Vector-based Models of Semantic Composition

Jeff Mitchell and Mirella Lapata

School of Informatics, University of Edinburgh
10 Chrichton Street, Edinburgh EH10 9AB, UK
[email protected], [email protected]

Abstract
Vector-based models of word meaning have become increasingly popular in cogni-
tive science. The appeal of these models lies in their ability to represent meaning
simply by using distributional information under the assumption that words occur-
ring within similar contexts are semantically similar. Despite their widespread use,
vector-based models are typically directed at representing words in isolation and
methods for constructing representations for phrases or sentences have received lit-
tle attention in the literature. This is in marked contrast to experimental evidence
(e.g., in sentential priming) suggesting that semantic similarity is more complex
than simply a relation between isolated words. This article proposes a framework
for representing the meaning of word combinations in vector space. Central to
our approach is vector composition which we operationalize in terms of additive
and multiplicative functions. Under this framework, we introduce a wide range of
composition models which we evaluate empirically on a phrase similarity task.

Introduction
The question of how semantic knowledge is acquired, organized, and ultimately used in lan-
guage processing and understanding has been a topic of great debate in cognitive science. This is
hardly surprising as the ability to retrieve and manipulate meaning influences many cognitive tasks
that go far and beyond language processing. Examples include memory retrieval (Deese, 1959;
Raaijmakers & Shiffrin, 1981), categorization (Estes, 1994; Nosofsky, 1984, 1986), problem solv-
ing (Holyoak & Koh, 1987; Ross, 1987, 1989a), reasoning (Rips, 1975; Heit & Rubinstein, 1994),
and learning (Gentner, 1989; Ross, 1984).
Previous accounts of semantic representation fall under three broad families, namely seman-
tic networks, feature-based models, and semantic spaces (for a fuller account of the different ap-
proaches and issue involved see e.g. Markman, 1998). Semantic networks (Collins & Quillian,
1969) represent concepts as nodes in a graph. Edges in the graph denote semantic relationships
between concepts (e.g., DOG IS - A MAMMAL, DOG HAS TAIL) and word meaning is expressed by
the number and type of connections to other words. In this framework, word similarity is a func-
tion of path length semantically related words are expected to have shorter paths between them
(e.g., poodle will be more similar to dog than animal ). Semantic networks constitute a some-
what idealized representation that abstracts away from real word usage they are traditionally
 VECTOR-BASED MODELS OF SEMANTIC COMPOSITION 2

hand-coded by modelers who a priori decide which relationships are most relevant in representing
meaning. More recent work (Steyvers & Tenenbaum, 2005) creates a semantic network from word
association norms (Nelson, McEvoy, & Schreiber, 1999), however these can only represent a small
fraction of the vocabulary of an adult speaker.
An alternative to semantic networks is the idea that word meaning can be described in terms
of feature lists (Smith & Medin, 1981). Theories tend to differ with respect to their definition
of features. In many cases these are created manually by the modeler (e.g., Hinton & Shallice,
1991). In other cases, the features are obtained by asking native speakers to generate attributes they
consider important in describing the meaning of a word (e.g., Andrews, Vigliocco, & Vinson, 2009;
McRae, de Sa, & Seidenberg, 1997). This allows the representation of each word by a distribution
of numerical values over the feature set. Admittedly, norming studies have the potential of revealing
which dimensions of meaning are psychologically salient. However, a number of difficulties arise
when working with such data (Murphy & Medin, 1985; Sloman & Rips, 1998). For example, the
number and types of attributes generated can vary substantially as a function of the amount of time
devoted to each word. There are many degrees of freedom in the way that responses are coded and
analyzed. And multiple subjects are required to create a representation for each word, which in
practice limits elicitation studies to a small-size lexicon.
A third popular tradition of studying semantic representation has been driven by the assump-
tion that word meaning can be learned from the linguistic environment. Words that are similar in
meaning e.g., boat and ship tend to occur in contexts of similar words, such as sail, sea, sailor, and
so on. Word meaning can be thus captured quantitatively in terms of simple co-occurrence statistics.
Semantic space models thus represent meaning as a vector in a high-dimensional space, where each
component corresponds to some contextual element in which the word is found. The contextual el-
ements can be words themselves (Lund & Burgess, 1996), larger linguistic units such as paragraphs
or documents (Landauer & Dumais, 1997), or even more complex linguistic representations such
as n-grams (Jones & Mewhort, 2007) and the argument slots of predicates (Grefenstette, 1994; Lin,
1998; Pado & Lapata, 2007). The advantage of taking such a geometric approach is that the simi-
larity of word meanings can be easily quantified by measuring their distance in the vector space, or
the cosine of the angle between them. A simplified example of a two-dimensional semantic space
is shown in Figure 1 (semantic spaces usually have hundreds of dimensions).
There are a number of well-known semantic space models in the literature. For example,
the Hyperspace Analog to Language model (HAL, Lund & Burgess, 1996) represents each word
by a vector where each element of the vector corresponds to a weighted co-occurrence value of
that word with some other word. Latent Semantic Analysis (LSA, Landauer & Dumais, 1997)
also derives a high-dimensional semantic space for words while using co-occurrence information
between words and the passages they occur in. LSA constructs a word-document co-occurrence
matrix from a large document collection. Matrix decomposition techniques are usually applied
to reduce the dimensionality of the original matrix thereby rendering it more informative. The
dimensionality reduction allows words with similar meaning to have similar vector representations,
even if they never co-occurred in the same document.
Probabilistic topic models (Griffiths, Steyvers, & Tenenbaum, 2007; Blei, Ng, & Jordan,
2003) offer an alternative to semantic spaces based on the assumption that words observed in a
corpus manifest some latent structure linked to topics. These models are similar in spirit to LSA,
they also operate on large corpora and derive a reduced dimensionality description of words and
documents. Crucially, words are not represented as points in a high-dimensional space but as a
 VECTOR-BASED MODELS OF SEMANTIC COMPOSITION 3

hand
arm
hair

difficulty
circumstance
problem situation
condition

Figure 1. In a semantic space words are represented as points, and proximity indicates semantic association.
Here, circumstance, situation and condition are similar to each other and different from hand, arm and hair.

probability distribution over a set of topics (corresponding to coarse-grained senses). Each topic is
a probability distribution over words, and the content of the topic is reflected in the words to which
it assigns high probability. Topic models are generative, they specify a probabilistic procedure by
which documents can be generated. So, to make a new document one first chooses a distribution
over topics. Then for each word in that document, one chooses a topic at random according to this
distribution, and selects a word from that topic. Under this framework, the problem of meaning
representation is expressed as one of statistical inference: given some data words in a corpus
infer the latent structure from which it was generated.
Semantic space models (and the related topic models) have been successful at simulating a
wide range of psycholinguistic phenomena including semantic priming (Lund & Burgess, 1996;
Landauer & Dumais, 1997; Griffiths et al., 2007), discourse comprehension (Landauer & Dumais,
1997; Foltz, Kintsch, & Landauer, 1998), word categorization (Laham, 2000), judgments of es-
say quality (Landauer, Laham, Rehder, & Schreiner, 1997a), synonymy tests (Landauer & Dumais,
1997; Griffiths et al., 2007) such as those included in the Test of English as Foreign Language
(TOEFL), reading times (McDonald, 2000; Griffiths et al., 2007), and judgments of semantic simi-
larity (McDonald, 2000) and association (Denhire & Lemaire, 2004; Griffiths et al., 2007).
Despite their widespread use, these models are typically directed at representing words in
isolation and methods for constructing representations for phrases or sentences have received little
attention in the literature. However, it is well-known that linguistic structures are compositional
(simpler elements are combined to form more complex ones). For example, morphemes are com-
bined into words, words into phrases, and phrases into sentences. It is also reasonable to assume
that the meaning of sentences is composed of the meanings of individual words or phrases. Much
experimental evidence also suggests that semantic similarity is more complex than simply a relation
between isolated words. For example, Duffy, Henderson, and Morris (1989) showed that priming
of sentence terminal words was dependent not simply on individual preceding words but on their
combination, and Morris (1994) later demonstrated that this priming also showed dependencies on
 VECTOR-BASED MODELS OF SEMANTIC COMPOSITION 4

the syntactic relations in the preceding context. Additional evidence comes from experiments where
target words in sentences are compared to target words in lists or scrambled sentences. Changes in
the temporal order of words in a sentence decrease the strength of the related priming effect (Foss,
1982; Masson, 1986; OSeaghdha, 1989; Simpson, Peterson, Casteel, & Brugges, 1989). For ex-
ample, Simpson et al. (1989) found relatedness priming effects for words embedded in grammatical
sentences (The auto accident drew a large crowd of people) but not for words in scrambled stimuli
(Accident of large the drew auto crowd a people). These findings highlight the role of syntactic
structure in modulating priming behavior. They also suggest that models of semantic similarity
should ideally handle the combination of semantic content in a syntactically aware manner.
Composition operations can be naturally accounted for within logic-based semantic frame-
works (Montague, 1974). Freges principle of compositionality states that the meaning of a com-
plete sentence must be explained in terms of the meanings of its subsentential parts, including those
of its singular terms. In other words, each syntactic operation of a formal language should have a
corresponding semantic operation. Problematically, representations in terms of logical formulas are
not well suited to modeling similarity quantitatively (as they are based on discrete symbols). On
the other hand, semantic space models can naturally measure similarity but are not compositional.
In fact, the commonest method for combining the vectors is to average them. While vector averag-
ing has been effective in some applications such as essay grading (Landauer & Dumais, 1997) and
coherence assessment (Foltz et al., 1998), it is unfortunately insensitive to word order, and more
generally syntactic structure, giving the same representation to any constructions that happen to
share the same vocabulary. This is illustrated in the example below taken from Landauer, Laham,
Rehder, and Schreiner (1997b). Sentences (1-a) and (1-b) contain exactly the same set of words but
their meaning is entirely different.

(1) a. It was not the sales manager who hit the bottle that day, but the office worker with the
serious drinking problem.
b. That day the office manager, who was drinking, hit the problem sales worker with the
bottle, but it was not serious.

The relative paucity of compositional models in the semantic space literature is in marked
contrast with work in the connectionist tradition where much effort has been devoted to problem of
combining or binding high-dimensional representations. The construction of higher-level structures
from low-level ones is fundamental not only to language but many aspects of human cognition such
as analogy retrieval and processing (Plate, 2000; Eliasmith & Thagard, 2001), memory (Kanerva,
1988), and problem solving (Ross, 1989b). Indeed, the issue of how to represent compositional
structure in neural networks has been a matter of great controversy (Fodor & Pylyshyn, 1988).
While neural networks can readily represent single distinct objects, in the case of multiple objects
there are fundamental difficulties in keeping track of which features are bound to which objects.
For the hierarchical structure of natural language this binding problem becomes particularly acute.
For example, simplistic approaches to handling sentences such as John loves Mary and Mary loves
John typically fail to make valid representations in one of two ways. Either there is a failure to
distinguish between these two structures, because the network fails to keep track of the fact that
John is subject in one and object in the other, or there is a failure to recognize that both structures
involve the same participants, because John as a subject has a distinct representation from John as
an object. The literature is littered with solutions to the binding problem (see the following section
for a detailed overview). These include tensor products (Smolensky, 1990), recursive distributed
 VECTOR-BASED MODELS OF SEMANTIC COMPOSITION 5

representations (RAAMS, Pollack, 1990), spatter codes (Kanerva, 1988), holographic reduced rep-
resentations (Plate, 1995), and convolution (Metcalfe, 1990).
In this article, we attempt to bridge the gap in the literature by developing models of semantic
composition that can represent the meaning of word combinations as opposed to individual words.
Our models are narrower in scope compared to those developed in earlier connectionist work. Our
vectors represent words, they are high-dimensional but relatively structured, every component cor-
responds to a predefined context in which the words are found. We take it as a defining property
of the vectors we consider that the values of their components are derived from event frequencies
such as the number of times a given word appears in a given context.1 Having this in mind, we
present a general framework for vector-based composition which allows us to consider different
classes of models. Specifically, we formulate composition as a function of two vectors and intro-
duce models based on addition and multiplication. The similarity between two complex expressions
can be thus naturally expressed using a geometric measure such as cosine or Euclidean distance. We
also investigate how the choice of the underlying semantic representation interacts with the choice
of composition function by comparing a spatial model that represents words as vectors in a high-
dimensional space against a probabilistic model that represents words as topic distributions. We
assess the performance of these models directly on a similarity task. We elicit similarity ratings for
pairs of adjective-noun, noun-noun and verb-object constructions and examine the strength of the
relationship between similarity ratings and the predictions of our models.
In the remainder, we review previous research on semantic composition and vector bind-
ing models. Next, we describe our modeling framework, present our elicitation experiments, and
discuss our results.

