Connecting frames and constructions: a case study of ‘eat’ and ‘feed’

William Croft, University of New Mexico

ABSTRACT

Constructional analysis of corpus data can contribute to the analysis of a semantic

frame, as demonstrated by a small corpus study of ‘eat’ and ‘feed’. The
‘eat’/‘feed’ frame forms part of a taxonomy of frames including the superordinate
frame of ‘consume’ and subordinate frames of human vs. animal eating;
constructional and metaphor data in the corpus shows that English covertly
distinguishes human and animal eating. The EAT frame includes three phases
(intake, process, and ingest), differentiated lexicogrammatically. The EAT frame
also includes three domains in its domain matrix: physical, biological (nutritional)
and social, all clearly differentiated by distinct constructions in the corpus. An
examination of metaphors with eat and feed in the corpus demonstrate that the
target domain contributes image-schematic structure to the metaphorical mapping,
contrary to the Invariance Hypothesis.

1. Introduction

Frame semantics and construction grammar complement each other. Frame semantics
is a theory of how concepts are organized and represented in the mind. Basically, human
experience provides frames of interconnected concepts, such as an automobile and its
operation, a wedding, or a bodily experience such as eating. Concepts relate to the
experiential frames as a whole: ‘by the term “frame” I have in mind any system of
concepts related in such a way that to understand any one of them you have to understand
the whole structure in which it fits’ (Fillmore 1982:111). A semantic frame is essentially
identical to what other cognitive linguists have called a domain: ‘semantic units are
characterized relative to cognitive domains, and…any concept or knowledge system can
function as a domain for this purpose’ (Langacker 1987:63).

A simple example of the importance of frames to semantic analysis can be illustrated

with words related to the semantic frame that will be examined in this article. The words
flesh and meat appear to denote the same substance. They differ semantically in the frame
in which they are defined: flesh is defined in the domain of a human or animal BODY,
while meat is defined in the domain of FOOD. Conversely, a single word may be
ambiguous because it has slightly different denotations in two different semantic domains
in which it is used. For example, chicken and lamb denote a whole, usually live, animal in
the ANIMAL frame, but a kind of meat (i.e., just that part of the slaughtered animal) in
the FOOD frame. The combination of a lexical contrast (as with flesh vs. meat) and a
slight difference in denotation (as with a chicken vs. chicken) is found in the lexical pairs
deer/venison and cow/beef.

Construction grammar, in contrast, is a theory of how grammatical structures are

represented in the mind. Constructions are symbolic units, a pairing of form and meaning,
where both form and meaning are construed broadly: the former including morphology,
syntax and even phonology and prosody, and the latter including semantics, information
structure/discourse function, and also social parameters of use. Construction grammar
was motivated by the fact that many grammatical structures are idiosyncratic with respect
to the highly general rules of syntax posited by generative grammar; and yet these
structures, such as the X let alone Y construction (Fillmore et al. 1988), and their semantic
interpretations are productive and rule-governed in their own distinctive right. Indeed, all
grammatical units can be represented as constructions, from words and morphemes
(atomic, substantive constructions) to the most general syntactic rules and associated
rules of semantic interpretation (complex, totally schematic constructions).

The emergence of construction grammar has led to a number of studies of different

grammatical constructions from a variety of perspectives. Some of the studies make
reference to frame semantics. Relatively few however provide us with a detailed
description of the semantic frame that underlies a particular construction; Fillmore and
Atkins’ (1992) case study of the argument structures of risk and the associated RISK
semantic frame are a notable exception. This article will develop an analysis of the
semantic frames associated with eat and feed, by examining the constructions with which
these two verbs are associated. Much of the analysis will be based on data from an
approximately 2 million word corpus of UPI top news stories from the late 1980s and
early 1990s.1 This corpus yielded 245 tokens of eat and 177 tokens of feed.

The goal is to demonstrate how a constructional analysis of syntax, in particular one

based on a corpus of naturally occurring data, even a small corpus such as this one, can
contribute to our understanding of semantic frames. One can exploit syntactic regularities
in order to identify the most salient, prototypical or essential properties of a particular
concept and its frame/domain. In particular, for verbs such as eat and feed, not only their
arguments but their optional modifiers (adjuncts) can provide much information about
frame semantic structure. The use of corpus data allows one to make fine-grained
generalizations about the syntactic relations between a verb, its arguments and its
adjuncts. Finally, the examination of this data raises issues about the nature of
frame/domain structures and relationships, in particular metaphorical relations between
frames/domains.

2. Taxonomic relations of the EAT frame

The verbs eat and feed are taken here to belong to a single semantic frame, the EAT
frame—though what is meant by a ‘single’ semantic frame is actually an important and
problematic issue. We will begin by observing that eating and feeding involve the same
physiological activity, which will be described in greater detail in the following sections,

1
I am grateful to Stanley Peters for making this corpus available to me and supporting
the initial research on this topic during a sabbatical at Stanford University in 1993. I am
also grateful to audiences at the Department of Linguistics and the Institute of Cognitive
Science at the University of Colorado at Boulder, and at the Department of English at the
University of Pisa, for their comments on earlier versions of the material presented here. I
remain solely responsible for any errors in the analysis or conclusions.
although feed in some of its argument structure constructions also involves a causative
and/or transfer relation not presupposed in the scenes evoked by eat (see §5).

