Published as a conference paper at ICLR 2017

S TRUCTURED ATTENTION N ETWORKS

Yoon Kim∗ Carl Denton∗ Luong Hoang Alexander M. Rush
{yoonkim@seas,carldenton@college,lhoang@g,srush@seas}.harvard.edu
School of Engineering and Applied Sciences
Harvard University
Cambridge, MA 02138, USA

A BSTRACT
arXiv:1702.00887v3 [cs.CL] 16 Feb 2017

Attention networks have proven to be an effective approach for embedding cat-

egorical inference within a deep neural network. However, for many tasks we
may want to model richer structural dependencies without abandoning end-to-end
training. In this work, we experiment with incorporating richer structural distri-
butions, encoded using graphical models, within deep networks. We show that
these structured attention networks are simple extensions of the basic attention
procedure, and that they allow for extending attention beyond the standard soft-
selection approach, such as attending to partial segmentations or to subtrees. We
experiment with two different classes of structured attention networks: a linear-
chain conditional random field and a graph-based parsing model, and describe
how these models can be practically implemented as neural network layers. Ex-
periments show that this approach is effective for incorporating structural biases,
and structured attention networks outperform baseline attention models on a va-
riety of synthetic and real tasks: tree transduction, neural machine translation,
question answering, and natural language inference. We further find that mod-
els trained in this way learn interesting unsupervised hidden representations that
generalize simple attention.

1 I NTRODUCTION

Attention networks are now a standard part of the deep learning toolkit, contributing to impressive
results in neural machine translation (Bahdanau et al., 2015; Luong et al., 2015), image captioning
(Xu et al., 2015), speech recognition (Chorowski et al., 2015; Chan et al., 2015), question answering
(Hermann et al., 2015; Sukhbaatar et al., 2015), and algorithm-learning (Graves et al., 2014; Vinyals
et al., 2015), among many other applications (see Cho et al. (2015) for a comprehensive review).
This approach alleviates the bottleneck of compressing a source into a fixed-dimensional vector by
equipping a model with variable-length memory (Weston et al., 2014; Graves et al., 2014; 2016),
thereby providing random access into the source as needed. Attention is implemented as a hidden
layer which computes a categorical distribution (or hierarchy of categorical distributions) to make a
soft-selection over source elements.
Noting the empirical effectiveness of attention networks, we also observe that the standard attention-
based architecture does not directly model any structural dependencies that may exist among the
source elements, and instead relies completely on the hidden layers of the network. While one might
argue that these structural dependencies can be learned implicitly by a deep model with enough data,
in practice, it may be useful to provide a structural bias. Modeling structural dependencies at the
final, output layer has been shown to be important in many deep learning applications, most notably
in seminal work on graph transformers (LeCun et al., 1998), key work on NLP (Collobert et al.,
2011), and in many other areas (Peng et al., 2009; Do & Artiéres, 2010; Jaderberg et al., 2014; Chen
et al., 2015; Durrett & Klein, 2015; Lample et al., 2016, inter alia).
In this work, we consider applications which may require structural dependencies at the attention
layer, and develop internal structured layers for modeling these directly. This approach generalizes
categorical soft-selection attention layers by specifying possible structural dependencies in a soft
∗
Equal contribution.

1
Published as a conference paper at ICLR 2017

manner. Key applications will be the development of an attention function that segments the source
input into subsequences and one that takes into account the latent recursive structure (i.e. parse tree)
of a source sentence.
Our approach views the attention mechanism as a graphical model over a set of latent variables. The
standard attention network can be seen as an expectation of an annotation function with respect to a
single latent variable whose categorical distribution is parameterized to be a function of the source.
In the general case we can specify a graphical model over multiple latent variables whose edges
encode the desired structure. Computing forward attention requires performing inference to obtain
the expectation of the annotation function, i.e. the context vector. This expectation is computed over
an exponentially-sized set of structures (through the machinery of graphical models/structured pre-
diction), hence the name structured attention network. Notably each step of this process (including
inference) is differentiable, so the model can be trained end-to-end without having to resort to deep
policy gradient methods (Schulman et al., 2015).
The differentiability of inference algorithms over graphical models has previously been noted by
various researchers (Li & Eisner, 2009; Domke, 2011; Stoyanov et al., 2011; Stoyanov & Eisner,
2012; Gormley et al., 2015), primarily outside the area of deep learning. For example, Gormley
et al. (2015) treat an entire graphical model as a differentiable circuit and backpropagate risk through
variational inference (loopy belief propagation) for minimium risk training of dependency parsers.
Our contribution is to combine these ideas to produce structured internal attention layers within
deep networks, noting that these approaches allow us to use the resulting marginals to create new
features, as long as we do so a differentiable way.
We focus on two classes of structured attention: linear-chain conditional random fields (CRFs) (Laf-
ferty et al., 2001) and first-order graph-based dependency parsers (Eisner, 1996). The initial work
of Bahdanau et al. (2015) was particularly interesting in the context of machine translation, as the
model was able to implicitly learn an alignment model as a hidden layer, effectively embedding
inference into a neural network. In similar vein, under our framework the model has the capacity
to learn a segmenter as a hidden layer or a parser as a hidden layer, without ever having to see a
segmented sentence or a parse tree. Our experiments apply this approach to a difficult synthetic re-
ordering task, as well as to machine translation, question answering, and natural language inference.
We find that models trained with structured attention outperform standard attention models. Analy-
sis of learned representations further reveal that interesting structures emerge as an internal layer of
the model. All code is available at http://github.com/harvardnlp/struct-attn.

2 BACKGROUND : ATTENTION N ETWORKS

A standard neural network consist of a series of non-linear transformation layers, where each layer
produces a fixed-dimensional hidden representation. For tasks with large input spaces, this paradigm
makes it hard to control the interaction between components. For example in machine translation,
the source consists of an entire sentence, and the output is a prediction for each word in the translated
sentence. Utilizing a standard network leads to an information bottleneck, where one hidden layer
must encode the entire source sentence. Attention provides an alternative approach.1 An attention
network maintains a set of hidden representations that scale with the size of the source. The model
uses an internal inference step to perform a soft-selection over these representations. This method
allows the model to maintain a variable-length memory and has shown to be crucially important for
scaling systems for many tasks.
Formally, let x = [x1 , . . . , xn ] represent a sequence of inputs, let q be a query, and let z be a
categorical latent variable with sample space {1, . . . , n} that encodes the desired selection among
these inputs. Our aim is to produce a context c based on the sequence and the query. To do so, we
assume access to an attention distribution z ∼ p(z | x, q), where we condition p on the inputs x and
a query q. The context over a sequence is defined as expectation, c = Ez∼p(z | x,q) [f (x, z)] where
f (x, z) is an annotation function. Attention of this form can be applied over any type of input,
however, we will primarily be concerned with “deep” networks, where both the annotation function

1
Another line of work involves marginalizing over latent variables (e.g. latent alignments) for sequence-to-
sequence transduction (Kong et al., 2016; Lu et al., 2016; Yu et al., 2016; 2017).

2
Published as a conference paper at ICLR 2017

and attention distribution are parameterized with neural networks, and the context produced is a
vector fed to a downstream network.
For example, consider the case of attention-based neural machine translation (Bahdanau et al., 2015).
Here the sequence of inputs [x1 , . . . , xn ] are the hidden states of a recurrent neural network (RNN),
running over the words in the source sentence, q is the RNN hidden state of the target decoder
(i.e. vector representation of the query q), and z represents the source position to be attended to
for translation. The attention distribution p is simply p(z = i | x, q) = softmax(θi ) where θ ∈ Rn
is a parameterized potential typically based on a neural network, e.g. θi = MLP([xi ; q]). The
annotation function is defined to simply return the selected hidden state, f (x, z) = xz . The context
vector can then be computed using a simple sum,
n
X
c = Ez∼p(z | x,q) [f (x, z)] = p(z = i | x, q)xi (1)
i=1

Other tasks such as question answering use attention in a similar manner, for instance by replacing
source [x1 , . . . , xn ] with a set of potential facts and q with a representation of the question.
In summary we interpret the attention mechanism as taking the expectation of an annotation function
f (x, z) with respect to a latent variable z ∼ p, where p is parameterized to be function of x and q.

