# sparsemap_differentiable_sparse_structured_inference__81767416.pdf

Sparse MAP: Differentiable Sparse Structured Inference

Vlad Niculae 1 André F. T. Martins 2 Mathieu Blondel 3 Claire Cardie 1

Structured prediction requires searching over a combinatorial number of structures. To tackle it, we introduce Sparse MAP: a new method for sparse structured inference, and its natural loss function. Sparse MAP automatically selects only a few global structures: it is situated between MAP inference, which picks a single structure, and marginal inference, which assigns nonzero probability to all structures, including implausible ones. Sparse MAP can be computed using only calls to a MAP oracle, making it applicable to problems with intractable marginal inference, e.g., linear assignment. Sparsity makes gradient backpropagation efficient regardless of the structure, enabling us to augment deep neural networks with generic and sparse structured hidden layers. Experiments in dependency parsing and natural language inference reveal competitive accuracy, improved interpretability, and the ability to capture natural language ambiguities, which is attractive for pipeline systems.

1. Introduction

Structured prediction involves the manipulation of discrete, combinatorial structures, e.g., trees and alignments (Bakır et al., 2007; Smith, 2011; Nowozin et al., 2014). Such structures arise naturally as machine learning outputs, and as intermediate representations in deep pipelines. However, the set of possible structures is typically prohibitively large. As such, inference is a core challenge, often sidestepped by greedy search, factorization assumptions, or continuous relaxations (Belanger & Mc Callum, 2016).

1Cornell University, Ithaca, NY 2Unbabel & Instituto de Telecomunicações, Lisbon, Portugal 3NTT Communication Science Laboratories, Kyoto, Japan. Correspondence to: Vlad Niculae <vlad@vene.ro>, André F. T. Martins <andre.martins@unbabel.com>, Mathieu Blondel <mathieu@mblondel.org>, Claire Cardie <cardie@cs.cornell.edu>.

Proceedings of the 35 th International Conference on Machine Learning, Stockholm, Sweden, PMLR 80, 2018. Copyright 2018 by the author(s).

argmax (1, 0, 0)

softmax (.5, .3, .2)

sparsemax (.6, .4, 0)

Figure 1. Left: in the unstructured case, softmax and sparsemax can be interpreted as regularized, differentiable arg max approximations; softmax returns dense solutions while sparsemax favors sparse ones. Right: in this work, we extend this view to structured inference, which consists of optimizing over a polytope M, the convex hull of all possible structures (depicted: the arborescence polytope, whose vertices are trees). We introduce Sparse MAP as a structured extension of sparsemax: it is situated in between MAP inference, which yields a single structure, and marginal inference, which returns a dense combination of structures.

In this paper, we propose an appealing alternative: a new inference strategy, dubbed Sparse MAP, which encourages sparsity in the structured representations. Namely, we seek solutions explicitly expressed as a combination of a small, enumerable set of global structures. Our framework departs from the two most common inference strategies in structured prediction: maximum a posteriori (MAP) inference, which returns the highest-scoring structure, and marginal inference, which yields a dense probability distribution over structures. Neither of these strategies is fully satisfactory: for latent structure models, marginal inference is appealing, since it can represent uncertainty and, unlike MAP inference, it is continuous and differentiable, hence amenable for use in structured hidden layers in neural networks (Kim et al., 2017). It has, however, several limitations. First, there are useful problems for which MAP is tractable, but marginal inference is not, e.g., linear assignment (Valiant, 1979; Taskar, 2004). Even when marginal inference is available, case-bycase derivation of the backward pass is needed, sometimes producing fairly complicated algorithms, e.g., second-order expectation semirings (Li & Eisner, 2009). Finally, marginal inference is dense: it assigns nonzero probabilities to all structures and cannot completely rule out irrelevant ones. This can be statistically and computationally wasteful, as well as qualitatively harder to interpret.

Sparse MAP: Differentiable Sparse Structured Inference

In this work, we make the following contributions:

1. We propose Sparse MAP: a new framework for sparse structured inference ( 3.1). The main idea is illustrated in Figure 1. Sparse MAP is a twofold generalization: first, as a structured extension of the sparsemax transformation (Martins & Astudillo, 2016); second, as a continuous yet sparse relaxation of MAP inference. MAP yields a single structure and marginal inference yields a dense distribution over all structures. In contrast, the Sparse MAP solutions are sparse combinations of a small number of often-overlapping structures.

2. We show how to compute Sparse MAP effectively, requiring only a MAP solver as a subroutine ( 3.2), by exploiting the problem s sparsity and quadratic curvature. Noticeably, the MAP oracle can be any arbitrary solver, e.g., the Hungarian algorithm for linear assignment, which permits tackling problems for which marginal inference is intractable.

3. We derive expressions for gradient backpropagation through Sparse MAP inference, which, unlike MAP, is differentiable almost everywhere ( 3.3). The backward pass is fully general (applicable to any type of structure), and it is efficient, thanks to the sparsity of the solutions and to reusing quantities computed in the forward pass.

4. We introduce a novel Sparse MAP loss for structured prediction, placing it into a family of loss functions which generalizes the CRF and structured SVM losses ( 4). Inheriting the desirable properties of Sparse MAP inference, the Sparse MAP loss and its gradients can be computed efficiently, provided access to MAP inference.

