49

Border Conicts: FrameNet Meets Construction Grammar

Charles J. Fillmore
University of California, Berkeley &
International Computer Science Institute, Berkeley
1. Te problem
I count myself among the linguists who believe in a continuity between grammar
and lexicon (Fillmore et al. 1988, Joshi 1983), and I entertain the common image
that each lexical item carries with it instructions on how it ts into a larger
semantic-syntactic structure, or, alternatively, on how semantic-syntactic structures
are to be built around it. My remarks here specically concern an ongoing eort
to describe and to annotate instances of, non-core syntactic structures, and to see
how the products of this work can be integrated with the existing lexical resource,
called FrameNet (FN), which is a set of procedures, and a growing database for
recording the meanings and the semantic and syntactic combinatorial properties of
lexical units. Te FrameNet project, which I have directed since 1997, has recently
begun exploring ways of creating a constructicon, a record of English grammatical
constructions, annotating sentences by noting which parts of them are licensed by
which specic constructions.
Te grammatical constructions that belong in the larger constructiconthat is, in
a construction-based grammarinclude those that cover the basic and familiar
patterns of predication, modication, complementation, and determination, but the
new project is concentrating on constructions that ordinary parsers are not likely to
notice, or that grammar checkers are likely to question. Some of them involve purely
grammatical patterns with no reference to any lexical items that participate in them,
some involve descriptions of enhanced demands that certain lexical units make on
their surroundings, and some are mixtures of the two.
2. Te work, the product, and the limitations of FrameNet
Since many features of the new resource are modeled on FrameNet, I think it useful
to review FNs goals and activities, and the features of its database (Baker et al. 2003,
Fillmore et al. 2003). FrameNet research amounts to
1. describing lexical units (LUs) in terms of the semantic frames they evoke, and
describing those frames (i.e., the situation types, etc., knowledge of which is
necessary for interpreting utterances in the language),
2. dening the frame elements (FEs) of each frame that are essential for a full
understanding of the associated situation type (the frame elements are the props,
participants, situation features that need to be identied or taken for granted in
sentences for which the frame is relevant),
30
Charles J. Fillmore
3. extracting from a very large corpus example sentences which contain each LU
targeted for analysis (FN has worked mainly with the British National Corpus),
4. selecting from the extracted sentences representative samples that cover the
range of combinatorial possibilities, and preparing annotations of them as layered
segmentation of the sentences, where the segments are labeled according to the
FEs they express, as well as the basic syntactic properties of the phrases bearing
the FE,
3. displaying the results in lexical entries which summarize the discovered combinatorial
aordances, both semantic and syntactic, as valence patterns, and creating links
from these patterns to the annotated sentences that evidence them, and
6. dening a network of frame-to-frame relations and the graphical means of
displaying these, that will show how some frames depend on or are elaborations
of other frames.
2.1. Te frames
Te frames developed in FrameNet are the conceptual structures against which the
LUs in the FN lexicon are understood and dened (Fillmore 1982, Fillmore & Atkins
1992, 1994). Tese can be as general as the location of some entity in an enclosure,
or as specic as interest on investment.
One FN frame that is simple enough to describe completely, and just complex
enough to be interesting, is the so-called Revenge frame, the nature of which
requires understanding a kind of history. In that history, one person (we call him
the Ovviuiv) did something to harm another person (what he did we call the
Ovvisi and his victim we call the I,Uviu_v.v1v); reacting to that act, someone
(the Aviuiv, possibly the same individual as the I,Uviu_v.v1v) acts so as to do
harm to the Ovviuiv, and what he does we call the PUisumi1. Tus, we have the
frame Revenge, and the frame elements Aviuiv, Ovviuiv, Ovvisi, I,Uviu_
v.v1v, and PUisumi1. Other features of the Revenge frame include the fact that
this kind of pay-back is independent of any judicial system. Tere is a very large set
of verbs, adjectives and nouns that evoke this frame, by which we mean that when
users of the language understand these words, their understanding includes all of
the elements of that scenario. Among the verbs that evoke this frame are avenge and
revenge, the nouns include vengeance and retribution, there are phrasal verbs like pay
back and get even, adjectives like vengeful and vindictive, support constructions like
take revenge on, wreak vengeance on, and exact retribution against, plus prepositional
adverbials like in retribution, or in revenge.
FrameNet has developed descriptions of over 800 frames to date, and nobody is
ready to estimate how many there are altogether. Te list from the time of the last
o cial release can be found at http://framenet.icsi.berkeley.edu.
31
Border Conicts: FrameNet Meets Construction Grammar
2.2. Te frame elements
Te frame elements (FEs) are somewhat analogous to the deep cases of early Fillmore
(Fillmore 1968, 1971), thematic roles in various generativist writings (Jackendo
1990), actants and circonstants in the Tesnire tradition (Tesnire 1939). Tere
are good reasons for not tying the frame elements into any of the familiar lists of
semantic roles (agent, patient, theme, experiencer, instrument, etc.). Since annotators
are asked to nd expressors of frame elements in actual sentences, FE names that
are memorable in respect to the frame itself will facilitate such identications. Tus
to take the case of the arguments of replace in a sentence like
[I] replaced [my stolen bicycle] [with a much cheaper one],
it makes more sense to refer to the phrases introducing the two bicycles as the Oiu
and the Niw than to try to gure out how well these roles can be accommodated in
the standard lists. (Te missing bicycle, in fact, is not a participant in the event
described by the sentence but is a necessary element of its meaning.) Te recognition
of FE commonalities across frames is made possibly by the system of frame-to-frame
relations.
