics.


Sarthak Garg, Stephan Peitz, Udhyakumar Nallasamy,
[and Matthias Paulik. 2019. Jointly learning to align](https://doi.org/10.18653/v1/D19-1453)
[and translate with transformer models. In](https://doi.org/10.18653/v1/D19-1453) _Proceed-_
_ings of the 2019 Conference on Empirical Methods_
_in Natural Language Processing and the 9th Inter-_
_national Joint Conference on Natural Language Pro-_
_cessing (EMNLP-IJCNLP)_, pages 4453–4462, Hong
Kong, China. Association for Computational Linguistics.


Marjan Ghazvininejad, Omer Levy, Yinhan Liu, and
[Luke Zettlemoyer. 2019. Mask-predict: Parallel de-](https://doi.org/10.18653/v1/D19-1633)
[coding of conditional masked language models. In](https://doi.org/10.18653/v1/D19-1633)
_Proceedings of the 2019 Conference on Empirical_
_Methods in Natural Language Processing and the_
_9th International Joint Conference on Natural Lan-_
_guage Processing (EMNLP-IJCNLP)_, pages 6112–
6121, Hong Kong, China. Association for Computational Linguistics.


Eva Hasler, Adri`a de Gispert, Gonzalo Iglesias, and
[Bill Byrne. 2018. Neural machine translation decod-](https://doi.org/10.18653/v1/N18-2081)
[ing with terminology constraints. In](https://doi.org/10.18653/v1/N18-2081) _Proceedings of_
_the 2018 Conference of the North American Chap-_
_ter of the Association for Computational Linguistics:_
_Human Language Technologies, Volume 2 (Short Pa-_
_pers)_, pages 506–512, New Orleans, Louisiana. Association for Computational Linguistics.


Jungo Kasai, James Cross, Marjan Ghazvininejad, and
[Jiatao Gu. 2020. Non-autoregressive machine trans-](http://proceedings.mlr.press/v119/kasai20a.html)
[lation with disentangled context transformer. In](http://proceedings.mlr.press/v119/kasai20a.html) _Pro-_
_ceedings of the 37th International Conference on_
_Machine Learning, ICML 2020, 13-18 July 2020,_
_Virtual Event_, volume 119 of _Proceedings of Ma-_
_chine Learning Research_, pages 5144–5155. PMLR.


Goro Kobayashi, Tatsuki Kuribayashi, Sho Yokoi, and
Kentaro Inui. 2020. Attention module is not only a
weight: Analyzing transformers with vector norms.
In _Proceedings of the 2020 Conference on Empiri-_
_cal Methods in Natural Language Processing and_
_the 10th International Joint Conference on Natural_
_Language Processing (EMNLP-IJCNLP)_ .


Philipp Koehn, Franz J. Och, and Daniel Marcu. 2003.

[Statistical phrase-based translation. In](https://www.aclweb.org/anthology/N03-1017) _Proceedings_
_of the 2003 Human Language Technology Confer-_
_ence of the North American Chapter of the Associa-_
_tion for Computational Linguistics_, pages 127–133.


Xintong Li, Guanlin Li, Lemao Liu, Max Meng, and
[Shuming Shi. 2019. On the word alignment from](https://doi.org/10.18653/v1/P19-1124)
[neural machine translation. In](https://doi.org/10.18653/v1/P19-1124) _Proceedings of the_
_57th Annual Meeting of the Association for Com-_
_putational Linguistics_, pages 1293–1303, Florence,
Italy. Association for Computational Linguistics.


[Percy Liang, Ben Taskar, and Dan Klein. 2006. Align-](https://www.aclweb.org/anthology/N06-1014)
[ment by agreement. In](https://www.aclweb.org/anthology/N06-1014) _Proceedings of the Human_
_Language Technology Conference of the NAACL,_
_Main Conference_, pages 104–111, New York City,
USA. Association for Computational Linguistics.


Chunyang Liu, Yang Liu, Maosong Sun, Huanbo Luan,
and Heng Yu. 2015. [Generalized agreement for](https://doi.org/10.18653/v1/D15-1210)
[bidirectional word alignment.](https://doi.org/10.18653/v1/D15-1210) In _Proceedings of_
_the 2015 Conference on Empirical Methods in Nat-_
_ural Language Processing_, pages 1828–1836, Lisbon, Portugal. Association for Computational Linguistics.


Yang Liu, Qun Liu, and Shouxun Lin. 2005. [Log-](https://doi.org/10.3115/1219840.1219897)
[linear models for word alignment.](https://doi.org/10.3115/1219840.1219897) In _Proceed-_
_ings of the 43rd Annual Meeting of the Association_
_for Computational Linguistics (ACL’05)_, pages 459–
466, Ann Arbor, Michigan. Association for Computational Linguistics.


[Robert C. Moore. 2004. Improving IBM word align-](https://doi.org/10.3115/1218955.1219021)
[ment model 1.](https://doi.org/10.3115/1218955.1219021) In _Proceedings of the 42nd An-_
_nual Meeting of the Association for Computational_
_Linguistics (ACL-04)_, pages 518–525, Barcelona,
Spain.



[Franz Josef Och and Hermann Ney. 2000. Improved](https://doi.org/10.3115/1075218.1075274)
[statistical alignment models. In](https://doi.org/10.3115/1075218.1075274) _Proceedings of the_
_38th Annual Meeting of the Association for Com-_
_putational Linguistics_, pages 440–447, Hong Kong.
Association for Computational Linguistics.


