The Role of an Abstract Ontology in the Computational

Interpretation of Creative Cross-modal Metaphor

Sylvia Weber Russell

Department of Computer Science, University of New Hampshire

Durham, NH 03824, U.S.A.

Abstract. Various approaches to computational metaphor interpretation are based

on pre-existing similarities between source and target domains and/or are based

on metaphors already observed to be prevalent in the language. This paper ad-

dresses similarity-creating cross-modal metaphoric expressions. The described

approach depends on the imposition of abstract ontological components, which

represent source concepts, onto target concepts. The challenge of such a system

is to represent both denotative and connotative components which are extensible,

together with a framework of all general domains between which such exten-

sions can conceivably occur. An existing ontology of this kind is outlined. It is

suggested that the use of such an ontology is well adapted to the interpretation of

both conventional and unconventional metaphor that is similarity-creating.

1 Introduction

The last couple of decades have seen an increasing number of computational approaches

to processing metaphor.By interdisciplinary consensus, this research has generally been

implemented as processes that map an expression in a source domain (domain of the

metaphorically used concept) to an interpretation in a topic- or target domain (domain

of intended meaning). (Within-) physical-domainmetaphor,as in the war horse The ship

plowed [through] the sea/waves, has received attention early on [1]. Treatments that

focus solely on physical-domain metaphor include those by Wilks [2] [3] and Fass and

Wilks [4] and are discussed in [5]. Fass [6], while presenting an extensive treatment of

metaphor in the context of literal, metonymic and anomalous expressions, also focuses

mainly on physical-domain metaphor.

This paper argues for the role of an abstract ontology in the interpretation of cross-

modal metaphor, with special attention to similarity-creating metaphor. Cross-modal

metaphor extends across ”conceptual domains” (modes, levels), as in The country leapt

to prosperity, which involves extension from a physical to a control (of wealth) domain,

or as in Encyclopedias are gold mines, which involves extension from a control (of

wealth) domain to a mental domain.

The discussion begins with an indication of what is meant by similarity-creating

metaphor and how some of the major research on metaphor does not address it. Some

notes on ontologies follow, as well as observations on abstraction, including mathemat-

ical abstraction and its potential contribution to an abstract natural-language ontology.

Weber Russell S. (2008).

The Role of an Abstract Ontology in the Computational Interpretation of Creative Cross-modal Metaphor.

In Proceedings of the 5th International Workshop on Natural Language Processing and Cognitive Science, pages 52-63

DOI: 10.5220/0001736900520063

 SciTePress

A program that relies on an abstract ontology to address similarity-creating metaphor

is then outlined, with an explanation of components of the ontology. This description is

followed by brief illustrations of relevant aspects of interpretations of verbal and nomi-

nal metaphor. The paper concludes by noting that some other researchers have found it

necessary to extend their metaphor processing systems with (at least implicit) abstract

ontological components, suggesting the need for attention to an abstract ontology in a

metaphor processing system, at least if the system is claimed to be explanatory.

2 Similarity-creating Metaphor

With respect to metaphor, the word ‘creative’ is often used interchangeably with the

word ‘novel.’ By ‘novel metaphor,’ some researchers refer simply to metaphors that

their systems–or perhaps evenhumans–havenot previously encountered.Such metaphors

may be based on representations that capture similarities existing prior to use of the

metaphor. By contrast, Indurkhya[7] presents evidence of the signiﬁcant role of similarity-

creating metaphors in cognition.

As Indurkhya points out, in many metaphoric expressions where verbs are used

unconventionally–some would say in a novel way, there is a pre-existing similarity. For

example, in The sky is crying, there is an easily recognizable similarity between crying

(tears falling) and one of the few things that fall from the sky (rain drops); the metaphor

can be analyzed by comparison, though it also suggests sadness. Similarity-creating

metaphor, on the other hand, is characterized as change of representation: ”In instanti-

ating the source concept network in the target realm, parts of the realm are ‘grouped’

together and made to correspond to the concepts of the concept network. In this pro-

cess, the target realm is given a new ontology, and its structure, as seen from the more

abstract concept network layer, is changed” (p. 254).