Our experiments demonstrate that Sparse MAP is useful both for predicting structured outputs, as well as for learning latent structured representations. On dependency parsing ( 5.1), structured output networks trained with the Sparse MAP loss yield more accurate models with sparse, interpretable predictions, adapting to the ambiguity (or lack thereof) of test examples. On natural language inference ( 5.2), we learn latent structured alignments, obtaining good predictive performance, as well as useful natural visualizations concentrated on a small number of structures.1

Notation. Given vectors a Rm, b Rn, [a; b] Rm+n denotes their concatenation; given matrices A Rm k, B Rn k, we denote their row-wise stacking as [A; B] R(m+n) k. We denote the columns of a matrix A by aj; by extension, a slice of columns of A is denoted AI for a set of indices I. We denote the canonical simplex by d := {y Rd : y 0, Pd i=1 yi = 1}, and the indicator function of a predicate p as I[p] = {1 if p, 0 otherwise }.

1 General-purpose dynet and pytorch implementations available at https://github.com/vene/sparsemap.

2. Preliminaries

2.1. Regularized Max Operators: Softmax, Sparsemax

As a basis for the more complex structured case, we first consider the simple problem of selecting the largest value in a vector θ Rd. We denote the vector mapping

arg max(θ) := arg max y d θ y.

When there are no ties, arg max has a unique solution ei peaking at the index i of the highest value of θ. When there are ties, arg max is set-valued. Even assuming no ties, arg max is piecewise constant, and thus is ill-suited for direct use within neural networks, e.g., in an attention mechanism. Instead, it is common to use softmax, a continuous and differentiable approximation to arg max, which can be seen as an entropy-regularized arg max

softmax(θ) := arg max y d θ y+H(y) = exp θ Pd i=1 exp θi (1)

where H(y) = P

i yi ln yi, i.e. the negative Shannon entropy. Since exp > 0 strictly, softmax outputs are dense.

By replacing the entropic penalty with a squared ℓ2 norm, Martins & Astudillo (2016) introduced a sparse alternative to softmax, called sparsemax, given by

sparsemax(θ) := arg max y d θ y 1

= arg min y d y θ 2 2 . (2)

Both softmax and sparsemax are continuous and differentiable almost everywhere; however, sparsemax encourages sparsity in its outputs. This is because it corresponds to an Euclidean projection onto the simplex, which is likely to hit its boundary as the magnitude of θ increases. Both mechanisms, as well as variants with different penalties (Niculae & Blondel, 2017), have been successfully used in attention mechanisms, for mapping a score vector θ to a d-dimensional normalized discrete probability distribution over a small set of choices. The relationship between arg max, softmax, and sparsemax, illustrated in Figure 1, sits at the foundation of Sparse MAP.

2.2. Structured Inference

In structured prediction, the space of possible outputs is typically very large: for instance, all possible labelings of a length-n sequence, spanning trees over n nodes, or oneto-one alignments between two sets. We may still write optimization problems such as max D s=1 θs, but it is impractical to enumerate all of the D possible structures and, in turn, to specify the scores for each structure in θ.

Sparse MAP: Differentiable Sparse Structured Inference

Instead, structured problems are often parametrized through structured log-potentials (scores) θ := A η, where A Rk D is a matrix that specifies the structure of the problem, and η Rk is lower-dimensional parameter vector, i.e., k D. For example, in a factor graph (Kschischang et al., 2001) with variables U and factors F, θ is given by

i U ηU,i(si) + X

f F ηF,f(sf),

where ηU and ηF are unary and higher-order log-potentials, and si and sf are local configurations at variable and factor nodes. This can be written in matrix notation as θ = M ηU +N ηF for suitable matrices {M, N}, fitting the assumption above with A = [M; N] and η = [ηU; ηF ].

We can then rewrite the MAP inference problem, which seeks the highest-scoring structure, as a k-dimensional problem, by introducing variables [u; v] Rk to denote configurations at variable and factor nodes:2

MAPA(η) := arg max u:=My y D θ y

= arg max u: [u;v] MA η U u + η F v, (3)

where MA := {[u; v] : u = My, v = Ny, y D} is the marginal polytope (Wainwright & Jordan, 2008), with one vertex for each possible structure (Figure 1). However, as previously said, since it is equivalent to a D-dimensional arg max, MAP is piecewise constant and discontinuous.

Negative entropy regularization over y, on the other hand, yields marginal inference,

Marginal A(η) := arg max u:=My y D θ y + H(y)

= arg max u: [u;v] MA η U u + η F v + HA(u, v).

(4) Marginal inference is differentiable, but may be more difficult to compute; the entropy HA(u, v) = H(y) itself lacks a closed form (Wainwright & Jordan, 2008, 4.1.2). Gradient backpropagation is available only to specialized problem instances, e.g. those solvable by dynamic programming (Li & Eisner, 2009). The entropic term regularizes y toward more uniform distributions, resulting in strictly dense solutions, just like in the case of softmax (Equation 1).

Interesting types of structures, which we use in the experiments described in Section 5, include the following.

2We use the notation arg maxu: [u;v] M to convey that the maximization is over both u and v, but only u is returned. Separating the variables as [u; v] loses no generality and allows us to isolate the unary posteriors u as the return value of interest.

Sequence tagging. Consider a sequence of n items, each assigned one out of a possible m tags. In this case, a global structure s is a joint assignment of tags (t1, , tn). The matrix M is nm-by-mn dimensional, with columns ms {0, 1}nm := [et1, ..., etn] indicating which tag is assigned to each variable in the global structure s. N is nm2-bymn dimensional, with ns encoding the transitions between consecutive tags, i.e., ns(i, a, b) := I[ti 1 = a & ti = b]. The Viterbi algorithm provides MAP inference and forwardbackward provides marginal inference (Rabiner, 1989).