We wanted to think of the frame elements as representing the kinds of information
that could be expressed in the sentences and phrases in which the frame is active,
and we wanted to be able to discover which parts of a sentence reveal information
about which frame element. Tere is an important constraint on this task,
distinguishing it from annotation practices that seek to learn everything about
each event in a continuous text. Since the information we record is supposed to be
relevant to the syntactic description of a given lexical unit, we require that the frame
elements we attend to are in grammatical construction with the lexical unit being
described. Annotators will ignore event-relevant information elsewhere in the text.
We make a distinction between core and peripheral FEs. Te core FEs are those
that are conceptually necessary in any realization of the frame by the nature of
that frame; the peripheral frame elements are the adjuncts that t the familiar
description time, place, and manner, etc., especially the etc. (the core/periphery
distinction can vary across frames; for verbs like reside, elapse, and behave, the
locative, temporal and manner components, respectively, are not peripheral). A
characteristic of the peripheral FEs is that they have essentially the same meaning
and the same syntactic marking wherever they appear; whatever distributional
limitations they have are explained by the fact that frames about happenings can
take time and place modication, frames about intentional acts can take instrument
and purpose modication, and so on. A third kind of frame element is what we refer
to as extrathematic: these are expressions (like benefactives or phrases like in revenge
or in return) that have the eect of situating the event signaled by the targets frame
in some larger or coterminous situation.
Te goal of FrameNet lexical descriptions is, for each frame-bearing word, to match
the words semantic combinatorial requirements with the manner of their syntactic
realization. Reversing the point of view, we seek to recognize in the syntactic nature
32
Charles J. Fillmore
of the phrases around a given frame-bearing lexical unit, information about the
participants in situation that is an instance of the frame. Te resulting pairing of
semantic and syntactic roles constitutes the valence description of the item.
2.3. Example sentences
Te goal in providing examples was to have, for each lexical unit, a full set of
illustrations of its basic combinatorial properties, and we preferred sentences whose
content was clearly relevant to the meaning of the word being exhibited. If we were
looking for an illustration of knife, we would prefer the butcher sharpened his knife
than the poet photographed a knife. Tese example-selecting decisions were made in
resistance to several kinds of pressure. Some members of the research community
wanted to see sentences of the most frequent type; but for many verbs, the most
frequent examples had mainly pronouns (I risked it). Some wanted us to include
complex and distorted sentences as well as the simplest type; some wanted us to
make sure we include creative uses of a word wherever we found them, scolding us
for neglecting metaphor and other gurative uses: our view echoes that of Patrick
Hanks (MS), namely, that we had the obligation to produce clear descriptions of the
norm, leaving it to some auxiliary research to explore the ways in which speakers
exploit the norm for creative expression. Where a metaphorical use was lexicalized,
the LU resulting from that lexicalization was included in its appropriate frame.
2.4. Te annotation
Te original mission of FN was purely lexicographic: to annotate a variety of typical
uses of each target LU and to seek to cover a wide range of relevant contexts for
the LU (i.e., all of its valence possibilities and representative samples of its semantic
collocates), and this meant creating a collection of sentences in which each was
annotated with respect to one word in it. Tus a sentence like
She smiled when we told her that her daughter had been nominated to receive
an important award.
might be annotated for the verb smile alone, as a member of the Make_faces frame,
where it belongs in the set frown, grimace, grin, pout, scowl, smile, smirk.
As the size of the lexicon increased, it became clear that there were sentences for
which FN was prepared to describe many of the words in it, and ultimately we
received a subcontract to look into the possibility of producing full text annotations.
Tat meant annotating each word in the sentencethat is, each frame-evoking
word. For the above example, that would mean showing the frame structure of the
words smile, tell, daughter, nominate, receive, important and award. For our purely
lexicographic purposes, we would have no reason to annotate the word told in this
sentencewe already have more than enough examples of the lemmabut it would
have to be done here again in order to prepare the semantic structure of the sentence
as a whole. Obviously this need increased our eagerness to nd ways of automating
parts of the annotation process.
33
Border Conicts: FrameNet Meets Construction Grammar
FrameNet has to date annotated a growing number of texts, some of them viewable
on the FN website. Most of them are only partially annotated, partly because they
contain lexical material FN has not yet worked through, and partly because they
contain meaningful grammatical patterns that FN annotation has not been prepared
to capture.
1
Te annotations themselves are presented in layered stand-o representation in
multiple layers. For lexicographic annotations, one layer identied the target LU
and its frame; another represented the FEs in the phrases that serve as its valents;
one indicated the phrase types of the constituents so identied; one indicated the
grammatical function of each valent; and a few other layers were dedicated to
special features associated with individual parts of speech. Te FEs were annotated
manually, the GF and the PT labels were attached automatically and checked
manually. Annotations viewable on the FrameNet website show only the frame
element labeling, as in Figure 1.
[
Fluid
Te River Liey] FLOWS
Target
[
Source
from west] [
Goal
to east] [
Area
through
the center of the city] [
Goal
to Dublin Bay].
Figure 1: FE annotation of a sentence
Full text annotations consist of sets of layers, each corresponding to one target
LU. It is virtually impossible to get a view of the full annotation of a long sentence,
but there is some experimental work being done to derive dependency trees from
these, with the nodes indicating lexical heads and their frames, the branches labeled
according to the frame element represented by the dependent nodes.
One special feature of FN annotation is the recording of FEs that are conceptually
present but syntactically missing. Tese are sorted into constructional null, such
as the missing subject of an imperative sentence; indenite null, such as the object
of intransitivized eat, sew, bake, etc.; and denite nulls (zero anaphora), entailing
that the missing element has to be recoverable in the context, such as the missing
object of we won (what is understood but unexpressed is the contestnot the prize),
the missing preposition phrase in she arrived (where the destination has to be
known) or mine is similar (where the unexpressed comparand has to be part of the
conversation), and so on. Te last of these plays an important role in construction
annotation as well. Such information is associated with the annotation of the LU
that licenses the omission.