Indurkhya acknowledges that the difference between suggestive similarity-based

metaphor and similarity-creating metaphor may be a matter of degree, i.e., degree

of participation of the target and source domains. That is, the closer the metaphor

is to being similarity-creating, the more the source ontology is imposed and the less

the pre-existing target ontology remains. In this paper it is assumed that cross-modal

metaphor (which Indurkhya does not focus on) is in a sense always similarity-creating,

because the real-world details of source and target will always differ. For example, in

the metaphor, Encyclopedias are gold mines, there is little physical similarity between

encyclopedias and gold mines, or between reading and mining.

3 Approaches to Metaphor

Through metaphor, different source concepts may be used to structure the target in

different ways. Lakoff and Johnson [8] recognize both the ”conceptual metaphors” (or

”metaphor themes”) LIFE IS A JOURNEY and LIFE IS A GAME, and perhaps other

”life” metaphors. Thus similarity can be created by re-conceptualizations. A problem

with Lakoff’s metaphors, however, is that they are categories, without speciﬁcations of

which components of the source domain are extensible.

The early approach of Carbonell [9] [10] is based on the stored conceptual metaphors

of Lakoff and Johnson. However, systems that rely only on stored conceptual metaphors

cannot interpret linguistic metaphors that do not ﬁt any stored conceptual metaphor.

Also, the metaphoric nature of the transferred properties themselves is not addressed.

For example, the phrase ‘ﬁrmly supported,’ used in his example of the MORE IS UP

conceptual metaphor, is simply applied to both source and target domains without se-

mantic analysis.

Hobbs [11] addresses metaphor without recourse to stored metaphors, using infer-

ences to express linguistic relationships. In his illustration he matches ‘send (a bill)’

in a Congress schema to ‘pitch (a ball)’ in a baseball schema, and proves the corre-

spondences between roles in the two schemata. This metaphor is certainly novel, but

Hobbs’s interpretation process is based on existing similarities between the source and

target schemata.

Approaches of other researchers that show some potential to address similarity-

creating metaphors are discussed at the end of this paper.

4 Ontologies

The term ”abstract ontology” might be seen as an oxymoron, and it is, if an ontology is

that which purportsto describe reality. Wilks [12], in dismissing the distinction between

traditional/classical and modern/AI-type meanings of ”ontology” as unimportant for

AI/NLP purposes, also rejects any claims that ”cleaning up” given ontologies will result

in any notable advances in the ﬁeld. This view (which I accept) is mentioned in order

to emphasize that the focus in this paper is only on the role that abstracted components

can play in a computational metaphor interpretation system with attention to presumed

cognitive components, and on what types of components are needed and are peculiar

to metaphor interpretation. While the ontology is explained below, the intent is not to

justify the exact form the individual components take. It is important, however, that the

ontology, being abstract, be relatively small and transparent, for purposes of evaluation

and revision.

A cross-modal metaphor-relevant ontology is based not on any objective reality,

but on a certain unconventional view of reality through language, which is itself con-

ceptualized from reality. A perhaps noncontroversial observation of Quine [13] on the

ontology of language would seem to apply to abstractions from language (i.e., to an

abstract ontology) as well–namely, that differences between one person’s ontology and

another may depend simply on how the ontologies are ”sliced” or how components are

grouped; correspondences between ontologies will probably not be one-to-one (cf. also

Whorf [14]. There is no claim in this paper, then, that the components of the abstract

ontology are universal, uniquely ”correct,” or language-free; there is merely an appeal

to a consensus of ”reasonableness” by speakers of the same language (and others that

are related to some extent). Neither is there speculation on the source of the given on-

tology in developmental or evolutionary terms. Relying on any such ontology may raise

the ”mentalism” criticism; however, as Quine says, ”I know no better.”

In cross-modal metaphor, any perceived or imposed similarity as mediated by the

ontology is abstract (in the conventional use of that word); some considerations of

abstraction follow.