Non-projective dependency parsing. Consider a sentence of length n. Here, a structure s is a dependency tree: a rooted spanning tree over the n2 possible arcs (for example, the arcs above the sentences in Figure 3). Each column ms {0, 1}n2 encodes a tree by assigning a 1 to its arcs. N is empty, MA is known as the arborescence polytope (Martins et al., 2009). MAP inference may be performed by maximal arborescence algorithms (Chu & Liu, 1965; Edmonds, 1967; Mc Donald et al., 2005), and the Matrix Tree theorem (Kirchhoff, 1847) provides a way to perform marginal inference (Koo et al., 2007; Smith & Smith, 2007).

Linear assignment. Consider a one-to-one matching (linear assignment) between two sets of n nodes. A global structure s is a n-permutation, and a column ms {0, 1}n2

can be seen as a flattening of the corresponding permutation matrix. Again, N is empty. MA is the Birkhoff polytope (Birkhoff, 1946), and MAP inference can be performed by, e.g., the Hungarian algorithm (Kuhn, 1955) or the Jonker-Volgenant algorithm (Jonker & Volgenant, 1987). Noticeably, marginal inference is known to be #P-complete (Valiant, 1979; Taskar, 2004, Section 3.5). This makes it an open problem how to use matchings as latent variables.

3. Sparse MAP

Armed with the parallel between structured inference and regularized max operators described in 2, we are now ready to introduce Sparse MAP, a novel inference optimization problem which returns sparse solutions.

3.1. Definition

We introduce Sparse MAP by regularizing the MAP inference problem in Equation 3 with a squared ℓ2 penalty on the returned posteriors, i.e., 1

2 u 2 2. Denoting, as above, θ := A η, the result is a quadratic optimization problem,

Sparse MAPA(η) := arg max u:=My y D θ y 1

= arg max u: [u,v] MA η U u + η F v 1

Sparse MAP: Differentiable Sparse Structured Inference

1 100 200 300 400 500 number of iterations

1 100 200 300 400 500 y 0

vanilla CG away-step CG

pairwise CG active set

Figure 2. Comparison of solvers on the Sparse MAP optimization problem for a tree factor with 20 nodes. The active set solver converges much faster and to a much sparser solution.

The quadratic penalty replaces the entropic penalty from marginal inference (Equation 4), which pushes the solutions to the strict interior of the marginal polytope. In consequence, Sparse MAP favors sparse solutions from the faces of the marginal polytope MA, as illustrated in Figure 1. For the structured prediction problems mentioned in Section 2.2, Sparse MAP would be able to return, for example, a sparse combination of sequence labelings, parse trees, or matchings. Moreover, the strongly convex regularization on u ensures that Sparse MAP has a unique solution and is differentiable almost everywhere, as we will see.

3.2. Solving Sparse MAP

We now tackle the optimization problem in Equation 5. Although Sparse MAP is a QP over a polytope, even describing it in standard form is infeasible, since enumerating the exponentially-large set of vertices is infeasible. This prevents direct application of, e.g., the generic differentiable QP solver of Amos & Kolter (2017). We instead focus on Sparse MAP solvers that involve a sequence of MAP problems as a subroutine this makes Sparse MAP widely applicable, given the availability of MAP implementations for various structures. We discuss two such methods, one based on the conditional gradient algorithm and another based on the active set method for quadratic programming. We provide a full description of both methods in Appendix A.

Conditional gradient. One family of such solvers is based on the conditional gradient (CG) algorithm (Frank & Wolfe, 1956; Lacoste-Julien & Jaggi, 2015), considered in prior work for solving approximations of the marginal inference problem (Belanger et al., 2013; Krishnan et al., 2015). Each step must solve a linearized subproblem. Denote by f the Sparse MAP objective from Equation 5,

f(u, v) := η U u + η F v 1

The gradients of f with respect to the two variables are

uf(u , v ) = ηU u , vf(u , v ) = ηV .

A linear approximation to f around a point [u ; v ] is

ˆf(u, v) := ( uf) u+( vf) v = (ηU u ) u+η F v.

Minimizing ˆf over M is exactly MAP inference with adjusted variable scores ηU u . Intuitively, at each step we seek a high-scoring structure while penalizing sharing variables with already-selected structures Vanilla CG simply adds the new structure to the active set at every iteration. The pairwise and away-step variants trade off between the direction toward the new structure, and away from one of the already-selected structures. More sophisticated variants have been proposed (Garber & Meshi, 2016) which can provide sparse solutions when optimizing over a polytope.

Active set method. Importantly, the Sparse MAP problem in Equation 5 has quadratic curvature, which the general CG algorithms may not optimally leverage. For this reason, we consider the active set method for constrained QPs: a generalization of Wolfe s min-norm point algorithm (Wolfe, 1976), also used in structured prediction for the quadratic subproblems by Martins et al. (2015). The active set algorithm, at each iteration, updates an estimate of the solution support by adding or removing one constraint to/from the active set; then it solves the Karush Kuhn Tucker (KKT) system of a relaxed QP restricted to the current support.