Composition
Compositionality allows languages to construct complex meanings from combinations of
simpler elements. This property is often captured in the following principle: the meaning of a whole
is a function of the meaning of the parts (Partee, 1995, p. 313). Therefore, whatever approach we
take to modeling semantics, representing the meanings of complex structures will involve modeling
the way in which meanings combine. Let us express the composition of two constituents, u and v,
in terms of a function acting on those constituents:
p = f (u, v) (1)
Partee (1995, p. 313) suggests a further refinement of the above principle taking the role of syntax
into account: the meaning of a whole is a function of the meaning of the parts and of the way they
are syntactically combined. We thus modify the composition function in (1) to account for the fact
that there is a syntactic relation R between constituents u and v:
p = f (u, v, R) (2)
Unfortunately, even this formulation may not be fully adequate. Lakoff (1977, p. 239), for
example, suggests that the meaning of the whole is greater than the meaning of the parts. The
1 For example, vectors offer a convenient representation for encoding features in machine learning, however the values

of these vectors are not always derived from event frequencies. Graphs are also often represented by an adjacency matrix,
a matrix with rows and columns labeled by graph vertices v, with a 1 or 0 in position (vi , v j ) according to whether vi
and v j are adjacent or not. This does not imply that an adjacency matrix is a vector-based model, as the values of the
elements in the matrix do not correspond to event frequencies.
 VECTOR-BASED MODELS OF SEMANTIC COMPOSITION 6

implication here is that language users are bringing more to the problem of constructing complex
meanings than simply the meaning of the parts and their syntactic relations. This additional in-
formation includes both knowledge about the language itself and also knowledge about the real
world. Thus, full understanding of the compositional process involves an account of how novel
interpretations are integrated with existing knowledge. Again, the composition function needs to
be augmented to include an additional argument, K, representing any knowledge utilized by the
compositional process:
p = f (u, v, R, K) (3)
The difficulty of defining compositionality is highlighted by Frege (1884, p. x) himself who
cautions never to ask for the meaning of a word in isolation but only in the context of a statement. In
other words, it seems that the meaning of the whole is constructed from its parts, and the meaning
of the parts is derived from the whole. Moreover, compositionality is a matter of degree rather than
a binary notion. Linguistic structures range from fully compositional (e.g., black hair), to partly
compositional syntactically fixed expressions, (e.g., take advantage), in which the constituents can
still be assigned separate meanings, and non-compositional idioms (e.g., kick the bucket) or multi-
word expressions (e.g., by and large), whose meaning cannot be distributed across their constituents
(Nunberg, Sag, & Wasow, 1994).
Despite the foundational nature of compositionality to language, there are significant ob-
stacles to understanding what exactly it is and how it operates. Most significantly, there is the
fundamental difficulty of specifying what sort of function of the meanings of the parts is involved
in semantic composition (Partee, 2004, p. 153). Fodor and Pylyshyn (1988) attempt to characterize
this function by appealing to the notion of systematicity. They argue that the ability to understand
some sentences is intrinsically connected to the ability to understand certain others. For example,
no-one who understands John loves Mary fails to understand Mary loves John. Therefore, the se-
mantic content of a sentence is systematically related to the content of its constituents and the ability
to recombine these according to a set of rules. In other words, if one understands some sentence and
the rules that govern its construction, one can understand a different sentence made up of the same
elements according to the same set of rules. In a related proposal, Holyoak and Hummel (2000)
claim that in combining parts to form a whole, the parts remain independent and maintain their
identities. This entails that John has the same independent meaning in both John loves the girl and
The boy hates John.
Aside from the philosophical difficulties of precisely determining what systematicity means
in practice (Pullum & Scholz, 2007; Spenader & Blutner, 2007; Doumas & Hummel, 2005), it is
worth noting that semantic transparency, the idea that words have meanings which remain unaf-
fected by their context, contradicts Freges (1884) claim that words only have definite meanings
in context. Consider for example the adjective good whose meaning is modified by the context in
which it occurs. The sentences John is a good neighbor and John is a lawyer do not imply John is a
good lawyer. In fact, we might expect that some of the attributes of a good lawyer are incompatible
with being a good neighbor, such as nit-picking over details, or not giving an inch unless required
by law. More generally, the claims of Fodor and Pylyshyn (1988) and Holyoak and Hummel (2000)
arise from a preconception of cognition as being essentially symbolic in character. While it is true
that the concatenation of any two symbols (e.g., g and l), will compose into an expression (e.g., gl),
within which both symbols maintain their identities, we cannot always assume that the the meaning
of a phrase is derived by simply concatenating the meaning of its constituents. Although the phrase
good lawyer is constructed by concatenating the symbols good and lawyer, the meaning of good
 VECTOR-BASED MODELS OF SEMANTIC COMPOSITION 7

will vary depending on the nouns it modifies.

Interestingly, Pinker (1994, p. 84) discusses the types of functions that are not involved in
semantic composition while comparing languages, which he describes as discrete combinatorial
systems, against blending systems. He argues that languages construct an unlimited number of
completely distinct combinations with an infinite range of properties. This is made possible by
creating novel, complex meanings which go beyond those of the individual elements. In contrast,
for a blending system the properties of the combination lie between the properties of its elements,
which are lost in the average or mixture. To give a concrete example, a brown cow does not identify
a concept intermediate between brown and cow (Kako, 1999, p. 2). Thus, composition based on
averaging or blending would produce greater generality rather than greater specificity.

Logic-based View Within symbolic logic, compositionality is accounted for elegantly by assum-
ing a tight correspondence between syntactic expressions and semantic form (Montague, 1974;
Blackburn & Bos, 2005). In this tradition, the meaning of a phrase or sentence is its truth-conditions
which are expressed in terms of truth relative to a model.2 In classical Montague grammar, for each
syntactic category there is a uniform semantic type (e.g., sentences express propositions, nouns
and adjectives express properties of entities, and verbs express properties of events). Most lexical
meanings are left unanalyzed and treated as primitive. In this framework, the proper noun John is
represented by the logical symbol JOHN denoting a specific entity, whereas a verb like wrote, is rep-
resented by a function from entities to propositions, expressed in lambda calculus as x.WROTE(x).
Applying this function to the entity JOHN yields the logical formula WROTE(JOHN) as a repre-
sentation of the sentence John wrote. It is worth noting that the entity and predicate within this
formula are represented symbolically, and that the connection between a symbol and its meaning is
an arbitrary matter of convention.
On one hand, the symbolic nature of logical representations is advantageous as it allows com-
position to be carried out syntactically. The laws of deductive logic in particular can be defined as
syntactic processes which act irrespective of the meanings of the symbols involved. On the other
hand, abstracting away from the actual meanings may not be fully adequate for modeling semantic
composition. For example, adjective-noun phrases are represented in terms of predicate conjunc-
tion, e.g., male lawyer corresponds to MALE(x) LAWYER(x). This approach cannot, however,
handle the context sensitive adjectives discussed above. John is a good lawyer is not equivalent to
the conjunction of John is good and John is a lawyer. More generally, modeling semantic com-
position means modeling the way in which meanings combine, and this requires that words have
representations which are richer than single, arbitrary symbols.

Connectionism Connectionist models of cognition (see among others Elman et al., 1996; Rumel-
hart, McClelland, & the PDP Research Group, 1986) can be seen as a response to the limitations of
traditional symbolic models. The key premise here is that knowledge is represented not as discrete
symbols that enter into symbolic expressions, but as patterns of activation distributed over many
processing elements. These representations are distributed in the sense that any single concept is
2 The structure common to all of the models in which a given language is interpreted reflects certain basic presuppo-
sitions about the structure of the world that are implicit in the language. In predicate logic, a model consists of the set
of truth-values {0,1}, a domain D which is some set of entities, and some n-ary relations on this set. The model also
consists of an interpretation function which assigns semantic values to all constants.
 VECTOR-BASED MODELS OF SEMANTIC COMPOSITION 8

v3 u1 v3 u2 v3 u3 v3

v v2 u1 v2 u2 v2 u3 v2

v1 u1 v1 u2 v1 u3 v1

u1 u2 u3

Figure 2. The tensor product of two three-dimensional vectors u and v.

represented as a pattern, i.e., vector, of activation over many elements (nodes or units) that are
typically assumed to correspond to neurons or small collections of neurons.
Much effort in the literature has been invested in enhancing the representational capabilities
of connectionist models with the means to combine a finite number of symbols into a much larger,
possibly infinite, number of specific structures. The key property of symbolic representations that
connectionist models attempt to emulate is their ability to bind one representation to another. The
fundamental operation underlying binding in symbolic systems is the concatenation of symbols ac-
cording to certain syntactic processes. And crucially the results of this operation can be broken
down into their original constituents. Thus, connectionists have sought ways of constructing com-
plex structures by binding one distributed representation to another in a manner that is reversible.
Smolensky (1990), for example, proposed the use of tensor products as a means of binding
one vector to another to produce structured representations. The tensor product u v is a matrix
whose components are all the possible products ui v j of the components of vectors u and v. Figure (1)
illustrates the tensor product for two three-dimensional vectors (u1 , u2 , u3 ) (v1 , v2 , v3 ). A major
difficulty with tensor products is their dimensionality which grows exponentially in size as more
constituents are composed (precisely, the tensor product has dimensionality m n).
To overcome this problem, other techniques have been proposed in which the binding of two
vectors results in a vector which has the same dimensionality as its components. Holographic re-
duced representations (Plate, 1991) are one implementation of this idea where the tensor product
is projected onto the space of the original vectors, thus avoiding any dimensionality increase. The
projection is defined in terms of circular convolution a mathematical function that compresses the
tensor product of two vectors. The compression is achieved by summing along the transdiagonal
elements of the tensor product. Noisy versions of the original vectors can be recovered by means
of circular correlation which is the approximate inverse of circular convolution. The success of cir-
cular correlation crucially depends on the components of the n-dimensional vectors u and v being
real numbers and randomly distributed with mean 0 and variance n1 . Binary spatter codes (Kanerva,
1988, 2009) are a particularly simple form of holographic reduced representation. Typically, these
vectors are random bit strings or binary N-vectors (e.g., N = 10, 000). Compositional representa-
tions are synthesized from parts or chunks. Chunks are combined by binding which is the same as
 VECTOR-BASED MODELS OF SEMANTIC COMPOSITION 9

taking the exclusive or (XOR) of two vectors. Here, only the transdiagonal elements of the tensor
product of two vectors are kept, and the rest are discarded.
From a computational perspective, both spatter codes and holographic reduced representa-
tions can be implemented efficiently3 and the dimensionality of the resulting vector does not change.
The downside is that operations like circular convolution are a form of lossy compression that intro-
duces noise into the representation. To retrieve the original vectors from their bindings a clean-up
memory process is usually employed where the noisy vector is compared to all component vectors
in order to find the closest one.
Tensors and their relatives can indeed represent relations (e.g., love(x,y) ) and role-filler bind-
ings (e.g., in loves(John, Mary) the lover role is bound to John and the beloved role is bound
to Mary) in a distributed fashion. However, Holyoak and Hummel (2000) claim that this form
of binding violates role-filler independence. In a truly compositional system, complex structures
gain meaning from the simpler parts from which they are formed and the simpler components re-
main independent, i.e., preserve their meaning (Doumas & Hummel, 2005; Doumas, Hummel, &
Sandhofer, 2008). Doumas and Hummel (2005) propose a model of role-filler binding based on
synchrony of neural firing. Vectors representing relational roles fire in synchrony with vectors rep-
resenting their fillers and out of synchrony with other role-filler bindings. These ideas are best
captured in LISA, a neural network that implements symbolic structures in terms of distributed rep-
resentations. Crucially, words and relations are represented by features (e.g., human, adult, male)
which albeit more informative than binary vectors, raise issue regarding their provenance and the
scalability of the models based on them (see the discussion in the Introduction).

Semantic Spaces The idea of representing word meaning in a geometrical space dates back to
Osgood, Suci, and Tannenbaum (1957), who used elicited similarity judgments to construct seman-
tic spaces. Subjects rated concepts on a series of scales whose endpoints represented polar oppo-
sites (e.g., happysad ); these ratings were further processed with factor analysis, a dimensionality
reduction technique, to uncover latent semantic structure. In this study, meaning representations
were derived directly from psychological data, thereby allowing the analysis of differences across
subjects. Unfortunately, multiple subject ratings are required to create a representation for each
word, which in practice limits the semantic space to a small number of words.
Building on this work and the well-known vector space model in information re-
trieval (Salton, Wong, & Yang, 1975; Deerwester, Dumais, Landauer, Furnas, & Harshman, 1990),
more recent semantic space models, such as LSA (Landauer & Dumais, 1997) and HAL lund-
burgess1996, overcome this limitation by constructing semantic representations indirectly from real
language corpora. A variety of such models have been proposed and evaluated in the literature.
Despite their differences, they are all based on the same premise: words occurring within similar
contexts are semantically similar (Harris, 1968). Semantic space models extract from a corpus a set
of counts representing the occurrences of a target word t in the specific context c of choice and then
map these counts into the components of a vector in some space. For example, Bullinaria and Levy
(2007) consider a range of component types, the simplest being to transform the raw frequencies
into conditional probabilities, p(ci |t). They also consider components based on functions of these
probabilities, such as the ratio of the conditional probability of the context to its overall probabil-
ity, or the pointwise mutual information between context and target. An issue here concerns the
3 Binary spatter codes have a runtime complexity of O(N) for a vector of length N and holographic reduced represen-

tations can be implemented using the Fast Fourier Transform, which is O(N log N).
 VECTOR-BASED MODELS OF SEMANTIC COMPOSITION 10

number of components the vectors should have, or which contexts should be used in constructing
the vectors. Often, the most frequent contexts are used, as rarer contexts yield unreliable counts.
Dimensionality reduction techniques can be also used to project high dimensional vectors onto a
lower dimensional space (Landauer & Dumais, 1997; Hofmann, 2001; Blei et al., 2003).
Semantic space models resemble the representations used in the connectionist literature.
Words are represented as vectors and their meaning is distributed across many dimensions. Cru-
cially, the vector components are neither binary nor randomly distributed. They correspond to
co-occurrence counts and it is assumed that differences in meaning arise from differences in the
distribution of these counts across contexts. That is not to say that high-dimensional randomly dis-
tributed representations are incompatible with semantic spaces. Kanerva, Kristoferson, and Holst
(2000) propose the use of random indexing as an alternative to the computationally costly singular
value decomposition employed in LSA. The procedure also builds a word-document co-occurrence
matrix, except that each document no longer has its own column. Instead, it is assigned a small
number of columns at random (the documents random index). So, each time a word occurs in the
document, the documents random index vector is added to the row corresponding to that word.
Random vectors have also been employed in an attempt to address a commonly raised criti-
cism against semantic space models, namely that they are inherently agnostic to the linguistic struc-
ture of the contexts in which a target word occurs. In other words, most of these models treat these
contexts as a structureless bag-of-words. Jones and Mewhort (2007) propose a model that makes
use of the linear order of words in a context. Their model represents words by high-dimensional
holographic vectors. Each word is is assigned a random4 environmental vector. Contextual infor-
mation is stored in a lexical vector which is computed with the aid of the environmental vectors.
Specifically, a words lexical vector is the superposition of the environmental vectors corresponding
to its co-occurring words in a sentence. Order information is the sum of all n-grams that include the
target word. The n-grams are encoded with the aid of a place-holder environmental vector and
circular convolution (Plate, 1995). The order vector is finally added to the lexical vector in order to
jointly represent structural and contextual information. Despite the fact that these vectors contain
information about multi-word structures in the contexts of target words, they are, nonetheless, still
fundamentally representations of individual isolated target words. Circular convolution is only used
to bind environmental vectors, which being random contain no semantic information. To make a
useful semantic representation of a target word, the vectors representing its contexts are summed
over, producing a vector which is no longer random and for which circular convolution is no longer
optimal.
Sahlgren, Host, and Kanerva (2008) provide and alternative to convolution by showing that
order information can also be captured by permuting the vector coordinates. Other models im-
plement more sophisticated versions of context that go beyond the bag-of-words model, without
however resorting to random vectors. For example, by defining context in terms of syntactic depen-
dencies (Grefenstette, 1994; Lin, 1998; Pado & Lapata, 2007) or by taking into account relational
information about how roles and fillers combine to create specific factual knowledge (Dennis, 2007).
So far the discussion has centered on the creation of semantic representations for individual
words. As mentioned earlier, the composition of vector-based semantic representations has received
relatively little attention. An alternative is not to compose at all but rather create semantic represen-
tations for phrases in addition to words. If a phrase is frequent enough, then it can be treated as a