5 Abstraction

In a sense, any representation, which is a mapping between reality and symbols, or

between those symbols and higher-level symbols, is abstract. In the context of mathe-

matics learning, Kaput [15] deﬁnes four interacting types of representation - 1) cogni-

tive and perceptual, 2) explanatory representation involving models, 3) representation

within mathematics and 4) external symbolic representation, such as a chip, which can

be instantiated by many different objects and can thus be a generalization or abstrac-

tion for cookies, baseball cards, dollars, etc. In natural language, similarly, the concept

underlying the word ‘object’ can be thought of as a generalization or abstraction for the

mentioned items; it is plausible that the cognitive components which relate to mathe-

matical abstraction are (or overlap with) those which structure linguistic metaphor. As

it is being argued for an abstract ontology for metaphor, a consideration of relation-

ships between mathematical and abstract linguistic components that might be included

in such an ontology follow.

5.1 Mathematical Language

There are frequent references to the power of mathematics to account for many analo-

gous situations through its abstract language. It is often difﬁcult to characterize mathe-

matical language and natural language independently in discourse, since mathematical

concepts can be embedded in natural language, not only in mathematical word prob-

lems, but in our everyday language about situations. English can embed both explicitly

numerical references, such as ‘ten’ and ‘a dozen,’ and expressions that are mathemat-

ically relevant but not necessarily so intended, such as ‘the rest of them,’ ‘a slice of

pizza,’ ‘altogether,’ ‘join,’ ‘more than,’ etc. [16]. The meshing of these languages cor-

responds to Kaput’s interaction between cognitive/perceptual and mathematical repre-

sentations and suggests common ontological components. For example, an interlingual

”PART” concept can be realized in both mathematical and nonmathematical language.

5.2 Reiﬁcation as a Basis for Spatial Structuring

In mathematical language, arithmetic equations represent structures with numbers as

abstractions not only of sets of objects, but also of non-object concepts (as in ‘he fell

twice’), and with operators that relate these sets; the abstraction to numbers estab-

lishes the basis for the equation. Similarly, in linguistic metaphoric extensions from

the physical domain, nonphysical concepts may become abstract ”objects,” allowing

verbal concepts to ”operate on” them. Reiﬁcation (or ”nominalization”)–treating an ac-

tion, relation or attribute as an ”abstract object” in the form of a noun–is thus a ﬁrst step

in the creation of this kind of metaphor. Expressed syntactically, reiﬁcation is an in-

stance of the ”abstract concept as object” metaphor [17], which has become integrated

into (some) natural languages as dead metaphors, i.e., usually thought of as literal.

Mathematical language and much of metaphoric language thus share spatial ground-

ing, suggesting that not only physical-domain verbs/attributes, but also such concepts

in nonphysical domains might be analytically based on simple spatial, i.e., existential

and relational, structures.

To illustrate, the physical action underlying the verb construct ’chase out/away (e.g.,

mosquitoes)’ can be extended to apply to conceptually different types of objects. In

‘chase away an idea,’ the ‘idea,’ which is a reiﬁcation in a mental domain, is ”taken

away” from the thinkers of the idea; mathematically, to ‘chase away six mosquitoes,’

as in a word problem, may mean to subtract or ”take away” 6. In both cases, symbols

are mapped from one domain to another through an abstraction representing ”leaving a

state” (of thought or of the presence of the six mosquitoes). Reiﬁcation, then, enables

the natural-language extension similarly as quantiﬁcation enables the mathematical ex-

tension.

Thus if we settle on a set of abstract components in terms of states of objects, exis-

tence and relations, and use them as the basis of abstract verb deﬁnitions in any domain,

then these components can be considered to be extended in metaphor and to contribute

to its interpretation (see Section 8).

6 An Ontology-based Metaphor Analysis Program

MAP, a computational metaphor paraphrase program [1] [5] [18], interprets an isolated

simple-sentence metaphor in terms of a roughly equivalent paraphrase conventionally

considered as ”literal.” The most critical aspect of the program resides in the (abstract)

lexicon, where verbs and nominals are represented by components of an abstract on-

tology. For verb-based metaphor, components representing a verb which serves as a

metaphoric source concept are interpreted in the target domain as indicated by the nom-

inal concepts or ”objects” with which the verbal concept is used. Thus for ”She chased

away the thought,” a mental domain is indicated by ”thought,” and the primitives un-

derlying ”chased away” lead to a paraphrase including a phrase such as ”voluntarily

stopped thinking about.” For nominal metaphor, the primitives underlying salient prop-

erties or predicates [5] of the source nominal are transferred to the target representation.