Comparison. Both algorithms enjoy global linear convergence with similar rates (Lacoste-Julien & Jaggi, 2015), but the active set algorithm also exhibits exact finite convergence this allows it, for instance, to capture the optimal sparsity pattern (Nocedal & Wright, 1999, Ch. 16.4 & 16.5). Vinyes & Obozinski (2017) provide a more in-depth discussion of the connections between the two algorithms. We perform an empirical comparison on a dependency parsing instance with random potentials. Figure 2 shows that active set substantially outperforms all CG variants, both in terms of objective value as well as in the solution sparsity, suggesting that the quadratic curvature makes Sparse MAP solvable in very few iterations to high accuracy. We therefore use the active set solver in the remainder of the paper.

3.3. Backpropagating Gradients through Sparse MAP

In order to use Sparse MAP as a neural network layer trained with backpropagation, one must compute products of the Sparse MAP Jacobian with a vector p. Computing the Jacobian of an optimization problem is an active research topic known as argmin differentiation, and is generally difficult. Fortunately, as we show next, argmin differentiation is always easy and efficient in the case of Sparse MAP.

Proposition 1 Denote a Sparse MAP solution by y and its support by I := {s : ys > 0}. Then, Sparse MAP is

Sparse MAP: Differentiable Sparse Structured Inference

differentiable almost everywhere with Jacobian

η = MD(I)A , where D(I) = D(I) given by

( I 1 1T Z1Z11T zs, s I 0 s / I ,

Z := (MI MI) 1.

The proof, given in Appendix B, relies on the KKT conditions of the Sparse MAP QP. Importantly, because D(I) is zero outside of the support of the solution, computing the Jacobian only requires the columns of M and A corresponding to the structures in the active set. Moreover, when using the active set algorithm discussed in 3.2, the matrix Z is readily available as a byproduct of the forward pass. The backward pass can, therefore, be computed in O(k|I|).

Our approach for gradient computation draws its efficiency from the solution sparsity and does not depend on the type of structure considered. This is contrasted with two related lines of research. The first is unrolling iterative inference algorithms, for instance belief propagation (Stoyanov et al., 2011; Domke, 2013) and gradient descent (Belanger et al., 2017), where the backward pass complexity scales with the number of iterations. In the second, employed by Kim et al. (2017), when inference can be performed via dynamic programming, backpropagation can be performed using secondorder expectation semirings (Li & Eisner, 2009) or more general smoothing (Mensch & Blondel, 2018), in the same time complexity as the forward pass. Moreover, in our approach, neither the forward nor the backward passes involve logarithms, exponentiations or log-domain classes, avoiding the slowdown and stability issues normally incurred.

In the unstructured case, since M = I, Z is also an identity matrix, uncovering the sparsemax Jacobian (Martins & Astudillo, 2016). In general, structures are not necessarily orthogonal, but may have degrees of overlap.

4. Structured Fenchel-Young Losses and the Sparse MAP Loss

With the efficient algorithms derived above in hand, we switch gears to defining a Sparse MAP loss function. Structured output prediction models are typically trained by minimizing a structured loss measuring the discrepancy between the desired structure (encoded, for instance, as an indicator vector y = es) and the prediction induced by the log-potentials η. We provide here a general family of structured prediction losses that will make the newly proposed Sparse MAP loss arise as a very natural case. Below, we let Ω: RD R denote a convex penalty function and denote

by Ω its restriction to D RD, i.e.,

( Ω(y), y D; , y / D.

The Fenchel convex conjugate of Ω is

Ω (θ) := sup y RD θ y Ω (y) = sup y D θ y Ω(y).

We next introduce a family of structured prediction losses, named after the corresponding Fenchel-Young duality gap.

Definition 1 (Fenchel-Young losses) Given a convex penalty function Ω: RD R, and a (k D)-dimensional matrix A = [M; N] encoding the structure of the problem, we define the following family of structured losses:

ℓΩ,A(η, y) := Ω (A η) + Ω (y) η Ay. (6)

This family, studied in more detail in(Blondel et al., 2018), includes the commonly-used structured losses:

Structured perceptron (Collins, 2002): Ω 0;

Structured SVM (Taskar et al., 2003; Tsochantaridis et al., 2004): Ω ρ( , y) for a cost function ρ, where y is the true output;

CRF (Lafferty et al., 2001): Ω H;

Margin CRF (Gimpel & Smith, 2010): Ω H + ρ( , y).

This leads to a natural way of defining Sparse MAP losses, by plugging the following into Equation 6:

Sparse MAP loss: Ω(y) = 1

Margin Sparse MAP: Ω(y) = 1

2 My 2 2 + ρ(y, y).

It is well-known that the subgradients of structured perceptron and SVM losses consist of MAP inference, while the CRF loss gradient requires marginal inference. Similarly, the subgradients of the Sparse MAP loss can be computed via Sparse MAP inference, which in turn only requires MAP. The next proposition states properties of structured Fenchel Young losses, including a general connection between a loss and its corresponding inference method.

Proposition 2 Consider a convex Ωand a structured model defined by the matrix A Rk D. Denote the inference objective fΩ(y) := η Ay Ω(y), and a solution y := arg max y D fΩ(y). Then, the following properties hold:

1. ℓΩ,A(η, y) 0, with equality when fΩ(y) = fΩ(y );

2. ℓΩ,A(η, y) is convex, ℓΩ,A(η, y) A(y y);

3. ℓtΩ,A(η, y) = tℓΩ(η/t, y) for any t R, t > 0.

Sparse MAP: Differentiable Sparse Structured Inference

Table 1. Unlabeled attachment accuracy scores for dependency parsing, using a bi-LSTM model (Kiperwasser & Goldberg, 2016). Sparse MAP and its margin version, m-Sparse MAP, produce the best parser on 4/5 datasets. For context, we include the scores of the Co NLL 2017 UDPipe baseline, which is trained under the same conditions (Straka & Straková, 2017).