4 Vector components are sampled at random from a Gaussian distribution with = 0 and = 1 where D = 2, 048.
D
 VECTOR-BASED MODELS OF SEMANTIC COMPOSITION 11

single target unit, and its occurrence across a range of contexts can be constructed in the same man-
ner as described above. Baldwin, Bannard, Tanaka, and Widdows (2003) apply this method to model
the decomposability of multi-word expressions such as noun compounds and phrasal verbs. Tak-
ing a similar approach, Bannard, Baldwin, and Lascarides (2003) develop a vector space model for
representing the meaning of verb-particle constructions. In the limit, such an approach is unlikely
to work as semantic representations for constructions that go beyond two-words will be extremely
sparse.
Vector addition or averaging (which are equivalent under the cosine measure) is the most
common form of vector combination (Landauer & Dumais, 1997; Foltz et al., 1998) However, vec-
tor addition is not a suitable model of composition for at least two reasons. Firstly, it is insensitive
to syntax and word order. Because vector addition is commutative, it assigns the same represen-
tation to any sentence containing the same constituents irrespective of their syntactic relations. It
is therefore a bag-of-words model of composition. In contrast, there is ample empirical evidence
that syntactic relations across and within sentences are crucial for sentence and discourse process-
ing (Neville, Nichol, Barss, Forster, & Garrett, 1991; West & Stanovich, 1986). Secondly, addition
simply blends together the content of all words involved to produce something in between them all.
Ideally, we would like a model of semantic composition that generates novel meanings by select-
ing and modifying particular aspects of the constituents participating in the composition. Kintsch
(2001) attempts to achieve this in his predication algorithm by modeling how the meaning of a pred-
icate (e.g., run) varies depending on the arguments it operates upon (e.g, the horse ran vs. the color
ran). The idea is to add not only the vectors representing the predicate and its argument but also the
neighbors associated with both of them. The neighbors, Kintsch argues, can strengthen features of
the predicate that are appropriate for the argument of the predication.
Tensor products have been recently proposed as an alternative to vector addition. (Aerts &
Czachor, 2004; Clark & Pulman, 2007; Widdows, 2008). However, as illustrated in Figure (1),
these representations grow exponentially as more vectors are combined. This fact undermines not
only their tractability in an artificial computational setting but also their plausibility as models of
human concept combination. Interestingly, Clark, Coecke, and Sadrzadeh (2008) try to construct a
tensor product based model of vector composition which makes an explicit connection to models
of linguistic composition. In particular, they show how vector-based semantics can be unified with
a compositional theory of grammatical types. Central to their approach is the association of each
grammatical type with a particular rank of tensor. So, for example, if we take nouns as being asso-
ciated with simple vectors, then an adjective as a noun modifier would be associated with a matrix,
i.e. a vector transformation. Clark et al. (2008) do not suggest concrete methods for constructing
or estimating the various tensors involved in their model. Instead, they are more interested in its
formal properties and do not report any empirical tests of this approach.
Unfortunately, comparisons across vector composition models have been few and far be-
tween. The merits of different approaches are illustrated with special purpose examples and large
scale evaluations are uniformly absent. For instance, Kintsch (2001) demonstrates how his own
composition algorithm works intuitively on a few hand selected examples but does not provide a
comprehensive test set (see Frank, Koppen, Noordman, & Vonk, 2007 for a criticism of Kintschs
2001 evaluation standards). In a similar vein, Widdows (2008) explores the potential of vector prod-
uct operations for modeling compositional phenomena in natural language, again on a small number
of hand picked examples.
Our work goes beyond these isolated proposals; we present a framework for vector com-
 VECTOR-BASED MODELS OF SEMANTIC COMPOSITION 12

music solution economy craft reasonable

practical 0 6 2 10 4
difficulty 1 8 4 4 0
Figure 3. A hypothetical semantic space for practical and difficulty.

position which allows us to explore a range of potential composition functions, their properties,
and relations. Under this framework, we reconceptualize existing composition models as well as
introduce novel ones. Our experiments make use of conventional semantic vectors built from co-
occurrence data. However, our compositional models are not tied to a specific representation and
could be used with the holographic vectors proposed in Jones and Mewhort (2007) or with random
indexing, however we leave this to future work. Within the general framework of co-occurrence-
based models we investigate how the choice of semantic representation interacts with the choice of
composition model. Specifically, we compare a spatial model that represents words as vectors in a
high-dimensional space against a probabilistic model (akin to LSA) that represents words as topic
distributions. We evaluate these models empirically on a phrase similarity task, using a rigorous
evaluation methodology.

Composition Models
Our aim is to construct vector representations for phrases and sentences. We assume that
constituents are represented by vectors which subsequently combine in some way to produce a
new vector. It is worth emphasizing that the problem of combining semantic vectors to make a
representation of a multi-word phrase, is different to the problem of how to incorporate information
about multi-word contexts into a distributional representation for a single target word. Whereas
Jones and Mewhort (2007) test this ability to memorize the linear structure of contexts in terms
of predicting a target word correctly given a context, our composition models will be evaluated in
terms of their ability to model semantic properties of simple phrases.
In this study we focus on small phrases, consisting of a head and a modifier or complement,
which form the building blocks of larger units. If we cannot model the composition of basic phrases,
there is little hope that we can construct compositional representations for sentences or even doc-
uments (we return to this issue in our Discussion section). So, given a phrase such as practical
difficulty and the vectors u and v representing the constituents practical and difficulty, respectively,
we wish to produce a representation p of the whole phrase. Hypothetical vectors for these con-
stituents are illustrated in Figure 3. This simplified semantic space5 will serve to illustrate examples
of the composition functions we consider in this paper.
In our earlier discussion, we defined p, the composition of vectors u and v, representing a
pair of words which stand in some syntactic relation R, given some background knowledge K as:

p = f (u, v, R, K) (4)

The expression above defines a wide class of composition functions. To derive specific models from
this general framework requires the identification of appropriate constraints that narrow the space of
5 The space has only five dimensions; the matrix cells denote the co-occurrence of the words practical and difficulty

with music, solution, and so on. We also experiment with an alternative semantic representation denoting the distribution
of words over topics. We refer the reader to our modeling experiments for details.
 VECTOR-BASED MODELS OF SEMANTIC COMPOSITION 13

functions being considered. To begin with, we will ignore K so as to explore what can be achieved
in the absence of any background or world knowledge. While background knowledge undoubtedly
contributes to the compositional process, and resources like WordNet (Fellbaum, 1998) may be
used to provide this information, from a methodological perspective it is preferable to understand
the fundamental processes of how representations are composed before trying to understand the
interaction between existing representations and those under construction. As far as the syntactic
relation R is concerned, we can proceed by investigating one such relation at a time, thus removing
any explicit dependence on R, but allowing the possibility that we identify distinct composition
functions for distinct syntactic relations.
Another particularly useful constraint is to assume that p lies in the same space as u and v.
This essentially means that all syntactic types have the same dimensionality. The simplification may
be too restrictive as it assumes that verbs, nouns and adjectives are substantially similar enough to
be represented in the same space. Clark et al. (2008) suggest a scheme in which the structure of
a representation depends on its syntactic type, such that, for example, if nouns are represented by
plain vectors then adjectives, as modifiers of nouns, are represented by matrices. More generally, we
may question whether representations in a fixed space are flexible enough to cover the full expres-
sivity of language. Intuitively, sentences are more complex than individual phrases and this should
be reflected in the representation of their meaning. In restricting all representations within a space
of fixed dimensions, we are implicitly imposing a limit on the complexity of structures which can
be fully represented. Nevertheless, the restriction renders the composition problem computationally
feasible. We can use a single method for constructing representations, rather than different methods
for different syntactic types. In particular, constructing a vector of n elements is easier than con-
structing a matrix of n2 elements. Moreover, our composition and similarity functions only have to
apply to a single space, rather than a set of spaces of varying dimensions.
Given these simplifying assumptions, we can now begin to identify specific mathematical
types of functions. For example, if we wish to work with linear composition functions, there are
two ways to achieve this. We may assume that p is a linear function of the Cartesian product of u
and v, giving an additive class of composition functions:

p = Au + Bv (5)

where A and B are matrices which determine the contributions made by u and v to p.
Or, we can assume that p is a linear function of the tensor product of u and v, giving a
multiplicative class of composition functions:

p = Cuv (6)

where C is a tensor of rank 3, which projects the tensor product of u and v onto the space of p. (For
readers unfamiliar with vector and tensor algebra we provide greater detail in Appendix C).
Linearity is very often a useful assumption because it constrains the problem considerably.
However, this usually means that the solution arrived at is an approximation to some other, non-
linear, structure. Going beyond the linear class of multiplicative functions, we will also consider
some functions which are quadratic in u, having the general form:

p = Duuv (7)

where D is now a rank 4 tensor which projects the product uuv onto the space of p.
 VECTOR-BASED MODELS OF SEMANTIC COMPOSITION 14

Within the additive model class (equation (5)), the simplest composition function is vector
addition:
p = u+v (8)
So, according to equation (8), the addition of the two vectors representing practical and difficulty
(see Figure 3) would be practical + difficulty = [1 14 6 14 4]. This model assumes that composi-
tion is a symmetric function of the constituents; in other words, the order of constituents essentially
makes no difference. While this might be reasonable for certain structures, a list perhaps, a model
of composition based on syntactic structure requires some way of differentiating the contributions
of each constituent.
Kintsch (2001) attempts to model the composition of a predicate with its argument in a man-
ner that distinguishes the role of these constituents, making use of the lexicon of semantic repre-
sentations to identify the features of each constituent relevant to their combination. Specifically, he
represents the composition in terms of a sum of predicate, argument and a number of neighbors of
the predicate.
p = u + v + ni (9)
i
Considerable latitude is allowed in selecting the appropriate neighbors. Kintsch (2001) consid-
ers only the m most similar neighbors to the predicate, from which he subsequently selects k,
those most similar to its argument. So, if in the composition of practical with difficulty, the
chosen neighbor is problem, with problem = [2 15 7 9 1], then this produces the representation
practical + difficulty + problem = [3 29 13 23 5].
This composition model draws inspiration from the construction-integration model (Kintsch,
1988), which was originally based on symbolic representations, and introduces a dependence on
syntax by distinguishing the predicate from its argument. In this process the selection of relevant
neighbors for the predicate plays a role similar to the integration of a representation with existing
background knowledge in the original construction-integration model. Here, background knowl-
edge takes the form of the lexicon from which the neighbors drawn.
A simpler approach to introducing dependence on the syntactic relation, R, is to weight the
constituents differentially in the summation.

p = v + u (10)

This makes the composition function asymmetric in u and v allowing their distinct syntactic
roles to be recognized. For instance, we could give greater emphasis to heads than other con-
stituents. As an example, if we set to 0.4 and to 0.6, then 0.4 practical = [0 2.4 0.8 4 1.6]
and 0.6 difficulty = [0.6 4.8 2.4 2.4 0], and practical difficulty is represented by their sum
0.4 practical + 0.6 difficulty = [0.6 5.6 3.2 6.4 1.6].
An extreme form of this differential in the contribution of constituents is where one of the
vectors, say u, contributes nothing at all to the combination:6

p=v (11)

In this case practical difficulty would be simply represented by difficulty = [1 8 4 4 0]. Admittedly
the model in (11) is impoverished and rather simplistic, however it can serve as a simple baseline
against which to compare more sophisticated models.
6 The model in (11) is equivalent to setting = 0.
 VECTOR-BASED MODELS OF SEMANTIC COMPOSITION 15

So far we have considered solely additive composition models. These models blend together
the content of the constituents being composed. The contribution of u in equation (8) is unaffected
by its relation to v. It might be preferable to scale each component of u with its relevance to v,
namely to pick out the content of each representation that is relevant to their combination. This can
be achieved by using a multiplicative function instead:

p = uv (12)

where the symbol  represents multiplication of the corresponding components:

pi = ui vi (13)

For this model, our example vectors would combine to give:

practical  difficulty = [0 48 8 40 0].
Note that the multiplicative function in (12) is still a symmetric function and thus does not
take word order or syntax into account. However, equation (12) is a particular instance of the more
general class of multiplicative functions (equation (6)), which allows the specification of asymmetric
syntax-sensitive functions. For example, the tensor product is an instance of this class with C being
the identity matrix.
p = uv (14)
Where the symbol stands for the operation of taking all pairwise products of the components of
u and v:
pi, j = ui v j (15)
So, the tensor product representation of practical difficulty is:

0 0 0 0 0
6 48 24 24 0
practical difficulty = 2 16 8 8 0 (16)
10 80 40 40 0
4 32 16 16 0

Circular convolution is also a member of this class:

p = u~v (17)

where the symbol ~ stands for a compression of the tensor product based on summing along its
transdiagonal elements:
pi = u j v(i j)modn (18)
j

