Abstract

Distributional semantics has been extended to phrases and sentences by means of composition operations. We look at how these operations affect similarity measurements, showing that similarity equations of an important class of composition methods can be decomposed into operations performed on the subparts of the input phrases. This establishes a strong link between these models and convolution kernels.

1. Introduction

Distributional semantics approximates word meanings with vectors tracking co-occurrence in corpora (Turney and Pantel 2010). Recent work has extended this approach to phrases and sentences through vector composition (Clark 2015). Resulting compositional distributional semantic models (CDSMs) estimate degrees of semantic similarity (or, more generally, relatedness) between two phrases: A good CDSM might tell us that green bird is closer to parrot than to pigeon, useful for tasks such as paraphrasing.

We take a mathematical look1 at how the composition operations postulated by CDSMs affect similarity measurements involving the vectors they produce for phrases or sentences. We show that, for an important class of composition methods, encompassing at least those based on linear transformations, the similarity equations can be decomposed into operations performed on the subparts of the input phrases, and typically factorized into terms that reflect the linguistic structure of the input. This establishes a strong link between CDSMs and convolution kernels (Haussler 1999), which act in the same way. We thus refer to our claim as the “Convolution Conjecture.”

We focus on the models in Table 1. These CDSMs all apply linear methods, and we suspect that linearity is a sufficient (but not necessary) condition to ensure that the Convolution Conjecture holds. We will first illustrate the conjecture for linear methods, and then briefly consider two nonlinear approaches: the dual space model of Turney (2012), for which it does, and a representative of the recent strand of work on neural-network models of composition, for which it does not.

Table 1

Compositional Distributional Semantic Models: , , and are distributional vectors representing the words a, b, and c, respectively; matrices X, Y, and Z are constant across a phrase type, corresponding to syntactic slots; the matrix A and the third-order tensor B represent the predicate words a in the first phrase and b in the second phrase, respectively.

2-word phrase3-word phrasereference
Multiplicative   Mitchell and Lapata (2008
Full Additive   Guevara (2010), Zanzotto et al. (2010
2-word phrase3-word phrasereference
Multiplicative   Mitchell and Lapata (2008
Full Additive   Guevara (2010), Zanzotto et al. (2010

2. Mathematical Preliminaries

Vectors are represented as small letters with an arrow and their elements are ai, matrices as capital letters in bold A and their elements are Aij, and third-order or fourth-order tensors as capital letters in the form A and their elements are Aijk or Aijkh. The symbol ⊡ represents the element-wise product and ⊗ is the tensor product. The dot product is and the Frobenius product—that is, the generalization of the dot product to matrices and high-order tensors—is represented as 〈A, BF and 〈A, BF. The Frobenius product acts on vectors, matrices, and third-order tensors as follows:
A simple property that relates the dot product between two vectors and the Frobenius product between two general tensors is the following:
where I is the identity matrix. The dot product of and can be rewritten as:
Let A and B be two third-order tensors and , , , four vectors. It can be shown that:
where is a non-standard way to indicate the tensor contraction of the tensor product between two third-order tensors. In this particular tensor contraction, the elements Ciknm of the resulting fourth-order tensor C are . The elements Diknm of the tensor are Diknm = xiykancm.

3. Formalizing the Convolution Conjecture

Structured Objects. In line with Haussler (1999), a structured objectxX is either a terminal object that cannot be furthermore decomposed, or a non-terminal object that can be decomposed into n subparts. We indicate with one such decomposition, where the subparts xiX are structured objects themselves. The set X is the set of the structured objects and TXX is the set of the terminal objects. A structured object x can be anything according to the representational needs. Here, x is a representation of a text fragment, and so it can be a sequence of words, a sequence of words along with their part of speech, a tree structure, and so on. The set is the set of decompositions of x relevant to define a specific CDSM. Note that a given decomposition of a structured object x does not need to contain all the subparts of the original object. For example, let us consider the phrase x = tall boy We can then define . This set contains the three possible decompositions of the phrase: , , and .

Recursive formulation of CDSM. A CDSM can be viewed as a function f that acts recursively on a structured object x. If x is a non-terminal object
where is the set of relevant decompositions, is a repeated operation on this set, γ is a function defined on f(xi) where xi are the subparts of a decomposition of x. If x is a terminal object, f(x) is directly mapped to a tensor. The function f may operate differently on different kinds of structured objects, with tensor degree varying accordingly. The set and the functions f, γ, and depend on the specific CDSM, and the same CDSM might be susceptible to alternative analyses satisfying the form in Equation (5). As an example, under Additive, x is a sequence of words and f is
where . The repeated operation corresponds to summing and γ is identity. For Multiplicative we have
where (a single trivial decomposition including all subparts). With a single decomposition, the repeated operation reduces to a single term; and here γ is the product (it will be clear subsequently, when we apply the Convolution Conjecture to these models, why we are assuming different decomposition sets for Additive and Multiplicative).