1
Te textschosen because other researchers are examining them as wellwere taken from
the Wall Street Journal section of the Penn TreeBank, the Nuclear Text Initiative website, and
a selection of Berlitz Travel Guides that have been made available to the American National
Corpus.
34
Charles J. Fillmore
2.6. Te entries
Each LU is identied by lemma, part of speech, and frame name. Te LUs were
chosen because of their membership in one of the frames being covered by
FrameNet, and what that means is that in many cases the most common use of a
lemma is not to be found: FN researchers have not reached that frame yet. Almost all
features of the lexical entry are produced automatically: handmade features include
a simple denition.
2
For valence-bearing words, the entry contains a table showing
the ways in which each frame element can match a phrase type, and a separate table
showing the variety of ways in which combinations of FEs and PTs make up the
valence exhibited by individual sentences. Viewers of the valence descriptions can
toggle between core FEs only, or all FEs found in the sentencescore, peripheral,
and extrathematic.
Te entries for nouns that designate events or states of aairs also include
information about the existence of support verbs and support prepositions; access
to the sentences will reveal which FEs are represented among the arguments of the
LUs verbal or prepositional support.
3

2.7. Frame-to-frame relations
Since frames can dier from each other in granularity, and some frames are clearly
related to other frames, it has proved necessary to create an ontology of frames,
linked to each other by several kinds of relations. Figure 2 is a display of the frame
relations centered on Commercial_transaction:
Figure 2: Frame-to-frame relations centered on Commercial_Transaction
Several dierent kinds of relations can be seen in this diagram. Commercial_
transaction has two components (related to the mother node by a Part_of relation
2
Te purpose of the denition is purely mnemonic, to aid the user in knowing which sense of
a word is being analyzed in a given entry. Where appropriate the denitions were taken from
the Concise Oxford Dictionary 10, with permission from Oxford University Press. Others
were in-house.
3
Te current database shows no way of classifying support constructions along the line of
the lexical functions of the MTT model of Igor Meluk and his colleagues, though various
researchers are seeking to derive such information automatically from the FN annotations.
(Rambow et al., MS, Bouveret & Fillmore, MS)
33
Border Conicts: FrameNet Meets Construction Grammar
as indicated by the broken line), and these are Commerce_goods_transfer and
Commerce_money_transfer. Each of these is a type of (=has an Inherits relation
to) the frame Transfer. Te low frames Commerce_buy and Commerce_sell have
separate Perspective_on relations to Commerce_goods_transfer, and the frames
Commerce_pay and Commerce_collect have Perspective_on relations to Commerce_
money_transfer. Tus, a commercial transaction is an instance of Reciprocality,
involving two co-occurring reciprocal transfers, one of goods and one of money.
Buying and Selling are perspective-varying instances of goods-transfer, diering
from the point of view of the buyer and the seller; and similarly with paying and
collecting (=charging) and their relation to money-transfer.
3. FrameNet treatment of multiwords so far
Te constructicon-building work concerns itself with linguistic knowledge that goes
beyond simple grammar and simple words, and hence it will include various kinds
of idioms and other multiwords. Tere are many kinds of multiwords that already
fall within the scope of FrameNet work.
4
Among the multiwords covered by current
FrameNet
3
we nd
1. phrasal verbs, with particles, which are simply treated as two-part verbs
that take a specic particle as a syntactic valent; the particle is more or less
motivated, but cant be understood as simply contributing its own meaning
a. Intransitive: pick up (increase), take o (start ying)
b. Transitive: take up (consider), take o (remove)
2. words with selected prepositional complements, listed with preposition,
syntactically selects P-headed phrase
a. Verbs: depend on, object to, cope with
b. Adjectives: fond of, proud of, interested in
c. Nouns: fondness for, pride in, interest in
3. support constructionssyntactically separate, treated as evoking a frame
linked to the noun rather than the verb
a. Verbal heads: take comfort in, take pride in, put emphasis on
b. Prepositional heads: at risk, in danger, under arrest
4. combinationscombining selected prepositional complement with particle
or noun
a. put up with (tolerate), break in on (interrupt)
b. take comfort in, place emphasis on
c. take into possession, take under consideration
4
Josef Ruppenhofer delivered a paper on this topic at an earlier Euralex meeting (Ruppenhofer
et al. 2002).
3
FN treatment of compound words has more or less awaited the capability of constructional
annotation. In the current databases, there are compounds that are simply treated as single
unanalyzed units, and there are others in which the head is a frame-bearing word and the
modier is labeled as an FE in the heads frame. FN has lacked the means of describing a
compound word both as a unit on its own and as having an internal structure.
36
Charles J. Fillmore
3. transparent nounsthe rst noun in [N of N] structures signifying
types, aggregates, portions, units, measures, epithets, etc.; the motivation
for recording these is to be able to recognize selectional or collocational
relations between the context and the second noun
a. my gem of a wife, in a part of the room, on this part of the shelf,
wreak this kind of havoc.
4. Full-text annotation and the confrontation with constructions
In carrying out full-text annotation the goal was to end up with structures which
could be the basis of the semantic integration of the whole sentence. Working with
one of those linguist-invented sentences like
Te Secretary ordered the Committee to consider selling its holdings to the
members
we should be able to identify straightforwardly the participants in the ordering event:
the Secretary gave the order, the Committee received the order, and to consider selling
its holdings to the members, species the order. For the verb consider, the entity that
was to do the considering was the Committee, and selling its holdings to the members
was to be the content of such considerations; and the three participants in the selling
event are to be the Committee as seller, the members as buyer, and the holdings as the
asset destined to change ownership. Te words Secretary, Committee and members
are all relational nouns used without any indication of what the other term of the
relation is, and thats possible if that other entity is understood in the context. A
simple frame-annotated dependency tree will fairly well capture the meaning of
the whole, with word-frame pairs making up the node labels, the branches labeled
according to the semantic role, and with the missing entities in the relational nouns
marked with the possibility of indexing them to contextually given entities.