Circular convolution compresses the matrix in (16) into the vector

practical ~ difficulty = [116 50 66 62 80].
One reason for choosing such multiplicative functions is that the magnitudes of u and v
can only affect the magnitude of p, not its direction. In contrast, in additive models, the relative
magnitudes of u and v, can have a considerable effect on both the magnitude and direction of p. This
can lead to difficulties when working with the cosine similarity measure, which is itself insensitive
to the magnitudes of vectors. For example, if vector definitions are optimized by comparing the
 VECTOR-BASED MODELS OF SEMANTIC COMPOSITION 16

y
v

u
x

Figure 4. Vector v is decomposed into x, a component parallel to u and y, a component orthogonal to u.

predictions from the cosine similarity measure to some gold standard, then it is the directions of
the vectors which are optimized, not their magnitudes. Utilizing vector addition as the composition
function makes the product of the composition dependent on an aspect of the vectors which has not
been optimized, namely their magnitude. Multiplicative combinations avoid this problem, because
effects of the magnitudes of the constituents only show up in the magnitude of the product, which
has no effect on the cosine similarity measure.
The multiplicative class of functions also allows us to think of one representation as mod-
ifying the other. This idea is fundamental in logic-based semantic frameworks (Montague, 1974)
where different syntactic structures are given different function types. To see how the vector u can
be thought of as something which modifies v, consider the partial product of C with u, producing a
matrix which we shall call U.
p = Cuv = Uv (19)

Here, the composition function can be thought of as the action of a matrix, U, representing one con-
stituent, on a vector, v, representing the other constituent. This is essentially Clark et al.s (2008)
approach to adjective-noun composition. In their scheme, nouns would be represented by vectors
and adjectives by matrices which map the original noun representation to the modified representa-
tion. In our approach all syntactic types are simply represented by vectors; nevertheless, we can
make use of their insight. Equation (19) demonstrates how a multiplicative composition tensor,
C, allows us to map a constituent vector, u, onto a matrix, U, while representing all words with
vectors.
Putting the simple multiplicative model (see equation (12)) into this form yields a matrix, U,
whose off-diagonal elements are zero and whose diagonal elements are equal to the components
of u.
Ui j = 0,Uii = ui (20)

The action of this matrix on v is a type of dilation, in that it stretches and squeezes v in various
directions. Specifically, v is scaled by a factor of ui along the ith basis.
One drawback of this process is that its results are dependent on the basis used. Ideally, we
would like to have a basis independent composition, i.e., one which is based solely on the geometry
 VECTOR-BASED MODELS OF SEMANTIC COMPOSITION 17

of u and v.7 One way to achieve basis independence is by dilating v along the direction of u,
rather than along the basis directions. We thus decompose v into a component parallel to u and
a component orthogonal to u, and then stretch the parallel component to modulate v to be more
like u. Figure 4 illustrates this decomposition of v where x is the parallel component and y is the
orthogonal component. These two vectors can be expressed in terms of u and v as follows:
uv
x= u (21)
uu
uv
y = vx = v u (22)
uu
Thus, if we dilate x by a factor , while leaving y unchanged, we produce a modified vector,
v0 , which has been stretched to emphasize the contribution of u:
uv uv uv
v0 = x + y = u+v = ( 1) u+v (23)
uu uu uu
However, since the cosine similarity function is insensitive to the magnitudes of vectors, we can
multiply this vector by any factor we like without essentially changing the model. In particular,
multiplying through by u u makes this expression easier to work with:

p = (u u)v + ( 1)(u v)u (24)

In order to apply this model to our example vectors, we must first calculate the dot products
practical practical = 156 and practical difficulty = 96. Then, assuming is 2, the result of the
composition is 96 difficulty + 156 practical = [96 1704 696 1944 624]. This is now an asym-
metric function of u and v, where v is stretched by a factor in the direction of u. However, it is
also a more complex type of function, being quadratic in u (equation (7)).
Again, we can think of the composition of u with v, for this function (equation (24)), in terms
of a matrix U which acts on v.
Ui, j = ( 1)ui u j (25)
Ui,i = ( uk uk ) + ( 1)ui ui (26)
k
Where i, j and k range over the dimensions of the vector space.
The matrix U has one eigenvalue which is larger by a factor of than all the other eigenvalues,
with the associated eigenvector being u. This corresponds to the fact that the action of this matrix
on v is a dilation which stretches v differentially in the direction of u. Intuitively, this seems like an
appropriate way to try to implement the idea that the action of combining two words can result in
specific semantic aspects becoming more salient.

Collecting Similarity Ratings for Phrases

Vector-based models aimed at representing the meaning of individual words are commonly
evaluated against human similarity judgments. For example, the benchmark dataset collected by
Rubenstein and Goodenough (1965) consists of 65 noun-pairs ranging from highly synonymous
(gem-jewel ) to semantically unrelated (noon-string). For each pair, subjects gave similarity ratings
7 This would allow, for example, the same composition function to be applied both to original vectors and to dimen-

sionality reduced versions, without worrying about how to match the bases of these two spaces.
 VECTOR-BASED MODELS OF SEMANTIC COMPOSITION 18

(on a scale of 0 to 4) whose average represents an estimate of the perceived similarity of the two
words. Analogously, to evaluate the different composition models introduced above, we first had
to elicit similarity judgments for phrases. Although such elicitation studies are less common in the
literature, there is evidence that humans can reliably judge whether any two phrases are similar.
For example, Lapata and Lascarides present an experiment where participants rate whether
adjective-noun combinations and their paraphrases have similar meanings, whereas other work
(Lapata & Lascarides, 2003; Li, McLean, Bandar, OShea, & Crockett, 2006; Mitchell & Lap-
ata, 2008) elicits similarity judgments for sentence pairs. In all cases, humans agree in their rat-
ings, although the agreement tends to be lower compared to ratings assigned to isolated word pairs.
Moreover, in computational linguistics, similarity judgments for phrases and sentences are rou-
tinely obtained as a means to evaluate the ability of an automatic system to generate paraphrases.
Specifically, paraphrase pairs are presented to judges who are asked to decide whether they are se-
mantically equivalent, i.e., whether they can be generally substituted for one another in the same
context without great information loss (Barzilay & Lee, 2003; Bannard & Callison-Burch, 2005).
Participants are usually asked to rate the paraphrase pairs using a nominal scale (e.g., definitely
similar, sometimes similar, never similar).
In our experiments, we collected similarity judgments for adjective-noun, noun-noun, and
verb-object combinations using a rating scale. Following previous work (Bullinaria & Levy, 2007;
Pado & Lapata, 2007; McDonald, 2000), we then used correlation analysis to examine the rela-
tionship between the human ratings and their corresponding vector-based similarity values. In this
section we describe our method for assembling the set of experimental materials and eliciting simi-
larity ratings for these stimuli.

Materials and Design

We evaluated the predictions of our composition models against similarity ratings which we
obtained for three types of phrases, adjective-nouns, noun-nouns, and verb-objects. These phrases
were selected from the British National Corpus (BNC), as we wanted to ensure they represented
genuine usage. We also chose relatively frequent phrases to avoid confounding effects from infre-
quent or unfamiliar constructions. Ideally, we would also like our phrase pairs to be representative
of the full variation in semantic similarity. A potential caveat here concerns the difficulty of the
task. For example, requiring fine grained similarity judgments would fail to yield ratings with ro-
bust inter-subject agreement, a prerequisite for reliably evaluating the models. On the other hand,
the relatively easy task of simply discriminating high similarity from low similarity items, will result
in all models achieving high scores and thus failing to clearly distinguish their performance.
Our approach was to collect pairs representative of three similarity bands (High, Medium,
and Low) applying a simple method that randomly samples and pairs phrases from a corpus
yields items that are mostly semantically unrelated, with only a few showing weak similarity. Ini-
tially, we extracted all adjective-noun, noun-noun, and verb-object combinations attested in the
corpus. The latter was parsed with RASP (Briscoe & Carroll, 2002), a broad coverage syntactic
analyzer. Our high similarity items were compiled from phrases occurring at least 100 times in
the BNC. For each grammatical construction, any two phrases were considered highly similar if
swapping their heads resulted into two new phrases which were also attested in the BNC at least
100 times. For example, practical difficultyeconomic problem is a candidate high similarity item,
because practical problem and economic difficulty are also high frequency phrases. Our hypoth-
esis was that the phrases resulting from this recombination process must exhibit some semantic
 VECTOR-BASED MODELS OF SEMANTIC COMPOSITION 19

overlap, especially if they appear often in the BNC. This procedure resulted in 11,476 candidate
adjective-noun, 366 noun-noun, and 1,004 verb-object pairs.
In order to reduce the set of items to a more manageable size and more importantly to guar-
antee that the phrases were indeed semantically similar, we resorted to WordNet (Fellbaum, 1998).
We used a well-known dictionary-based similarity measure, originally proposed by Lesk (1986), to
rank the candidate phrase pairs. According to this measure, the semantic relatedness of two words
is proportional to the extent of overlap of their dictionary definitions8 (glosses in WordNet). We
computed the similarity of two phrases, as the sum of the similarities of their constituents. The
36 highest ranking phrase pairs (for each grammatical structure) on this measure formed our high-
similarity items (e.g., vast amountlarge quantity, telephone numberphone call, start workbegin
career). These 36 phrase pairs (72 phrases in total) were subsequently recombined to produce the
items in the medium and low similarity bands. This was done in order to eliminate any confound-
ing effects relating to the vocabulary of the individual phrases. By choosing the same same set of
phrases to construct all three bands, differences between bands cannot be attributed to lexical choice
but instead to their actual similarity relations.
Specifically, the high similarity phrases were first randomly split into the three groups, and
then candidate items for the remaining bands were constructed by pairing phrases from each of these
groups. So, each phrase was used three times in our materials: once in a high similarity pair, once
in a medium pair and once in a low pair. For example, practical difficulty from the first group was
paired with effective way from the third group to produce the item practical difficultyeffective way.
The Lesk similarity for each of these pairs was calculated as above and the 36 highest ranking items
on this measure were selected, subject to the constraint that each phrase was only used once in each
group. This produced a set of Medium similarity items, which, while they scored reasonably highly
on the WordNet-based measure, did not have the recombination property described above (e.g., so-
cial activityeconomic condition, market leaderboard member, discuss issuepresent problem). A
further 36 items were selected from the same set of candidate items, though in this case by choosing
the lowest ranking items. This produced a set of Low similarity items (e.g., practical difficultycold
air, phone call state benefit, drink wateruse test). The entire list of experimental stimuli is given
in Appendix A.
Thus, in our experimental design, the subject ratings and model predictions were the depen-
dent variables, and the bands and groups acted as blocking factors with a 3 3 structure. For each
phrase type (i.e., adjective-noun, noun-noun, and verb-object) we collected 108 items, 12 for each
band by group cell. The selected verb-object pairs were converted into a simple sentence by adding
a subject and articles or pronouns where appropriate. All verbs were in the past tense. The sentential
subjects were familiar proper names (BNC corpus frequency > 30 per million) balanced for gender.
Procedure
The elicitation studies were conducted online using Webexp (Keller, Gunasekharan, Mayo,
& Corley, 2009), an interactive software package for administering web-based psychological exper-
iments. Subjects took part in an experimental session that lasted approximately 20 minutes. The
experiment was self-paced, and response times were recorded to allow the data to be screened for
anomalies. Subjects accessed the experiment using their web browser, which established an Internet
connection to the experimental server running WebExp 2.1
8 We used the implementation provided in the WordNet Similarity package (Pedersen, Patwardhan, & Michelizzi,

2004).
 VECTOR-BASED MODELS OF SEMANTIC COMPOSITION 20

Table 1: Descriptive statistics for similarity experiments (adjective-noun, noun-noun, and verb-object), by
subjects.
Adjective-Noun Noun-Noun Verb-Object
Mean SD SE Mean SD SE Mean SD SE
High 3.76 1.926 0.093 4.13 1.761 0.085 3.91 2.031 0.098
Medium 2.50 1.814 0.087 3.04 1.732 0.083 2.85 1.775 0.085
Low 1.99 1.353 0.065 2.80 1.529 0.074 2.38 1.525 0.073

Subjects were given instructions that explained the task and provided examples (our instruc-
tions for the adjective-noun similarity experiment are reproduced in Appendix B). They were asked
to judge the similarity of phrases using a seven point rating scale where a high number indicates
higher similarity. To familiarize subjects with the similarity rating task, the experiment consisted of
a practice phase (of five items), followed by the experimental phase. In both phases, the participants
saw one phrase pair at a time and rated its similarity by clicking on one of seven buttons displaying
the numbers 1 to 7. The set of practice and experimental items was presented in random order.

Subjects
The experiment was completed by unpaid volunteers, all self-reported native speakers of En-
glish. Subjects were recruited by postings to local email lists; they had to be linguistically naive,
neither linguists nor students of linguistics were allowed to participate. The adjective-noun exper-
iment was completed by 88 participants; 69 subjects took part in the noun-noun experiment and
91 in the verb-object experiment. 14 participants were eliminated because they were non-native
English speakers. The data of 30 subjects was excluded after inspection of their responses revealed
anomalies in their ratings. For example, they were pressing buttons randomly, alternately, or rated
all phrase pairs uniformly. This left 204 subjects for analysis, 72 for the adjective-noun, 56 for
the noun-noun, and 76 for the verb-object experiment. 35 participants were male and 73 female,
94 were right-handed, and 14 left-handed. The subject ages ranged from 17 to 66, the mean was 31.
Participants were randomly allocated to a development set, used for optimizing model parameters,
and a test set on which the final evaluation of all models was carried out. For each experiment the
test set contained 36 participants, and the development set contained 18.