[Franz Josef Och and Hermann Ney. 2003. A systematic](https://doi.org/10.1162/089120103321337421)
[comparison of various statistical alignment models.](https://doi.org/10.1162/089120103321337421)
_Computational Linguistics_, 29(1):19–51.


Santanu Pal, Sudip Kumar Naskar, Mihaela Vela, Qun
Liu, and Josef van Genabith. 2017. [Neural auto-](https://www.aclweb.org/anthology/E17-2056)
[matic post-editing using prior alignment and rerank-](https://www.aclweb.org/anthology/E17-2056)
[ing. In](https://www.aclweb.org/anthology/E17-2056) _Proceedings of the 15th Conference of the_
_European Chapter of the Association for Computa-_
_tional Linguistics: Volume 2, Short Papers_, pages
349–355, Valencia, Spain. Association for Computational Linguistics.


Jan-Thorsten Peter, Arne Nix, and Hermann Ney.
2017. Generating alignments using target foresight in attention-based neural machine translation.
_The Prague Bulletin of Mathematical Linguistics_,
108(1):27–36.


Sara Sabour, Nicholas Frosst, and Geoffrey E. Hinton.
[2017. Dynamic routing between capsules. In](https://proceedings.neurips.cc/paper/2017/hash/2cad8fa47bbef282badbb8de5374b894-Abstract.html) _Ad-_
_vances in Neural Information Processing Systems_
_30: Annual Conference on Neural Information Pro-_
_cessing Systems 2017, December 4-9, 2017, Long_
_Beach, CA, USA_, pages 3856–3866.


Rico Sennrich, Barry Haddow, and Alexandra Birch.
2016. [Neural machine translation of rare words](https://doi.org/10.18653/v1/P16-1162)
[with subword units. In](https://doi.org/10.18653/v1/P16-1162) _Proceedings of the 54th An-_
_nual Meeting of the Association for Computational_
_Linguistics (Volume 1: Long Papers)_, pages 1715–
1725, Berlin, Germany. Association for Computational Linguistics.


Ilya Sutskever, Oriol Vinyals, and Quoc V. Le. 2014.

[Sequence to sequence learning with neural networks.](https://proceedings.neurips.cc/paper/2014/hash/a14ac55a4f27472c5d894ec1c3c743d2-Abstract.html)
In _Advances in Neural Information Processing Sys-_
_tems 27: Annual Conference on Neural Informa-_
_tion Processing Systems 2014, December 8-13 2014,_
_Montreal, Quebec, Canada_, pages 3104–3112.


Akihiro Tamura, Taro Watanabe, and Eiichiro Sumita.
2014. [Recurrent neural networks for word align-](https://doi.org/10.3115/v1/P14-1138)
[ment model.](https://doi.org/10.3115/v1/P14-1138) In _Proceedings of the 52nd An-_
_nual Meeting of the Association for Computational_
_Linguistics (Volume 1: Long Papers)_, pages 1470–
1480, Baltimore, Maryland. Association for Computational Linguistics.


Zhaopeng Tu, Zhengdong Lu, Yang Liu, Xiaohua Liu,
and Hang Li. 2016. [Modeling coverage for neu-](https://doi.org/10.18653/v1/P16-1008)
[ral machine translation. In](https://doi.org/10.18653/v1/P16-1008) _Proceedings of the 54th_
_Annual Meeting of the Association for Computa-_
_tional Linguistics (Volume 1: Long Papers)_, pages
76–85, Berlin, Germany. Association for Computational Linguistics.


Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob
Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz


[Kaiser, and Illia Polosukhin. 2017. Attention is all](https://proceedings.neurips.cc/paper/2017/hash/3f5ee243547dee91fbd053c1c4a845aa-Abstract.html)
[you need. In](https://proceedings.neurips.cc/paper/2017/hash/3f5ee243547dee91fbd053c1c4a845aa-Abstract.html) _Advances in Neural Information Pro-_
_cessing Systems 30: Annual Conference on Neural_
_Information Processing Systems 2017, December 4-_
_9, 2017, Long Beach, CA, USA_, pages 5998–6008.


Jesse Vig and Yonatan Belinkov. 2019. [Analyzing](https://doi.org/10.18653/v1/W19-4808)
[the structure of attention in a transformer language](https://doi.org/10.18653/v1/W19-4808)
[model. In](https://doi.org/10.18653/v1/W19-4808) _Proceedings of the 2019 ACL Workshop_
_BlackboxNLP: Analyzing and Interpreting Neural_
_Networks for NLP_, pages 63–76, Florence, Italy. Association for Computational Linguistics.


Nan Yang, Shujie Liu, Mu Li, Ming Zhou, and Neng[hai Yu. 2013. Word alignment modeling with con-](https://www.aclweb.org/anthology/P13-1017)
[text dependent deep neural network.](https://www.aclweb.org/anthology/P13-1017) In _Proceed-_
_ings of the 51st Annual Meeting of the Association_
_for Computational Linguistics (Volume 1: Long Pa-_
_pers)_, pages 166–175, Sofia, Bulgaria. Association
for Computational Linguistics.


Thomas Zenkel, Joern Wuebker, and John DeNero.
2019. Adding interpretable attention to neural translation models improves word alignment. _arXiv_
_preprint arXiv:1901.11359_ .


Thomas Zenkel, Joern Wuebker, and John DeNero.
2020. [End-to-end neural word alignment outper-](https://doi.org/10.18653/v1/2020.acl-main.146)
[forms GIZA++. In](https://doi.org/10.18653/v1/2020.acl-main.146) _Proceedings of the 58th Annual_
_Meeting of the Association for Computational Lin-_
_guistics_, pages 1605–1617, Online. Association for
Computational Linguistics.