One doesnt have to look far to nd sentences containing structures that do not lend
themselves to such simple treatment. Here are the rst three sentences of a leader
from the Economist newspaper of June 17, 2007, with comments on those features
that go beyond simple lexicon and simple grammar.
For all the disappointments, posterity will look more kindly on Tony Blair than
Britons do today. Few Britons, it seems, will shed a tear when Tony Blair leaves
the stage on June 27th afer a decade as prime minister, as he nally announced
this week he would do. Opinion polls have long suggested that he is unpopular.
1. for all the disappointments:
for all X is a concessive structure with a meaning like in spite of X; seems
to be restricted to denite objects; not best treated as a complex preposition
2. look kindly on:
a phrasal verb with the meaning judge positively
37
Border Conicts: FrameNet Meets Construction Grammar
3. [posterity] will look more kindly on Tony Blair than [Britons] do [today]:
a comparative structure with a double-focus comparand[Britons] [today],
each accented, requiring the semantic unpacking of posterity as something
like [the world] [in the future] (a contestable interpretation)
4. few Britons:
not a vague indication of cardinality like a few Britons, semantically a
negator (= not many), creating a negative polarity context (see item 6)
3. it seems:
an epistemic parenthesis, bearing no structural relation to the rest of the
sentence but limited in the positions that would welcome it
6. shed a tear:
a VP collocation of the minimizer type, appropriate to the negative polarity
context created by few; similar in this respect to drink a drop, lif a nger,
give a damn, eat a bite
7. leave the stage:
metaphor, referring here to leaving the PM-ship
8. on June 17th:
use of the preposition on with day-level temporal units (cf. in March, at
noon, in the morning)
9. June 27th:
one of various ways of pairing a date with a month name
10. as prime minister:
as selecting role name; requires context implying service in a role
11. as he announced he would do:
relativizer as (consider replacing as with which)
12. would do:
the form of VP ellipsis (including do afer a modal) found in BrE missing or
rare in AmE (as he announced he would)
13. this week:
an expression in which the rst element is taken from the list this/next/last
and the second is a calendric unit name like week, month, year, but not day
14. have long suggested:
the use of long in the meaning for a long time has numerous contextual
constraints, di cult to pin down; here both (a) the position between have
and the participle and (b) restriction to certain classes of verb meanings
seem necessary (compare I have long known that ... with I long knew that...
and I have long lived in California.)
5. Constructions and the new constructicon
Section 3 oered a number of ways in which the behavior of multiword expressions
can be incorporated into the FN lexicon and into FN-style annotations, that is,
where the information recorded is mainly limited to a small number of requirements
38
Charles J. Fillmore
that lexical items impose on their immediate grammatical environment. Stepping
outside of that is a denite new challenge.
5.1. Te annotation challenge
How did FrameNet become concerned with such matters: First, with our eorts
in full text annotation, we became interested in the possibilities of making better
coverage of all of the linguistic properties of texts, not just those involving simple
predicates and their valence structures. Second, it seems clear that while with
support constructions we moved slightly beyond standard valence projections, the
view of syntactic structure within which we explained the syntactic concomitants
of lexical selection needs to be expanded. Tird, the community in Berkeley that
got started with FrameNet is also a community that has an interest in the broader
theory of grammatical constructions. Fourth, and most importantly, it seemed likely
that the same data structure and annotation sofware devised for lexical annotation
could be assigned to the treatment of constructions.
In 2007 FrameNet received a small grant for doing exploratory research on
designing a constructicon, an inventory of minor grammatical constructions, and
to demonstrate a means of annotatng instances of them. Te parallels to ordinary
FN lexical annotation were triking, as can be seen in Table 1.
Lexical FrameNet Constructicon
Frame descriptions describe the frames
and their components, set up FE names for
annotation, and specify frame-to-frame
relations; lexical entries are linked to frames,
valence descriptions show combinatory
possibilities, entries link valence patterns to
sets of annotated sentences.
Constructicon entries describe the
constructions and their components, set up
construction elements (CEs, the syntactic
elements that make up a construct),
explain the semantic contribution of
the construction, specify construction-
to-construction relations, and link
construction descriptions with annotated
sentences that exhibit their type.
Te FEs are given names according to their
role in the frame, and provide labels for
the phrases in the annotations that give
information about the FE.
Te CEs are named according to their
function in the constructs, they provide the
labels on words and phrases in annotated
sentences.
Te syntactic propertiesgrammatical
functions and phrase typesare identied
for all constituents that realize frame
elements.
Phrase types are identied for constituents
that serve as CEs in a construct; for
constructions that are headed by lexical
units, grammatical function labels will also
be relevant.
Example sentences are selected that illustrate
the use of the lexical units described.
Example sentences are selected and
annotated for the ways they illustrate the
use of the construction.
Annotations identify the LU, the FEs, and
the GFs and PTs of the segments marked
o.
Annotations contain labels for the CEs and
identify, for lexically marked constructions,
the relevant lexical material.
39
Border Conicts: FrameNet Meets Construction Grammar
Lexical FrameNet Constructicon
Valence patterns are identied, and linked
to the annotations.
Varieties of construct patterns are identied
and linked to the annotations.
Frame-to-frame relationships are
documented and displayed in a separate
resource.