3 S TRUCTURED ATTENTION
Attention networks simulate selection from a set using a soft model. In this work we consider gener-
alizing selection to types of attention, such as selecting chunks, segmenting inputs, or even attending
to latent subtrees. One interpretation of this attention is as using soft-selection that considers all pos-
sible structures over the input, of which there may be exponentially many possibilities. Of course,
this expectation can no longer be computed using a simple sum, and we need to incorporate the
machinery of inference directly into our neural network.
Define a structured attention model as being an attention model where z is now a vector of discrete
latent variables [z1 , . . . , zm ] and the attention distribution is p(z | x, q) is defined as a conditional
random field (CRF), specifying the independence structure of the z variables. Formally, we assume
an undirected graph structure with m vertices. The CRF is parameterized with clique (log-)potentials
θC (zC ) ∈ R, where the zC indicates the subset of z given by P clique C. Under this definition, the
attention probability is defined as, p(z | x, q; θ) = softmax( C θC (zC )), whereP for symmetry we
use softmax in a general sense, i.e. softmax(g(z)) = Z1 exp(g(z)) where Z = z0 exp(g(z 0 )) is
the implied partition function. In practice we use a neural CRF, where θ comes from a deep model
over x, q.
assume that the annotation function f factors (at least) into clique
In structured attention, we also P
annotation functions f (x, z) = C fC (x, zC ). Under standard conditions on the conditional inde-
pendence structure, inference techniques from graphical models can be used to compute the forward-
pass expectations and the context:
X
c = Ez∼p(z | x,q) [f (x, z)] = Ez∼p(zC | x,q) [fC (x, zC )]
C

3.1 E XAMPLE 1: S UBSEQUENCE S ELECTION

Suppose instead of soft-selecting a single input, we wanted to explicitly model the selection of con-
tiguous subsequences. We could naively apply categorical attention over all subsequences, or hope
the model learns a multi-modal distribution to combine neighboring words. Structured attention
provides an alternate approach.
Concretely, let m = n, define z to be a random
Pn vector z = [z1 , . . . , zn ] with zi ∈ {0, 1}, and define
our annotation function to be, f (x, z) = i=1 fi (x, zi ) where fi (x, zi ) = 1{zi = 1}xi . The
explicit expectation is then,
n
X
Ez1 ,...,zn [f (x, z)] = p(zi = 1 | x, q)xi (2)
i=1

3
Published as a conference paper at ICLR 2017

q q q

z1 z1 z2 z3 z4 z1 z2 z3 z4

x1 x2 x3 x4 x1 x2 x3 x4 x1 x2 x3 x4

(a) (b) (c)

Figure 1: Three versions of a latent variable attention model: (a) A standard soft-selection attention network,
(b) A Bernoulli (sigmoid) attention network, (c) A linear-chain structured attention model for segmentation.
The input and query are denoted with x and q respectively.

Equation (2) is similar to equation (1)—both are a linear combination of the input representations
where the scalar is between [0, 1] and represents how much attention should be focused on each
input. However, (2) is fundamentally different in two ways: (i) it allows for multiple inputs (or no
inputs) to be selected for a given query; (ii) we can incorporate structural dependencies across the
zi ’s. For instance, we can model the distribution over z with a linear-chain CRF with pairwise edges,
n−1
!
X
p(z1 , . . . , zn | x, q) = softmax θi,i+1 (zi , zi+1 ) (3)
i=1

where θk,l is the pairwise potential for zi = k and zi+1 = l. This model is shown in Figure 1c.
Compare this model to the standard attention in Figure 1a, or to a simple Bernoulli (sigmoid) selec-
tion method, p(zi = 1 | x, q) = sigmoid(θi ), shown in Figure 1b. All three of these methods can
use potentials from the same neural network or RNN that takes x and q as inputs.
In the case of the linear-chain CRF in (3), the marginal distribution p(zi = 1 | x) can be calculated
efficiently in linear-time for all i using message-passing, i.e. the forward-backward algorithm. These
marginals allow us to calculate (2), and in doing so we implicitly sum over an exponentially-sized
set of structures (i.e. all binary sequences of length n) through dynamic programming. We refer to
this type of attention layer as a segmentation attention layer.
Note that the forward-backward algorithm is being used as parameterized pooling (as opposed to
output computation), and can be thought of as generalizing the standard attention softmax. Crucially
this generalization from vector softmax to forward-backward is just a series of differentiable steps,2
and we can compute gradients of its output (marginals) with respect to its input (potentials). This
will allow the structured attention model to be trained end-to-end as part of a deep model.

3.2 E XAMPLE 2: S YNTACTIC T REE S ELECTION

This same approach can be used for more involved structural dependencies. One popular structure
for natural language tasks is a dependency tree, which enforces a structural bias on the recursive
dependencies common in many languages. In particular a dependency tree enforces that each word
in a source sentence is assigned exactly one parent word (head word), and that these assignments do
not cross (projective structure). Employing this bias encourages the system to make a soft-selection
based on learned syntactic dependencies, without requiring linguistic annotations or a pipelined
decision.
A dependency parser can be partially formalized as a graphical model with the following cliques
(Smith & Eisner, 2008): latent variables zij ∈ {0, 1} for all i 6= j, which indicates that the i-th
word is the parent of the j-th word (i.e. xi → xj ); and a special global constraint that rules out
configurations of zij ’s that violate parsing constraints (e.g. one head, projectivity).
The parameters to the graph-based CRF dependency parser are the potentials θij , which reflect
the score of selecting xi as the parent of xj . The probability of a parse tree z given the sentence
2
As are other dynamic programming algorithms for inference in graphical models, such as (loopy and non-
loopy) belief propagation.

4
Published as a conference paper at ICLR 2017

procedure F ORWARD BACKWARD(θ) procedure BACKPROP F ORWARD BACKWARD(θ, p, ∇L p)

α[0, hti] ← 0 ∇L L
α ← log p ⊗ log ∇p ⊗ β ⊗ −A
β[n + 1, hti] ← 0 ∇L L
β ← log p ⊗ log ∇p ⊗ α ⊗ −A
n; c ∈ C do
for i = 1, . . . ,L α̂[0, hti] ← 0
α[i, c] ← y α[i − 1, y] ⊗ θi−1,i (y, c)
β̂[n + 1, hti] ← 0
, 1; c ∈ C do
for i = n, . . . L for i = n, . . . 1; c ∈ C do
β[i, c] ← y β[i + 1, y] ⊗ θi,i+1 (c, y) β̂[i, c] ← ∇L α [i, c] ⊕
L
y θi,i+1 (c, y) ⊗ β̂[i + 1, y]
A ← α[n + 1, hti] for i = 1, . . . , n; c ∈ C do
for i = 1, . . . , n; c ∈ C do α̂[i, c] ← ∇L
L
β [i, c] ⊕ y θi−1,i (y, c) ⊗ α̂[i − 1, y]
p(zi = c | x) ← exp(α[i, c] ⊗ β[i, c]
⊗ − A) for i = 1, . . . , n; y, c ∈ C do
∇Lθi−1,i (y,c) ← signexp(α̂[i, y] ⊗ β[i + 1, c]
return p
⊕ α[i, y] ⊗ β̂[i + 1, c]
⊕ α[i, y] ⊗ β[i + 1, c] ⊗ −A)
return ∇L
θ

Figure 2: Algorithms for linear-chain CRF: (left) computation of forward-backward tables α, β, and marginal
probabilities p from potentials θ (forward-backward algorithm); (right) backpropagation of loss gradients with
respect to the marginals ∇Lp . C denotes the state space and hti is the special start/stop state. Backpropagation
log p
uses the identity ∇L
log p = p ∇L L L
p to calculate ∇θ = ∇log p ∇θ , where is the element-wise multiplication.
Typically the forward-backward with marginals is performed in the log-space semifield R ∪ {±∞} with binary
operations ⊕ = logadd and ⊗ = + for numerical precision. However, backpropagation requires working with
the log of negative values (since ∇L p could be negative), so we extend to a field [R ∪ {±∞}] × {+, −} with
special +/− log-space operations. Binary operations applied to vectors are implied to be element-wise. The
signexp function is defined as signexp(la ) = sa exp(la ). See Section 3.3 and Table 1 for more details.

x = [x1 , . . . , xn ] is,
 

p(z | x, q) = softmax 1{z is valid} 1{zij = 1}θij 

X
(4)
i6=j

where z is represented as a vector of zij ’s for all i 6= j. It is possible to calculate the marginal
probability of each edge p(zij = 1 | x, q) for all i, j in O(n3 ) time using the inside-outside algorithm
(Baker, 1979) on the data structures of Eisner (1996).
Pn
The parsing contraints ensure that each word has exactly one head (i.e. i=1 zij = 1). Therefore if
we want to utilize the soft-head selection of a position j, the context vector is defined as:
n n
1{zij = 1}xi
X X
fj (x, z) = cj = Ez [fj (x, z)] = p(zij = 1 | x, q)xi
i=1 i=1

Note that in this case the annotation function has the subscript j to produce a context vector for
each word in the sentence. Similar types of attention can be applied for other tree properties (e.g.
soft-children). We refer to this type of attention layer as a syntactic attention layer.