The first problem in frame semantic analysis that we encounter is the taxonomic
definition of frames. The problem can be illustrated with the semantic range of words
used for the basic process in the EAT frame. English contrasts eat with drink. This lexical
distinction suggests that there are two distinct semantic frames, EAT and DRINK.
Bengali, on the other hand, has a single verb which is used for EAT, DRINK and also
SMOKE. German, on the other hand, has two verbs for EAT, one for animals (fressen)
and one for humans (essen), as well as DRINK (trinken). The German animal verb can be
used with humans subject referents, to describe a particular way of eating (more
repulsive, sloppy, etc.). This suggests that there are two different frames for EAT in
German, which we can describe as A-EAT and H-EAT.

How many frames are there? There is a taxonomic semantic relationship between the
range of activities covered by the verbs in the different languages:

Figure 1. Taxonomic relations of the EAT frame

CONSUME (Bengali)

EAT DRINK (English)

A-EAT H-EAT DRINK (German)

Taxonomic relations between frames are different from the relation between a
concept and its frame, although any concept can function as a frame. For example, the
concept RADIUS is understood in the frame CIRCLE; but CIRCLE itself is also a
concept which must be understood in the frame SPACE (or TWO-DIMENSIONAL
SPACE). The relation between RADIUS and CIRCLE and between CIRCLE and SPACE
is what Langacker describes as the profile-base relation (Langacker 1987:183; this is the
same as the concept-frame relation). Any concept can function as either a profile (concept
in the usual sense) or a base (i.e., a frame). The profile-case or concept-frame relation is
not a taxonomic one: as the quotations from Fillmore and Langacker above indicate, the
base/frame for a profile/concept is the set or system of concepts that are presupposed by
the profile/concept. EAT and DRINK are not presupposed by CONSUME; they are
simply subtypes of consumption. In other words, concepts are organized in at least two
different ways in the mind: in a profile/base relation (concept/frame), and in a taxonomic
relation (hypernym/hyponym).

The crosslinguistic pattern diagrammed in Figure 1 might lead one to believe that
each of the three languages has a different set of semantic frames (one, two, or three) for
this family of human activities. In fact, however, there is evidence from constructions that
English, at least, has at least the two lower taxonomic levels for the EAT frame. (I expect
that evidence from constructions will also show that the highest level CONSUME frame
exists in English as well, but this corpus study does not include DRINK.)

Certain constructions, notably the transitive argument structure construction, are

found with both human and animal eating (the corpus examples were extracted by a
KWIC concordancer with a 160-character or 200-character window):

(1) EaterSBJ eat FoodOBJ

WASHINGTON (UPI) -- Ralph Nader admits he ATE a sugar-loaded ice

cream bar, but the uncompromising consumer advocate claims he was
entrapped. Vanity

y rooted in the jaw bone to let the bear kill and EAT seals. Removing it took
a lot of digging and the tooth came out in several pieces, Kotz said. The
operation

The transitive construction, which has a more or less uniform meaning across human
and animal eating, implies the taxonomic level of the EAT frame implied by the English
word eat. The participant roles for eat are defined in terms of the EAT frame, that is,
Eater and Food, not Agent and Patient. This level of precision is found in both the frame
semantic and construction grammar literature on argument linking (e.g., Fillmore and
Atkins 1992; Goldberg 1995; see also Croft 1991), and will be necessary for
differentiating the constructional patterns for the EAT frames observed in the corpus.

In contrast to the transitive construction in (1), some constructions with eat are found
only with animal eaters:

(2) FoodOBJ-eating A-EaterNOM

ied, crew members said. ``It is the biggest meat-EATING animal that ever
walked on this earth,'' museum Curator Jack Horner said while strolling
area

a government officials on whether they wanted oil-EATING microbes to be

airlifted to the Persian Gulf. ``We're just on a holding pattern. We're
making

(3) A-EaterSBJ feed (on FoodOBL/off of FoodOBL)

but what you're going to have is birds FEEDING and moving and coming in
contact with it,'' he said. The Audubon Society also was gathering lists of
 mercial fisheries. The perch thrived by FEEDING ON the 300 species of
mostly indigenous cichlids, leaving an estimated 150 species endangered.
In June 1988,

e fish, the clams. Everything out there FEEDS OFF OF that.'' Of immediate
concern was a nesting colony of tens of thousands of laughing gulls on
Pelican Island

(4) Time-feeding A-Eater

indoors to reduce exposure to the night-FEEDING mosquitos that carry the

virus. Florida has recorded 65 confirmed cases and three deaths since the

Constructing clauses or phrases using the constructions in (2)-(4) with human eaters
leads to odd results: although we would not reject them as outright unacceptable, they are
not anything a speaker would say, or they strike one as funny or satirical, not unlike the
use of German fressen with a human subject referent:

(5) a. pizza-eating teenagers

b. vegetable-eating children
(6) I saw that the Joneses were feeding, so I didn’t ring the doorbell.
(7) night-feeding graduate students

Example (8), on the other hand, is acceptable, which suggests that the human eating
frame really depends more on conventions and manner of eating rather than the species of
the eater:

(8) The baby was feeding.

The same appears to be true with the difference between EAT and DRINK. Levin
writes, ‘eat involves ingesting solids and drink liquids’ (Levin 1993:214). However,
some things that one eats in English are not very solid (oatmeal, jello), and other items
can be eaten or drunk: one can eat soup or eat yogurt, or drink soup or drink yogurt.
Again, it appears to be conventions and manner of ingestion rather than solely the
physical state of the thing ingested. Nevertheless, manner of ingestion does correlate with
the physical state of the thing ingested: it would be a challenge to drink a steak or eat
milk.