Thus for Political movements are glaciers, extended predicates include components rep-

resenting slow change.

MAP treats dead (frozen, assimilated) metaphors and novel metaphors (whether

similarity-based or similarity-creating) in the same way, though of course dead metaphors

and even some metaphors that are ”alive” but conventional could be deﬁned directly in

the lexicon for efﬁciency purposes. The focus of this discussion, however, is on MAP’s

ability to interpret similarity-creating metaphors.

7 MAP’s Ontology

A recognition that natural language and mathematics share spatial structure, i.e., struc-

ture in terms of objects and relations, suggests that 1) a small number of abstract de-

scriptors that overlap with those of mathematics in being spatially based reﬂect some

intuitive consensus of speakers of the language with respect to the design of an ontol-

ogy and 2) such spatially based structures provide a framework for additional, quali-

fying primitives, some of which can also be drawn from mathematics. The ontology

described below consists of extensible components including spatial structures, which

represent the potential similarities between source and target, and domains, which rep-

resent the differences.

7.1 Abstract Extensible Components

The task of determining a set of extensible components of verbal concepts entails con-

sidering which concepts speakers of a given language recognize in a literal meaning of

a verb that allows them to understand a metaphoric use of that verb, even if they have

never heard it before. If much of our language is spatially structured, we should be able

to see (though not prove) some cognitive basis for components in the abstract domain

of mathematics (arithmetic, calculus, logic) and its application in physics. The follow-

ing structures and features either have a math-physical counterpart and/or have a broad

linguistic consensus.

Structures. The basic structure assigned to all verbs is a STATE, the beginning or

end of a STATE, or transition through a STATE, all of which correspond to boundary

points in mathematical functions or to the space inbetween. The STATE itself may be

either existential (OBJECT BE), existential with a (static or dynamic) attribute (OB-

JECT BE ¡attribute¿), or relational (OBJECT AT LOCATION). Any of these structures

can be negated with the component NOT. In addition, any of the above structures can

be caused, i.e., have an AGENT.

These abstract structures can be thought of as unary

or binary abstract case structures,

either of which can be operated on by an AGENT.

Features. It is qualiﬁers and connotations that are often the reason for a metaphor.

These are represented as abstract, conceptual features (more ﬂexible than explicit cat-

egories), with polarity or magnitude speciﬁcations as appropriate. As qualiﬁers of ac-

tions, features for action verbs correlate with function-relevant mathematical descrip-

tors: CONTINUITY, REPETITION (frequency) and SPEED (rate). Verbs with quan-

titative attributes (‘grow’), may have MAGNITUDE and GREATER/LESS-than. VO-

LITION is a feature describing an actor. Responses of an experiencer of the metaphor

have EVALUATION values and FORCE magnitude. EVALUATION and FORCE cor-

respond to Osgood’s [21] ”evaluative”and ”potency” factors–two of the three nonstruc-

tural factors (the other being ”activity,” reﬁned in the action features above) he empiri-

cally determined to be extended in metaphoric usage (see also Aarts and Calbert [22].

Various emotions are also incorporated. Emotional states are not abstract in themselves;

however, they are abstracted rather than literal; the fear experienced when one’s hope

is torpedoed is probably not the same as that when one’s boat is literally torpedoed.

In the case of an agentive verb, it is the ”effect” or ”result” of the action that is considered

of primary salience and receives the domain speciﬁcation. Verbs involving other higher-level

primitives, such as PURPOSE, are formalizable in the ontology but have not yet been included

in MAP’s lexicon.

Abstract case structures are simpler than traditional case structures, as Fillmore’s [19] dative

and locative cases or Schank’s [20] ”conceptual” cases are combined in (abstract) LOCATION.

7.2 Conceptual Domains

Conceptual domains are orthogonal to the extensible portion of the ontology. For cross-

modal metaphor, the domains are only those general, Aristotelian-like domains which,

along with the PHYSICAL (animate and inanimate), are thought of as human facul-

ties: MENTAL, with subdomains of intellect, attitude and volition; SENSORY, with

sense-speciﬁc subdomains linking PHYSICAL and MENTAL; and CONTROL, with

subdomains intrinsic (e.g., ‘talent’) and extrinsic (e.g., ‘wealth,’ ‘rights’). This taxon-

omy within the ontology is obviously breadth- rather than depth-oriented. Every verb

in the lexicon is assigned the conceptual domain in which it is thought to be literal. The

model allows a concept in any conceptual domain to be a source, though the source is

more often PHYSICAL. It is the difference between the conceptual domains of a verb

and its object or subject that triggers cross-modal metaphor recognition.