3.3 E ND - TO -E ND T RAINING

Graphical models of this form have been widely used as the final layer of deep models. Our contri-
bution is to argue that these networks can be added within deep networks in place of simple attention
layers. The whole model can then be trained end-to-end.
The main complication in utilizing this approach within the network itself is the need to backprop-
agate the gradients through an inference algorithm as part of the structured attention network. Past
work has demonstrated the techniques necessary for this approach (see Stoyanov et al. (2011)), but
to our knowledge it is very rarely employed.
Consider the case of the simple linear-chain CRF layer from equation (3). Figure 2 (left) shows the
standard forward-backward algorithm for computing the marginals p(zi = 1 | x, q; θ). If we treat the
forward-backward algorithm as a neural network layer, its input are the potentials θ, and its output

5
Published as a conference paper at ICLR 2017

after the forward pass are these marginals.3 To backpropagate a loss through this layer we need to
compute the gradient of the loss L with respect to θ, ∇L
θ , as a function of the gradient of the loss with
respect to the marginals, ∇Lp . 4
As the forward-backward algorithm consists of differentiable steps,
this function can be derived using reverse-mode automatic differentiation of the forward-backward
algorithm itself. Note that this reverse-mode algorithm conveniently has a parallel structure to the
forward version, and can also be implemented using dynamic programming.
However, in practice, one cannot simply
use current off-the-shelf tools for this task.
⊕ ⊗
For one, efficiency is quite important for sa sb la+b sa+b la·b sa·b
these models and so the benefits of hand-
optimizing the reverse-mode implementa- + + la + log(1 + d) + la + lb +
tion still outweighs simplicity of automatic + − la + log(1 − d) + la + lb −
− + la + log(1 − d) − la + lb −
differentiation. Secondly, numerical pre-
− − la + log(1 + d) − la + lb +
cision becomes a major issue for struc-
tured attention networks. For computing
Table 1: Signed log-space semifield (from Li & Eis-
the forward-pass and the marginals, it is im- ner (2009)). Each real number a is represented as a pair
portant to use the standard log-space semi- (la , sa ) where la = log |a| and sa = sign(a). Therefore
field over R ∪ {±∞} with binary opera- a = sa exp(la ). For the above we let d = exp(lb − la ) and
tions (⊕ = logadd, ⊗ = +) to avoid un- assume |a| > |b|.
derflow of probabilities. For computing the
backward-pass, we need to remain in log-
space, but also handle log of negative values (since ∇Lp could be negative). This requires extending
to the signed log-space semifield over [R ∪ {±∞}] × {+, −} with special +/− operations. Ta-
ble 1, based on Li & Eisner (2009), demonstrates how to handle this issue, and Figure 2 (right)
describes backpropagation through the forward-backward algorithm. For dependency parsing, the
forward pass can be computed using the inside-outside implementation of Eisner’s algorithm (Eis-
ner, 1996). Similarly, the backpropagation parallels the inside-outside structure. Forward/backward
pass through the inside-outside algorithm is described in Appendix B.

4 E XPERIMENTS
We experiment with three instantiations of structured attention networks on four different tasks: (a)
a simple, synthetic tree manipulation task using the syntactic attention layer, (b) machine translation
with segmentation attention (i.e. two-state linear-chain CRF), (c) question answering using an n-
state linear-chain CRF for multi-step inference over n facts, and (d) natural language inference with
syntactic tree attention. These experiments are not intended to boost the state-of-the-art for these
tasks but to test whether these methods can be trained effectively in an end-to-end fashion, can yield
improvements over standard selection-based attention, and can learn plausible latent structures. All
model architectures, hyperparameters, and training details are further described in Appendix A.

4.1 T REE T RANSDUCTION

The first set of experiments look at a tree-transduction task. These experiments use synthetic data
to explore a failure case of soft-selection attention models. The task is to learn to convert a random
formula given in prefix notation to one in infix notation, e.g.,
( ∗ ( + ( + 15 7 ) 1 8 ) ( + 19 0 11 ) ) ⇒ ( ( 15 + 7 ) + 1 + 8 ) ∗ ( 19 + 0 + 11 )
The alphabet consists of symbols {(, ), +, ∗}, numbers between 0 and 20, and a special root symbol
$. This task is used as a preliminary task to see if the model is able to learn the implicit tree structure
on the source side. The model itself is an encoder-decoder model, where the encoder is defined
below and the decoder is an LSTM. See Appendix A.2 for the full model.
3
Confusingly, “forward” in this case is different than in the forward-backward algorithm, as the marginals
themselves are the output. However the two uses of the term are actually quite related. The forward-backward
algorithm can be interpreted as a forward and backpropagation pass on the log partition function. See Eisner
(2016) for further details (appropriately titled “Inside-Outside and Forward-Backward Algorithms Are Just
Backprop”). As such our full approach can be seen as computing second-order information. This interpretation
is central to Li & Eisner (2009).
4
In general we use ∇ab to denote the Jacobian of a with respect to b.

6
Published as a conference paper at ICLR 2017

Figure 3: Visualization of the source self-attention distribution for the simple (left) and structured (right)
attention models on the tree transduction task. $ is the special root symbol. Each row delineates the distribution
over the parents (i.e. each row sums to one). The attention distribution obtained from the parsing marginals are
more able to capture the tree structure—e.g. the attention weights of closing parentheses are generally placed
on the opening parentheses (though not necessarily on a single parenthesis).

Training uses 15K prefix-infix pairs where the maximum nesting depth is set to be between 2-4 (the
above example has depth 3), with 5K pairs in each depth bucket. The number of expressions in each
parenthesis is limited to be at most 4. Test uses 1K unseen sequences with depth between 2-6 (note
specifically deeper than train), with 200 sequences for each depth. The performance is measured
as the average proportion of correct target tokens produced until the first failure (as in Grefenstette
et al. (2015)).
For experiments we try using different forms of self -attention over embedding-only encoders. Let
xj be an embedding for each source symbol; our three variants of the source representation x̂j are:
(a) no atten, just symbol embeddings by themselves, i.e. x̂j = xj ; (b) simplePn attention, symbol
embeddings and soft-pairing for each symbol, i.e. x̂j = [xj ; cj ] where cj = i=1 softmax(θij )xi
is calculated using soft-selection;
Pn (c) structured attention, symbol embeddings and soft-parent, i.e.
x̂j = [xj ; cj ] where cj = i=1 p(zij = 1 | x)xi is calculated using parsing marginals, obtained
from the syntactic attention layer. None of these models use an explicit query value—the potentials
come from running a bidirectional LSTM over the source, producing hidden vectors hi , and then
computing
θij = tanh(s> tanh(W1 hi + W2 hj + b))
where s, b, W1 , W2 are parameters (see Appendix A.1).
The source representation [x̂1 , . . . , x̂n ] are attended
over using the standard attention mechanism at each
Depth No Atten Simple Structured
decoding step by an LSTM decoder.5 Additionally,
2 7.6 87.4 99.2 symbol embedding parameters are shared between the
3 4.1 49.6 87.0 parsing LSTM and the source encoder.
4 2.8 23.3 64.5
5 2.1 15.0 30.8
6 1.5 8.5 18.2 Results Table 2 has the results for the task. Note
that this task is fairly difficult as the encoder is quite
Table 2: Performance (average length to fail- simple. The baseline model (unsurprisingly) performs
ure %) of models on the tree-transduction task. poorly as it has no information about the source order-
ing. The simple attention model performs better, but
is significantly outperformed by the structured model
with a tree structure bias. We hypothesize that the
model is partially reconstructing the arithmetic tree. Figure 3 shows the attention distribution for the
simple/structured models on the same source sequence, which indicates that the structured model is
able to learn boundaries (i.e. parentheses).
5
Thus there are two attention mechanisms at work under this setup. First, structured attention over the
source only to obtain soft-parents for each symbol (i.e. self-attention). Second, standard softmax alignment
attention over the source representations during decoding.

7
Published as a conference paper at ICLR 2017

4.2 N EURAL M ACHINE T RANSLATION

Our second set of experiments use a full neural machine translation model utilizing attention over
subsequences. Here both the encoder/decoder are LSTMs, and we replace standard simple attention
with a segmentation attention layer. We experiment with two settings: translating directly from
unsegmented Japanese characters to English words (effectively using structured attention to perform
soft word segmentation), and translating from segmented Japanese words to English words (which
can be interpreted as doing phrase-based neural machine translation). Japanese word segmentation
is done using the KyTea toolkit (Neubig et al., 2011).
The data comes from the Workshop on Asian Translation (WAT) (Nakazawa et al., 2016). We
randomly pick 500K sentences from the original training set (of 3M sentences) where the Japanese
sentence was at most 50 characters and the English sentence was at most 50 words. We apply the
same length filter on the provided validation/test sets for evaluation. The vocabulary consists of all
tokens that occurred at least 10 times in the training corpus.
The segmentation attention layer is a two-state CRF where the unary potentials at the j-th decoder
step are parameterized as

hi Whj , k = 1
θi (k) =
0, k=0
Here [h1 , . . . , hn ] are the encoder hidden states and h0j is the j-th decoder hidden state (i.e. the
query vector). The pairwise potentials are parameterized linearly with b, i.e. all together
θi,i+1 (zi , zi+1 ) = θi (zi ) + θi+1 (zi+1 ) + bzi ,zi+1
Therefore the segmentation attention layer requires just 4 additional parameters. Appendix A.3
describes the full model architecture.

Pn experiment with three attention configurations: (a) standard simple attention, i.e. cj =
We
i=1 softmax(θ
Pn i )hi ; (b) sigmoid attention: multiple selection with Bernoulli random variables,
i.e. cj = i=1 sigmoid(θi )hi ; (c) structured attention, encoded with normalized CRF marginals,
n n
X p(zi = 1 | x, q) 1X
cj = hi γ= p(zi = 1 | x, q)
i=1
γ λ i=1

The normalization term γ is not ideal but we found it to be helpful for stable training.6 λ is a
hyperparameter (we use λ = 2) and we further add an l2 penalty of 0.005 on the pairwise potentials
b. These values were found via grid search on the validation set.