Definition 1 (Convolution Conjecture)

Definition 1 (Convolution Conjecture)
For every CDSM f along with its set, there exist functions K,Ki and a function g such that:

The Convolution Conjecture postulates that the similarity K(f (x), f (y)) between the tensors f (x) and f (y) is computed by combining operations on the subparts, that is, Ki(f (xi), f (yi)), using the function g. This is exactly what happens in convolution kernels (Haussler 1999). K is usually the dot product, but this is not necessary: We will show that for the dual-space model of Turney 2012 K turns out to be the fourth root of the Frobenius tensor.

4. Comparing Composed Phrases

We illustrate now how the Convolution Conjecture (CC) applies to the considered CDSMs, exemplifying with adjective–noun and subject–verb–object phrases. Without loss of generality we use tall boy and red cat for adjective–noun phrases and goats eat grass and cows drink water for subject–verb–object phrases.

Additive Model.K and Ki are dot products, g is the identity function, and f is as in Equation (6). The structure of the input is a word sequence (i.e., x = (w1w2)) and the relevant decompositions consist of these single words, . Then

The CC form of Additive shows that the overall dot product can be decomposed into dot products of the vectors of the single words. Composition does not add any further information. These results can be easily extended to longer phrases and to phrases of different length.

Multiplicative Model.K, g are dot products, Ki the component-wise product, and f is as in Equation (7). The structure of the input is x = (w1w2), and we use the trivial single decomposition consisting of all subparts (thus summation reduces to a single term):
This is the dot product between an indistinct chain of element-wise products and a vector of all ones or the product of two separate element-wise products, one on adjectives , and one on nouns . In this latter CC form, the final dot product is obtained in two steps: first separately operating on the adjectives and on the nouns; then taking the dot product of the resulting vectors. The comparison operations are thus reflecting the input syntactic structure. The results can be easily extended to longer phrases and to phrases of different lengths.
Full Additive Model. The input consists of a sequence of (label,word) pairs x = ((L1w1), … , (Lnwn)) and the relevant decomposition set includes the single tuples, that is, . The CDSM f is defined as
The repeated operation here is summation, and γ the matrix-by-vector product. In the CC form, K is the dot product, g the Frobenius product, K1(f (x), f (y)) = f (x)Tf (y), and K2(f (x), f (y)) = f (x)f (y)T. We have then for adjective–noun composition (by using the property in Equation (3)):
The CC form shows how Full Additive factorizes into a more structural and a more lexical part: Each element of the sum is the Frobenius product between the product of two matrices representing syntactic labels and the tensor product between two vectors representing the corresponding words. For subject–verb–object phrases ((Sw1) (Vw2) (Ow3)) we have
Again, we observe the factoring into products of syntactic and lexical representations.

By looking at Full Additive in the CC form, we observe that when XTYI for all matrix pairs, it degenerates to Additive. Interestingly, Full Additive can also approximate a semantic convolution kernel (Mehdad, Moschitti, and Zanzotto 2010), which combines dot products of elements in the same slot. In the adjective–noun case, we obtain this approximation by choosing two nearly orthonormal matrices A and N such that AAT = NNTI and ANT0 and applying Equation (2): .

This approximation is valid also for three-word phrases. When the matrices S, V, and O are such that XXTI with X one of the three matrices and YXT0 with X and Y two different matrices, Full Additive approximates a semantic convolution kernel comparing two sentences by summing the dot products of the words in the same role, that is,

Results can again be easily extended to longer and different-length phrases.

Lexical Function Model. We distinguish composition with one- vs. two argument predicates. We illustrate the first through adjective–noun composition, where the adjective acts as the predicate, and the second with transitive verb constructions. Although we use the relevant syntactic labels, the formulas generalize to any construction with the same argument count. For adjective–noun phrases, the input is a sequence of (label, word) pairs (x = ((A,w1),(N,w2))) and the relevant decomposition set again includes only the single trivial decomposition into all the subparts: . The method itself is recursively defined as
Here, K and g are, respectively, the dot and Frobenius product, K1(f (x), f (y)) = f (x)Tf (y), and K2(f (x), f (y)) = f (x)f (y)T. Using Equation (3), we have then
The role of predicate and argument words in the final dot product is clearly separated, showing again the structure-sensitive nature of the decomposition of the comparison operations. In the two-place predicate case, again, the input is a set of (label, word) tuples, and the relevant decomposition set only includes the single trivial decomposition into all subparts. The CDSM f is defined as
K is the dot product and g(x, y, z) = 〈x, yzF, —that is, the tensor contraction2 along the second index of the tensor product between f (x) and f (y)—and K2(f (x), f (y)) = K3(f (x), f (y)) = f (x) ⊗ f (y) are tensor products. The dot product of and is (by using Equation (4))

We rewrote the equation as a Frobenius product between two fourth-order tensors. The first combines the two third-order tensors of the verbs and the second combines the vectors representing the arguments of the verb, that is: . In this case as well we can separate the role of predicate and argument types in the comparison computation.