3. The basic eating process

The basic eating process in the EAT frame involves three phases: intake, processing
and ingestion. There does not seem to be specialized verbs in English for simply putting
something in one’s mouth; eat and drink are general verbs that appear to cover the entire
process (see below). Levin (1993) gives a subclassification of verbs of ingestion (in the
broad sense) based on their distribution in English argument structure alternations.
Levin’s Chew verbs profile the processing phases of the EAT frame (Levin 1993:214):
(9) Chew Verbs: chew, chomp, crunch, gnaw, lick, munch, nibble, pick, peck, sip,
slurp, suck

Levin’s Gobble and Devour verbs profile the ingestion phase of the EAT frame
(Levin 1993:214-15; the subclasses differ in that former can occur with the particle down
and have zero-derived nominals in some cases):

(10) Gobble Verbs: bolt, gobble, gulp, guzzle, quaff, swig, swallow, swill, wolf
(11) Devour Verbs: consume, devour, imbibe, ingest, swill

Levin notes that Chew verbs, as well as eat, allow the Conative alternation while the
Gobble and Devour verbs do not (Levin 1993:213-15):

(12) Cynthia ate at the peach.

(13) Cynthia nibbled at/on the carrot.
(14) *Cynthia gobbled at/on the pizza.
(15) *Cynthia devoured at/on the pizza.

The difference in argument structure is due to the fact that the initial phases (intake
and processing) can be done incrementally, so that part of the Food has not begun the
process; but ingestion is accomplished after the processing phase and generally is
achieved completely.

There are distinct English verbs for the inverse of the processing and ingestion
phases: spit out and throw up respectively. This intuition is supported by the following
introspective judgements:

Processing verbs:

(16) a. He chewed (on) the meat but then spit it out.

b. *He chewed on the meat but then threw it up.

Ingestion verbs:

(17) a. *He gobbled the meat (down) but then spit it up.
b. He gobbled the meat (down) but then threw it up.
(18) a. *She devoured the meat but then spit it out.
b. She devoured the meat but then threw it up.

The behavior of eat in these contexts is more problematic, at least to this speaker:

(19) ?He ate the meat but then spit it out.

(20) She ate the meat but then threw it up.

Example (20) fits with the meaning of eat as covering the entire process. Example
(19), if acceptable, would support an ambiguity in which eat also profiles only the intake
phase. Example (19) is sufficiently acceptable to me, and (21) seems fully acceptable:
(21) She drank the juice but then spit it out.

So it appears that eat and drink may also profile the initial intake phase as well as the
whole process.

4. The domain matrix of eating

Probably the most important complication with regard to frame semantic analysis are
domain matrices. Langacker (who uses the term ‘domain’ rather than ‘frame’) points out
that concepts simultaneously inhabit multiple domains, the combintion of which he calls
a domain matrix (Langacker 1987:147). For example a BOOK is both a physical object
and an entity with meaningful content (Cruse [2000] describes these as different facets of
the meaning of BOOK). Some domain matrices are more tightly bound together than
others. Langacker argues that the domain matrix for colors includes the three
psychophysical domains of HUE, BRIGHTNESS and SATURATION (Langacker
1987:150-51). Domains that are this tightly bound together are called dimensions by
Langacker (1987:150-52). Nevertheless even these dimensions are separable, in that
BRIGHT and DULL are profiled against the BRIGHTNESS dimension or domain, and
LIGHT and DARK are profiled against the SATURATION domain. Some domains entail
other domains. A CIRCLE as an abstract geometric concept is profiled against SPACE
alone. But any physical object is profiled against a domain matrix of both SPACE and
MATTER; moreover, MATTER only exists in SPACE. Careful analysis of domains or
frames reveals that most concepts are profiled against a domain matrix of multiple
domains that are otherwise separable (Croft 1993; Clausner and Croft 1999).

The EAT frame consists of at least three clearly distinguishable domains (frames).
The first is the purely physical frame which was described in the phases above: intake,
processing, and ingestion which includes the destruction of the object. This domain in the
EAT matrix involves both spatial concepts (movement into the body) and material
concepts (the material breaking down and destruction of the FOOD).

The second domain is the biological domain, in particular the nutritional cycle. Food
is nutrition: it nourishes the person or animal. However, nutrition is used up and so the
nutritional process must be iterated regularly.

The third domain is the domain of social activity. There are certain culturally defined
combinations of food. Eating is typically performed at particular times (i.e., meals).
Eating is also often performed with other people. Finally, eating is generally associated
with certain places (restaurants, dining areas).