Results Results for the translation task on the test

set are given in Table 3. Sigmoid attention outper-
Simple Sigmoid Structured
forms simple (softmax) attention on the character-to-
C HAR 12.6 13.1 14.6 word task, potentially because it is able to learn many-
W ORD 14.1 13.8 14.3 to-one alignments. On the word-to-word task, the op-
posite is true, with simple attention outperforming sig-
Table 3: Translation performance as mea- moid attention. Structured attention outperforms both
sured by BLEU (higher is better) on character- models on both tasks, although improvements on the
to-word and word-to-word Japanese-English word-to-word task are modest and unlikely to be sta-
translation for the three different models.
tistically significant.
For further analysis, Figure 4 shows a visualization of
the different attention mechanisms on the character-to-word setup. The simple model generally
focuses attention heavily on a single character. In contrast, the sigmoid and structured models are
able to spread their attention distribution on contiguous subsequences. The structured attention
learns additional parameters (i.e. b) to smooth out this type of attention.
6 Pn
With standard expectation (i.e. cj = i=1 p(zi = 1 | x, q)hi ) we empirically observed the marginals
to quickly saturate. We tried various strategies to overcome this, such as putting an l2 penalty on the unary
potentials and initializing with a pretrained sigmoid attention model, but simply normalizing the marginals
proved to be the most effective. However, this changes the interpretation of the context vector as the expectation
of an annotation function in this case.

8
Published as a conference paper at ICLR 2017

Figure 4: Visualization of the source attention distribution for the simple (top left), sigmoid (top right), and
structured (bottom left) attention models over the ground truth sentence on the character-to-word translation
task. Manually-annotated alignments are shown in bottom right. Each row delineates the attention weights
over the source sentence at each step of decoding. The sigmoid/structured attention models are able learn an
implicit segmentation model and focus on multiple characters at each time step.

4.3 Q UESTION A NSWERING

Our third experiment is on question answering (QA) with the linear-chain CRF attention layer for
inference over multiple facts. We use the bAbI dataset (Weston et al., 2015), where the input is a set
of sentences/facts paired with a question, and the answer is a single token. For many of the tasks
the model has to attend to multiple supporting facts to arrive at the correct answer (see Figure 5 for
an example), and existing approaches use multiple ‘hops’ to greedily attend to different facts. We
experiment with employing structured attention to perform inference in a non-greedy way. As the
ground truth supporting facts are given in the dataset, we are able to assess the model’s inference
accuracy.
The baseline (simple) attention model is the End-To-End Memory Network (Sukhbaatar et al.,
2015) (MemN2N), which we briefly describe here. See Appendix A.4 for full model details. Let
x1 , . . . , xn be the input embedding vectors for the n sentences/facts and let q be the query embed-
ding. In MemN2N, zk is the random variable for the sentence to select at the k-th inference step
(i.e. k-th hop), and thus zk ∈ {1, . . . , n}. The probability distribution over
Pnzk is given by p(zk =
i | x, q) = softmax((xki )> qk ), and the context vector is given by ck = i=1 p(zk = i | x, q)oki ,
where xki , oki are the input and output embedding for the i-th sentence at the k-th hop, respectively.
The k-th context vector is used to modify the query qk+1 = qk + ck , and this process repeats for
k = 1, . . . , K (for k = 1 we have xki = xi , qk = q, ck = 0). The K-th context and query vectors
are used to obtain the final answer. The attention mechanism for a K-hop MemN2N network can
therefore be interpreted as a greedy selection of a length-K sequence of facts (i.e. z1 , . . . , zK ).
For structured attention, we use an n-state, K-step linear-chain CRF.7 We experiment with two
different settings: (a) a unary CRF model with node potentials
θk (i) = (xki )> qk
7
Note that this differs from the segmentation attention for the neural machine translation experiments de-
scribed above, which was a K-state (with K = 2), n-step linear-chain CRF.

9
Published as a conference paper at ICLR 2017

MemN2N Binary CRF Unary CRF

Task K Ans % Fact % Ans % Fact % Ans % Fact %
TASK 02 - T WO S UPPORTING FACTS 2 87.3 46.8 84.7 81.8 43.5 22.3
TASK 03 - T HREE S UPPORTING FACTS 3 52.6 1.4 40.5 0.1 28.2 0.0
TASK 07 - C OUNTING 3 83.2 − 83.5 − 79.3 −
TASK 08 - L ISTS S ETS 3 94.1 − 93.3 − 87.1 −
TASK 11 - I NDEFINITE K NOWLEDGE 2 97.8 38.2 97.7 80.8 88.6 0.0
TASK 13 - C OMPOUND C OREFERENCE 2 95.6 14.8 97.0 36.4 94.4 9.3
TASK 14 - T IME R EASONING 2 99.9 77.6 99.7 98.2 90.5 30.2
TASK 15 - BASIC D EDUCTION 2 100.0 59.3 100.0 89.5 100.0 51.4
TASK 16 - BASIC I NDUCTION 3 97.1 91.0 97.9 85.6 98.0 41.4
TASK 17 - P OSITIONAL R EASONING 2 61.1 23.9 60.6 49.6 59.7 10.5
TASK 18 - S IZE R EASONING 2 86.4 3.3 92.2 3.9 92.0 1.4
TASK 19 - PATH F INDING 2 21.3 10.2 24.4 11.5 24.3 7.8
AVERAGE − 81.4 39.6 81.0 53.7 73.8 17.4

Table 4: Answer accuracy (Ans %) and supporting fact selection accuracy (Fact %) of the three QA models
on the 1K bAbI dataset. K indicates the number of hops/inference steps used for each task. Task 7 and 8 both
contain variable number of facts and hence they are excluded from the fact accuracy measurement. Supporting
fact selection accuracy is calculated by taking the average of 10 best runs (out of 20) for each task.

and (b) a binary CRF model with pairwise potentials

θk,k+1 (i, j) = (xki )> qk + (xki )> xk+1

j + (xk+1
j )> qk+1

The binary CRF model is designed to test the model’s ability P to perform sequential reasoning. For
both (a) and (b), a single context vector is computed: c = z1 ,...,zK p(z1 , . . . , zK | x, q)f (x, z)
(unlike MemN2N which computes K context vectors). Evaluating c requires summing over all nK
possible sequences of length K, which may not be practical for large values of K. However, if
PK
f (x, z) factors over the components of z (e.g. f (x, z) = k=1 fk (x, zk )) then one can rewrite the
PK Pn
above sum in terms of marginals: c = k=1 i=1 p(zk = i | x, q)fk (x, zk ). In our experiments,
we use fk (x, zk ) = okzk . All three models are described in further detail in Appendix A.4.

Results We use the version of the dataset with 1K questions for each task. Since all models reduce
to the same network for tasks with 1 supporting fact, they are excluded from our experiments. The
number of hops (i.e. K) is task-dependent, and the number of memories (i.e. n) is limited to be
at most 25 (note that many question have less than 25 facts—e.g. the example in Figure 5 has 9
facts). Due to high variance in model performance, we train 20 models with different initializations
for each task and report the test accuracy of the model that performed the best on a 10% held-out
validation set (as is typically done for bAbI tasks).
Results of the three different models are shown in Table 4. For correct answer seletion (Ans %),
we find that MemN2N and the Binary CRF model perform similarly while the Unary CRF model
does worse, indicating the importance of including pairwise potentials. We also assess each model’s
ability to attend to the correct supporting facts in Table 4 (Fact %). Since ground truth supporting
facts are provided for each query, we can check the sequence accuracy of supporting facts for each
model (i.e. the rate of selecting the exact correct sequence of facts) by taking the highest probability
sequence ẑ = argmax p(z1 , . . . , zK | x, q) from the model and checking against the ground truth.
Overall the Binary CRF is able to recover supporting facts better than MemN2N. This improvement
is significant and can be up to two-fold as seen for task 2, 11, 13 & 17. However we observed that
on many tasks it is sufficient to select only the last (or first) fact correctly to predict the answer,
and thus higher sequence selection accuracy does not necessarily imply better answer accuracy (and
vice versa). For example, all three models get 100% answer accuracy on task 15 but have different
supporting fact accuracies.
Finally, in Figure 5 we visualize of the output edge marginals produced by the Binary CRF model
for a single question in task 16. In this instance, the model is uncertain but ultimately able to select
the right sequence of facts 5 → 6 → 8.

10
Published as a conference paper at ICLR 2017

Figure 5: Visualization of the attention distribution over supporting fact sequences for an example question
in task 16 for the Binary CRF model. The actual question is displayed at the bottom along with the correct
answer and the ground truth supporting facts (5 → 6 → 8). The edges represent the marginal probabilities
p(zk , zk+1 | x, q), and the nodes represent the n supporting facts (here we have n = 9). The text for the
supporting facts are shown on the left. The top three most likely sequences are: p(z1 = 5, z2 = 6, z3 =
8 | x, q) = 0.0564, p(z1 = 5, z2 = 6, z3 = 3 | x, q) = 0.0364, p(z1 = 5, z2 = 2, z3 = 3 | x, q) = 0.0356.

4.4 NATURAL L ANGUAGE I NFERENCE

The final experiment looks at the task of natural language inference (NLI) with the syntactic atten-
tion layer. In NLI, the model is given two sentences (hypothesis/premise) and has to predict their
relationship: entailment, contradiction, neutral.
For this task, we use the Stanford NLI dataset (Bowman et al., 2015) and model our approach off
of the decomposable attention model of Parikh et al. (2016). This model takes in the matrix of
word embeddings as the input for each sentence and performs inter-sentence attention to predict the
answer. Appendix A.5 describes the full model.
As in the transduction task, we focus on modifying the input representation to take into account soft
parents via self-attention (i.e. intra-sentence attention). In addition to the three baselines described
for tree transduction (No Attention, Simple, Structured), we also explore two additional settings: (d)
hard pipeline parent selection, i.e. x̂j = [xj ; xhead(j) ], where head(j) is the index of xj ’s parent8 ;
(e) pretrained structured attention: structured attention where the parsing layer is pretrained for one
epoch on a parsed dataset (which was enough for convergence).