Loss en zh vi ro ja

Structured SVM 87.02 81.94 69.42 87.58 96.24 CRF 86.74 83.18 69.10 87.13 96.09 Sparse MAP 86.90 84.03 69.71 87.35 96.04 m-Sparse MAP 87.34 82.63 70.87 87.63 96.03

UDPipe baseline 87.68 82.14 69.63 87.36 95.94

Proof is given in Appendix C. Property 1 suggests that pminimizing ℓΩ,A aligns models with the true label. Property 2 shows how to compute subgradients of ℓΩ,A provided access to the inference output [u ; v ] = Ay Rk. Combined with our efficient procedure described in Section 3.2, it makes the Sparse MAP losses promising for structured prediction. Property 3 suggests that the strength of the penalty Ωcan be adjusted by simply scaling η. Finally, we remark that for a strongly-convex Ω, ℓΩ,A can be seen as a smoothed perceptron loss; other smoothed losses have been explored by Shalev-Shwartz & Zhang (2016).

5. Experimental Results

In this section, we experimentally validate Sparse MAP on two natural language processing applications, illustrating the two main use cases presented: structured output prediction with the Sparse MAP loss ( 5.1) and structured hidden layers ( 5.2). All models are implemented using the dynet library v2.0.2 (Neubig et al., 2017).

5.1. Dependency Parsing with the Sparse MAP Loss

We evaluate the Sparse MAP losses against the commonly used CRF and structured SVM losses. The task we focus on is non-projective dependency parsing: a structured output task consisting of predicting the directed tree of grammatical dependencies between words in a sentence (Jurafsky & Martin, 2018, Ch. 14). We use annotated Universal Dependency data (Nivre et al., 2016), as used in the Co NLL 2017 shared task (Zeman et al., 2017). To isolate the effect of the loss, we use the provided gold tokenization and part-of-speech tags. We follow closely the bidirectional LSTM arc-factored parser of Kiperwasser & Goldberg (2016), using the same model configuration; the only exception is not using externally pretrained embeddings. Parameters are trained using Adam (Kingma & Ba, 2015), tuning the learning rate on the grid {.5, 1, 2, 4, 8} 10 3, expanded by a factor of 2 if the best model is at either end.

We experiment with 5 languages, diverse both in terms of

Table 2. Test accuracy scores for natural language inference with structured and unstructured variants of ESIM. In parentheses: the percentage of pairs of words with nonzero alignment scores.

ESIM variant Multi NLI SNLI

softmax 76.05 (100%) 86.52 (100%) sequential 75.54 (13%) 86.62 (19%) matching 76.13 (8%) 86.05 (15%)

family and in terms of the amount of training data (ranging from 1,400 sentences for Vietnamese to 12,525 for English). Test set results (Table 1) indicate that the Sparse MAP losses outperform the SVM and CRF losses on 4 out of the 5 languages considered. This suggests that Sparse MAP is a good middle ground between MAP-based and marginalbased losses in terms of smoothness and gradient sparsity.

Moreover, as illustrated in Figure 4, the Sparse MAP loss encourages sparse predictions: models converge towards sparser solutions as they train, yielding very few ambiguous arcs. When confident, Sparse MAP can predict a single tree. Otherwise, the small set of candidate parses returned can be easily visualized, often indicating genuine linguistic ambiguities (Figure 3). Returning a small set of parses, also sought concomittantly by Keith et al. (2018), is valuable in pipeline systems, e.g., when the parse is an input to a downstream application: error propagation is diminished in cases where the highest-scoring tree is incorrect (which is the case for the sentences in Figure 3). Unlike K-best heuristics, Sparse MAP dynamically adjusts its output sparsity, which is desirable on realistic data where most instances are easy.

5.2. Latent Structured Alignment for Natural Language Inference

In this section, we demonstrate Sparse MAP for inferring latent structure in large-scale deep neural networks. We focus on the task of natural language inference, defined as the classification problem of deciding, given two sentences (a premise and a hypothesis), whether the premise entails the hypothesis, contradicts it, or is neutral with respect to it.

We consider novel structured variants of the state-of-the-art ESIM model (Chen et al., 2017). Given a premise P of length m and a hypothesis H of length n, ESIM:

1. Encodes P and H with an LSTM.

2. Computes alignment scores G Rm n; with gij the inner product between the P word i and H word j.

3. Computes P-to-H and H-to-P alignments using row-wise, respectively column-wise softmax on G.

4. Augments P words with the weighted average of its aligned H words, and vice-versa.

5. Passes the result through another LSTM, then predicts.

Sparse MAP: Differentiable Sparse Structured Inference

They did a vehicle wrap for my Toyota Venza that looks amazing .

1.0 1.0 1.01.0 1.0 1.01.01.0

the broccoli looks browned around the edges .

Figure 3. Example of ambiguous parses from the UD English validation set. Sparse MAP selects a small number of candidate parses (left: three, right: two), differing from each other in a small number of ambiguous dependency arcs. In both cases, the desired gold parse is among the selected trees (depicted by the arcs above the sentence), but it is not the highest-scoring one.

Training Validation

1 5 10 15 20 25 30 35 40 45 50 Epoch

1 5 10 15 20 25 30 35 40 45 50 Epoch

Figure 4. Distribution of the tree sparsity (top) and arc sparsity (bottom) of Sparse MAP solutions during training on the Chinese dataset. Shown are respectively the number of trees and the average number of parents per word with nonzero probability.