All of these aspects of these three domains in the domain matrix of EAT are
manifested in the distributional patterns of EAT constructions and concepts in the corpus
used in this article. The most obvious manifestation of the distinct domains is in the
words used to describe the Food participant role. These words can be described as falling
into the categories of a Simple Food Item, used for the physical process (22); a Nutrient,
used for the biological cycle (23); and a Meal, used for the social activity (24):
(22) Simple Food Item:

rsing Home in Palatka, where more than 200 people ATE ICE CREAM and
BIRTHDAY CAKE under a tent at an afternoon lawn party. Mayor Tim
Smith proclaimed

(23) Nutrient:

lements, the report said. Lactating women should EAT at least 1,800
CALORIES a day and no less than 1,500 calories daily and should be
discouraged from

t, no matter where they get their lunch, students EAT too much FAT.
``Pizza, hamburgers and fried chicken dominate the main course
options, and french fries,

(24) Meal:

ave her jewels and furs at home. She will have to EAT standard jail FARE:
frozen TV DINNERS. ``We haven't won any awards, but our meals have
been approved by the county Department of Health

In addition, Levin’s subclassification of Ingestion Verbs reveals verb classes defined

by their argument structure distributions that are specific to each of the domains. I
described the verb classes specific to the physical domain in §3. Levin’s Gorge verbs are
specific to the biological nutritive cycle (Levin 1993:216):

(25) Gorge Verbs: exist, feed, flourish, gorge, live, prosper, survive, thrive

(26) a. Cynthia gorged on peaches.

b. They lived on potatoes for years.
c. They thrived on a diet of ice cream and chocolate.

The Gorge verbs describe different degrees of supplying oneself with nutrition and
the degree of success in obtaining enough nutrition. Levin notes that they require a
second argument describing the Nutrient (diet, in her terms), as well as the Eater, and
they do not occur in the Conative construction characteristic of the initial phases of the
physical process.

Levin’s Dine verbs are specific to the social activity of eating (Levin 1993:215-16):

(27) Dine Verbs: banquet, breakfast, brunch, dine, feast, graze, lunch, luncheon, nosh,
picnic, snack, sup
(28) a. They snacked on fresh strawberries.
b. They lunched on scallops, crab and Johannisberg Riesling.

Levin notes that the Dine verbs are zero-related to nouns describing meals or other
eating events (e.g. snack). They can be used intransitively, simply describing partaking of
the meal, though they also allow a prepositional phrase argument that describes the food
items making up the meal, which usually are constrained by cultural convention (or are
explicitly noted violations of that convention).

Most attention to frame semantic analyses of verbal concepts focus on arguments.

However, adjuncts can be at least as informative as arguments in defining the salient
properties of a semantic frame. Adjuncts are often lumped together as being optional, and
occurring multiply. But if we examine a corpus such as the one used in this study, we
find that adjuncts, while optional, are specific in meaning and highlight relevant
properties of the semantic frame. Multiple adjuncts each describe different and
complementary semantic structures in the frame.

For example, the corpus contained a number of adjuncts referring to the nutritional
cycle of eating, illustrated in (29)-(32) below:

Nutritional quality

(29) Adjective eatingNOUN Noun: careless eating habits, eating disorders, healthy
eating guidelines

(30) eat QualityAdverb: eating right, eating healthy, eat well

Frequency/regularity

(31) eatingNOUN Noun: eating habits, eating patterns, eating schedule

(32) eat FrequencyAdverb: eating regularly

These adverbs refer to the two salient aspects of the biological nutrition domain/frame
in the EAT matrix mentioned above: eating provides nutrition to the eater, and the
nutritional process must be repeated on a regular basis.

There were two apparent exceptions to the constructions found with eating and
nutrition in the corpus. In (33), a Meal word, characteristic of the social activity domain
in the EAT matrix, is used in combination with a Frequency Adverb, characteristic of the
biological nutrition domain in the matrix:

(33) a six-month supply of food for a force of 400,000 EATING THREE MEALS A
DAY. The contracts were let this week to seven different food suppliers
throughout
 However, the context clearly indicates that the domain of the discourse is the
nutritional needs of the soldiers (in the first Gulf war): the number of meals indicates an
adequate amount of nutrition.

The second appaarent exception is the use of a Rate Adverb, which is unexpected in a
nutritional context:

(34) little more time with nurses because he is still EATING FAR MORE SLOWLY
THAN his sisters. Should the triplets have a sweet tooth, they're in luck.
They've got a dad

The rate of eating is not usually relevant to the nutritional domain in the EAT matrix.
In this case, however, the rate of eating is crucial to the nutrition for the baby, unlike, say,
an adult eating slowly at the dinner table on one particular occasion.

In addition to these examples of particular adjuncts, there are clausal constructions

that appear to be specific to the biological and social domains in the EAT matrix. One
construction specific to the biological domain in the matrix is a construction with diet as
a direct object but requiring a modifier indicating a nutrient or nutritional value:

(35) EaterSBJ eat NutrientMOD diet

ur cancer risk, and at worst, you are going to be EATING a HEALTHY DIET,''
he said. The 88,751 female nurses in the study ranged in age from 34
and 59 at

nefit. During the study, both twins in each pair ATE DIETS THAT PROVIDED
AN AVERAGE OF ABOUT 900 MILLIGRAMS OF CALCIUM DAILY. However, tablet

(36) a. *Danielle at a diet.

b. ?*Danielle ate the diet. [must refer to a previously-specified diet to be
acceptable at all]

There are several constructions in the corpus that are specific to the social activity of
eating in the EAT matrix. The comitative adjunct, naming a person with whom one eats,
occurred in the corpus only in constructions when the direct object was a Meal
expression, not a Food Item or Nutrient expression:

(37) EaterSBJ eat MealOBJ with Co-EaterOBL

Seife lied to investigators when he said he never ATE meals WITH drug
makers. Prosecutors said Seife often enjoyed lunches, dinners and drinks
at the