Construction-to-construction relationships
are identied and (will eventually be)
displayed
Table 1: Lexical and Constructional Description and Annotation Compared
Te questions to ask for setting up an annotation system for constructions include:
What is the constituent (the construct) within which a construction operates: What
needs to be tagged within a construct: What are the functions of the elements of the
construction: What if anything reveals to the reader/listener that theres anything
special about the sentence:
In FN lexicographic annotation, we describe a frame and its components or
participants, we annotate sentences by identifying the target lexical item and
bracketing o the valents and labeling them with frame element names. In
constructional annotation, then, we should be able to describe a construction and
name the parts of sentences that are the constituents of the constructs licensed by
the construction, and then to bracket o those components and assign them labels
assigned to the elements of the construction. One important dierence is that ofen
there is no target LU to link the construction to.
Figures 3 and 4 show the similarity of lexical and constructional annotations, as
they appear in the annotation tool. Te lexical example represents the clause one of
them accused Mr Wisson of kidnapping; the constructional example represents the
sentence None of these arguments is notably strong, let alone conclusive. Te list of
labels at the bottom of each is the list appropriate to a single level: the FE level in the
lexical example, the CE level in the construction example.
Figure 3: Lexical annotation of the verb accuse in the Judgment_Communication frame
60
Charles J. Fillmore
Figure 4: Constructional annotation of a phrase built around the conjunction let alone
5.2. Te varieties of constructions needing annotation
Te assumption that it would be easy to adapt the FrameNet annotation tool to
construction annotation turned out to be false. Essentially the rst half of the year
of this grant passed by before a proper annotation tool was ready. Finally, in the
spring semester, there are two graduate students working on the project, Russell Lee-
Goldman and Russell Rhodes, with strong backup by Michael Ellsworth and Project
Manager Collin Baker. By the time of the Euralex meeting, I expect to be able to give
a coherent report on our accomplishments and their signicance. In the meantime,
however, I oer some hastily gathered notes on the types of constructions we need
to cover. In the nal report almost all of the construction descriptions will include
references to the relevant literature, omitted here with apologies, including names
like Boas, Borsley, Crof, Goldberg, Jackendo, Kay, Lako, Lambrecht, McCawley,
Michaelis, OConnor, Pullum, Pustejovsky, Sag, Wierzbicka, Zwicky.
3.2.1. Lexical constructions
For an important class of cases, the grammar allows words with one meaning to be
paired with the combinatory aordances that are common to a semantically dened
class of words (in the case of verbs, this amounts to valence patterns; for nouns, the
dierence between proper and common nouns, or that between count and non-
count nouns; for adjectives the dierence between scalar and non-scalar adjectives).
Te word coercion is sometimes used to cover such relationship.
We can distinguish the words that are at home with these aordances from
the words that are their guests. Tere is an obvious problem for a corpus-based
lexicon-building eort like FrameNet, since there is no automatic way of telling the
dierence: should the derived behavior of frequent guests be listed in the lexicon
or merely recognized in context as an instance of the construction: Its a problem for
lexicography in general, since the decisions that need to be made one way or another
are not always clearcut.
61
Border Conicts: FrameNet Meets Construction Grammar
EXAMPLES include the phenomena in much of the literature on Argument
Structure Constructions, especially in the work of Adele Goldberg. Te meanings
created by these constructions involve specied relations between the meaning of
the guest and the semantic expectations of the host pattern: slipping someone a
banknote is using a slipping action to give someone a banknote, wriggling into the
swimsuit is entering the swimsuit (putting it on) with a wriggling motion; an event
of sneezing the napkin o the table is one in which the air current created by a sneeze
has motive force. With nouns, examples like we had beaver for dinner show the use
of the name of an animal with the grammar of a mass noun, coercing a construal as
the esh of the animal prepared for human consumption.
6
3.2.2. Verbs with contextual requirements outside of their phrasal projection
For the kinds of examples we have in mind under this category it should be possible
simply to specify the greater context as part of the combinatory aordancesbut
there is no familiar formal way to do this within theories of valence. Te most
common cases are words that t negative polarity contexts, contexts including
negation straight on or other sources of general irrealis contexts, like questions,
conditional clauses, and dozens of others (since we are mainly interested in
identifying cases and annotating them, the kinds of careful formulation that a true
grammar would need can be glossed over). Verbs that require contexts that involve
both ability and negation allow various ways of expressing those contexts.
EXAMPLES include cant stand, cant aord, cant tell, cant seem to..., cant help.
Te contexts can be expressed in dierent ways: in were you ever able to aord such
luxuries? the polarity is not triggered by a negative morpheme, and the ability is
expressed by an adjective rather than a modal. In its too dark to tell what theyre
doing, the semantics of not + able is entailed in the meaning of too. In the case
of the verb brook a rst impression might be that its required negation is local
i.e., in the determiner of the direct objectbut the negation can be presented by an
external negation with any replacing the no in the determiner position: I will brook
no interruption, I am too busy to brook any distraction.
3.2.3. Templatic constructions
Some constructions seem to require a pattern of xed positions with strict
requirements on what can ll those positions: such is the case of the linguistic way of
expressing proportions of the kind A:B=C:D; it is su cient to think of the sentences
as providing ways of pronouncing the symbols in such a representation.
EXAMPLES are ofen found in lower-grades test questions: Six is to three as four is
to two; blood is to red as snow is to white.
7
6
Te construction does not merely convert the animal name into the name of a continuous
substance. A sentence like the neighborhood fox likes beaver is not licensed by this
construction.
7
Tese sentences could be given a somewhat tortured parse, involving the extraposition of
the as-phrase: if we think of as four is to two as identical to what four is to two, and as naming
a particular relation, then we can see the pattern by putting things back: Six is [what four is
to two] to three.
