Professional Documents
Culture Documents
Multiword Expressions. An Extremist Approach (Fillmore, Charles J.) (2003)
Multiword Expressions. An Extremist Approach (Fillmore, Charles J.) (2003)
Multiword Expressions. An Extremist Approach (Fillmore, Charles J.) (2003)
1
Dependency Representation The graphs
ISH's knowledge is about • There are other, perhaps more standard, ways of showing
– unitary words and dependency graphs. A common one is to show the
– word-to-word relations. dependency relations with simple points and then to show
That can be represented in dependency diagrams in the linearization with lines pointing down to a text.
– which each node is a word and · ·· ·
– each word-to-word link, i.e., each branch, · · ·
• stands for one of the basic grammatical ·
relations and
• is capable of bearing a frame-based semantic
· · ·
relation to the governor. ·
My parents gave me a copy of that fascinating book about frogs
2
The blue lines Complementation
His parents gave me a copy of that
• Each blue line in the graphs which follow identify the kind
of dependency relation named in the slide's heading. fascinating book about frogs.
Specification Modification
His parents gave me a copy of that His parents gave me a copy of that
fascinating book about frogs. fascinating book about frogs.
gave gave
parents me copy parents me copy
his a of his a of
Actually his can also be book book
thought of as satisfying
a frame requirement of that fascinating about that fascinating about
the relational noun parents. frogs frogs
© 2003 Charles Fillmore 11 / 92 © 2003 Charles Fillmore 12 / 92
3
So ... Where the ISH idealization fails
The study of MWEs can begin with an 1. Some apparent MWEs are best analyzed as
examination of meaning units of the single words, occupying one node.
language that do not lend themselves 2. Some MWEs can be represented as
to such a simple treatment. dependency subgraphs (not "just" word
strings, or collocate sets).
3. Some MWEs are the product of "non-core"
semi-independent mini-grammars (e.g.,
personal names, number words or the
phrases we use for telling clock time).
© 2003 Charles Fillmore 13 / 92 © 2003 Charles Fillmore 14 / 92
• The highlighted passage in the next slide - a repetition of the preceding 1. Some apparent MWEs are best analyzed as
slide - is in reference to work on multiword entities that has not been
sensitive to grammatical relations. Early on the search for simple single words, occupying one node.
contiguous strings of words was seen as inadequate (many MWEs can
be interrupted) and attention was given rather to collocations, words
2. Some MWEs can be represented as
that co-occur in statistically significant ways. In such work the dependency subgraphs (not "just" word
emphasis was on choosing a "span" within which significant collocates strings, or collocate sets).
could be reliably found, in general recognizing that they at least have
to be in the same sentence. This paper emphasizes the specific kinds of 3. Some MWEs are the product of "non-core"
grammatical relations that collocates have to have with each other to semi-independent mini-grammars (e.g.,
be linguistically significant.
personal names, number words or the
phrases we use for telling clock time).
© 2003 Charles Fillmore 15 / 92 © 2003 Charles Fillmore 16 / 92
4
Theoretical Issues 1. “Runs"
Headedness There are things that look like MWEs (that are
Relation to the rest of the grammar written as sequences of words), but they have no
How many of them are there? (Jackendoff; Mel'cuk) internal variation and may just as well be thought of
How do we count them? as long words with spaces in them.
Language acquisition Examples
Language typology – used to, let alone, of course, all of a sudden, first off
Role in parole Some are easily mislearned:
How are they represented? – by and large > by in large
– to all intents and purposes > to all intensive purposes
– an arm and a leg > a nominal egg
"Runs" But,
• By "runs" is meant strings of words that have no interruptable I won't talk about certain other kinds of
substructure - they are just run through from beginning to end. Nobody
would interrupt a phrase like "of course" ("of y'know course") or "all semantically penetrable "long words" that
of a sudden" ("all of, well, a sudden") or, in general, any of the other identify very specific entities or concepts;
expressions listed here.
e.g.,
• The fact that many complex runs get reanalyzed means that they are
not learned by building up structures out of their parts, but as wholes; – names of organizations
but, that said, they nevertheless have to be given some kind of analysis, The American Society for the Prevention of
if only because English simply doesn't have many really long words, Cruelty to Animals. (ASPCA)
and also because they need some kind of syntactic structuring for us to
know how to apply stress rules to them. – names of officially designated crimes
assaulting a federal officer with a lethal weapon
5
2. Dependency Subgraphs 2. Dependency Subgraphs
Some MWEs can be analyzed in terms of sets of Some MWEs can be analyzed in terms of sets of nodes
nodes connected to each other through syntactic connected to each other through syntactic head-dependent
relations.
head-dependent relations.
x x x x
x x
y y z y A y y A z y
A z
z
6
2. Dependency Subgraphs
Some MWEs can be analyzed in terms of sets of
nodes bearing direct or indirect syntactic head-
dependent relations to each other.