Also, an adjunct phrase indicating an Eating Place only occurred in the corpus when
the direct object was a Meal expression:
(38) EaterSBJ eat (MealOBJ) at Eating-PlaceOBL

his staff. Cost, $301.04. On Oct. 28, 1990, Frank ATE dinner AT the Ritz
Carlton, Dearborn, Mich., WITH Ford, David Geiss, Ford's administrative
assistant,

son said. ``We could have saved $90 billion if we ATE those meals AT
home.'' He said restaurant food tends to be high in fat and sodium, as
does processed

Two more substantive (i.e. not schematic) expressions, eat out and eat in peace, are
apparently specific to the social activity domain in the EAT matrix (note that the latter
example has a Meal expression as direct object):

(39) Asked how she was coping, Amati replied, ``We're EATING OUT. We just
feel very isolated.'' Hundreds of Consolidated Edison workers swarmed
through Fulton

(40) you'd long for the good old days, when you could EAT your lunch IN PEACE.
But Jose can't hide. Which brings us to the World Series that the A's have
so far

These cooccurrence patterns are not thought of as grammatically obligatory. Most

English-speaking linguists would not reject a sentence such as I ate a pizza with Carol as
unacceptable. But it appears that one would rarely if ever actually say such a thing. This
is because of the clash between the semantics of the word chosen for the Food role and
the social activity domain in the EAT matrix that is selected by use of the comitative
construction. Grammatical unacceptability is only a weak indicator of the actual
grammatical patterns in language use that tell us about the semantics of the words and
constructions of the language.

5. Feeding and the EAT frame

The verb feed is a three-argument verb. Although it can be semantically analyzed as a

causative verb of eat (and was included in this study for this reason), its syntax resembles
that of a transfer verb such as give or send. Feed occurs in the Dative alternation (cf.
Levin 1993: 217, exx. 665a-b):

(40) refully check all baby food jars before FEEDING it TO their children for
evidence of tampering and to make sure that the vacuum seal has not
been broken.

(41) are also a sure-fire fattener and that FEEDING their pets small servings of
low-fat pig chow is the only way to keep them trim. The North American
Potbellied
 However, unlike transfer verbs but like its anticausative counterpart eat, feed allows
unspecified object drop (cf. Levin 1993:216, ex. 664):

(42) irlift. ``Instead of one city, we must FEED people scattered through some
of the least accessible, most remote points on earth,'' Rosenblatt told the
senators.

(43) o at the queen's disposal. Elizabeth declined to EAT during the reception,
but did drink a Beefeater martini, said Kim May, a waitress who worked
the private

(44) *Vera gave/sent Fred. [unacceptable when Fred is interpreted as the recipient]

This last fact makes feed look more like a causative of eat. Thus, the syntactic
distribution of feed partly partakes of the semantics of transfer, as well as partaking of the
semantics of eat as the causative of eat.

6. Metaphors with eat and feed

The preceding sections offer a detailed frame-semantic analysis of the literal (source
domain) concepts EAT and FEED. By using the constructions in which eat and feed
occur, and examining both arguments and adjuncts that occur with eat and feed, we were
able to identify the taxonomic structure of the domain/frame, the temporal phrases of the
central process in the domain/frame, and the domain/frame matrix structure. Armed with
this fine-structured analysis of the source domain, we can come up with a more fine-
grained analysis of what is going on with metaphorical mappings of EAT and FEED that
occurred in the corpus.

In §2, I argued that certain constructions revealed a covert distinction between

Animal-Eating and Person-Eating in English, although both are described with a single
lexical item eat. The constructions used to distinguish Animal-Eating from Person-Eating
occur in certain metaphors, and therefore demonstrate that certain metaphors have
Animal-Eating as their source domain, not Person-Eating or Eating in general. Two
Animal-Eating metaphors found in the corpus are given in (45)-(46):

(45) EXPLOITATION IS ANIMAL-EATING

Clinton said. ``These people have been FEEDING ON poor folks for a
decade now and they have no intention of doing anything for them.'' He
said his

(46) STIMULATION OF EMOTION IS ANIMAL-EATING

going to take it any more. Their rage FEEDS ON the budget debate that
threatens to shut down the government. It has the rich quaking in their
Guccis.
 The corpus also contained metaphorical examples with aspectual particles such as eat
away at and eat through that are not possible with human or animal eating, or at least not
as easily acceptable:

(47) remained in effect through Sunday. ``The ocean's EATING AWAY AT the
bottom of houses now. If you ask me, the whole coastal area should be a
state park for birds, not people,'' Scura said.

(48) Challenger was destroyed after 5,000-degree flame ATE THROUGH an O-

ring joint in the ship's right-side solid-fuel booster, creating a deadly, out-
of-control

(49) a. *?Penny ate away at the spaghetti.

b. Penny ate through the whole plate of spaghetti [requires or prefers an adjective
of completeness, unnecessary in the metaphorical use]

However, as was pointed out by an anonymous audience member at a presentation of

an earlier version of the content of this article, there is another subtype of literal eating,
which we might call Tiny Animal Swarm Eating, not represented in the corpus, in which
the constructions eat away at and eat through are acceptable:

(50) a. Caterpillars ate away at the foliage of the trees at the edge of the meadow.
b. Termites ate through the timbers holding up the roof.

Hence, (47)-(48) are examples of GRADUAL DESTRUCTION IS TINY ANIMAL

SWARM EATING. The absence of the literal (source domain) uses of eat away at and
eat through is a consequence of using a relatively small corpus that is restricted to one
genre (news stories).