Results
We first performed a series of Kruskal-Wallis rank sum tests to examine the relationship
between our similarity bands and the elicited similarity ratings. Within each experiment, the sub-
ject ratings were significantly different (p < 0.01) across all bands, and also between each pair of
bands. Furthermore, the statistics in Table 1 demonstrate that the mean ratings show the correct
ordering (High > Medium > Low) and that there is substantial overlap between each band. These
results confirm that our procedure for generating the materials produced items with a wide range of
similarities.
We further examined how well the participants agreed in their similarity judgments for
adjective-noun, noun-noun, and verb-object combinations. Inter-subject agreement gives an upper
bound for the task and allows us to interpret how well our models are doing in relation to humans.
 VECTOR-BASED MODELS OF SEMANTIC COMPOSITION 21

To calculate inter-subject agreement we used leave one-out resampling. The technique is a spe-
cial case of n-fold cross-validation (Weiss & Kulikowski, 1991) and has been previously used for
measuring how well humans agree on judging semantic similarity (Resnik & Diab, 2000; Resnik,
1999). For each subject group we divided the set of the subjects responses with size m into a set
of size m 1 (i.e., the response data of all but one subject) and a set of size one (i.e., the response
data of a single subject). We then correlated the ratings of the former set with the ratings of the
latter using Spearmans correlation coefficient. This was repeated m times. For the adjective-noun
experiment, the average inter-subject agreement was .52 (Min = 0.35, Max = 0.73, SD = 0.12),
for the noun-noun experiment .51 (Min = 0.36, Max = 0.58, SD = 0.06), and for the verb-object
experiment 0.55 (Min = 0.45, Max = 0.65, SD = 0.06). These results indicate that the participants
found the similarity rating task relatively difficult, though still produced ratings with a reasonable
level of consistency.

Modeling Experiments
Semantic Representation
Irrespectively of their form, all composition models discussed here rely on vector-based rep-
resentations for individual words. Our experiments examined two such representations and their
impact on composition. Our first model is a simple and popular (McDonald, 2000; Bullinaria &
Levy, 2007; Lowe, 2000) semantic space that associates each vector component with a particu-
lar context word, and assigns it a value based on the strength of its co-occurrence with the target
(i.e., the word for which a semantic representation is being constructed). This model has the benefits
of simplicity and also of being largely free of any additional theoretical assumptions over and above
the distributional approach to semantics.
For our experiments, we built the semantic space on a lemmatized version of the BNC. Fol-
lowing previous work (Bullinaria & Levy, 2007), we optimized its parameters on a word-based
semantic similarity task. The task involves examining the degree of correlation between the hu-
man judgments for two individual words and vector-based similarity values. We experimented with
a variety of dimensions (ranging from 50 to 500,000), vector component definitions (e.g., point-
wise mutual information or log likelihood ratio) and similarity measures (e.g., cosine or confusion
probability). We used WordSim353, a benchmark dataset (Finkelstein et al., 2002), consisting of
relatedness judgments (on a scale of 0 to 10) for 353 word pairs.
We obtained best results with a model using a context window of five words on either side
of the target word and 2,000 vector components. The latter were the most common context words
(excluding a list of stop words). These components were set to the ratio of the probability of the
context word given the target word to the probability of the context word overall:

p(ci |t) f reqci ,t f reqtotal

vi (t) = = (27)
p(ci ) f reqt f reqci

Where f reqci ,t , f reqtotal , f reqt and f reqci are the frequencies of the context word ci with the target
word t, the total count of all word tokens, the frequency of the context word ci and the frequency of
the target word t, respectively.
This configuration gave high correlations with the WordSim353 similarity judgments using
the cosine measure ( = 0.42). In addition, Bullinaria and Levy (2007) found that these parameters
perform well on a number of other tasks such as the synonymy task from the Test of English as
 VECTOR-BASED MODELS OF SEMANTIC COMPOSITION 22

a Foreign Language (TOEFL). We compute the similarity between two vectors v(t1 ) and v(t2 )
representing target words t1 and t2 , respectively as:

v(t1 ) v(t2 )
sim(t1 ,t2 ) = cos(v(t1 ), v(t2 )) = (28)
|v(t1 )||v(t2 )|

Probabilistic topic models offer an alternative to semantic spaces. Although several variants
have been proposed in the literature (e.g., Griffiths et al., 2007; Blei et al., 2003) they are all based
on the same fundamental idea: documents are mixtures of topics where a topic is a probability
distribution over words. And the content of a topic is expressed by the probabilities of the words
within that topic. A topic model is a generative model specifying a specific process of how to gen-
erate a document. Our experiments are based on the Latent Dirichlet Allocation (LDA, Blei et al.,
2003) topic model where the generative process for a document d is as follows. We first draw the
mixing proportion over topics d from a Dirichlet prior9 with parameters . Next, for each of the
Nd words wdn in document d, a topic zdn is first drawn from a multinomial distribution with param-
eters dn . The probability of a word token w taking on value i given that topic z = j is parametrized
using a matrix with bi j = p(w = i|z = j). Integrating out d s and zdn s, gives P(D|, ), the
probability of a corpus (or document collection):
!
ZM Nd
P(d |) P(zdn |d )P(wdn |zdn , ) dd (29)
d=1 n=1 zdn

The central computational problem in topic modeling is to obtain the posterior distribu-
tion P(, z|w, , ) of the hidden variables given a document w = (w1 , w2 , . . . , wN ). Although this
distribution is intractable in general, a variety of approximate inference algorithms have been pro-
posed in the literature including expectation maximization, variational expectation maximization,
expectation propagation, several forms of Markov chain Monte Carlo (MCMC), and variational
inference. Our model adopts the Gibbs sampling procedure discussed in Griffiths et al. (2007).
Under this model, constructing a semantic representation for a target word amounts to esti-
mating the topic proportions for that word. We therefore select the number of topics, K, and train
the LDA algorithm on a document collection to obtain the parameters, where represents the
probability of a word wi given a topic z j , p(wi |z j ) = i j . The meaning of wi is thus extracted from
and is a K-element vector, whose components correspond to the probability of wi given each topic
assumed to have generated the document collection. Figure 5 gives an example of the semantic
representations extracted by the LDA model. Similarity in this model can be also measured as the
cosine of the angle between the topic vectors representing any two words.
For our experiments, we trained an LDA model on the BNC corpus.10 We optimized the
models parameters in terms of correlation on the same WordSim353 dataset used for the simpler
semantic space model. We varied the number of topics from 10 to 1000 and obtained best results
9 The 1
Dirichlet distribution is a commonly used prior for multinomials P() = n a 1
where a1 . . . , an are
B(a1 ...,an ) i i
i=1
the parameters of the prior and the normalizing constant B(a1 . . . , an ) is the n-dimensional Beta function. One important
reason for the use of the Dirichlet prior in the case of multinomial parameters is its mathematical expedience. It is a
conjugate prior, i.e., of the same functional from as the likelihood function for the multinomial. This means that the prior
and the likelihood can easily combine according to Bayes law to specify the posterior distribution P(|c1 . . . , ck ) where
c1 . . . , ck are the counts for each outcome.
10 The implementation we used is available at http://gibbslda.sourceforge.net/.
 VECTOR-BASED MODELS OF SEMANTIC COMPOSITION 23

Figure 5. Semantic representations obtained from the LDA topic model

Topic 2
difficulty
Topic 1 Topic 2 Topic 3 Topic n
problem
practical 0.39 0.02 0.32 ...
situation
difficulty 0.03 0.44 0.11 ...
crisis
produce 0.06 0.17 0.09 ...
hardship
(a) Meaning representations obtained from the LDA model; rows
are target words, columns correspond to P(w|z), the probability of
a target word w given a topic z. (b) Topic 2 represented as a probability dis-
tribution over words (only the five words that
have the highest probability under this topic are
shown.

Table 2: Parameters for Kintschs composition model.

Semantic Space LDA
m k m k
Adjective-Noun 10 10 50 10
Noun-Noun 100 1 50 1
Verb-Object 50 10 100 1

with 100 topics ( = 0.48). The hyperparameters and were initialized to 0.1, and 0.01 respec-
tively.

Parameters for Composition Models

Using the semantic representations described above, our experiments assessed the perfor-
mance of the simple additive and multiplicative models (see equations (8) and (12), respectively),
Kintschs (2001) model, the tensor product (equation (14)) and circular convolution (equation (17)).
Note that Kintschs model has two free parameters, the m neighbors most similar to the head, and
the k of m neighbors closest to its dependent, which we optimized on the development set. Table 2
shows the best parameters for the semantic space and LDA topic models, respectively. In addition
to these models, we also considered two models based on the weighted sum of two values. These
were the weighted addition model (equation (10)) and the dilation model (equation (24)). We tuned
the parameters for these models on the development set. These include the weights and for
the additive models, the dilation factor for the dilation models and their direction11 . The optimal
parameters for these models are shown in Tables 3 and 4.
Finally, as a baseline, we considered two non-compositional models. The first simply uses the
vector representing the head of the phrase as the representation of the whole phrase (equation (11)),
and the second treats the phrase as a single target unit and thus extracts a vector representation for
the whole phrase. The latter baseline is only applicable to the standard semantic space. LDA derives
semantic representations for individual words rather than word combinations (although extensions
11 The dilation model can be applied in two ways, since the function is asymmetric in u and v. Either u can be used to

dilate v, or v can be used to dilate u.

 VECTOR-BASED MODELS OF SEMANTIC COMPOSITION 24

Table 3: Parameters for weighted addition composition models.

Semantic Space LDA

Adjective-Noun 0.88 0.12 0.65 0.35
Noun-Noun 0.32 0.68 0.34 0.66
Verb-Object 0.31 0.69 0.50 0.50

Table 4: Parameters for dilation models.

Semantic Space LDA
Direction Direction
Adjective-Noun 16.7 Adjective 2.2 Noun
Noun-Noun 8.3 Head Noun 7.1 Head Noun
Verb-Object 7.7 Verb 6.3 Verb

of the basic model have been proposed that take word order into account; see Wallach (2002) for an
example). Table 5 gives the details of the composition functions we evaluated, expressed in terms
of the vector components for each model.

Evaluation
We evaluated the proposed composition models via correlation analysis. Specifically, the
elicited similarity ratings were correlated with our models predictions using Spearmans corre-
lation coefficient.12 Given some composition function, f (, ), and two phrases a1 b1 and a2 b2 , we
applied f to the vectors u1 and v1 representing a1 and b1 , respectively, to produce a composite
representation, p1 . Analogously, vectors u2 and v2 yield p2 as a representation for a2 b2 . Under
this setup, we can calculate the similarity of two phrases by measuring their distance in semantic
space. A large number of such measures have been proposed in the literature (see Bullinaria and
Levy (2007) and Weeds (2003) for an overview). We opted for the widely used cosine measure
(see equation (28)) due to its simplicity and good performance in simulating word similarity ratings
(Bullinaria & Levy, 2007; McDonald, 2000; Griffiths et al., 2007).

Results
Table 6 shows the correlation of the subjects similarity ratings with the models predictions
when using a simple co-occurrence-based semantic space. All models are significantly correlated
with the human judgments (p < 0.01). The only exception is circular convolution when applied to
noun-noun combinations. Let us first consider, the simpler composition models based on vector ad-
dition (see Additive and Kintsch in the table). Within this class of models we observe that Kintschs
model fails to improve on the simple additive model and is significantly13 worse (p < 0.01) than the
12 We avoided correlating the model predictions with averaged participant judgments as this is inappropriate given the
ordinal nature of the scale of these judgments and also leads to a dependence between the number of participants and the
magnitude of the correlation coefficient.
13 We examined whether the correlations achieved differ significantly using a t-test (Cohen & Cohen, 1983).
 VECTOR-BASED MODELS OF SEMANTIC COMPOSITION 25

Table 5: Composition functions considered in our experiments.

Model Function
Additive pi = ui + vi
Kintsch pi = ui + vi + ni
Multiplicative pi = ui vi
Tensor Product pi, j = ui v j
Circular Convolution pi = j u j vi j
Weighted Additive pi = vi + ui
Dilation pi = vi j u j u j + ( 1)ui j u j v j
Head Only pi = vi
Target Unit pi = vi (t1t2 )

Table 6: Correlation coefficients of model predictions with subject similarity ratings (Spearmans ) using a
simple semantic space.
Model Adjective-Noun Noun-Noun Verb-Object
Additive 0.36 0.39 0.30
Kintsch 0.32 0.22 0.29
Mutliplicative 0.46 0.49 0.37
Tensor Product 0.41 0.36 0.33
Convolution 0.09 0.05 0.10
Weighted Additive 0.44 0.41 0.34
Dilation 0.44 0.41 0.38
Target Unit 0.43 0.34 0.29
Head Only 0.43 0.17 0.24
Humans 0.52 0.49 0.55

standard additive model for the noun compounds.