62
Charles J. Fillmore
3.2.4. A mere ve dollars
Tere is a phrasing of numerical expressions that requires (a) the singular indenite
determiner, (b) an adjective that qualies a number, and (c) a number, such that the
combination demands a noun head that matches the number and can contradict the
singularity of the article a. Tat is, for something like a mere ve dollars, all three
elements are required: a ve dollars doesnt work, mere ve dollars doesnt work, a
mere dollars doesnt work. We see the construction as determining the prenominal
phrase only: in the manner of an ordinary cardinal number, the noun can be deleted
if its nature is understood in the contextas people or dollars, for example, in a
mere two million.
EXAMPLES show adjectives with minimizing, neutral and maximizing senses:
a paltry twenty cents, an additional thirty pages, a whopping seven billion dollars.
An expression like another $200 is a disguised instance of this construction, where
an+other is analogous to an+additional, and $200 is shown as two-hundred +dollars.
Te modifying adjectives that appear in constructs that instance this construction
make up an interesting class.
3.2.3. Presentative constructions
George Lako has discussed a family of constructions using here and there which
have important communicative functions. Formally, they begin with here or there,
they have a verb which most typically is be, come, go, sit, stand, or lie, with the
restriction that if the subject is a pronoun it precedes the verb but if it is a lexical
NP it follows the verb, and utterances of them have the function of announcing
something about the appearance or presence of something. In the complete
version, they include some kind of secondary predicate, that can be an adjective, a
preposition phrase, a participial phrase, or a with(out) clause.
EXAMPLES include here comes that old fool; there she stood, with her hands on her
hips; here comes Billy, crawling on his hands and knees; here I am, ready to serve.
3.2.6. Wherewithal
Tere is a construction which uses the determiner the and a noun construed as
naming a resource; it is followed by an indication of what the resource could be used
for, expressed as an innitival VP or a for-PP; and its governing context identies
someone as a Posissov (or not) of a su cient supply of the resources to carry out
the purpose represented by the nouns complement. A parallel construction exists
with the word enough in place of the. Te name its been given is due to the fact that
the noun wherewithal occurs only in this construction!
EXAMPLES with physical resources include I dont have the resources to landscape
the garden, we lack the sta for such a project, who will provide me the wherewithal to
accomplish this, they denied me the funds to complete the job, do we have the fuel to
make it to the next town? Nouns that designate spiritual resources that t the same
construction include courage, spirit, will, guts, balls, and several others. Arguments
that this construction is needed include the observation that the combination of
63
Border Conicts: FrameNet Meets Construction Grammar
the nominal and the complement cannot serve as a self-standing NP: we spilled
the fuel to make it to the next town. Te purpose complement can be omitted in
contexts where it is understood: A sentence like where did you nd the cash? can
be an instance of this construction, addressed to someone who had just bought an
expensive car, or it can be used simply to refer to some until-now misplaced amount
of money. Te existence of the Wherewithal construction explains that ambiguity.
3.2.7. Gapping and Right Node Raising
Some constructions are purely organizational, and have no lexical components
beyond conjunctions or words that can function as conjunctions. Tose referred to
as Gapping and Right Node Raising (RNR) omit phrases whose meaning is shared
against elements that are in focal contrast.
EXAMPLES of RNR include John loves, but Mary hates, rock music, where comma
intonation separates the two truncated conjuncts from their common completion;
gapping is seen when the shared element is between the focal elements: John loves
peaches and Mary apples. Tose are obviously made-up sentences, chosen for
their brevity. An attested sentence that exemplies both of these constructions
simultaneously is Bears have become largely, and pandas entirely, noncarnivorous.
3.2.8. Let alone
Let alone is a conjunction whose combinatory potential and semantic-pragmatic
interpretation are discussed in Fillmore-Kay-OConnor 1988 and some discussions
following that. Briey, the pieces that are in focal contrast can be
8
assembled with
their surrounding contexts to form two propositions, one of these propositions is
responsive to the context (i.e., to some assumed or expressed context proposition),
the other is strongly asserted by the speaker, and it contextually entails the rst.
EXAMPLES include the sentence in Figure x, None of the arguments is notably
strong, let alone conclusive. Numerous examples of multiple foci are found in the
FKO article. Let alone sentences frequently exemplify RNR: I wouldnt touch, let
alone eat, anything that ugly (Made-up sentence).
3.2.9. Verb ones way
A much-studied construction is a way of providing motion verbs by inserting a verb
that indicates an action by which someone is able to move, or a path through which
8
For example:
Context proposition spoken by interlocutor: Can you give me a dollar?
Direct response to the context proposition: I wont give you a dollar.
Response that strongly entails the context-relevant response: I wouldnt lend my mother a
nickel.
Result: I wouldnt lend my mother a nickel, let alone give you a dollar.
Relevant scales for the triple contrasting foci: Im more likely to lend money to someone than
to give it away; Id be more generous to my mother than to you; a dollar is a lot more than a
nickel.
64
Charles J. Fillmore
the mover moves, or an activity on the movers part during which they moved. Te
structure is (a) verb plus (b) possessive pronoun coreferential to the moving entity
plus (c) the word way: VERB ones WAY. Te most neutral verb that is at home in
this construction is make (Lets start making our way home.) Te verb wend exists
only in this construction.
EXAMPLES that show the variety include She pushed her way through the crowd, the
river winds its way through the prairie, we dined our way through the south of France.