The ontology of MAP is thus based on a small, organized set of primitives. The

delineation of (a small set of) extensible and nonextensible components in a transparent

frameworkallows the management of the ontology and the representation of verbal con-

cepts in terms of that ontology to be feasible. Also, by deﬁning open-set words through

the abstract components, we can observe directly to what extent a metaphoric use of

that word is adequately interpreted (how well it conforms to human understanding),

and we can note which components, when imposed on the target domain, positively or

negatively affect the interpretation of similarity-creating metaphors.

8 Interpretation

8.1 Constraints

MAP does not compare a source representation with a target; it is not similarity-based.

Rather, the abstract source representation is imposed, i.e., directly projected onto the tar-

get. However,source representations cannot be imposed arbitrarily.There are some con-

straints on interpretations to assure (as far as possible) that the expression makes sense

metaphorically, i.e., is not ”anomalous,” indicating a probable mis-parse. For example,

when a transitive verb is used metaphorically with an object nominal in a different con-

ceptual domain, there are still some abstract constraints that the object must satisfy.

These constraints are realized in MAP as conceptual (abstract) features of nominals.

For cross-modal metaphor interpretation, these features are fewer than literal semantic

features of nominals, since many details of the nominal concept drop out of the pic-

ture. For example, PART (of), CONTAINED (in) and FIXED (to) features that might

apply to literal language merge, because the speciﬁc topographical features of literal

objects are not signiﬁcantly distinguished for abstract concepts. This feature set and its

application are discussed in detail in [18]. Current (binary-valued) features are: SHAPE

(vs. amorphous, mass), 1-DIMENSIONAL (linear-like), FIXED/PART/CONTAINED

(subordinate), COMPLEX (vs. elementary), FLUID, ANIMATE (dynamic).

That the meaning of a novel metaphor depends on its literal meaning does not necessarily

imply that literal meanings are accessed before metaphoric ones by humans.

8.2 Paraphrase

If there happens to be a verb in the target domain that has an abstract representation

in common with the source (at least the structure), then that verb can be included in

the paraphrase.

For the example news torpedo his hope

that verb might be ‘disap-

point,’ which has the same structure as the verb ‘torpedo,’ i.e., AGENT cause LEAVE-

STATE (OBJECT BE), where the OBJECT is in the MENTAL-ATTITUDE domain.

In an ”undoing” of reiﬁcation, the reiﬁed abstract OBJECT ‘hope’ from the input is

mapped to the verbal ‘hope’ as part of the paraphrase. If any (or all) components are

”left over” from the source representation, they are lexicalized directly; here, this would

be FORCE: HIGH, SPEED: HIGH. Lexicalization gives:

STRUCTURE: news cause he stop hope

CHARACTER of the ACTION: intensely, suddenly

If no target word with a similar abstract structure is found, all the abstract components

are lexicalized, together with an indication of the target domain. Abstraction necessar-

ily entails a loss of information, and the paraphrases produced often seem inadequate in

being too general, though ”literal” and not wrong. However, it was deemed important to

start with a broad, non-ad hoc framework, rather than to attend to target-domain detail.

The characterization of nominals for nominal metaphor interpretation is much more

open than for verbs, since objects can mean many things to many people. As the mean-

ing of even one sense of a nominal is less constrained than that of a verb, which has

inherent structure, there are more possibilities for similarity-creating metaphors. For

nominal metaphor, MAP transfers putative salient properties of source nominals [23]

[5], represented in terms of the described ontological components, to the target. One of

the most prominent properties of a nominal that enters into metaphoric interpretation is

its function (cf. Gibson’s ”affordances” [24] or typical action.