Within the class of multiplicative models (see Multiplicative, Tensor Product, and Circular
Convolution in Table 6), the simple multiplicative model significantly (p < 0.01) outperforms all
other models. Specifically, both tensor products and circular convolution are significantly worse
(p < 0.01). The multiplicative model is also significantly better than the Additive one (p < 0.01).
These results are observed across the board, with adjective-noun, noun-noun, and verb-object com-
binations. It is worth noting that circular convolution is the worst performing model. The tensor
product itself, from which circular convolution is derived, is significantly better (p < 0.01) in all
experiments. This indicates that the manner in which circular convolution projects the tensor prod-
uct down onto a lower dimensional space does not preserve any useful information the product may
have contained. In addition, the fact that the tensor product is significantly worse than the simple
multiplicative model indicates that the off diagonal elements of the product, which are discarded in
the simple multiplicative model, are probably not contributing much to the composition.
We next consider the Weighted Additive and Dilation models. Recall that these models are
parametrized; in dilation models the modifier dilates the head by a factor whereas the weighted
 VECTOR-BASED MODELS OF SEMANTIC COMPOSITION 26

Table 7: Correlation coefficients of model predictions with subject similarity ratings (Spearmans ) using
the LDA topic model.
Model Adjective-Noun Noun-Noun Verb-Object
Additive 0.37 0.45 0.40
Kintsch 0.30 0.28 0.33
Multiplicative 0.25 0.45 0.34
Tensor Product 0.39 0.43 0.33
Convolution 0.15 0.17 0.12
Weighted Additive 0.38 0.46 0.40
Dilation 0.38 0.45 0.41
Head Only 0.35 0.27 0.17
Humans 0.52 0.49 0.55

additive model weights the constituents in the summation differentially. As shown in Table 6 the two
models perform similarly. This is not entirely surprising, as both consist of a sum of the constituents
multiplied by scalar factors (see equations (10) and (24)). The performance of these models does
not differ significantly, except in the case of verb-object combinations where the dilation model
performs significantly better (p < 0.01). We conjecture that the dilation model is more accurate at
capturing selectional restrictions. This model also fares similarly to the multiplicative model. The
two models yield correlations that are not significantly different, except in the case of noun-noun
combinations, where the multiplicative model is better (p < 0.01).
The two non-compositional models, Target Unit and Head Only, perform worse than mul-
tiplicative composition, with this difference reaching significance (p < 0.01) for noun-noun and
verb-object combinations. In general, the target unit model performs better than the head only
model (it obtains significantly (p < 0.01) better correlations for noun-noun combinations). This
is not surprising, the target unit model may be non-compositional, but nevertheless represents the
semantics of the two words participating in the composition more faithfully, whereas the head only
model offers a more impoverished representation as it is based solely on the meaning of the head.
In sum, we find that the multiplicative, weighted additive and dilation models perform overall
best. The multiplicative model has a slight advantage as it has no parameters (other than the semantic
space representing the individual words), and is conceptually simpler than the other two models.
On the down side, it is does not take syntactic information into account, whereas the other two can
modulate the role of syntactic structure by tuning the appropriate weights. We should also note
that in all cases our compositional models fall behind the human upper bound (see the last row in
Table 6). The multiplicative model comes close when applied to noun-noun combinations.
We now turn our attention to the compositional models which employ the LDA topic model.
As can be seen in Table 7, Kintschs model remains worse than the simple additive model for all
constructions considered here (and the differences are statistically significant (p < 0.01)). Regard-
ing compositional models based on multiplication, we observe that tensor products and the simple
multiplicative model yield comparable performances for noun-noun and verb-object combinations.
They differ for adjective-nouns with the tensor product being significantly better (p < 0.01). Cir-
cular convolution remains the worst performing model. Not surprisingly, Weighted Additive and
Dilation models obtain almost identical performances. And they are not significantly different from
 VECTOR-BASED MODELS OF SEMANTIC COMPOSITION 27

the simple Additive model. The non-compositional model (Head Only) is significantly worse than
these models. Comparing the spatial and topic-based representations reveals that the multiplicative
composition model on the simple semantic space is significantly (p < 0.01) better than the dilation
model with LDA, except in the verb-object experiment, where there is no significant difference
between them.
In conclusion, we observe that dilation models perform consistently well across representa-
tions. This is not entirely unexpected as they are more flexible than other compositional models
due to their parametric nature. They can be tuned to model more faithfully specific syntactic con-
structions while being sensitive to the underlying semantic representation. Our results also indicate
that additive composition functions work best with the LDA topic model, whereas a multiplicative
composition function produced the most predictive similarity values with a simple semantic space.
We attribute the disparity in performance to the sparsity of the LDA representations. The simple
semantic space contains highly distributed representations, with the semantic content spread across
the great variety of contexts a target word occurs in. In contrast, topic models tend to produce
representations in which the vast majority of topics are inactive (i.e., zero) and when these top-
ics are multiplied by other topics, the result is zero. Thus, multiplicative combinations of sparse
representations tend to result in a loss of useful information.

Discussion
In this paper we presented a framework for vector-based semantic composition. We formu-
lated composition as a function of two vectors and introduced several models based on addition and
multiplication. These models were applied to vectors corresponding to distinct meaning represen-
tations: a simple semantic space based on word co-occurrences and a topic-based model built using
LDA. We compared the model predictions empirically on a phrase similarity task, using ratings
elicited from native speakers. Overall, we observe that dilation models perform consistently well
across semantic representations. A compositional model based on component-wise multiplication
performs best on the simple semantic space, whereas additive models are preferable with LDA.
Interestingly, we also find that the compositional approach to constructing representations outper-
forms a more direct non-compositional approach based on treating the phrases essentially as single
lexical units. This is not entirely surprising as our materials which were compiled so as to avoid a
high degree of lexicalization. Such an approach may be better suited to modeling non-compositional
structures that are lexicalized and frequently occurring (Baldwin et al., 2003; Bannard et al., 2003).
Despite this success, a significant weakness of many of the models considered here is their
insensitivity to syntax. The multiplicative model, in particular, is symmetric, and thus makes no
distinction between the constituents it combines. Yet, in spite of this, it is the strongest model for
the simple semantic space. And although the weighted addition and dilation models differentiate
between constituents, their dependence on syntax is rather limited, involving only a differential
weighting of the contribution of each constituent. Perhaps more importantly, none of the represen-
tations could be said to have any internal structure. Thus, they cannot be broken down into parts
which can be independently interpreted or operated upon. Symbolic representations, in contrast,
build complex structures by, for example, binding predicates to arguments. In fact, it is often ar-
gued that however composition is implemented it must exhibit certain features characteristic of this
symbolic binding (Fodor & Pylyshyn, 1988; Holyoak & Hummel, 2000).
Our results do not indicate that models which mimic symbolic binding (i.e., tensor products
and circular convolution) are better than those that dont (at least for the phrase similarity task and
 VECTOR-BASED MODELS OF SEMANTIC COMPOSITION 28

p q

u v w x

practical difficulties slowed progress

Figure 6. Example of composition operating over parse trees.

root practical difficulties slowed progress

u v w x

Figure 7. Example of composition operating over dependency structures.

the syntactic structures we examined). In particular, circular convolution is, across the board, the
worst performing model. One issue in the application of circular convolution is that it is designed
for use with random vectors, as opposed to the structured semantic vectors we assume here. A more
significant issue, however, concerns symbol binding in general which is somewhat distinct from
semantic composition. In modeling the composition of an adjective with a noun, it is not enough
to simply bind the representation of one to the representation of the other, we must instead model
the interaction between their meanings and their integration to form a whole. Circular convolution
is simply designed to allow a pair of vectors to be bound in a manner that allows the result to be
decomposed into its original constituents at a later time. This may well be adequate as a model for
syntactic operations on symbols, but, as our results show, it is not, by itself, enough to model the
process of semantic composition. Nevertheless, we anticipate further improvements to our vector-
based composition models will involve taking a more sophisticated approach to the structure of
representations, in particular with regard to predicate-argument structures. Our results also suggest
that assuming a single semantic representation may not be sufficient for all tasks. For instance, it is
not guaranteed that the same highly structured representations appropriate for deductive inference
will also provide a good model for semantic similarity. Semantics, covering such a wide range of
cognitive phenomena, might well be expected to involve multiple systems and processes, which
make use of quite distinct representations.
In this article, we have been concerned with modeling the similarity between simple phrases,
consisting of heads and their dependents. We have thus avoided the important question of how
vectors compose to create representations for larger phrases and sentences. It seems reasonable to
 VECTOR-BASED MODELS OF SEMANTIC COMPOSITION 29

assume that the composition process operates over syntactic representations such as binary parse
trees. A sentence will typically consist of several composition operations, each applied to a pair of
constituents u and v. Figure 5 depicts this composition process for the sentence practical difficulties
slowed progress. Initially, practical and difficulties are composed into p, and slowed and progress
into q. The final sentence representation, s, is the composition of the pair of phrase representations p
and q. Alternatively, composition may operate over dependency graphs representing words and their
relationship to syntactic modifiers using directed edges (see the example in Figure 6).
It is interesting then to consider which composition function would be best suited for repre-
senting sentences. For example, we could adopt different functions for different constructions. Our
experiments show that the simple multiplicative model performs best at modeling adjective-noun
and noun-noun combinations, whereas the dilation model is better for verb-object constructions.
Alternatively, we could adopt a single composition function that applies uniformly across all syn-
tactic relations. As discussed earlier, the simple multiplicative function is insensitive to syntax and
word order. The dilation model, however, remedies this. It is also based on a multiplicative com-
position function, but can take syntax into account by stretching one vector along the direction of
another one (see equation (24)).
Overall, we anticipate that more substantial correlations with human similarity judgments
can be achieved by implementing more sophisticated models from within the framework outlined
here. In particular, the general class of multiplicative models (see equation (6)) appears to be a
fruitful area to explore. Future directions include constraining the number of free parameters in
linguistically plausible ways and scaling to larger datasets. The applications of the framework
discussed here are many and varied. We intend to assess the potential of our composition models on
context sensitive semantic priming (Till, Mross, & Kintsch, 1988) and inductive inference (Heit &
Rubinstein, 1994). Another interesting application concerns sentence processing and the extent to
which the compositional models discussed here can explain reading times in eye-tracking corpora
(Pynte, New, & Kennedy, 2008; Demberg & Keller, 2008).

References
Aerts, D., & Czachor, M. (2004). Quantum aspects of semantic analysis and symbolic artificial intelligence.
Journal of Physics A-Mathematical and General, 37, L123L32.
Andrews, M., Vigliocco, G., & Vinson, D. (2009). Integrating experiential and distributional data to learn
semantic representations. Psychological Review(3), 463498.
Baldwin, T., Bannard, C., Tanaka, T., & Widdows, D. (2003). An empirical model of multiword expression
decomposability. In Proceedings of the ACL 2003 Workshop on Multiword Expressions (pp. 8996).
Morristown, NJ.
Bannard, C., Baldwin, T., & Lascarides, A. (2003). A statistical approach to the semantics of verb-particles.
In Proceedings of the ACL 2003 Workshop on Multiword Expressions (pp. 6572). Morristown, NJ.
Bannard, C., & Callison-Burch, C. (2005). Paraphrasing with bilingual parallel corpora. In Proceedings of
the 43rd annual meeting of the association for computational linguistics (pp. 597604). Ann Arbor.
Barzilay, R., & Lee, L. (2003). Learning to paraphrase: An unsupervised approach using multiple-sequence
alignment. In Proceedings of the human language technology conference and annual meeting of the
north american chapter of the association for computational linguistics (pp. 1623).
Blackburn, P., & Bos, J. (2005). Representation and inference for natural language: A first course in
computational semantics. Seattle, WA: Stanford: CSLI Press.
Blei, D. M., Ng, A. Y., & Jordan, M. I. (2003). Latent Dirichlet Allocation. Journal of Machine Learning
Research, 3, 9931022.
 VECTOR-BASED MODELS OF SEMANTIC COMPOSITION 30

Briscoe, E., & Carroll, J. (2002). Robust accurate statistical annotation of general text. In Proceedings of the
3rd International Conference on Language Resources and Evaluation (pp. 14991504). Las Palmas,
Canary Islands.
Bullinaria, J., & Levy, J. (2007). Extracting semantic representations from word co-occurrence statistics: A
computational study. Behavior Research Methods, 39, 510526.
Clark, S., Coecke, B., & Sadrzadeh, M. (2008). A compositional distributional model of meaning. In
Proceedings of the 2nd Symposium on Quantum Interaction (pp. 133140). Oxford, UK: College
Publications.
Clark, S., & Pulman, S. (2007). Combining symbolic and distributional models of meaning. In Proceedings
of the AAAI Spring Symposium on Quantum Interaction, Stanford, CA, 2007 (pp. 5255).
Cohen, J., & Cohen, P. (1983). Applied multiple regression/correlation analysis for the behavioral sciences.
NJ: Erlbaum: Hillsdale.
Collins, A. M., & Quillian, M. R. (1969). Retrieval time from semantic memory. Journal of Verbal Learning
and Verbal Behavior(8), 240-248.
Deerwester, S. C., Dumais, S. T., Landauer, T. K., Furnas, G. W., & Harshman, R. A. (1990). Indexing by
latent semantic analysis. Journal of the American Society of Information Science, 41(6), 391-407.
Deese, J. (1959). On the prediction of occurrence of particular verbal intrusions in immediate recall. Journal
of Experimental Psychology, 58, 1722.
Demberg, V., & Keller, F. (2008). Data from eye-tracking corpora as evidence for theories of syntactic
processing complexity. Cognition, 101(2), 193210.
Denhire, G., & Lemaire, B. (2004). A computational model of childrens semantic memory. In Proceedings
of the 26th Annual Meeting of the Cognitive Science Society (pp. 297302). Mahwah, NJ: Lawrence
Erlbaum Associates.
Dennis, S. (2007). Introducing word order in an lsa framework. In P. Press (Ed.), Latent semantic analysis: A
road to meaning (pp. 449464). Thomas K. Landauer and Danielle S. McNamara and Simon Dennis
and Walter Kintsch.
Doumas, L. A. A., & Hummel, J. E. (2005). Modeling human mental representations: What works and what
doesnt and why. In K. J. Holyoak & R. G. Morrison (Eds.), The cambridge handbook of thinking and
reasoning (pp. 7391). Cambridge, UK: Cambridge University Press.
Doumas, L. A. A., Hummel, J. E., & Sandhofer, C. M. (2008). A theory of the discovery and predication of
relational concepts. Psychological Review, 115, 143.
Duffy, S. A., Henderson, J. M., & Morris, R. K. (1989). Semantic facilitation of lexical access during sentence
processing. Journal of Experimental Psychology: Learning, Memory, and Cognition, 15, 791801.
Eliasmith, C., & Thagard, P. (2001). Integrating structure and meaning: A distributed model of analogical
mapping. Cognitive Science, 25(1), .
Elman, J., Bates, E., Johnson, M., Karmiloff-Smith, A., Parisi, D., & Plunkett, K. (1996). Rethinking
innateness: A connectionist perspective. Cambridge, MA: MIT Press/Bradford Books.
Estes, W. K. (1994). Classification and cognition. New York: Oxford University Press.
Fellbaum, C. (Ed.). (1998). Wordnet: An electronic lexical database (language, speech, and communication).
The MIT Press. Hardcover.
Finkelstein, L., Gabrilovich, E., Matias, Y., Rivlin, E., Solan, Z., Wolfman, G., et al. (2002). Placing search
in context: The concept revisited. ACM Transactions on Information Systems, 20(1), 116131.
Fodor, J., & Pylyshyn, Z. (1988). Connectionism and cognitive architecture: A critical analysis. Cognition,
28, 3-71.
Foltz, P. W., Kintsch, W., & Landauer, T. (1998). The measurement of textual coherence with latent semantic
analysis. Discourse Process, 15, 285307.
Foss, D. J. (1982). A discourse on semantic priming. Cognitive Psychology, 14, 590607.
Frank, S., Koppen, M., Noordman, L., & Vonk, W. (2007). World knowledge in computational models of
discourse comprehension. Discourse Processes. (In press)
Frege, G. (1884). Die Grundlagen der Arithmetik. Breslau: W. Koebner.
 VECTOR-BASED MODELS OF SEMANTIC COMPOSITION 31