In §4, I argued that the EAT frame is a matrix consisting of three component
domains/frames: the physical process, the biological nutrition cycle, and the social
activity. The following metaphorical use of feed is motivated by the nutritional domain in
the EAT matrix:

(51) REINFORCING CAUSATION IS EATING/NUTRITION

emarkable 60 percent of the white vote, FEEDING rumors that he will run for
governor next year, or for president in 1992. Some strategists believe he
could

New York. The market's uncertainty was FED by ``conflicting news'' on

the Persian Gulf crisis, Bentz said. U.N. Secretary General Javier Perez
de

The target domains here contain scalable entities (rumors, market uncertainty) whose
scale is caused to increase by information that is the causer/agent of metaphorical feed.
This causation of scalar increase meaning of metaphorical feed is probably motivated by
the fact that biological nutrition is also scalar and is boosted by the Nutrient supplied in
literal feeding.

The metaphor illustrated in (52) appears to be a combination of both the physical

domain and the biological domain of the EAT frame/domain matrix:

(52) A CHANNEL IS FEEDING/MOTION+NUTRITION

r what caused the fire. Major roadways FEEDING the City Hall area from
north, south, east and west and the northwest were closed to traffic.
Business in the

who was on the expedition. The lake is FED by some 336 rivers and
streams and is home to an estimated 600 plant species and more than
1,200 animal

These examples involve a complex metaphorical mapping. As with the metaphor in

(51), there is an increase in a scalar property of a scalable entity (traffic, volume of water)
that is caused by the causer/agent of feed. In addition, however, all the examples in the
corpus involve motion from outside the metaphorical Eater (the City Hall area, the lake)
into the Eater. This motion is part of the physical process of feeding in its literal source
domain but it is not highlighted at all in the source domain.

This last observation leads us to a significant conclusion we can draw for

metaphorical mappings from the detailed frame-semantic and constructional analysis of
eat and feed. One question in linguistic theories of metaphor has been: what sort of
properties of the source domain are mapped onto the target domain? Concrete properties
of objects and actions in a source domain are rarely if ever mapped to the target domain.
Instead, what is mapped is what is loosely called ‘abstract’ structure, which has been
analyzed in cognitive linguistics as image-schematic structure. Lakoff (1990) describes
this as the Invariance Hypothesis: ‘Metaphorical mappings preserve the cognitive
topology (that is, the image-schematic structure) of the source domain’ (Lakoff 1990:54).
For example, in the metaphor LOVE IS A JOURNEY (Lakoff 1990:47-51), the
expression We’ve hit a dead-end street maps the image-schematic concept of a path that
ends, not the more concrete concepts of the absence of asphalt at that point, or a signpost
stating “Dead End”.

Turner (in collaboration with Lakoff) provides a refinement of the original statement
of the Invariance Hypothesis, which recognizes that some image-schematic structure of
the source domain is not necessarily mapped into the target domain (Lakoff [1993:215]
has a similar reformulation):

In metaphoric mapping, for those components of the source and target domains
determined to be involved in the mapping, preserve the image-schematic structure
of the target, and import as much image-schematic structure from the source as is
consistent with that preservation (Turner 1990:254)
 Turner explains that this reformulation is intended to take care of two types of
situations. The first is that some image-schematic structure from the source domain is
simply not mapped into the target domain. He gives the example of our boss described as
a crab: although there is image-schematic structure in the physical image of the crab, it is
simply not mapped into the target domain (Turner 1990:253). Gentner et al. (2001:207)
argue that this is actually a serious weakening of the Invariance Hypothesis, since one
cannot predict what image-schematic structure is mapped into the target domain; but they
acknowledge that large-scale image-schematic structures are mapped into target domains,
as predicted by the mapping theory of conceptual metaphor, even if it is not all image-
schematic structure from the source domain.

The second is that if there is image-schematic structure in the target domain, the
metaphorical mapping must map “like to like”, or it is not mapped. For example, Turner
cites a passage from Auden in which a weeping man drooping his head is described as
‘helpless and ugly as an embryo chicken’ (Turner 1990:251); Turner notes that the spatial
orientation of the man’s head and the chicken embryo’s head (both hanging down) must
be preserved and mapped like onto like. If the target domain differs in such a way that it
is incompatible with image-schematic structure in the source domain, then that image-
schematic structure is not mapped into the target domain. Turner gives the example of
LIFE IS A JOURNEY. One can say that one has reached a fork in the path in their life.
Part of the image-schematic structure of the source domain is that a path is fixed and
lasting. Hence, in a journey, one can start down one fork, but then turn back and take the
other fork. However, in the target domain of life, one cannot reverse direction on the path
of life, due to the irreversibility of time and choices once they are made: ‘we cannot take
a step back and be again at the metaphoric fork in the road, because the fork doesn’t exist
anymore’ (Turner 1990:254). Hence the fixed and lasting property of the literal path does
not map into the metaphorical path of life.

The reformulation of the Invariance Hypothesis demonstrates, among other things,

that the presence of image-schematic structure in the source domain is not a sufficient
condition for it to be mapped into the target domain. Also, the image-schematic structure
in the source domain must be mapped into the same corresponding image-schematic
structure in the target domain, if that structure exists. And if the image-schematic
structure of the source domain is incompatible with the image-schematic structure of the
target domain, then it cannot be part of the metaphorical mapping.