We consider the following structured replacements for the independent row-wise and column-wise softmaxes (step 3):

Sequential alignment. We model the alignment of p to h as a sequence tagging instance of length m, with n possible tags corresponding to the n words of the hypothesis. Through transition scores, we enable the model to capture continuity and monotonicity of alignments: we parametrize transitioning from word t1 to t2 by binning the distance t2 t1 into 5 groups, { 2 or less, 1, 0, 1, 2 or more}. We similarly parametrize the initial alignment using bins {1, 2 or more} and the final alignment as { 2 or less, 1}, allowing the model to express whether an alignment starts at the beginning or ends on the final word of h; formally

ηF (i, t1, t2) :=

wbin(t2 t1) 0 < i < n, wstart bin(t2) i = 0,

wend bin(t1) i = n.

We align p to h applying the same method in the other direction, with different transition scores w. Overall, sequential alignment requires learning 18 additional scalar parameters.

Matching alignment. We now seek a symmetrical alignment in both directions simultaneously. To this end, we cast the alignment problem as finding a maximal weight bipartite matching. We recall from 2.2 that a solution can be found via the Hungarian algorithm (in contrast to marginal

inference, which is #P-complete). When n = m, maximal matchings can be represented as permutation matrices, and when n = m some words remain unaligned. Sparse MAP returns a weighted average of a few maximal matchings. This method requires no additional learned parameters.

We evaluate the two models alongside the softmax baseline on the SNLI (Bowman et al., 2015) and Multi NLI (Williams et al., 2018) datasets.3 All models are trained by SGD, with 0.9 learning rate decay at epochs when the validation accuracy is not the best seen. We tune the learning rate on the grid 2k : k { 6, 5, 4, 3} , extending the range if the best model is at either end. The results in Table 2 show that structured alignments are competitive with softmax in terms of accuracy, but are orders of magnitude sparser. This sparsity allows them to produce global alignment structures that are interpretable, as illustrated in Figure 5.

Interestingly, we observe computational advantages of sparsity. Despite the overhead of GPU memory copying, both training and validation in our latent structure models take roughly the same time as with softmax and become faster as the models grow more certain. For the sake of comparison, Kim et al. (2017) report a 5 slow-down in their structured attention networks, where they use marginal inference.

3We split the Multi NLI matched validation set into equal validation and test sets; for SNLI we use the provided split.

Sparse MAP: Differentiable Sparse Structured Inference

a gentleman overlooking

a neighborhood

a situation

a gentleman overlooking

a neighborhood

(a) softmax

a gentleman overlooking

a neighborhood

a situation

a gentleman overlooking

a neighborhood

(b) sequence

a gentleman overlooking a neighborhood situation .

a police officer watches a situation closely .

a situation

a gentleman overlooking

a neighborhood

(c) matching

Figure 5. Latent alignments on an example from the SNLI validation set, correctly predicted as neutral by all compared models. The premise is on the y-axis, the hypothesis on the x-axis. Top: columns sum to 1; bottom: rows sum to 1. The matching alignment mechanism yields a symmetrical alignment, and is thus shown only once. Softmax yields a dense alignment (nonzero weights are marked with a border). The structures selected by sequential alignment are overlayed as paths; the selected matchings are displayed in the top right.

6. Related Work

Structured attention networks. Kim et al. (2017) and Liu & Lapata (2018) take advantage of the tractability of marginal inference in certain structured models and derive specialized backward passes for structured attention. In contrast, our approach is modular and general: with Sparse MAP, the forward pass only requires MAP inference, and the backward pass is efficiently computed based on the forward pass results. Moreover, unlike marginal inference, Sparse MAP yields sparse solutions, which is an appealing property statistically, computationally, and visually.

K-best inference. As it returns a small set of structures, Sparse MAP brings to mind K-best inference, often used in pipeline NLP systems for increasing recall and handling uncertainty (Yang & Cardie, 2013). K-best inference can be approximated (or, in some cases, solved), roughly K times slower than MAP inference (Yanover & Weiss, 2004; Camerini et al., 1980; Chegireddy & Hamacher, 1987; Fromer & Globerson, 2009). The main advantages of Sparse MAP are convexity, differentiablity, and modularity, as Sparse MAP can be computed in terms of MAP subproblems. Moreover, it yields a distribution, unlike K-best, which does not reveal the gap between selected structures,

Learning permutations. A popular approach for differentiable permutation learning involves mean-entropic optimal transport relaxations (Adams & Zemel, 2011; Mena et al., 2018). Unlike Sparse MAP, this does not apply to general

structures, and solutions are not directly expressible as combinations of a few permutations.

Regularized inference. Ravikumar et al. (2010), Meshi et al. (2015), and Martins et al. (2015) proposed ℓ2 perturbations and penalties in various related ways, with the goal of solving LP-MAP approximate inference in graphical models. In contrast, the goal of our work is sparse structured prediction, which is not considered in the aforementioned work. Nevertheless, some of the formulations in their work share properties with Sparse MAP; exploring the connections further is an interesting avenue for future work.

7. Conclusion

We introduced a new framework for sparse structured inference, Sparse MAP, along with a corresponding loss function. We proposed efficient ways to compute the forward and backward passes of Sparse MAP. Experimental results illustrate two use cases where sparse inference is well-suited. For structured prediction, the Sparse MAP loss leads to strong models that make sparse, interpretable predictions, a good fit for tasks where local ambiguities are common, like many natural language processing tasks. For structured hidden layers, we demonstrated that Sparse MAP leads to strong, interpretable networks trained end-to-end. Modular by design, Sparse MAP can be applied readily to any structured problem for which MAP inference is available, including combinatorial problems such as linear assignment.