Results Results of our models are shown in Table 5. Simple attention improves upon the no
attention model, and this is consistent with improvements observed by Parikh et al. (2016) with
their intra-sentence attention model. The pipelined model with hard parents also slightly improves
upon the baseline. Structured attention outperforms both models, though surprisingly, pretraining
the syntactic attention layer on the parse trees performs worse than training it from scratch—it is
possible that the pretrained attention is too strict for this task.
We also obtain the hard parse for an example sentence by running the Viterbi algorithm on the
syntactic attention layer with the non-pretrained model:

$ The men are fighting outside a deli .

8
The parents are obtained from running the dependency parser of Andor et al. (2016), available at
https://github.com/tensorflow/models/tree/master/syntaxnet

11
Published as a conference paper at ICLR 2017

Model Accuracy %
Handcrafted features (Bowman et al., 2015) 78.2
LSTM encoders (Bowman et al., 2015) 80.6
Tree-Based CNN (Mou et al., 2016) 82.1
Stack-Augmented Parser-Interpreter Neural Net (Bowman et al., 2016) 83.2
LSTM with word-by-word attention (Rocktäschel et al., 2016) 83.5
Matching LSTMs (Wang & Jiang, 2016) 86.1
Decomposable attention over word embeddings (Parikh et al., 2016) 86.3
Decomposable attention + intra-sentence attention (Parikh et al., 2016) 86.8
Attention over constituency tree nodes (Zhao et al., 2016) 87.2
Neural Tree Indexers (Munkhdalai & Yu, 2016) 87.3
Enhanced BiLSTM Inference Model (Chen et al., 2016) 87.7
Enhanced BiLSTM Inference Model + ensemble (Chen et al., 2016) 88.3
No Attention 85.8
No Attention + Hard parent 86.1
Simple Attention 86.2
Structured Attention 86.8
Pretrained Structured Attention 86.5

Table 5: Results of our models (bottom) and others (top) on the Stanford NLI test set. Our baseline model has
the same architecture as Parikh et al. (2016) but the performance is slightly different due to different settings
(e.g. we train for 100 epochs with a batch size of 32 while Parikh et al. (2016) train for 400 epochs with a batch
size of 4 using asynchronous SGD.)

Despite being trained without ever being exposed to an explicit parse tree, the syntactic attention
layer learns an almost plausible dependency structure. In the above example it is able to correctly
identify the main verb fighting, but makes mistakes on determiners (e.g. head of The should be
men). We generally observed this pattern across sentences, possibly because the verb structure is
more important for the inference task.

5 C ONCLUSION

This work outlines structured attention networks, which incorporate graphical models to generalize
simple attention, and describes the technical machinery and computational techniques for backprop-
agating through models of this form. We implement two classes of structured attention layers: a
linear-chain CRF (for neural machine translation and question answering) and a more complicated
first-order dependency parser (for tree transduction and natural language inference). Experiments
show that this method can learn interesting structural properties and improve on top of standard mod-
els. Structured attention could also be a way of learning latent labelers or parsers through attention
on other tasks.
It should be noted that the additional complexity in computing the attention distribution increases
run-time—for example, structured attention was approximately 5× slower to train than simple at-
tention for the neural machine translation experiments, even though both attention layers have the
same asymptotic run-time (i.e. O(n)).
Embedding differentiable inference (and more generally, differentiable algorithms) into deep mod-
els is an exciting area of research. While we have focused on models that admit (tractable) exact
inference, similar technique can be used to embed approximate inference methods. Many optimiza-
tion algorithms (e.g. gradient descent, LBFGS) are also differentiable (Domke, 2012; Maclaurin
et al., 2015), and have been used as output layers for structured prediction in energy-based models
(Belanger & McCallum, 2016; Wang et al., 2016). Incorporating them as internal neural network
layers is an interesting avenue for future work.

ACKNOWLEDGMENTS
We thank Tao Lei, Ankur Parikh, Tim Vieira, Matt Gormley, André Martins, Jason Eisner, Yoav
Goldberg, and the anonymous reviewers for helpful comments, discussion, notes, and code. We
additionally thank Yasumasa Miyamoto for verifying Japanese-English translations.

12
Published as a conference paper at ICLR 2017

R EFERENCES
Daniel Andor, Chris Alberti, David Weiss, Aliaksei Severyn, Alessandro Presta, Kuzman Ganchev,
Slav Petrov, and Michael Collins. Globally Normalized Transition-Based Neural Networks. In
Proceedings of ACL, 2016.
Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural Machine Translation by Jointly
Learning to Align and Translate. In Proceedings of ICLR, 2015.
James K. Baker. Trainable Grammars for Speech Recognition. Speech Communication Papers for
the 97th Meeting of the Acoustical Society, 1979.
David Belanger and Andrew McCallum. Structured Prediction Energy Networks. In Proceedings of
ICML, 2016.
Samuel R. Bowman, Christopher D. Manning, and Christopher Potts. Tree-Structured Composition
in Neural Networks without Tree-Structured Architectures. In Proceedings of the NIPS workshop
on Cognitive Computation: Integrating Neural and Symbolic Approaches, 2015.
Samuel R. Bowman, Jon Gauthier, Abhinav Rastogi, Raghav Gupta, Christopher D. Manning, and
Christopher Potts. A Fast Unified Model for Parsing and Sentence Understanding. In Proceedings
of ACL, 2016.
William Chan, Navdeep Jaitly, Quoc Le, and Oriol Vinyals. Listen, Attend and Spell.
arXiv:1508.01211, 2015.
Liang-Chieh Chen, Alexander G. Schwing, Alan L. Yuille, and Raquel Urtasun. Learning Deep
Structured Models. In Proceedings of ICML, 2015.
Qian Chen, Xiaodan Zhu, Zhenhua Ling, Si Wei, and Hui Jiang. Enhancing and Combining Se-
quential and Tree LSTM for Natural Language Inference. arXiv:1609.06038, 2016.
Kyunghyun Cho, Aaron Courville, and Yoshua Bengio. Describing Multimedia Content using
Attention-based Encoder-Decoder Networks. In IEEE Transactions on Multimedia, 2015.
Jan Chorowski, Dzmitry Bahdanau, Dmitriy Serdyuk, Kyunghyun Cho, and Yoshua Bengio.
Attention-Based Models for Speech Recognition. In Proceedings of NIPS, 2015.
Ronan Collobert, Jason Weston, Leon Bottou, Michael Karlen, Koray Kavukcuoglu, and Pavel
Kuksa. Natural Language Processing (almost) from Scratch. Journal of Machine Learning Re-
search, 12:2493–2537, 2011.
Trinh-Minh-Tri Do and Thierry Artiéres. Neural Conditional Random Fields. In Proceedings of
AISTATS, 2010.
Justin Domke. Parameter Learning with Truncated Message-Passing. In Proceedings of CVPR,
2011.
Justin Domke. Generic methods for optimization-based modeling. In AISTATS, pp. 318–326, 2012.
John Duchi, Elad Hazan, and Yoram Singer. Adaptive Subgradient Methods for Online Learning
and Stochastic Optimization. Journal of Machine Learning Research, 12:2021–2159, 2011.
Greg Durrett and Dan Klein. Neural CRF Parsing. In Proceedings of ACL, 2015.
Jason M. Eisner. Three New Probabilistic Models for Dependency Parsing: An Exploration. In
Proceedings of ACL, 1996.
Jason M. Eisner. Inside-Outside and Forward-Backward Algorithms are just Backprop. In Proceed-
ings of Structured Prediction Workshop at EMNLP, 2016.
Matthew R. Gormley, Mark Dredze, and Jason Eisner. Approximation-Aware Dependency Parsing
by Belief Propagation. In Proceedings of TACL, 2015.
Alex Graves, Greg Wayne, and Ivo Danihelka. Neural Turing Machines. arXiv:1410.5401, 2014.