Gentner, D. (1989). The mechanisms of analogical learning. In S. Vosniadou & A. Ortony (Eds.), Similarity
and analogical reasoning (pp. 199241). Cambridge: Cambridge University Press.
Grefenstette, G. (1994). Explorations in automatic thesaurus discovery. Norwell, MA, USA: Kluwer
Academic Publishers.
Griffiths, T. L., Steyvers, M., & Tenenbaum, J. B. (2007). Topics in semantic representation. Psychological
Review, 114(2), 211244.
Harris, Z. (1968). Mathematical structures of language. New York: Wiley.
Heit, E., & Rubinstein, J. (1994). Similarity and property effects in inductive reasoning. Journal of Experi-
mental Psychology: Learning, Memory, and Cognition, 20, 411422.
Hinton, J., & Shallice, T. (1991). Lesioning an attractor network: Investigations of acquired dyslexia.
Psychological Review, 98, 7495.
Hofmann, T. (2001). Unsupervised learning by probabilistic latent semantic analysis. Machine Learning,
42(1-2), 177196.
Holyoak, K. J., & Hummel, J. E. (2000). The proper treatment of symbols in a connectionist architecture. In
E. Dietrich & A. Markman (Eds.), Cognitive dynamics: Conceptual change in humans and machines.
(pp. 229264). Cambridge, MA: MIT Press.
Holyoak, K. J., & Koh, K. (1987). Surface and structural similarity in analogical transfer. Memory and
Cognition, 15, 332340.
Jones, M., & Mewhort, D. (2007). Representing word meaning and order information in a composite holo-
graphic lexicon. Psychological Review, 114(1), 137.
Kako, E. (1999). Elements of syntax in the systems of three language-trained animals. Animal Learning and
Behavior, 27, 114.
Kanerva, P. (1988). Sparse distributed memory. Cambridge, MA: MIT Press.
Kanerva, P. (2009). Hyperdimensional computing: An introduction to computing in distributed representation
with high dimensional random vectors. Cognitive Computation, 1, 139159.
Kanerva, P., Kristoferson, J., & Holst, A. (2000). Random indexing of text samples for latent semantic
analysis. In L. R. Gleitman & A. K. Joshi (Eds.), Proceedings of proceedings of the 22nd annual
conference of the cognitive science society (p. 1036). Mahwah, New Jersey: Erlbaum.
Keller, F., Gunasekharan, S., Mayo, N., & Corley, M. (2009). Timing accuracy of web experiments: A case
study using the WebExp software package. Behavior Research Methods, 41(1), 112.
Kintsch, W. (1988, April). The role of knowledge in discourse comprehension: a construction-integration
model. Psychological Review, 95(2), 163182.
Kintsch, W. (2001). Predication. Cognitive Science, 25(2), 173202.
Laham, D. R. (2000). Automated content assessment of text using latent semantic analysis to simulate human
cognition. Unpublished doctoral dissertation, University of Colorado at Boulder.
Lakoff, G. (1977). Linguistic gestalts. In W. Beach, S. Fox, & S. Philosoph (Eds.), Papers from the 13th
Regional Meeting, Chicago Linguistic Society (pp. 236287). Chicago, Illinois: Chicago Linguistic
Society.
Landauer, T. K., & Dumais, S. T. (1997). A solution to platos problem: the latent semantic analysis theory
of acquisition, induction and representation of knowledge. Psychological Review, 104(2), 211240.
Landauer, T. K., Laham, D., Rehder, B., & Schreiner, M. E. (1997a). How well can passage meaning be
derived without using word order? a comparison of latent semantic analysis and humans. In Erlbaum
(Ed.), Proceedings of the nineteenth annual conference of the cognitive science society (pp. 412417).
Mahwah, NJ.
Landauer, T. K., Laham, D., Rehder, B., & Schreiner, M. E. (1997b). How well can passage meaning be
derived without using word order: A comparison of latent semantic analysis and humans. In Nineteenth
annual conference of the cognitive science society (pp. 412417). Stanford, CA: Lawrence Erlbaum.
Lapata, M., & Lascarides, A. (2003). A probabilistic account of logical metonymy. Computational Linguis-
tics, 29(2), 263317.
Lesk, M. (1986). Automatic sense disambiguation using machine readable dictionaries: How to tell a pine
cone from an ice cream cone. In Proc. of the 5th sigdoc (pp. 2426). New York, NY.
 VECTOR-BASED MODELS OF SEMANTIC COMPOSITION 32

Li, Y., McLean, D., Bandar, Z., OShea, J., & Crockett, K. (2006). Sentence similarity based on semantic nets
and corpus statistics. IEEE Transactions on Knowledge and Data Engineering, 18(8), 11381149.
Lin, D. (1998). Automatic Retrieval and Clustering of Similar Words. In Proceedings of the 17th Interna-
tional Conference on Computational Linguistics (pp. 768774).
Lowe, W. (2000). What is the dimensionality of human semantic space? In R. M. French & J. P. Sougne
(Eds.), Proceedings of the 6th neural computation and psychology workshop (p. 303-311). London:
Springer Verlag.
Lund, K., & Burgess, C. (1996). Producing high-dimensional semantic spaces from lexical co-occurrence.
Behavior Research Methods, Instruments & Computers, 28, 203208.
Markman, A. B. (1998). Knowledge representation. Lawrence Erlbaum Associates.
Masson, M. E. (1986). Comprehension of rapidly presented sentences: The mind is quicker than the eye.
Journal of MEmory and Language(25), 588604.
McDonald, S. (2000). Environmental determinants of lexical processing effort. Unpublished doctoral disser-
tation, University of Edinburgh.
McRae, K., de Sa, V. R., & Seidenberg, M. S. (1997). On the nature and scope of featural representations of
word meaning. Journal of Experimental Psychology, 126, 99130.
Metcalfe, E. J. (1990). A compositive holographic associative recall model. Psychological Review, 88,
627661.
Mitchell, J., & Lapata, M. (2008). Vector-based models of semantic composition. In Proceedings of acl-08:
Hlt (pp. 236244). Columbus, Ohio.
Montague, R. (1974). English as a formal language. In R. Montague (Ed.), Formal philosophy. New Haven,
CT: Yale University Press.
Morris, R. K. (1994). Lexical and message-level sentence context effects on fixation times in reading. Journal
of Experimental Psychology: Learning, Memory, and Cognition, 20, 92103.
Murphy, G. L., & Medin, D. L. (1985). The role of theories in conceptual coherence. Psychological Review,
92(289316).
Nelson, D. L., McEvoy, C. L., & Schreiber, T. (1999). The university of South Florida word association
norms. (http://www.usf.edu/Freeassociation)
Neville, H., Nichol, J. L., Barss, A., Forster, K. I., & Garrett, M. F. (1991). Syntactically based sentence
processing classes: evidence from event-related brain potentials. Journal of Cognitive Neuroscience,
3, 151165.
Nosofsky, R. (1984). Choice, similarity, and the context theory of classification. Journal of Experimental
Psychology: Learning, Memory, and Cognition, 10, 104114.
Nosofsky, R. (1986). Attention, similarity, and the identification-categorization relationship. Journal of
Experimental Psychology: General, 115, 3957.
Nunberg, G., Sag, I., & Wasow, T. (1994). Idioms. Language, 70, 491538.
OSeaghdha, P. G. (1989). The dependence of lexical relatedness effects on syntactic connectedness. Journal
of Experiment Psychology: Learning, Memory and Cognition, 15, 7387.
Osgood, C. E., Suci, G. J., & Tannenbaum, P. H. (1957). The measurement of meaning. Chicago: University
of Illinois Press.
Pado, S., & Lapata, M. (2007). Dependency-based construction of semantic space models. Computational
Linguistics, 33(2), 161199.
Partee, B. (1995). Lexical semantics and compositionality. In L. Gleitman & M. Liberman (Eds.), Invitation
to cognitive science part i: Language (pp. 311360). Cambridge, MA: MIT Press.
Partee, B. (2004). Compositionality in formal semantics. Oxford, UK: Blackwell Publishing.
Pedersen, T., Patwardhan, S., & Michelizzi, J. (2004). WordNet::similarity measuring the relatedness of
concepts. In HLT-NAACL 2004: Demonstration Papers (pp. 3841). Boston, MA.
Pinker, S. (1994). The language instinct: How the mind creates language. New York: HarperCollins.
Plate, T. A. (1991). Holographic reduced representations: Convolution algebra for compositional distributed
representations. In J. Mylopoulos & R. Reiter (Eds.), Proceedings of the 12th international joint
 VECTOR-BASED MODELS OF SEMANTIC COMPOSITION 33

conference on artificial intelligence, sydney, australia, august 1991 (pp. 3035). San Mateo, CA:
Morgan Kaufmann.
Plate, T. A. (1995, May). Holographic reduced representations [Paper]. IEEE Transactions on Neural
Networks, 6(3), 623641.
Plate, T. A. (2000). Analogy retrieval and processing with distributed vector representations. Expert Systems:
The Journal of Knowledge Engineering, 17(1), 2940.
Pollack, J. B. (1990). Recursive distributed representations. Artifical Intelligence, 46(12), 77105.
Pullum, G. K., & Scholz, B. C. (2007). Systematicity and natural language syntax. Croatian Journal of
Philosophy, 7(21), 375402.
Pynte, J., New, B., & Kennedy, A. (2008). A multiple regression analysis of syntactic and semantic influences
in reading normal text. Journal of Eye Movement Research, 2(1)(4), 111.
Raaijmakers, J. G. W., & Shiffrin, R. M. (1981). Search of associative memory. Psychological Review, 88,
93134.
Resnik, P. (1999). Semantic similarity in a taxonomy: An information-based measure and its application to
problems of ambiguity in natural language. Journal of Artificial Intelligence Research, 95-130.
Resnik, P., & Diab, M. (2000). Measuring verb similarity. In L. R. Gleitman & A. K. Joshi (Eds.), Pro-
ceedings of the 22nd Annual Conference of the Cognitive Science Society (p. 399-404). Mahwah, NJ:
Lawrence Erlbaum Associates.
Rips, L. J. (1975). Inductive judgments about natural categories. Journal of Verbal Learning and Verbal
Behaviour, 14, 665681.
Ross, B. H. (1984). Remindings and their effects in learning a cognitive skill. Cognitive Psychology,
16(371416).
Ross, B. H. (1987). This is like that: the use of earlier problems and the separation of similarity effects.
Journal of Experimental Psychology: Learning, Memory, and Cognition, 13, 629639.
Ross, B. H. (1989a). Distinguishing types of superficial similarities: Different effects on the access and
use of earlier problems. Journal of Experimental Psychology: Learning, Memory, and Cognition, 15,
456468.
Ross, B. H. (1989b). Distinguishing types of superficial similarities: Different effects on the access and use
of earlier problems. Journal of Experimental Psychology: Learning, Memory, and Cognition, 15(3),
456468.
Rubenstein, H., & Goodenough, J. B. (1965). Contextual correlates of synonymy. Communications of the
ACM, 8(10), 627633.
Rumelhart, D. E., McClelland, J. L., & the PDP Research Group. (1986). Parallel distributed processing:
Explorations in the microstructure of cognition (Vol. 1). Cambridge MA: MIT Press.
Sahlgren, M., Host, A., & Kanerva, P. (2008). Permutations as a means to encode order in word space. In
Proceedings of the 30th annual meeting of the cognitive science society (pp. 13001305). Mahwah,
New Jersey: Erlbaum.
Salton, G., Wong, A., & Yang, C. S. (1975). A vector space model for automatic indexing. Communications
of the ACM, 18(11), 613620.
Simpson, G. B., Peterson, R. R., Casteel, M. A., & Brugges, C. (1989). Lexical and context effects in word
recognition. Journal of Experimental Psychology: Learning, Memory and Cognition, 15(8897).
Sloman, S. A., & Rips, L. J. (1998). Similarity as an explanatory construct. Cognition, 65, 87101.
Smith, E. E., & Medin, D. L. (1981). Categories and concepts. Cambridge, MA: Harvard University Press.
Smolensky, P. (1990). Tensor product variable binding and the representation of symbolic structures in
connectionist systems. Artificial Intelligence, 46, 159216.
Spenader, J., & Blutner, R. (2007). Compositionality and systematicity. In G. Bouma, I. Krmer, & J. Zwarts
(Eds.), Cognitive foundations of interpretation (pp. 163174). Amsterdam: KNAW publications.
Steyvers, M., & Tenenbaum, J. B. (2005). The large-scale structure of semantic networks: Statistical analyses
and a model of semantic growth. Cognitive Science, 29, 4178.
Till, R. E., Mross, E. F., & Kintsch, W. (1988). Time course of priming for associate and inference words in
discourse context. Memory and Cognition, 16, 283299.
 VECTOR-BASED MODELS OF SEMANTIC COMPOSITION 34

Wallach, H. M. (2002). Structured topic models for language. Unpublished doctoral dissertation, University
of Cambridge.
Weeds, J. (2003). Measures and applications of lexical distributional similarity. Unpublished doctoral
dissertation, University of Sussex, Brighton.
Weiss, S. M., & Kulikowski, C. A. (1991). Computer systems that learn: Classification and prediction
methods from statistics, neural nets, machine learning, and expert systems. San Mateo, CA: Morgan
Kaufmann.
West, R. F., & Stanovich, K. E. (1986). Robust effects of syntactic structure on visual word processing.
Journal of Memory and Cognition, 14, 104112.
Widdows, D. (2008). Semantic vector products: Some initial investigations. In Proceedings of the second
symposium on quantum interaction (qi-2008). Oxford, UK: College Publications.
 VECTOR-BASED MODELS OF SEMANTIC COMPOSITION 35

Appendix A: Materials
Our experimental stimuli for adjective-noun, noun-noun, and verb-object combinations are
shown in Tables 810, respectively.