Extension of the Lexical Function to structured objects of different lengths is treated by using the identity element ε for missing parts. As an example, we show here the comparison between tall boy and cat where the identity element is the identity matrix I:
Dual Space Model. We have until now applied the CC to linear CDSMs with the dot product as the final comparison operator (what we called K). The CC also holds for the effective Dual Space model of Turney (2012), which assumes that each word has two distributional representations, wd in “domain” space and wf in “function” space. The similarity of two phrases is directly computed as the geometric average of the separate similarities between the first and second words in both spaces. Even though there is no explicit composition step, it is still possible to put the model in CC form. Take x = (x1, x2) and its trivial decomposition. Define, for a word w with vector representations wd and wf: . Define also , and g(a, b) to be . Then
A Neural-network-like Model. Consider the phrase (w1, w2, …, wn) and the model defined by , where σ(·) is a component-wise logistic function. Here we have a single trivial decomposition that includes all the subparts, and γ(x1,…,xn) is defined as σ(x1 + … + xn). To see that for this model the CC cannot hold, consider two two-word phrases (ab) and (cd)
We need to rewrite this as
But there is no possible choice of g, K1, and K2 that allows Equation (21) to be written as Equation (22). This example can be regarded as a simplified version of the neural-network model of Socher et al. (2011). The fact that the CC does not apply to it suggests that it will not apply to other models in this family.

5. Conclusion

The Convolution Conjecture offers a general way to rewrite the phrase similarity computations of CDSMs by highlighting the role played by the subparts of a composed representation. This perspective allows for a better understanding of the exact operations that a composition model applies to its input. The Convolution Conjecture also suggests a strong connection between CDSMs and semantic convolution kernels. This link suggests that insights from the CDSM literature could be directly integrated in the development of convolution kernels, with all the benefits offered by this well-understood general machine-learning framework.

Acknowledgments

We thank the reviewers for helpful comments. Marco Baroni acknowledges ERC 2011 Starting Independent Research Grant n. 283554 (COMPOSES).

Notes

1

Ganesalingam and Herbelot (2013) also present a mathematical investigation of CDSMs. However, except for the tensor product (a composition method we do not consider here as it is not empirically effective), they do not look at how composition strategies affect similarity comparisons.

2

Grefenstette et al. (2013) first framed the Lexical Function in terms of tensor contraction.

References

Clark
,
Stephen
.
2015
.
Vector space models of lexical meaning
. In
Shalom
Lappin
and
Chris
Fox
, editors,
Handbook of Contemporary Semantics, 2nd ed
.
Blackwell
,
Malden, MA
.
In press
.
Coecke
,
Bob
,
Mehrnoosh
, and
Stephen
Clark
.
2010
.
Mathematical foundations for a compositional distributional model of meaning
.
Linguistic Analysis
,
36
:
345
384
.
Ganesalingam
,
Mohan
and
Aurélie
Herbelot
.
2013
.
Composing distributions: Mathematical structures and their linguistic interpretation
.
Working paper, Computer Laboratory, University of Cambridge. Available at www.cl.cam.ac.uk/∼ah433/
.
Grefenstette
,
Edward
,
Georgiana
Dinu
,
Yao-Zhong
Zhang
,
Mehrnoosh
, and
Marco
Baroni
.
2013
.
Multi-step regression learning for compositional distributional semantics
.
Proceedings of IWCS
, pages
131
142
,
Potsdam
.
Guevara
,
Emiliano
.
2010
.
A regression model of adjective-noun compositionality in distributional semantics
. In
Proceedings of GEMS
, pages
33
37
,
Uppsala
.
Haussler
,
David
.
1999
.
Convolution kernels on discrete structures
.
Technical report USCS-CL-99-10, University of California at Santa Cruz
.
,
Yashar
,
Alessandro
Moschitti
, and
Fabio Massimo
Zanzotto
.
2010
.
Syntactic/semantic structures for textual entailment recognition
. In
Proceedings of NAACL
, pages
1,020
1,028
,
Los Angeles, CA
.
Mitchell
,
Jeff
and
Mirella
Lapata
.
2008
.
Vector-based models of semantic composition
. In
Proceedings of ACL
, pages
236
244
,
Columbus, OH
.
Socher
,
Richard
,
Eric
Huang
,
Jeffrey
Pennin
,
Andrew
Ng
, and
Christopher
Manning
.
2011
.
Dynamic pooling and unfolding recursive autoencoders for paraphrase detection
. In
Proceedings of NIPS
, pages
801
809
,
.
Turney
,
Peter
.
2012
.
Domain and function: A dual-space model of semantic relations and compositions
.
Journal of Artificial Intelligence Research
,
44
:
533
585
.
Turney
,
Peter
and
Patrick
Pantel
.
2010
.
From frequency to meaning: Vector space models of semantics
.
Journal of Artificial Intelligence Research
,
37
:
141
188
.
Zanzotto
,
Fabio Massimo
,
Ioannis
Korkontzelos
,
Francesca
Falucchi
, and
Suresh
Manandhar
.
2010
.
Estimating linear models for compositional distributional semantics
. In
Proceedings of COLING
, pages
1,263
1,271
,
Beijing
.

Author notes

*

Department of Enterprise Engineering, University of Rome “Tor Vergata,” Viale del Politecnico, 1, 00133 Rome, Italy. E-mail: fabio.massimo.zanzotto@uniroma2.it.