13
Published as a conference paper at ICLR 2017

Alex Graves, Greg Wayne, Malcolm Reynolds, Tim Harley, Ivo Danihelka, Agnieszka Grabska-
Barwinska, Sergio Gomez Colmenarejo, Edward Grefenstette, Tiago Ramalho, John Agapiou,
Adria Puigdomenech Badia, Karl Moritz Hermann, Yori Zwols, Georg Ostrovski, Adam Cain,
Helen King, Christopher Summerfield, Phil Blunsom, Koray Kavukcuoglu, and Demis Hassabis.
Hybrid Computing Using a Neural Network with Dynamic External Memory. Nature, October
2016.
Edward Grefenstette, Karl Moritz Hermann, Mustafa Suleyman, and Phil Blunsom. Learning to
Transduce with Unbounded Memory. In Proceedings of NIPS, 2015.
Karl Moritz Hermann, Tomas Kocisky, Edward Grefenstette, Lasse Espeholt, Will Kay, Mustafa
Suleyman, and Phil Blunsom. Teaching Machines to Read and Comprehend. In Proceedings of
NIPS, 2015.
Max Jaderberg, Karen Simonyan, Andrea Vedaldi, and Andrew Zisserman. Deep Structured Output
Learning for Unconstrained Text Recognition. In Proceedings of ICLR, 2014.
Diederik Kingma and Jimmy Ba. Adam: A Method for Stochastic Optimization. In Proceedings of
ICLR, 2015.
Eliyahu Kipperwasser and Yoav Goldberg. Simple and Accurate Dependency Parsing using Bidi-
rectional LSTM Feature Representations. In TACL, 2016.
Lingpeng Kong, Chris Dyer, and Noah A. Smith. Segmental Recurrent Neural Networks. In Pro-
ceedings of ICLR, 2016.
John Lafferty, Andrew McCallum, and Fernando Pereira. Conditional Random Fields: Probabilistic
Models for Segmenting and Labeling Sequence Data. In Proceedings of ICML, 2001.
Guillaume Lample, Miguel Ballesteros, Sandeep Subramanian, Kazuya Kawakami, and Chris Dyer.
Neural Architectures for Named Entity Recognition. In Proceedings of NAACL, 2016.
Yann LeCun, Leon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based Learning Applied
to Document Recognition. In Proceedings of IEEE, 1998.
Zhifei Li and Jason Eisner. First- and Second-Order Expectation Semirings with Applications to
Minimum-Risk Training on Translation Forests. In Proceedings of EMNLP 2009, 2009.
Liang Lu, Lingpeng Kong, Chris Dyer, Noah A. Smith, and Steve Renals. Segmental Recurrent
Neural Networks for End-to-End Speech Recognition. In Proceedings of INTERSPEECH, 2016.
Minh-Thang Luong, Hieu Pham, and Christopher D. Manning. Effective Approaches to Attention-
based Neural Machine Translation. In Proceedings of EMNLP, 2015.
Dougal Maclaurin, David Duvenaud, and Ryan P. Adams. Gradient-based Hyperparameter Opti-
mization through Reversible Learning. In Proceedings of ICML, 2015.
Lili Mou, Rui Men, Ge Li, Yan Xu, Lu Zhang, Rui Yan, and Zhi Jin. Natural language inference by
tree-based convolution and heuristic matching. In Proceedings of ACL, 2016.
Tsendsuren Munkhdalai and Hong Yu. Neural Tree Indexers for Text Understanding.
arxiv:1607.04492, 2016.
Toshiaki Nakazawa, Manabu Yaguchi, Kiyotaka Uchimoto, Masao Utiyama, Eiichiro Sumita, Sadao
Kurohashi, and Hitoshi Isahara. Aspec: Asian scientific paper excerpt corpus. In Nicoletta Calzo-
lari (Conference Chair), Khalid Choukri, Thierry Declerck, Marko Grobelnik, Bente Maegaard,
Joseph Mariani, Asuncion Moreno, Jan Odijk, and Stelios Piperidis (eds.), Proceedings of the
Ninth International Conference on Language Resources and Evaluation (LREC 2016), pp. 2204–
2208, Portoro, Slovenia, may 2016. European Language Resources Association (ELRA). ISBN
978-2-9517408-9-1.
Graham Neubig, Yosuke Nakata, and Shinsuke Mori. Pointwise Prediction for Robust, Adaptable
Japanese Morphological Analysis. In Proceedings of ACL, 2011.

14
Published as a conference paper at ICLR 2017

Ankur P. Parikh, Oscar Tackstrom, Dipanjan Das, and Jakob Uszkoreit. A Decomposable Attention
Model for Natural Language Inference. In Proceedings of EMNLP, 2016.
Jian Peng, Liefeng Bo, and Jinbo Xu. Conditional Neural Fields. In Proceedings of NIPS, 2009.
Jeffrey Pennington, Richard Socher, and Christopher D. Manning. GloVe: Global Vectors for Word
Representation. In Proceedings of EMNLP, 2014.
Tim Rocktäschel, Edward Grefenstette, Karl Moritz Hermann, Tomas Kocisky, and Phil Blunsom.
Reasoning about Entailment with Neural Attention. In Proceedings of ICLR, 2016.
John Schulman, Nicolas Heess, Theophane Weber, and Pieter Abbeel. Gradient estimation using
stochastic computation graphs. In Advances in Neural Information Processing Systems, pp. 3528–
3536, 2015.
David A. Smith and Jason Eisner. Dependency Parsing as Belief Propagation. In Proceedings of
EMNLP, 2008.
Veselin Stoyanov and Jason Eisner. Minimum-Risk Training of Approximate CRF-based NLP Sys-
tems. In Proceedings of NAACL, 2012.
Veselin Stoyanov, Alexander Ropson, and Jason Eisner. Empirical Risk Minimization of Graphical
Model Parameters Given Approximate Inference, Decoding, and Model Structure. In Proceedings
of AISTATS, 2011.
Sainbayar Sukhbaatar, Arthur Szlam, Jason Weston, and Rob Fergus. End-To-End Memory Net-
works. In Proceedings of NIPS, 2015.
Oriol Vinyals, Meire Fortunato, and Navdeep Jaitly. Pointer Networks. In Proceedings of NIPS,
2015.
Shenlong Wang, Sanja Fidler, and Raquel Urtasun. Proximal Deep Structured Models. In Proceed-
ings of NIPS, 2016.
Shuohang Wang and Jing Jiang. Learning Natural Language Inference with LSTM. In Proceedings
of NAACL, 2016.
Jason Weston, Sumit Chopra, and Antoine Bordes. Memory Networks. arXiv:1410.3916, 2014.
Jason Weston, Antoine Bordes, Sumit Chopra, Alexander M Rush, Bart van Merriënboer, Armand
Joulin, and Tomas Mikolov. Towards Ai-complete Question Answering: A Set of Prerequisite
Toy Tasks. arXiv preprint arXiv:1502.05698, 2015.
Kelvin Xu, Jimma Ba, Ryan Kiros, Kyunghyun Cho, Aaron Courville, Ruslan Salakhutdinov,
Richard Zemel, and Yoshua Bengio. Show, Attend and Tell: Neural Image Caption Generation
with Visual Attention. In Proceedings of ICML, 2015.
Lei Yu, Jan Buys, and Phil Blunsom. Online Segment to Segment Neural Transduction. In Proceed-
ings of EMNLP, 2016.
Lei Yu, Phil Blunsom, Chris Dyer, Edward Grefenstette, and Tomas Kocisky. The Neural Noisy
Channel. In Proceedings of ICLR, 2017.
Kai Zhao, Liang Huang, and Minbo Ma. Textual Entailment with Structured Attentions and Com-
position. In Proceedings of COLING, 2016.

15
Published as a conference paper at ICLR 2017

APPENDICES

A M ODEL D ETAILS

A.1 S YNTACTIC ATTENTION

The syntactic attention layer (for tree transduction and natural language inference) is similar to the
first-order graph-based dependency parser of Kipperwasser & Goldberg (2016). Given an input sen-
tence [x1 , . . . , xn ] and the corresponding word vectors [x1 , . . . , xn ], we use a bidirectional LSTM
to get the hidden states for each time step i ∈ [1, . . . , n],

hfwd
i = LSTM(xi , hfwd
i−1 ) hbwd
i = LSTM(xi , hbwd
i+1 ) hi = [hfwd bwd
i ; hi ]

where the forward and backward LSTMs have their own parameters. The score for xi → xj (i.e. xi
is the parent of xj ), is given by an MLP

θij = tanh(s> tanh(W1 hi + W2 hj + b))

These scores are used as input to the inside-outside algorithm (see Appendix B) to obtain the prob-
ability of each word’s parent p(zij = 1 | x), which is used to obtain the soft-parent cj for each word
xj . In the non-structured case we simply have p(zij = 1 | x) = softmax(θij ).

A.2 T REE T RANSDUCTION

Let [x1 , . . . , xn ], [y1 , . . . , ym ] be the sequence of source/target symbols, with the associated embed-
dings [x1 , . . . , xn ], [y1 , . . . , ym ] with xi , yj ∈ Rl . In the simplest baseline model we take the source
representation to be the matrix of the symbol embeddings. The decoder is a one-layer LSTM which
produces the hidden states h0j = LSTM(yj , h0j−1 ), with h0j ∈ Rl . The hidden states are combined
with the input representation via a bilinear map W ∈ Rl×l to produce the attention distribution used
to obtain the vector mi , which is combined with the decoder hidden state as follows,
n
exp xi Wh0j X
αi = Pn 0 mi = αi xi ĥj = tanh(U[mi ; h0j ])
k=1 exp xk Whj i=1

Here we have W ∈ Rl×l and U ∈ R2l×l . Finally, ĥj is used to to obtain a distribution over the next
symbol yj+1 ,
p(yj+1 | x1 , . . . , xn , y1 , . . . , yj ) = softmax(Vĥj + b)
For structured/simple models, the j-th source representation are respectively
" n
# " n
#
X X
x̂i = xi ; p(zki = 1 | x) xk x̂i = xi ; softmax(θki ) xk
k=1 k=1

where θij comes from the bidirectional LSTM described in A.1. Then αi and mi changed accord-
ingly,
n
exp x̂i Wh0j X
αi = Pn 0 mi = αi x̂i
k=1 exp x̂k Whj i=1

Note that in this case we have W ∈ R2l×l and U ∈ R3l×l . We use l = 50 in all our experiments.
The forward/backward LSTMs for the parsing LSTM are also 50-dimensional. Symbol embeddings
are shared between the encoder and the parsing LSTMs.
Additional training details include: batch size of 20; training for 13 epochs with a learning rate
of 1.0, which starts decaying by half after epoch 9 (or the epoch at which performance does not
improve on validation, whichever comes first); parameter initialization over a uniform distribution
U [−0.1, 0.1]; gradient normalization at 1 (i.e. renormalize the gradients to have norm 1 if the l2
norm exceeds 1). Decoding is done with beam search (beam size = 5).