3.2.10. In ones own right
A number of constructions depend on the extended reexive possessive pronoun
ones own: he nally has a room of his own, youre on your own now, but one we
have examined is the adjunct in ones own right. A typical background assumption
for its use is something like this: A is a liated with B in some way (a relative, an
assistant), B is already known for some property or accomplishment, the sentence
asserts that same property or accomplishment of A, and the construction conveys
the assumption that As accomplishments are not due to the a liation with B. Te
son of a poet can be a ne poet in his own right, the husband of a famous chemist can
be an accomplished chemist in his own right. It would sound odd to say of the wife
of right-wing radio commentator Rush Limbaugh that she is a major intellectual in
her own right, without invoking a belief that Mr. Limbaugh is a major intellectual. (I
dont even know if hes marriedthis is just an example).
3.2.11. Rate phrases
Te concept of rate is expressed in English with two adjacent NPs in which the
rst identies a quantity of units of some type and the second introduces a unit of
a dierent type across which the measurement applies, more or less as numerator
to denominator. Typically the second NP is marked with a or per, but other types
occur as well. Tese expressions express such notions as growth rate, frequency, fuel
e ciency, speed, and the like.
EXAMPLES include it grows four inches a day, but also four inches every three days;
my Hummer gets seven miles a gallon; our committee meets twice a week; we were
moving at 150 km per hour. Te type of rate can be calculated by comparing the two
kinds of units, and can be supported by making note of aspects of the governing
context, such as the items grow, meet, gets, and at of the examples.
3.2.12. Measurement phrases
Some scalar adjectives, but not all, support measurement qualiers that indicate a
quantity of units used for values on the scale.
EXAMPLES include ve meters long/wide/tall/thick, and seventeen years old.
Weight and cost values are expressed verbally, with the verbs weigh and cost; there
is no twenty pounds heavy or twenty dollars expensive. Comparative expressions,
however, can have measured gaps across the board: twenty pound heavier, twenty
dollars cheaper, three years older, etc.
63
Border Conicts: FrameNet Meets Construction Grammar
3.2.13. Deictically anchored calendar units
Te lexical set this-next-last occurs in several constructions dedicated to locating a
reference time to the present momentthe temporal deictic centerwith respect to
calendric time periods like week, month, and year. Tis makes reference to the period
containing now; next refers to the period following the period containing now;
and last refers to the period preceding the period containing now. Tese patterns
do not apply to days, however: at the day level the same functions are served by the
lexical items today, yesterday, tomorrow.
EXAMPLES illustrating one of the constructions, simply identifying a period, are
next year, last month, this week; a second construction uses these words to mark a
recurring point or subdivision of a larger unit and locates the event within the lower
unit with respect to whether the larger period is current, past, or future to now:
next Wednesday, last summer, this August; the third construction uses next and
last in a xed pattern where the word is understood as picking up the immediately
preceding mention of the time entity: the week afer next, the month before last, and
the summer afer next, the Christmas before last.
3.2.14. Te + Adjective
Expressions like the rich and the poor are usually thought of as showing these
adjectives being used as a noun. Instead of attributing a part-of-speech change to
the adjective, it would seem that a better analysis is that the combination THE +
Adjective-Phrase behaves like a full NP. How else could we understand the very rich,
the very young: Not as very modifying a noun, presumably. Te constraints seem to
be that the adjectives designate some categorizing property of humans; the resulting
phrase is human, generic, and plural. Certain adjectivespoor, rich, young, oldare
frequent guests of this construction, but the lexicographers decision to identify
them as actual nouns in those contexts does not seem helpful.
3.2.13. Adjective + and + Adjective
Tese same adjectives can be used, in roughly the same meaning, when they
surround and, as in he was beloved of rich and poor alike. In this case the denite
article is not needed, but the conjunction is necessary: he was beloved of poor does
not work.
3.2.16. Degree modiers of adjectives
Its di cult to decide how many constructions are needed for the intended family
of constructions, perhaps several, with constructional inheritance connecting
them. Some examples communicating su ciency or excess have extraposable
complements: too and enough go with an accompanying innitival VP, so goes with
a that-clause. Others question a scalar value posed in the context, require negative
polarity, are accented, and do not have an extraposed complement.
EXAMPLES include shes not that young, you cant be too hungry or you d help us
get dinner ready, youre too young to understand, hes so senile that he cant follow the
66
Charles J. Fillmore
conversation, I am hungry enough to eat a horse. For too and enough, the complement
can be omitted when the idea is contextually given: shes too young, shes not old
enough.
3.2.17. Adjective comparison
Comparison makes up a huge topic, that will not be conquered during the time of
this pilot study, but theyre included here because of some further constructions that
will include them. Te comparative markers also carry extraposable complements:
more/er- and less than; [not...] so and as as.
EXAMPLES include Shes much more intelligent than you said, are you as angry as
you seem, its less warm today than it was yesterday.
3.2.18. Comparative Negation with no rather than not
If I say that youre not more qualied for the job than I am, I could believe that we are
both well qualied, and that I should certainly be included among the candidates.
On the other hand, if I say that youre no more qualied for the job than I am, its
assumed that were both barely qualied, and (say) Im complaining that they had
no right to give you the job. Using this construction seems to suggest that both of
the things being compared are at the low end of the scale. Your puppy is no bigger
than a mouse!
3.2.19. NP-internal degree-modied adjectives
All of the adjective modiers weve just reviewed can be used predicatively, but there
is a construction that allows them to be used attributively, but only in the case of a
singular indenite count noun. Tose that have extraposed complements allow them
to be extraposed afer the noun. Te adjectival part precedes the indenite article.
(Compare [an] [intelligent] man with [too intelligent] [a] man.) A variant of the
construction has an intrusive of which sounds more natural in some contexts than
others. We have nothing to say about that just now.