Table 8: Materials for eliciting similarity judgments on adjective-noun combinations.

High
American countryEuropean state, industrial areawhole country, vast amountlarge quantity, new
bodywhole system, small houselittle room, early eveningprevious day, special circumstance
particular case, black hairdark eye, new informationfurther evidence, economic development
rural community, economic problempractical difficulty, new lawpublic building, general
principlebasic rule, central authoritylocal office, older manelderly woman, high pricelow cost,
different kindvarious form, old personelderly lady, better jobgood place, new lifeearly age, cer-
tain circumstanceeconomic condition, earlier workearly stage, federal assemblynational govern-
ment, effective wayefficient use, social activitypolitical action, similar resultgood effect, major
issuesocial event, different partnorthern region, important partsignificant role, new situation
present position, right handleft arm, general levelhigh point, large numbergreat majority, long
periodshort time, hot weathercold air, modern languagenew technology

Medium
new lifemodern language, good placehigh point, social activityeconomic condition, different
partvarious form, better jobgood effect, old personright hand, local officenew technology,
high priceshort time, social eventlow cost, early stagelong period, efficient usesignificant role,
national governmentcold air, large numbervast amount, economic problemnew situation, new
informationgeneral level, small houseimportant part, European statepresent position, political
actioneconomic development, large quantitygreat majority, dark eyeleft arm, northern region
industrial area, little roomsimilar result, major issueAmerican country, hot weatherfurther ev-
idence, new lawbasic rule, certain circumstanceparticular case, older mannew body, previous
dayearly age, earlier workearly evening, public buildingcentral authority, elderly womanblack
hair, different kindwhole system, effective waypractical difficulty, whole countrygeneral princi-
ple, rural communityfederal assembly, special circumstanceelderly lady

Low
new situationdifferent kind, effective wayimportant part, general levelfederal assembly, cen-
tral authoritypolitical action, major issueearlier work, older mangreat majority, large number
certain circumstance, general principlepresent position, similar resultbasic rule, northern region
early age, left armelderly woman, hot weatherelderly lady, new lawmodern language, previ-
ous daylong period, whole countrydifferent part, social activitywhole system, new technology
public building, high pointparticular case, social eventspecial circumstance, new bodysignificant
role, early eveninggood effect, black hairright hand, practical difficultycold air, short time
rural community, new lifeeconomic development, small houseold person, local officeindustrial
area, national governmentnew information, efficient uselittle room, various formEuropean state,
better jobeconomic problem, economic conditionAmerican country, early stagedark eye, large
quantitygood place, vast amounthigh price, further evidencelow cost
 VECTOR-BASED MODELS OF SEMANTIC COMPOSITION 36

Table 9: Materials for eliciting similarity judgments on noun-noun combinations.

High
development planaction programme, telephone numberphone call, marketing directorassistant manager,
support groupcomputer system, training programmeeducation course, training collegeeducation offi-
cer, planning committeeeducation authority, oil industrycomputer company, health servicecommunity
care, wage increasetax rate, environment secretarydefence minister, office workerhealth minister, tv set
television programme, party officialgovernment leader, state controlgovernment intervention, tax charge
interest rate, news agencyintelligence service, service departmentpersonnel manager, research work
development project, company directorassistant secretary, labour costcapital market, league match
football club, world economymarket leader, state benefithousing department, town councilcity centre,
party leaderopposition member, management structuredatum system, kitchen doorbedroom window,
committee meetingboard member, town hallcounty council, research contractfuture development, rail-
way stationbus company, care planbusiness unit, study groupmanagement skill, tax creditfamily al-
lowance, security policyhousing benefit

Medium
state controltown council, party officialopposition member, intelligence servicebus company, state
benefitcounty council, interest ratebusiness unit, government interventionparty leader, research work
city centre, capital marketfuture development, football clubtown hall, market leaderboard member, tv
setbedroom window, labour costhousing benefit, care planaction programme, management structure
computer system, datum systemsupport group, study groupcomputer company, research contract
training programme, security policydefence minister, family allowancetax rate, tax creditwage in-
crease, management skillplanning committee, committee meetingphone call, railway stationoil indus-
try, kitchen dooroffice worker, education authorityservice department, development plantelevision pro-
gramme, community caretax charge, assistant managercompany director, marketing directorpersonnel
manager, health serviceassistant secretary, education officerdevelopment project, education course
housing department, health ministergovernment leader, telephone numberleague match, environment
secretarynews agency, training collegeworld economy

Low
development projectcare plan, television programmeresearch contract, government leadersecurity pol-
icy, tax chargedatum system, news agencystudy group, world economymanagement structure, as-
sistant secretarycommittee meeting, company directortax credit, league matchfamily allowance, ser-
vice departmentrailway station, housing departmentkitchen door, personnel managermanagement
skill, bus companyhealth service, city centrecommunity care, business unitdevelopment plan, town
halleducation course, future developmenttelephone number, party leaderenvironment secretary, town
councileducation authority, board memberassistant manager, bedroom windoweducation officer, county
councilmarketing director, opposition memberhealth minister, housing benefittraining college, action
programmetv set, support groupinterest rate, tax ratemarket leader, training programmeresearch work,
defence ministergovernment intervention, office workerparty official, computer companyintelligence
service, computer systemstate control, oil industrycapital market, planning committeefootball club,
phone callstate benefit, wage increaselabour cost
 VECTOR-BASED MODELS OF SEMANTIC COMPOSITION 37

Table 10: Materials for eliciting similarity judgments on verb-object combinations.

High
produce effectachieve result, require attentionneed treatment, present problemface difficulty,
leave companyjoin party, satisfy demandmeet requirement, use powerexercise influence, shut
doorclose eye, sell propertybuy land, reduce amountincrease number, send messagereceive let-
ter, suffer losscause injury, use testpass time, write bookread word, start workbegin career,
reach levelachieve end, stress importanceemphasise need, use methoddevelop technique, hold
meetingattend conference, use knowledgeacquire skill, win matchplay game, like peopleask
man, follow roadcross line, help peopleencourage child, pose problemaddress question, raise
headlift hand, pay pricecut cost, leave housebuy home, wave handstretch arm, discuss issue
consider matter, provide helpoffer support, win battlefight war, remember namehear word, set
exampleprovide system, provide datumcollect information, pour teadrink water, share interest
express view

Medium
write bookhear word, address questionraise head, read wordremember name, follow road
set example, use methoddrink water, hold meetinglift hand, win matchfight war, play game
win battle, start workwave hand, achieve endexpress view, develop techniqueprovide help, at-
tend conferenceshare interest, provide systemuse power, cut costreduce amount, buy home
sell property, consider matterproduce effect, leave housebuy land, pay pricerequire attention,
collect informationreceive letter, offer supportneed treatment, discuss issuepresent problem,
stretch armclose eye, pour teajoin party, provide datumshut door, face difficultypose problem,
achieve resultreach level, exercise influenceuse knowledge, satisfy demandemphasise need, send
messageask man, use testacquire skill, meet requirementhelp people, leave companyencourage
child, pass timecross line, suffer lossbegin career, increase numberlike people, cause injury
stress importance

Low
use knowledgeprovide system, pose problemconsider matter, encourage childleave house, reach
levelprovide datum, ask manstretch arm, acquire skillbuy home, stress importancecut cost, be-
gin careerpay price, cross lineoffer support, help peoplediscuss issue, like peoplecollect infor-
mation, emphasise needpour tea, drink wateruse test, remember namepass time, share interest
exercise influence, hear wordsend message, wave handleave company, fight warincrease number,
provide helpsatisfy demand, raise headcause injury, lift handachieve result, set exampleface dif-
ficulty, express viewsuffer loss, win battlemeet requirement, buy landwrite book, receive letter
read word, produce effectstart work, present problemaddress question, use powerdevelop tech-
nique, sell propertyhold meeting, shut doorfollow road, join partyplay game, close eyeachieve
end, reduce amountwin match, need treatmentuse method, require attentionattend conference
 VECTOR-BASED MODELS OF SEMANTIC COMPOSITION 38

Appendix B: Experimental Instructions

In this experiment you will be shown a pair of noun phrases. Your task is to judge how similar
the two phrases are. You will make this judgement by choosing a rating from 1 (not very similar)
to 7 (very similar). The focus is on the similarity of the concepts named by the phrases, not any
association between the two phrases.
For example, if you were asked to make the following comparison:

(1) a. professional advice

b. expert opinion

you would give this a high similarity rating (e.g. 6 or 7). Both these phrases concern guidance or
instruction from a knowledgeable person and so have highly similar meanings. On the other hand,
if you were given the following comparison:

(2) a. social worker

b. wide range

you would probably choose a low similarity rating (e.g. 1 or 2), since one is an occupation and the
other is a magnitude. Likewise, for this comparison:

(3) a. increasing taxation

b. public protest

you would also choose a low similarity rating (e.g. 1 or 2), since they are different things, even
though they might be associated, in that the first could lead to the second. Of course, associated
phrases may also be similar.
Sometimes the two phrases will have meanings that are moderately different though still have
much in common. For instance, in this comparison:

(4) a. human behaviour

b. social activity

you would choose a middling rating (e.g., 3, 4 or 5) if you felt that the meanings of the two phrases
were reasonably different but also had some similarities. For instance both involve the interactions
of people, although the two phrases also invoke other distinct concepts.
There are no correct answers, so whatever choice seems appropriate to you is a valid re-
sponse. Simply try to rate how similar the meanings of the two phrases are. Base your judgment on
your first impression of what each phrase means. The whole experiment should take only about 10
minutes.
Remember:
Rate the similarity of the phrases not their association.
Base your judgment on your first impression of what each phrase means.
There are no correct answers.
At the start of the experiment you will be given a few examples to practice on.
 VECTOR-BASED MODELS OF SEMANTIC COMPOSITION 39

Appendix C: Simple Vector and Tensor Algebra

Formally, vectors are defined as objects within a vector space, which is itself defined in terms
of the abstract relations and properties of the objects it contains. Informally, we usually think of a
vector as something having magnitude and direction, such as an arrow in space. From the point of
view of computation, we can always represent vectors concretely in terms of some particular basis,
that is as a set of co-ordinates. Thus, a vector v is represented by its components vi in that basis.
An important relation between vectors is the dot product, defined as the sum of the products
of the components:
u v = ui vi (30)
i
where the index i ranges over all the components (i.e., the dimensions of the space). This allows us
to define the length of vectors:
|v| = v v (31)
The dot product also has a useful relation to the angle, , between the two vectors:

u v = |u||v|cos() (32)

This implies that the dot product of any two orthogonal vectors ( = 90 ) is zero. Equa-
tions (30)(32) allow us to calculate the cosine of the angle as:
i ui vi
cos() = (33)
i ui ui i vi vi
The two most basic transformations which operate on vectors are addition and multiplication
by a scalar:
p = u+v (34)
pi = ui + vi (35)
q = sv (36)
qi = svi (37)
Any vector v, can be expressed in terms of a component, vk , parallel to and a component, v ,
orthogonal to a second vector, u.
v = vk + v (38)
Taking the dot product of u on both sides of this equation yields:

v u = |vk ||u|Cos(0 ) + |v ||u|Cos(90 ) (39)

Since Cos(0 ) = 1 and Cos(90 ) = 0, this implies:

vu
|vk | = (40)
|u|
Thus, we can construct vk by normalizing u to give a unit vector which points in the right direction,
and then multiplying by the right magnitude |vk |.
vu u vu
vk = = u (41)
|u| |u| u u
 VECTOR-BASED MODELS OF SEMANTIC COMPOSITION 40

The orthogonal component, v , can then be calculated from the fact that the two components must
combine to give v.
vu
v = v u (42)
uu
Another important set of transformations of vectors are the linear transformations induced by
matrices.
w = Mv (43)
wi = Mi j v j (44)
j

A matrix, M, is represented by an array of values, Mi j , indexed by a pair of indices, i and j. The

element Mi j can be thought of as determining how much the jth component of the original vector, v,
contributes to the ith component of the transformed vector, w.
Alternatively, we can think of M in terms of its polar decomposition into a matrix, D, of
dilations, and a matrix, R, of rotations.
M = DR (45)
The action of D on a vector is to stretch it by various amounts in various directions, and the action
of R is to rotate it around various axes, without changing its length. The directions in which a
matrix dilates vectors without changing their direction are known as eigenvectors, and the amounts
by which they are stretched are known as eigenvalues.
The tensor product, , takes a pair of vectors and combines them to form a higher dimen-
sional vector.
t = uv (46)
In particular, if u and v have dimension n, then t has dimension n2 . The components of t are all the
pairwise products of the components of u and v.

ti j = ui v j (47)

It is possible to project such a product of vectors down onto a vector of the same dimension using a
rank 3 tensor.
r = Cuv (48)
Again this is a linear transformation. The tensor C has three indices, corresponding to the three
vectors r, u and v.
ri = Ci jk uk v j (49)
jk

A simple example of such a rank 3 tensor would be one in which Ci jk = 1 when i = j = k and 0
otherwise, which yields:
ri = ui vi (50)
which can be also written as:
r = uv (51)
A more complex example is the tensor, C, with components Ci jk = 1 when k = (i j)modn and 0
otherwise.
ri = u j v(i j)modn (52)
j
 VECTOR-BASED MODELS OF SEMANTIC COMPOSITION 41

also written as:

r = u~v (53)
Multiplying a single vector, u, by a rank 3 tensor, C, produces a matrix, U.

U = Cu (54)

Ui j = Ci jk uk (55)
k

Multiplication of v by this matrix, U, then results in the same vector, r, produced by the prod-
uct Cuv:
Uv = Cuv = r (56)

Ui j v j = Ci jk uk v j = ri (57)
j j k

It is also possible to define higher order tensors, such as a rank 4 tensor which acts on three vectors:

x = Duvy (58)

xi = Di jkl ul vk y j (59)
jkl