16
Published as a conference paper at ICLR 2017

A.3 N EURAL M ACHINE T RANSLATION

The baseline NMT system is from Luong et al. (2015). Let [x1 , . . . , xn ], [y1 , . . . , ym ] be the
source/target sentence, with the associated word embeddings [x1 , . . . , xn ], [y1 , . . . , ym ]. The en-
coder is an LSTM over the source sentence, which produces the hidden states [h1 , . . . , hn ] where
hi = LSTM(xi , hi−1 )
and hi ∈ R . The decoder is another LSTM which produces the hidden states h0j ∈ Rl . In the simple
l

attention case with categorical attention, the hidden states are combined with the input representation
via a bilinear map W ∈ Rl×l and this distribution is used to obtain the context vector at the j-th
time step,
n
X
θi = hi Wh0j cj = softmax(θi )hi
i=1

The Bernoulli attention network has the same θi but instead uses a sigmoid to obtain the weights of
the linear combination, i.e.,
n
X
cj = sigmoid(θi )hi
i=1

And finally, the structured attention model uses a bilinear map to parameterize one of the unary
potentials
hi Wh0j , k = 1

θi (k) =
0, k=0

θi,i+1 (zi , zi+1 ) = θi (zi ) + θi+1 (zi+1 ) + bzi ,zi+1

where b are the pairwise potentials. These potentials are used as inputs to the forward-backward
algorithm to obtain the marginals p(zi = 1 | x, q), which are further normalized to obtain the context
vector
n n
X p(zi = 1 | x, q) 1X
cj = hi γ= p(zi = 1 | x, q)
i=1
γ λ i

We use λ = 2 and also add an l2 penalty of 0.005 on the pairwise potentials b. The context vector
is then combined with the decoder hidden state
ĥj = tanh(U[cj ; h0j ])

and ĥj is used to obtain the distribution over the next target word yj+1

p(yj+1 | x1 , . . . , xn , y1 , . . . yj ) = softmax(Vĥj + b)

The encoder/decoder LSTMs have 2 layers and 500 hidden units (i.e. l = 500).
Additional training details include: batch size of 128; training for 30 epochs with a learning rate of
1.0, which starts decaying by half after the first epoch at which performance does not improve
on validation; dropout with probability 0.3; parameter initialization over a uniform distribution
U [−0.1, 0.1]; gradient normalization at 1. We generate target translations with beam search (beam
size = 5), and evaluate with multi-bleu.perl from Moses.9

A.4 Q UESTION A NSWERING

Our baseline model (MemN2N) is implemented following the same architecture as described in
Sukhbaatar et al. (2015). In particular, let x = [x1 , . . . , xn ] represent the sequence of n facts with
the associated embeddings [x1 , . . . , xn ] and let q be the embedding of the query q. The embeddings
9
https://github.com/moses-smt/mosesdecoder/blob/master/scripts/generic/
multi-bleu.perl

17
Published as a conference paper at ICLR 2017

are obtained by simply adding the word embeddings in each sentence or query. The full model with
K hops is as follows:
p(zk = i | x, q) = softmax((xki )> qk )
Xn
k
c = p(zk = i | x, q)oki
i=1
k+1
q = q k + ck
p(y | x, q) = softmax(W(qK + cK ))
where p(y | x, q) is the distribution over the answer vocabulary. At each layer, {xki } and {oki } are
computed using embedding matrices Xk and Ok . We use the adjacent weight tying scheme from
the paper so that Xk+1 = Ok , WT = OK . X1 is also used to compute the query embedding at the
first hop. For k = 1 we have xki = xi , qk = q, ck = 0.
For both the Unary and the Binary CRF models, the same input fact and query representations are
computed (i.e. same embedding matrices with weight tying scheme). For the unary model, the
potentials are parameterized as
θk (i) = (xki )> qk

and for the binary model we compute pairwise potentials as

θk,k+1 (i, j) = (xki )> qk + (xki )> xk+1
j + (xk+1
j )> qk+1
The qk ’s are updated simply with a linear mapping, i.e.
qk+1 = Qqk
In the case of the Binary CRF, to discourage the model from selecting the same fact again we
additionally set θk,k+1 (i, i) = −∞ for all i ∈ {1, . . . , n}. Given these potentials, we compute the
marginals p(zk = i, zk+1 = j | x, q) using the forward-backward algorithm, which is then used to
compute the context vector:
X K
X
c= p(z1 , . . . , zK | x, q)f (x, z) f (x, z) = fk (x, zk ) fk (x, zk ) = okzk
z1 ,...,zK k=1

Note that if f (x, z) factors over the components of z (as is the case above) then computing c only
requires evaluating the marginals p(zk | x, q).
Finally, given the context vector the prediction is made in a similar fashion to MemN2N:
p(y | x, q) = softmax(W(qK + c))

Other training setup is similar to Sukhbaatar et al. (2015): we use stochastic gradient descent with
learning rate 0.01, which is divided by 2 every 25 epochs until 100 epochs are reached. Capacity
of the memory is limited to 25 sentences. The embedding vectors are of size 20 and gradients are
renormalized if the norm exceeds 40. All models implement position encoding, temporal encoding,
and linear start from the original paper. For linear start, the softmax(·) function in the attention
layer is removed at the beginning and re-inserted after 20 epochs for MemN2N, while for the CRF
models we apply a log(softmax(·)) layer on the qk after 20 epochs. Each model is trained separately
for each task.

A.5 NATURAL L ANGUAGE I NFERENCE

Our baseline model/setup is essentially the same as that of Parikh et al. (2016). Let
[x1 , . . . , xn ], [y1 , . . . , ym ] be the premise/hypothesis, with the corresponding input representations
[x1 , . . . , xn ], [y1 , . . . , ym ]. The input representations are obtained by a linear transformation of
the 300-dimensional pretrained GloVe embeddings (Pennington et al., 2014) after normalizing the
GloVe embeddings to have unit norm.10 The pretrained embeddings remain fixed but the linear layer
10
We use the GloVe embeddings pretrained over the 840 billion word Common Crawl, publicly available at
http://nlp.stanford.edu/projects/glove/

18
Published as a conference paper at ICLR 2017

(which is also 300-dimensional) is trained. Words not in the pretrained vocabulary are hashed to one
of 100 Gaussian embeddings with mean 0 and standard deviation 1.
We concatenate each input representation with a convex combination of the other sentence’s input
representations (essentially performing inter-sentence attention), where the weights are determined
through a dot product followed by a softmax,
  " #
m n
X exp e ij
X exp e ij
eij = f (xi )> f (yj ) x̄i = xi ; Pm yj  ȳj = yj ; Pn xi
j=1 k=1 exp eik i=1 k=1 exp ekj

Here f (·) is an MLP. The new representations are fed through another MLP g(·), summed, combined
with the final MLP h(·) and fed through a softmax layer to obtain a distribution over the labels l,
n
X m
X
x̄ = g(x̄i ) ȳ = g(ȳj )
i=1 j=1

p(l | x1 , . . . , xn , y1 , . . . , ym ) = softmax(Vh([x̄; ȳ]) + b)

All the MLPs have 2-layers, 300 ReLU units, and dropout probability of 0.2. For structured/simple
models, we first employ the bidirectional parsing LSTM (see A.1) to obtain the scores θij . In the
structured case each word representation is simply concatenated with its soft-parent
" n
#
X
x̂i = xi ; p(zki = 1 | x)xk
k=1

and x̂i (and analogously ŷj ) is used as the input to the above model. In the simple case (which
closely corresponds to the intra-sentence attention model of Parikh et al. (2016)), we have
" n
#
X exp θki
x̂i = xi ; Pn xk
l=1 exp θli
k=1

The word embeddings for the parsing LSTMs are also initialized with GloVe, and the parsing layer
is shared between the two sentences. The forward/backward LSTMs for the parsing layer are 100-
dimensional.
Additional training details include: batch size of 32; training for 100 epochs with Adagrad (Duchi
et al., 2011) where the global learning rate is 0.05 and sum of gradient squared is initialized to
0.1; parameter intialization over a Gaussian distribution with mean 0 and standard deviation 0.01;
gradient normalization at 5. In the pretrained scenario, pretraining is done with Adam (Kingma &
Ba, 2015) with learning rate equal to 0.01, and β1 = 0.9, β2 = 0.999.