As nominal metaphor typically involves extension of verbal or attributive compo-

nents, a brief indication of nominal metaphor interpretation will serve to illustrate fur-

ther representational aspects of verbal metaphor as well. For the example Encyclope-

dias are gold mines: The FUNCTION of ‘(gold) mine’ (one takes gold out of it) has

the resultant STATE structure (‘*’ is a reference to the user of the concept having the

function)

(ENTER-STATE) (AT (OBJ: +PHYSICAL/gold LOC: *)),

which is in the CONTROL-EXTRINSIC (of +PHYSICAL) domain. A connotation is

that the OBJECT is very valuable (EVALUATION: HIGH). The abstract structure and

the EVALUATION are transferred to the FUNCTION predicate of ‘encyclopedia,’ (one

reads it), which has the resultant STATE structure

(ENTER-STATE) (AT (OBJ: +MENTAL-INTELL LOC: *))

in the MENTAL-INTELLECTUAL domain. The paraphrase is ‘One read encyclopedia

has result one has knowledge which-has high value.’

Such verbs in English often have Latin or Greek etymology, that itself metaphorically im-

poses physically derived components on otherwise inexpressible concepts. Syntax itself also

reﬂects this semantic structuring; direct objects can be non-PHYSICAL, i.e., not ”really” ob-

jects, showing an analogy between PHYSICAL and non-PHYSICAL usages.

Irrelevant grammar-related elements are ignored in input and output examples.

This example is of only average richness, but the added connotation of high value,

along with the lack of pre-existing literal similarity, makes it similarity-creating. A

metaphor that is perhaps more clearly similarity-creating is the PHYSICAL-domain

metaphor Dumps are gold mines. Here the entire FUNCTION structure of ‘dump’ (to

put things into it rather than literally or metaphorically take them out) is overridden;

the interpretation is that something of extreme value can be found in dumps. Of note is

that in both cases the property of ”high value,” along with other factors such as conno-

tations, are culturally based and constitute the kind of information that Indurkhya [7]

claims must be represented in meanings of objects if similarity-creating metaphors are

to be interpreted computationally. Nominal metaphor interpretations are considered to

be only likely, not deﬁnitive. However, metaphors that have more obscure interpreta-

tions usually require further elaboration, requiring multi-sentence analysis.

This approach appears to correspond with Indurkhya’s view of similarity-creating

metaphor; the source ontology is imposed onto as opposed to compared with the target

domain. Moreover, cultural and experiential factors–the imagined experience which In-

durkhya claims as missing from computational treatments of metaphor– are represented

symbolically as imposed concepts.

9 Comparative Evaluation

While most computational approaches to metaphor do not address similarity-creating

metaphors, as they are not based on a semantic analysis that allows ontological compo-

nents to shape the target domain, the following research has some aspects corresponding

to aspects of MAP.

Martin’s [25] system is similar to Carbonell’s earlier work (see Section 3), with

a much more comprehensive implementation. However, he has extended his system

through a recognition of the conceptual relationship between states and their beginnings

and endings. These correspond to MAP’s basic abstract structural components.

Carbonell and Minton [26] specify their method for metaphor interpretation in terms

of transfer of portions of a graph consisting of concepts (nodes) linked by relations.

Thus for X is a puppet of Y, the node CONTROL between the object ‘puppet’ and the

actor ‘puppeteer’ is transferred to the node between X and Y. This process and type

of representation is similar to that of MAP. However, a comprehensive representation

system does not appear to exist, and they do not incorporate affective or cultural com-

ponents.

The idea underlying the representations of the system of Suwa and Motoda [27]

is perhaps the most similar to that of MAP. Their ontology itself does not explicitly

distinguish domains as in MAP and thus is not as transparent as MAP’s, but they do

use a ﬁnite, relatively small ontology consisting of what they call abstract primitives.

These are only in the form of verb structures, through which they match source and

target verbs–a method which apparently succeeds in an interpretation only if such a

match exists. Experiential factors are not incorporated. Their system therefore does not

address similarity-creating metaphor as it stands; however, they discuss the addition of

”new” components to the target and could in theory achieve this, given their abstract

ontology.