Sparse MAP: Differentiable Sparse Structured Inference

Acknowledgements

We thank Tim Vieira, David Belanger, Jack Hessel, Justine Zhang, Sydney Zink, the Unbabel AI Research team, and the three anonymous reviewers for their insightful comments. This work was supported by the European Research Council (ERC St G Deep SPIN 758969) and by the Fundação para a Ciência e Tecnologia through contracts UID/EEA/50008/2013, PTDC/EEI-SII/7092/2014 (Learn Big), and CMUPERI/TIC/0046/2014 (Go Local).

Adams, R. P. and Zemel, R. S. Ranking via sinkhorn propagation. ar Xiv e-prints, 2011.

Amos, B. and Kolter, J. Z. Opt Net: Differentiable optimization as a layer in neural networks. In ICML, 2017.

Bakır, G., Hofmann, T., Schölkopf, B., Smola, A. J., Taskar, B., and Vishwanathan, S. V. N. Predicting Structured Data. The MIT Press, 2007.

Belanger, D. and Mc Callum, A. Structured prediction energy networks. In ICML, 2016.

Belanger, D., Sheldon, D., and Mc Callum, A. Marginal inference in MRFs using Frank-Wolfe. In NIPS Workshop on Greedy Opt., FW and Friends, 2013.

Belanger, D., Yang, B., and Mc Callum, A. End-to-end learning for structured prediction energy networks. In ICML, 2017.

Birkhoff, G. Tres observaciones sobre el algebra lineal. Univ. Nac. Tucumán Rev. Ser. A, 5:147 151, 1946.

Blondel, M., Martins, A. F., and Niculae, V. Learning classifiers with Fenchel-Young losses: Generalized entropies, margins, and algorithms. ar Xiv e-prints, 2018.

Bowman, S. R., Angeli, G., Potts, C., and Manning, C. D.

A large annotated corpus for learning natural language inference. In EMNLP, 2015.

Boyd, S. and Vandenberghe, L. Convex Optimization. Cambridge University Press, 2004.

Camerini, P. M., Fratta, L., and Maffioli, F. The k best spanning arborescences of a network. Networks, 10(2): 91 109, 1980.

Chegireddy, C. R. and Hamacher, H. W. Algorithms for finding K-best perfect matchings. Discrete Applied Mathematics, 18(2):155 165, 1987.

Chen, Q., Zhu, X., Ling, Z.-H., Wei, S., Jiang, H., and Inkpen, D. Enhanced LSTM for natural language inference. In ACL, 2017.

Chu, Y.-J. and Liu, T.-H. On the shortest arborescence of a directed graph. Science Sinica, 14:1396 1400, 1965.

Clarke, F. H. Optimization and Nonsmooth Analysis. SIAM, 1990.

Collins, M. Discriminative training methods for Hidden Markov Models: Theory and experiments with perceptron algorithms. In EMNLP, 2002.

Domke, J. Learning graphical model parameters with approximate marginal inference. IEEE T. Pattern. Anal., 35 (10):2454 2467, 2013.

Edmonds, J. Optimum branchings. J. Res. Nat. Bur. Stand., 71B:233 240, 1967.

Fenchel, W. On conjugate convex functions. Canad. J. Math, 1(73-77), 1949.

Frank, M. and Wolfe, P. An algorithm for quadratic programming. Nav. Res. Log., 3(1-2):95 110, 1956.

Fromer, M. and Globerson, A. An LP view of the M-best MAP problem. In NIPS, 2009.

Garber, D. and Meshi, O. Linear-memory and decomposition-invariant linearly convergent conditional gradient algorithm for structured polytopes. In NIPS, 2016.

Gimpel, K. and Smith, N. A. Softmax-margin CRFs: Training log-linear models with cost functions. In NAACL, 2010.

Jonker, R. and Volgenant, A. A shortest augmenting path algorithm for dense and sparse linear assignment problems. Computing, 38(4):325 340, 1987.

Jurafsky, D. and Martin, J. H. Speech and Language Processing (3rd ed.). draft, 2018.

Keith, K., Blodgett, S. L., and O Connor, B. Monte Carlo syntax marginals for exploring and using dependency parses. In NAACL, 2018.

Kim, Y., Denton, C., Hoang, L., and Rush, A. M. Structured attention networks. In ICLR, 2017.

Kingma, D. and Ba, J. Adam: A method for stochastic optimization. In ICLR, 2015.

Kiperwasser, E. and Goldberg, Y. Simple and accurate dependency parsing using bidirectional LSTM feature representations. TACL, 4:313 327, 2016.

Kirchhoff, G. Ueber die auflösung der gleichungen, auf welche man bei der untersuchung der linearen vertheilung galvanischer ströme geführt wird. Annalen der Physik, 148(12):497 508, 1847.

Koo, T., Globerson, A., Carreras Pérez, X., and Collins, M.

Structured prediction models via the matrix-tree theorem. In EMNLP, 2007.

Krishnan, R. G., Lacoste-Julien, S., and Sontag, D. Barrier Frank-Wolfe for marginal inference. In NIPS, 2015.

Kschischang, F. R., Frey, B. J., and Loeliger, H.-A. Factor graphs and the sum-product algorithm. IEEE T. Inform. Theory, 47(2):498 519, 2001.