These constraints are all necessary for the Invariance Hypothesis, and it has
apparently been assumed that those constraints are all that is necessary to maintain the
Invariance Hypothesis in some form. But there is another aspect of metaphorical mapping
that does not appear to be captured by this reformulation of the Invariance Hypothesis.
This is that image-schematic structure from the target domain comes to be used in the
metaphorical mapping, even if that image-schematic structure does not exist in the
source domain.

This conclusion can be drawn from the corpus because image-schematic structure is
manifested, among other places, in the argument structure constructions of a verb which
is metaphorically mapped. For example, the TIME IS MONEY metaphor includes a
transfer image-schema from the commercial transaction source frame/domain. This
schema is manifested in the Dative argument structure of a metaphorical expression such
as I can give you a half hour of my time. The argument structure provides the mapping of
the Money role into my freedom of action for the time period, the Recipient role into the
person who has control over what I do in that time period, and the Donor role into my
power to decide what to do in that time period. The argument structure encodes the
overall transfer image-schema that these roles are part of.

When we turn to the metaphors with eat and feed in the corpus, we find metaphors in
which the metaphorical mapping follows the Invariance Hypothesis as expected. These
include the examples in (45)-(48) and (51)-(52) above. For example, in EXPLOITATION
IS ANIMAL-EATING, the image-schematic structure of the Eater gaining nutrition and
the gradual destruction of the Food as a consequence is mapped straightforwardly into the
Exploiter (= Eater) gaining benefit or money through the gradual destruction of the assets
of the Exploited (= Food).

In addition to these metaphors, however, we find metaphors using constructions that

are not possible in the literal uses of eat and feed. For example, there is a container
directional (into) phrase that is found in various FEED metaphors that is not found in the
literal expressions:

(53) and Texas. Authorities say rabbits are FED INTO an underground network
that extends across the nation, from New Mexico and West Texas to
greyhound

(54) cording about 800 alligator bellows and FEEDING them INTO computers
programmed to analyze sound frequencies. ``So much energy is produced
from this

(55) Satan.'' ``The other influence, which FED an especially virulent strain of
anti-Semitism INTO the bloodstream of today's Klan was David Duke (of

(56) *I fed a doughnut into the baby.

The same family of metaphors allows an anticausative argument structure for

metaphorical feed, in which the direct object of the transitive construction illustrated in
(53)-(55), and also in the literal use feed in (40), is coded as the subject of an intransitive
feed:

(57) ub-surface porous reservoir, which then FEEDS inward radially TO a vent
where it's concentrated, picks up the dark particles (and) launches at a
velocity of abo

(58) engineering data that would allow us to FEED back INTO the space station
design process.'' The other was to ``pick out at least one of those types
of
 Again, this argument structure construction is not possible with the literal use of feed:

(59) *The ice cream fed into/to the boy in less than five minutes.

These argument structure constructions represent a image-schematic structure in the

target domain which is not found in the source domain, or at least cannot be highlighted
in the source domain by the use of these constructions. These constructions appear to be
sanctioned by the type of motion (sometimes literal motion, sometimes metaphorical
motion) that is found in the target domain in (53)-(55) and (57)-(58). But the sort of
motion in the target domain is of a different type, or is at least more general, than the type
of motion found in the intake phase of the EAT frame (see also (52)). Hence it licenses
constructions such as the container directional into and the anticausative that are found in
(53)-(55) and (57)-(58), but not in literal uses of feed such as (56) and (59). The container
directional into is disallowed in the literal meaning of feed even though the Food does
move into a container (the human or animal body), because the literal meaning employs a
transfer image-schema, not a motion image-schema.

These examples indicate that image-schematic structure from the source domain is
not a necessary part of a metaphorical mapping: image-schematic structure from the
target domain can be part of a metaphorical mapping. This is a stronger challenge to the
Invariance Hypothesis as originally formulated than Turner’s examples that led to his
reformulation. In Turner’s examples, image-schematic structure in the target domain
blocked the mapping of image-schematic structure from the source domain, but it did not
itself participate in the mapping.

The last example I present from the corpus indicates that the constraint that requires
mapping “like to like” does not always apply either. In (60), a compound participial
construction is used with metaphorical feed to describe the target domain entity that the
Food role in the source domain is mapped into, namely rain:

(60) Metaphorical: “Food”-fedADJ

week's earthquakes in California, RAIN-FED floods that put parts of Texas

under water for weeks, and the mudslides in Puerto Rico that alone
swallowed $6 milli

However, when the same compound participial construction is used in the source
domain of feed, the role that is found in the construction is the Source:

(61) Literal: Source-fedADJ

r Milk.'' A British study found BREAST-FED children had ``IQs an average

of 10 points higher'' than children fed cow's milk as babies, he said.
Barnard

In other words, the image-schematic structure represented by the compound

participial construction in the source domain does not map over to the image-schematic
structure represented by the compound participial construction in the target domain. It is
likely that these are two different image-schematic structures in the two domains that
happen to be expressible by the same (compound participial) construction.

These metaphors show that the relationship between source and target domain is more
complex than even the revised version of the Invariance Hypothesis indicates. The
structure of the target domain is intertwined with the source domain metaphorical
expressions more closely than the Invariance Hypothesis indicates: its image-schematic
structure can be imposed on the source domain concept even when the source domain
lacks that image-schematic structure. This fact may be more easily accommodated in the
blending theory model of metaphor (Fauconnier and Turner 2002), in which the metaphor
is more of a blend of the two domains than a one-way mapping from source to target
domain. Nevertheless, all of the metaphorical sentences denote states of affairs in the
target domain, not the source domain. It is not clear how blending theory captures this
fundamental asymmetry in metaphor. It appears that a complete theory of metaphor will
have to accommodate the deep contribution of the target domain to a metaphorical
mapping that is captured by blending theory and also the fundamentally asymmetric
denotation of the target domain by metaphorical expressions that is captured by the
mapping theory.