In the recent work of Barnden et al. [28] [29] and Agerri et. al. [30], it is acknowl-

edged that many metaphoric usages are not adequately covered by Lakoff’s conceptual

metaphors. They present ”view-neutral mapping adjuncts” (VNMAs), which ‘trans-

fer those aspects that are not part of any speciﬁc metaphorical view’ or conceptual

metaphor [30]. VNMAs appear to correlate with the structural metaphoric extensions

of MAP, and are applied as ”default rules.”

The metaphor theory and attendant hypotheses underlying the system of Narayanan

[31] also have signiﬁcant similarities with the described system (though his model dif-

fers in his neural-like implementation). As aspects of his theory in part applies to Barn-

den et al.’s and Agerri et al.’s work as well, it will be discussed in somewhat greater de-

tail. Narayanan’s treatment of nominals, verbs and adverbs in verbal metaphor in terms

of invariant components corresponds with that of MAP in at least two ways. First, the

prevalence of spatio-temporal structures as extensible to other domains is incorporated.

(Narayanan proceeds further to establish correspondences between motion verbs ex-

pressing such structures and possibilities as part of a sequence of actions and inferences

leading to a goal in a target domain.) Second, from looking at databases, Narayanan

has concluded the invariance of certain ”parameters” which correspond to MAP’s ad-

verbial features expressing evaluation, agent attitude/intent and other (nongrammatical)

aspects. His determination can be viewed as corroborating support for the inclusion of

such features.

The ﬁne granularity of Narayanan’s representation of his two target domains, e.g.,

economic policy and politics is appealing, though limited in breadth. While Narayanan

works out speciﬁc mappingsto his targetdomains, MAP deals generally with metaphoric

extension between domains in a proposed domain ontology. MAP thus reveals how a

source domain, which may sometimes not be the spatial/physical domain (and may even

be an ”abstract” domain, such as the economic domain), can structure any domain.

Another differenceconcerns the way in which source concepts are projected metaphor-

ically onto the target. In Narayanan’s system, entities and actions are projected di-

rectly through pre-established ”conceptual metaphors” in the sense of Lakoff, such as

MOVERS ARE ACTORS or OBSTACLES ARE DIFFICULTIES, which must be stored.

From the point of view of language understanding, MAP shows how a metaphoric

usage might be understood in terms of perceived or imposed similarities represented

by semantic components of literally understood lexical items, whether the metaphor

is conventional or creative, and whether stored or not. Apart from these explanatory

differences, Narayanan’s system for projecting verbal concepts has similarities in con-

cept to MAP, with more detailed paraphrases for the two domains he treats. The dif-

ferences perhaps reﬂect the differing intended tasks–narratives within a speciﬁc topic

domain/discipline in the case of Narayanan’s system, and spontaneous references to

metaphor in open discourse in the case of MAP.

10 Conclusions

Some metaphor programs other than MAP produce more detailed interpretations as

a result of being similarity-based or being restricted to certain domains. MAP on the

other hand was designed for scope rather than detail, without regard to any speciﬁc

examples or domains, and its focus on extensible components based on the seman-

tics of the metaphorically used concept enables it to at least minimally ”understand”

similarity-creating metaphors. The described ontology accounts for both similarities

(through extensible components) and differences (between conceptual domains) under-

lying cross-modal metaphor.Extensible components include not only structures but also

connotations and stereotypic experience, imposition of which is offered as an example

of what Indurkhya calls a re-structuring by projection of the source concept network

onto the target realm. It would seem that the computational interpretation of similarity-

creating metaphors with cognitive relevance requires either an abstract ontology of the

type presented or some implicit incorporation of its elements into the method.

References

1. Russell, S.W.: Computer understanding of metaphorically used verbs. American Journal of

Computational Linguistics Microﬁche 44 (1976)

2. Wilks, Y.: Knowledge structures and language boundaries. In: IJCAI. (1977) 151–157

3. Wilks, Y.: Making preferences more active. Artiﬁcial Intelligence 11 (1978) 197–223

4. Fass, D., Wilks, Y.: Preference semantics, ill-formedness, and metaphor. American Journal

of Computational Linguistics 9 (1983) 178–187

5. Russell, S.W.: Information and experience in metaphor: A perspective from computer anal-

ysis. Metaphor and Symbolic Activity 1 (1986) 227–270

6. Fass, D.: Processing Metonymy and Metaphor. Ablex, Greenwich, CT (1997)

7. Indurkhya, B.: Metaphor and Cognition. Kluwer, Dordrecht (1992)

8. Lakoff, G., Johnson, M.: Metaphors We Live by. Chicago University Press, Chicago (1980)

9. Carbonell, J.: Metaphor: A key to extensible semantic analysis. In: ACL. (1980) 17–21

10. Carbonell, J.: Metaphor: An inescapable phenomenon in natural-language comprehension.

In Lehnert, W., Ringle, M., eds.: Stragegies for Natural Language Processing. Erlbaum,