EXAMPLES include youre too intelligent a man to act like that, thats much bigger
of a house than we need, thats as sensible a solution as we can expect, is it really that
big of a problem, thats no bigger a problem than others we had in the past, thats so
big a problem that we ll never be able to deal with it, is this big enough of a box? Te
limitation to indenite singular count nouns is striking: its not that hot of soup,
theyre no older of people than my parents.
3.2.20. Ones every something
I once proposed that a particular expression with every was dedicated to talk about
indulgence fantasies, but have learned from corpus data that it is also frequent in
paranoid talk.
EXAMPLES of the former kind include we are here to meet your every need, you will
obey my every command, my every dream has been fullled, Ive satised my every
67
Border Conicts: FrameNet Meets Construction Grammar
wish; but the other kinds include why are you dogging my every step, they watch my
every move, he records my every gesture. And there are neutral expressions as well,
so it probably requires no more than a sense of extreme attentiveness. Whatever it
is, the relationship between the Possissov and the noun has to be agentive in some
wayit cannot be one of simple possession: they stole my every donut doesnt seem
to work.
3.2.21. Plural-noun reciprocals as predicates
Some plural undetermined nominals can occur as predicates indicating a
symmetrical social relation between two people. We were best friends in high school
can be expressed from one members point of view: I was best friends with him in
high school. If the subject is singular, a with is needed to identify the other member
of the relationship. Tis only works with nominals that indicate some kind of
social relation that inherently is (like cousin or friend) or can be (like brother or
sister) symmetrical: were siblings can stand alone as a predicate, were sons requires
mention of the second term of the relationship, I was foreigners with him in Japan
doesnt work: foreigner isnt a relation between two people
EXAMPLES include we were colleagues in the post o ce, she is cousins with a
very rich man, and, from the web, my theory is that Harrys mother is siblings with
Voldemort.
6. Opportunities for a construction-expanded FrameNet
Te decision to enter constructional information and lexical information in the same
database turns out to have many advantages. In particular, its seldom necessary to
worry about whether were dealing with a lexical or a grammatical structure. Some
products of a construction are simply lexical units in essentially every way, except in
that they are generated rather than requiring individual listing in a dictionarys
wordlist: this is true of the products of argument structure constructions as well
as a number of derivational patterns, morphological or zero derivation. Te
lexicographer might now have a principled way of deciding whether a frequent
guest deserves inclusion in the lexicons standing wordlist. Some constructs behave
like ordinary lexical items in their external environment, and can then be annotated
as equivalent to single LUs in their own right: the reciprocal best friends can be
annotated as an ordinary symmetric predicate of the kind that permits both joint
and disjoint expression of the paired participants. Te phrase to push ones way in
its external syntax works just like an ordinary motion verb and acquires the valence
expectations shared by ordinary motion verbs and can be annotated as such. Many
of the constructions produce constituents that t their environment in normal ways
requiring nothing special: a rate expression classied as indicating FviqUicv, or
Sviiu, or Ui1_vvici, or W.uis, can combine with whatever marking goes with the
governing predicate and nd its place in the annotations for that predicate. Te zero
anaphora facts that FrameNet has encountered in preparing lexical descriptions are
similar to those that occur with constructions as well, and pose similar challenges to
68
Charles J. Fillmore
theories of anaphora. Tus, to take a sentence like otherwise most members wouldnt
have the funds, a search for cohesion with preceding texts would have to include the
condition implied by otherwise, the organization presupposed by members, and the
purpose-indicating complement of the Wherewithal construction that the funds are
needed for.
Whether parsers can recognize (and interpret) instances of special constructions
will remain to be seen. Its possible that a very large sample of construction-
annotated texts could provide the learning corpus for statistics-based parsers. An
apparent number agreement failure could lead to interpretations that permit such
possibilities: she is friends with the president, a mere twenty pages. In many cases
there are overt markers of a construction that could initiate specic steps to nd the
components (the phrase let alone). A comma before a conjunction in will trigger a
search for discontinuities permitted by RNR and Gapping structures. And in some
cases the failure to nd, in the immediate context, a needed valent of a verb or head
of a modier should guide the search for explanations: the hanging largely in the
sentence bears have become largely and pandas entirely noncarnivorous should serve
as a clue.
References
Fillmore, C. J. (1982). Frame semantics. In Linguistics in the Morning Calm. Seoul:
Hanshin Publishing Co. 111-137.
Fillmore, C. J.; Atkins, B. T. S. (1992). Towards a frame-based organization of the
lexicon: Te semantics of RISK and its neighbors. In Lehrer, A.; Kittay, E. (eds.).
Frames, Fields, and Contrast: New Essays in Semantics and Lexical Organization.
Hillsdale: Lawrence Erlbaum Associates. 73-102.
Fillmore, C. J.; Atkins, B. T. S. (1994). Starting where the dictionaries stop: Te challenge for
computational lexicography. In Atkins, B. T. S.; Zampolli, A. (eds.). Computational
Approaches to the Lexicon. Oxford: Oxford University Press. 349-393.
Fillmore, C. J.; Kay, P.; OConnor, M. C. (1988). Regularity and idiomaticity in
grammatical constructions. Language 64 (3). 301-338.
Goldberg, A. (1993). Constructions. A Construction Grammar approach to argument
structure. Chicago: University of Chicago Press.
Jackendo, R. (1990). Semantic Structures. Cambridge: MIT Press.
Joshi, A. K. (1983). Tree-adjoining grammars: How much context sensitivity is required
to provide reasonable structural descriptions:. In Dowty, D.; Karttunen, L.; Zwicky,
A. (eds.). Natural Language Parsing. Cambridge: Cambridge University Press. 206-
230.
Lako, G. (1987). Women, Fire, and Dangerous Tings. Chicago: University of Chicago
Press.
Tesnire, L. (1939). Elements de syntaxe structurale. Paris: Klincksieck.