7. Conclusions

This article has proposed an analysis of the EAT semantic frame, using a combination
of corpus data and introspective judgement. The analysis builds on existing theories of
frames, domains and metaphors, and uses a constructional approach to the syntactic data
in the corpus to illuminate the structure of the semantic frame for EAT. In addition to the
specifics of the EAT frame, some general conclusions about frames and constructions can
be drawn from the analysis.

Frames (domains) are related to one another in several different ways. The basic
relationship is one of profile to base, as in RADIUS-CIRCLE-SPACE. But frames also
exist in taxonomic hierarchies, and constructions may apply at different levels in the
hierarchy (as in EAT vs. ANIMAL-EAT). Frames involving actions such as eating also
can be divided into phases of a sequential action (e.g. intake, processing, ingestion). Most
frames or domains can be broken down into a matrix of domains (e.g. physical process,
biological nutritional cycle, and social activity). The domain matrix functions as the
frame for the whole activity (EAT), but individual domains in the matrix are pulled out or
at least highlighted by particular verbs or constructions. Likewise, single subtypes of a
domain, or a single domain in a matrix, may be the source domain for certain metaphors.

The examination of metaphors in the corpus indicates that the Invariance Hypothesis
associated with the mapping theory of conceptual metaphor does not capture the full
complexity of the possible ways in which image-schematic structure participates in a
metaphor. Not only does image-schematic structure from the source domain play a role in
metaphor (when it is not incompatible with the target domain); image-schematic structure
from the target domain, manifested in the constructions in which the metaphorical words
occur, also becomes part of the metaphor.
 Finally, and most important, this study demonstrates that a careful analysis of the
different constructions used in a naturally-occurring corpus with the words whose frames
are of interest can be extremely revealing. Not only does the corpus analysis confirm our
intuitions about the semantic frame; it reveals a number of subtle distinctions that enrich
the analysis. This is only possible by examining not only the argument structure
constructions with verbs such as eat and feed, but also the semantic subtypes of the
argument expressions, the semantic types of adjuncts, and even the modifiers found with
arguments as well as adverbs modifying the verb. Verbal constructions are more than
argument structure, and the semantic frames of verbs can be greatly illuminated by a
corpus-based examination of all constructions involving them.

References

Clausner, Timothy C. & William Croft. 1999. Domains and image-schemas. Cognitive
Linguistics 10.1-31.
Croft, William. 1991. Syntactic categories and grammatical relations: The cognitive
organization of information. Chicago: University of Chicago Press.
Croft, William. 1993. The role of domains in the interpretation of metaphors and
metonymies. Cognitive Linguistics 4.335-70. (Revised version printed in Metaphor
and metonymy in comparison and contrast, ed. René Dirven & Ralf Pörings, 161-
205. Berlin: Mouton de Gruyter, 2002.)
Cruse, D. Alan. 2000. Aspects of the micro-structure of word meanings. Polysemy:
Theoretical and Computational Approaches, ed. Yael Ravin & Claudia Leacock, 30-
51. Oxford: Oxford University Press.
Fauconnier, Gilles & Mark Turner. 2002. The way we think. New York: Basic Books.
Fillmore, Charles J. 1982. Frame semantics. Linguistics in the morning calm, ed. The
Linguistic Society of Korea, 111-137. Seoul: Hanshin.
Fillmore, Charles J. & Beryl T. Atkins. 1992. Toward a frame-based lexicon: the
semantics of RISK and its neighbors. Frames, fields and contrasts: New essays in
semantic and lexical organization, ed. Adrienne Lehrer and Eva Feder Kittay, 75-
102. Hillsdale, N. J.: Lawrence Erlbaum Associates.
Fillmore, Charles J., Paul Kay and Mary Catherine O’Connor. 1988. Regularity and
idiomaticity in grammatical constructions: the case of let alone. Language 64.501-
538.
Gentner, Dedre, Brian F. Bowdle, Phillip Wolff and Consuelo Boronat. 2001. Metaphor
is like analogy. The analogical mind: perspectives from cognitive science, ed. Dedre
Gentner, Brian F. Bowdle, Phillip Wolff and Consuelo Boronat, 199-253. Cambridge,
Mass.: MIT Press.
Goldberg, Adele E. 1995. Constructions: A construction grammar approach to argument
structure. Chicago: University of Chicago Press.
Lakoff, George. 1990. The Invariance Hypothesis: is abstract reason based on image-
schemas? Cognitive Linguistics 1.39-74.
Lakoff, George. 1993. The contemporary theory of metaphor. Metaphor and thought, ed.
Andrew Ortony, 202-251. Cambridge: Cambridge University Press.
Langacker, Ronald W. 1987. Foundations of Cognitive Grammar, vol. I: theoretical
prerequisites. Stanford: Stanford University Press.
Levin, Beth. 1993. English verb classes and alternations. Chicago: University of Chicago
Press.
Turner, Mark. 1990. Aspects of the Invariance Hypothesis. Cognitive Linguistics 1, 247-
255.