B F ORWARD /BACKWARD THROUGH THE I NSIDE -O UTSIDE A LGORITHM

Figure 6 shows the procedure for obtaining the parsing marginals from the input potentials. This
corresponds to running the inside-outside version of Eisner’s algorithm (Eisner, 1996). The inter-
mediate data structures used during the dynamic programming algorithm are the (log) inside tables
α, and the (log) outside tables β. Both α, β are of size n × n × 2 × 2, where n is the sentence length.
First two dimensions encode the start/end index of the span (i.e. subtree). The third dimension
encodes whether the root of the subtree is the left (L) or right (R) index of the span. The fourth
dimension indicates if the span is complete (1) or incomplete (0). We can calculate the marginal
distribution of each word’s parent (for all words) in O(n3 ) using this algorithm.
Backward pass through the inside-outside algorithm is slightly more involved, but still takes O(n3 )
time. Figure 7 illustrates the backward procedure, which receives the gradient of the loss L with
respect to the marginals, ∇L p , and computes the gradient of the loss with respect to the potentials
∇Lθ . The computations must be performed in the signed log-space semifield to handle log of negative
values. See section 3.3 and Table 1 for more details.

19
Published as a conference paper at ICLR 2017

procedure I NSIDE O UTSIDE(θ)

α, β ← −∞ . Initialize log of inside (α), outside (β) tables
for i = 1, . . . , n do
α[i, i, L, 1] ← 0
α[i, i, R, 1] ← 0
β[1, n, R, 1] ← 0
for k = 1, . . . , n do . Inside step
for s = 1, . . . , n − k do
t←s+k L
α[s, t, R, 0] ← u∈[s,t−1] α[s, u, R, 1] ⊗ α[u + 1, t, L, 1] ⊗ θst
L
α[s, t, L, 0] ← u∈[s,t−1] α[s, u, R, 1] ⊗ α[u + 1, t, L, 1] ⊗ θts
L
α[s, t, R, 1] ← u∈[s+1,t] α[s, u, R, 0] ⊗ α[u, t, R, 1]
L
α[s, t, L, 1] ← u∈[s,t−1] α[s, u, L, 1] ⊗ α[u, t, L, 0]
for k = n, . . . , 1 do . Outside step
for s = 1, . . . , n − k do
t←s+k
for u = s + 1, . . . , t do
β[s, u, R, 0] ←⊕ β[s, t, R, 1] ⊗ α[u, t, R, 1]
β[u, t, R, 1] ←⊕ β[s, t, R, 1] ⊗ α[s, u, R, 0]
if s > 1 then
for u = s, . . . , t − 1 do
β[s, u, L, 1] ←⊕ β[s, t, L, 1] ⊗ α[u, t, L, 0]
β[u, t, L, 0] ←⊕ β[s, t, L, 1] ⊗ α[s, u, L, 1]
for u = s, . . . , t − 1 do
β[s, u, R, 1] ←⊕ β[s, t, R, 0] ⊗ α[u + 1, t, L, 1] ⊗ θst
β[u + 1, t, L, 1] ←⊕ β[s, t, R, 0] ⊗ α[s, u, R, 1] ⊗ θst
if s > 1 then
for u = s, . . . , t − 1 do
β[s, u, R, 1] ←⊕ β[s, t, L, 0] ⊗ α[u + 1, t, L, 1] ⊗ θts
β[u + 1, t, L, 1] ←⊕ β[s, t, L, 0] ⊗ α[s, u, R, 1] ⊗ θts
A ← α[1, n, R, 1] . Log partition
for s = 1, . . . , n − 1 do . Compute marginals. Note that p[s, t] = p(zst = 1 | x)
for t = s + 1, . . . , n do
p[s, t] ← exp(α[s, t, R, 0] ⊗ β[s, t, R, 0] ⊗ −A)
if s > 1 then
p[t, s] ← exp(α[s, t, L, 0] ⊗ β[s, t, L, 0] ⊗ −A)
return p

Figure 6: Forward step of the syntatic attention layer to compute the marginals, using the inside-outside
algorithm (Baker, 1979) on the data structures of Eisner (1996). We assume the special root symbol is the first
element of the sequence, and that the sentence length is n. Calculations are performed in log-space semifield
with ⊕ = logadd and ⊗ = + for numerical precision. a, b ← c means a ← c and b ← c. a ←⊕ b means
a ← a ⊕ b.

20
Published as a conference paper at ICLR 2017

procedure BACKPROP I NSIDE O UTSIDE(θ, p, ∇L p)

log p
for s, t = 1, . . . , n; s 6= t do . Backpropagation uses the identity ∇L θ = (p ∇Lp )∇θ
L
δ[s, t] ← log p[s, t] ⊗ log ∇p [s, t] . δ = log(p ∇L p)
L L L
∇α , ∇β , log ∇θ ← −∞ L L
. Initialize inside (∇α ), outside (∇β ) gradients, and log of ∇Lθ
for s = 1, . . . , n − 1 do . Backpropagate δ to ∇L α and ∇β
L

for t = s + 1, . . . , n do
∇L L
α [s, t, R, 0], ∇β [s, t, R, 0] ← δ[s, t]
L
∇α [1, n, R, 1] ←⊕ −δ[s, t]
if s > 1 then
∇L L
α [s, t, L, 0], ∇β [s, t, L, 0] ← δ[t, s]
L
∇α [1, n, R, 1] ←⊕ −δ[s, t]
for k = 1, . . . , n do . Backpropagate through outside step
for s = 1, . . . , n − k do
t←s+k
ν ← ∇L β [s, t, R, 0] ⊗ β[s, t, R, 0] . ν, γ are temporary values
for u = t, . . . , n do
∇L L
β [s, u, R, 1], ∇α [t, u, R, 1] ←⊕ ν ⊗ β[s, u, R, 1] ⊗ α[t, u, R, 1]
if s > 1 then
ν ← ∇L β [s, t, L, 0] ⊗ β[s, t, L, 0]
for u = 1, . . . , s do
∇L L
β [u, t, L, 1], ∇α [u, s, L, 1] ←⊕ ν ⊗ β[u, t, L, 1] ⊗ α[u, s, L, 1]
L
ν ← ∇β [s, t, L, 1] ⊗ β[s, t, L, 1]
for u = t, . . . , n do
∇L L
β [s, u, L, 1], ∇α [t, u, L, 0] ←⊕ ν ⊗ β[s, u, L, 1] ⊗ α[t, u, L, 1]
for u = 1, . . . , s − 1 do
γ ← β[u, t, R, 0] ⊗ α[u, s − 1, R, 1] ⊗ θut
∇L L L
β [u, t, R, 0], ∇α [u, s − 1, R, 1], log ∇θ [u, t] ←⊕ ν ⊗ γ
γ ← β[u, t, L, 0] ⊗ α[u, s − 1, R, 1] ⊗ θtu
∇L L L
β [u, t, L, 0], ∇α [u, s − 1, R, 1], log ∇θ [t, u] ←⊕ ν ⊗ γ
L
ν ← ∇β [s, t, R, 1] ⊗ β[s, t, R, 1]
for u = 1, . . . , s do
∇L L
β [u, t, R, 1], ∇α [u, s, R, 0] ←⊕ ν ⊗ β[u, t, R, 1] ⊗ α[u, s, R, 0]
for u = t + 1, . . . , n do
γ ← β[s, u, R, 0] ⊗ α[t + 1, u, L, 1] ⊗ θsu
∇L L L
β [s, u, R, 0], ∇α [t + 1, u, L, 1], log ∇θ [s, u] ←⊕ ν ⊗ γ
γ ← β[s, u, L, 0] ⊗ α[t + 1, u, L, 1] ⊗ θus
∇L L L
β [s, u, L, 0], ∇α [t + 1, u, L, 1], log ∇θ [u, s] ←⊕ ν ⊗ γ
for k = n, . . . , 1 do . Backpropagate through inside step
for s = 1, . . . , n − k do
t←s+k
ν ← ∇L α [s, t, R, 1] ⊗ α[s, t, R, 1]
for u = s + 1, . . . , t do
∇L L
α [u, t, R, 0], ∇α [u, t, R, 1] ←⊕ ν ⊗ α[s, u, R, 0] ⊗ α[u, t, R, 1]
if s > 1 then
ν ← ∇L α [s, t, L, 1] ⊗ α[s, t, L, 1]
for u = s, . . . , t − 1 do
∇L L
α [s, u, L, 1], ∇α [u, t, L, 0] ←⊕ ν ⊗ α[s, u, L, 1] ⊗ α[u, t, L, 0]
ν ← ∇L α [s, t, L, 0] ⊗ α[s, t, L, 0]
for u = s, . . . , t − 1 do
γ ← α[s, u, R, 1] ⊗ α[u + 1, t, L, 1] ⊗ θts
∇L L L
α [s, u, R, 1], ∇α [u + 1, t, L, 1], log ∇θ [t, s] ←⊕ ν ⊗ γ
L
ν ← ∇α [s, t, R, 0] ⊗ α[s, t, R, 0]
for u = s, . . . , t − 1 do
γ ← α[s, u, R, 1] ⊗ α[u + 1, t, L, 1] ⊗ θst
∇L L L
α [s, u, R, 1], ∇α [u + 1, t, L, 1], log ∇θ [s, t] ←⊕ ν ⊗ γ
L
return signexp log ∇θ . Exponentiate log gradient, multiply by sign, and return ∇L θ

Figure 7: Backpropagation through the inside-outside algorithm to calculate the gradient with respect to the
input potentials. ∇ab denotes the Jacobian of a with respect to b (so ∇L
θ is the gradient with respect to θ).
a, b ←⊕ c means a ← a ⊕ c and b ← b ⊕ c.

21