Hillsdale, NJ (1982) 415–434

11. Hobbs, J.: Metaphor and abduction. In Ortony, A., Slack, J., Stock, O., eds.: Communication

from an Artiﬁcial Intelligence Perspective: Theoretical and Applied Issues. Springer, Berlin

(1992) 35–58

12. Wilks, Y.: Ontotherapy: Or, how to stop worrying about what there is. In: NLPCS. (2007)

3–22

13. Quine, W.: Ontological Relativity and Other Essays (The John Dewey Essays in Philosophy).

Columbia University Press, New York (1969)

14. Whorf, B.: Language and logic. In Carroll, J.B., ed.: Language, Thought, and Reality:

Selected Papers of Benjamin Lee Whorf. MIT Press, Cambridge, MA (1956) 233–245

15. Kaput, J.: Representation systems and mathematics. In Janvier, C., ed.: Problems of Rep-

resentation in the Teaching and Learning of Mathematics. Erlbaum, Hillsdale, NJ (1989)

19–26

16. LeBlanc, M., Weber-Russell, S.: A computer model of the role of text integration in the

solution of arithmetic word problems. Cognitive Science 20 (1996) 357–408

17. Russell, S.W.: Verbal concepts as abstract structures: The most basic conceptual metaphor?

Metaphor and Symbolic Activity 4 (1989) 55–60

18. Russell, S.W.: Metaphoric coherence: Distinguishing verbal metaphor from anomaly. Com-

putational Intelligence 8 (1992) 553–574

19. Fillmore, C.: The case for case. In Bach, E., Harms, R., eds.: Universals in Linguistic Theory.

Holt, Rinehart and Winston, New York (1968) 1–88

20. Schank, R.: Conceptual Information Processing. North Holland, Amsterdam (1975)

21. Osgood, C.: The cognitive dynamics of synesthesia and metaphor. In Honeck, R., Hoffman,

R., eds.: Cognition and Figurative Language. Erlbaum, Hillsdale, NJ (1980) 203–238

22. Aarts, J., Calbert, J.: Metaphor and Non-metaphor: The semantics of Adjective-Noun Com-

binations. Max Niemayer, T¨ubingen, Germany (1979)

23. Winston, P.: Learning by creating and justifying transfer frames. Artiﬁcial Intelligence 10

(1978) 147–172

24. Gibson, J.: The theory of affordances. In Shaw, R., Bransford, J., eds.: Perceiving, Acting

and Knowing. Erlbaum, Hillsdale, NJ (1977) 67–82

25. Martin, J.: A Computational Model of Metaphor Interpretation. Academic Press, New York

(1990)

26. Carbonell, J., Minton, S.: Metaphor and Common-Sense Reasoning, Rep. No. CMU-CS-83-

110. Carnegie-Mellon University, Pittsburgh, PA (1983)

27. Suwa, M., Motoda, H.: Learning metaphorical relationships between concepts based on

semantic representation using abstract primitives. In: IJCAI-CANL. (1991) 123–131

28. Barnden, J., Glasbey, S., Lee, M., Wallington, A.M.: Domain-transcending mappings in a

system for metaphorical reasoning. In: EACL. (2003) 57–61

29. Barnden, J., Glasbey, S., Lee, M., Wallington, A.M.: Varieties and directions of inter-domain

inﬂuence in metaphor. Metaphor and Symbol 19 (2004) 1–30

30. Agerri, R., Barnden, J., Lee, M., Wallington, A.: Invariant mappings and contexts in a com-

putational approach to metaphor interpretation. In: IJCAI-MRCS. (2007)

31. Narayanan, S.: Moving right along: A computational model of metaphoric reasoning about

events. In: AAAI. (1999)