Sparse MAP: Differentiable Sparse Structured Inference

Kuhn, H. W. The Hungarian method for the assignment problem. Nav. Res. Log., 2(1-2):83 97, 1955.

Lacoste-Julien, S. and Jaggi, M. On the global linear convergence of Frank-Wolfe optimization variants. In NIPS, 2015.

Lafferty, J. D., Mc Callum, A., and Pereira, F. C. N. Conditional Random Fields: Probabilistic models for segmenting and labeling sequence data. In ICML, 2001.

Li, Z. and Eisner, J. First-and second-order expectation semirings with applications to minimum-risk training on translation forests. In EMNLP, 2009.

Liu, Y. and Lapata, M. Learning structured text representations. TACL, 6:63 75, 2018.

Martins, A. F. and Astudillo, R. F. From softmax to sparsemax: A sparse model of attention and multi-label classification. In ICML, 2016.

Martins, A. F., Smith, N. A., and Xing, E. P. Concise integer linear programming formulations for dependency parsing. In ACL-IJCNLP, 2009.

Martins, A. F., Figueiredo, M. A., Aguiar, P. M., Smith, N. A., and Xing, E. P. AD3: Alternating directions dual decomposition for MAP inference in graphical models. JMLR, 16(1):495 545, 2015.

Mc Donald, R., Crammer, K., and Pereira, F. Online largemargin training of dependency parsers. In ACL, 2005.

Mena, G., Belanger, D., Linderman, S., and Snoek, J. Learning latent permutations with Gumbel-Sinkhorn networks. In ICLR, 2018.

Mensch, A. and Blondel, M. Differentiable dynamic programming for structured prediction and attention. In ICML, 2018.

Meshi, O., Mahdavi, M., and Schwing, A. Smooth and strong: MAP inference with linear convergence. In NIPS, 2015.

Neubig, G., Dyer, C., Goldberg, Y., Matthews, A., Ammar, W., Anastasopoulos, A., Ballesteros, M., Chiang, D., Clothiaux, D., Cohn, T., et al. Dy Net: The dynamic neural network toolkit. preprint ar Xiv:1701.03980, 2017.

Niculae, V. and Blondel, M. A regularized framework for sparse and structured neural attention. In NIPS, 2017.

Nivre, J., de Marneffe, M.-C., Ginter, F., Goldberg, Y., Hajic, J., Manning, C. D., Mc Donald, R. T., Petrov, S., Pyysalo, S., Silveira, N., et al. Universal Dependencies v1: A multilingual treebank collection. In LREC, 2016.

Nocedal, J. and Wright, S. Numerical Optimization. Springer New York, 1999.

Nowozin, S., Gehler, P. V., Jancsary, J., and Lampert, C. H.

Advanced Structured Prediction. MIT Press, 2014.

Rabiner, L. R. A tutorial on Hidden Markov Models and

selected applications in speech recognition. P. IEEE, 77 (2):257 286, 1989.

Ravikumar, P., Agarwal, A., and Wainwright, M. J. Messagepassing for graph-structured linear programs: Proximal methods and rounding schemes. JMLR, 11:1043 1080, 2010.

Shalev-Shwartz, S. and Zhang, T. Accelerated proximal stochastic dual coordinate ascent for regularized loss minimization. Math. Program., 155(1):105 145, 2016.

Smith, D. A. and Smith, N. A. Probabilistic models of nonprojective dependency trees. In EMNLP, 2007.

Smith, N. A. Linguistic Structure Prediction. Synthesis Lectures on Human Language Technologies. Morgan and Claypool, May 2011.

Stoyanov, V., Ropson, A., and Eisner, J. Empirical risk minimization of graphical model parameters given approximate inference, decoding, and model structure. In AISTATS, 2011.

Straka, M. and Straková, J. Tokenizing, POS tagging, lemmatizing and parsing UD 2.0 with UDPipe. In Co NLL Shared Task, 2017.

Taskar, B. Learning Structured Prediction Models: A Large Margin Approach. Ph D thesis, Stanford University, 2004.

Taskar, B., Guestrin, C., and Koller, D. Max-Margin Markov Networks. In NIPS, 2003.

Tsochantaridis, I., Hofmann, T., Joachims, T., and Altun, Y.

Support vector machine learning for interdependent and structured output spaces. In ICML, 2004.

Valiant, L. G. The complexity of computing the permanent. Theor. Comput. Sci., 8(2):189 201, 1979.

Vinyes, M. and Obozinski, G. Fast column generation for atomic norm regularization. In AISTATS, 2017.

Wainwright, M. J. and Jordan, M. I. Graphical models, exponential families, and variational inference. Found. Trends Mach. Learn., 1(1 2):1 305, 2008.

Williams, A., Nangia, N., and Bowman, S. R. A broadcoverage challenge corpus for sentence understanding through inference. In NAACL, 2018.

Wolfe, P. Finding the nearest point in a polytope. Mathematical Programming, 11(1):128 149, 1976.

Yang, B. and Cardie, C. Joint inference for fine-grained opinion extraction. In ACL, 2013.

Yanover, C. and Weiss, Y. Finding the M most probable configurations using loopy belief propagation. In NIPS, 2004.

Zeman, D., Popel, M., Straka, M., Hajic, J., Nivre, J., Ginter, F., Luotolahti, J., Pyysalo, S., Petrov, S., Potthast, M., et al. Co NLL 2017 shared task: Multilingual parsing from raw text to universal dependencies. Co NLL, 2017.