The linearization of the elements of this kind of MWE
may be constrained, or it may be subject to ordinary
syntactic principles.
There is no need to see the elements of such a
Subcategorization Details
subgraph as making up a "single constituent" in a
corresponding phrase-structure representation.
7
Prepositional Verbs -
In the Old Days ...
Intransitive
V
About half a century ago it was generally believed Verb > preposition is the
that in Deep Structure, phrases like pick up, take off, lexical unit.
etc., started out as single constituents, and a Particle X prep Exx:
Movement Transformation allowed the extraction of look for ('seek'), object to
Y ('oppose'), look into
the particle so that it could follow the direct object. ('investigate')
Interruptible:
[take off] [your shoes] >> [take] [your shoes] [off] look I looked long and hard for
the perfect wife.
We objected strenuously to
A dependency subgraph can recognize the unity of X for her proposal.
the two-word block without worrying about phrasal Comment:
constituency. Y Some PPs are omissible,
some aren't. look (for), look
into
© 2003 Charles Fillmore 29 / 92 © 2003 Charles Fillmore 30 / 92
Omissibility PP Omissibility
• It may seem paradoxical that the real form of the lexical unit is Omissible Non-omissible
something like look for but that the for-phrase can be omitted. This can
be more straightforwardly represented in the Construction Grammar (under conditions of
framework - but that's another story. zero anaphora)
8
Prepositional Verbs -
Particle-&-Preposition Verbs
Transitive
V
Verb > preposition is the V Verb > {part,prep} is
lexical unit. the lexical unit.
X Y prep Exx: Exx:
talk into ('persuade'), rid of X part prep put up with ('tolerate'), look
Z Comment: PP is sometimes up to ('respect'), break in on
omissible: Y ('interrupt')
The judge cleared me (of all Not generally interruptible, I
clear charges). think (haven't checked
They tried to talk me *(into put corpus data).
quitting my job).
X of Who will rid me *(of this (But: This is the sort of errant
Y nonsense up with which I
meddlesome priest)? X up with
shall not put.
Z Winston Churchill)
Y
© 2003 Charles Fillmore 33 / 92 © 2003 Charles Fillmore 34 / 92
9
V+N+P Verbs Other Parts of Speech
V Verb > /N,prep/ is the
Adjectives can have prepositional and
lexical unit.
Exx: clausal complements:
X N prep take advantage of ('exploit'),
take part in ('participate in'), – fond of cats; interested in math; similar to mud
Y take charge of
Comments:
Nouns can have prepositional and
take N can be modified; N can be causal complements:
passive subject:
Considerable advantage was – top of the tower; friend to the poor; journey into the
X part in taken of this opportunity. jungle; copy of the book
Pseudo-passive:
Y They were cruelly taken
advantage of.
These are getting close to support
N does not take a determiner.
verb constructions. © 2003 Charles Fillmore 37 / 92 © 2003 Charles Fillmore 38 / 92
Support Constructions
10
Support Verbs with Subject N
11
SVs can resolve polysemy. A common test of SVs:
12
Beyond "light verbs" Examples
Simple cases: the verb has essential no meaning Simple, active:
except to reveal that its subject is necessarily a
participant in the event named by the noun. – he made a complaint
– a. active role
Nuanced:
– b. passive role
More nuanced cases: the verb contributes – he registered a complaint
information about register, attitude, aktionsart,
or the like.
More extended cases: the verb identifies its subject
as a participant in the larger scenario associated with
the event named by the verb.
13
Examples Examples
14
For the full story,
and then some, see ...
Support Verbs with Adjective
V
Mel'cuk, Igor' (1995), Phrasemes in language and phraseology Verb > A is the lexical unit,
in linguistics. In M. Everaert et al., Idioms: Structural and A is semantic head, V is
Psychological Perspectives. Lawrence Erlbaum Associates. X A support verb, A may have
Mel'cuk, Igor' (1996), Lexical functions: a tool for the its own complements (e.g.,
description of lexical relations in a lexicon. In Leo Wanner, ed., rid of).
Lexical Functions in Lexicography and Natural Language
Processing. John Benjamins. Exx:
be + any predicate adjective;
Mel'cuk, Igor' (1998), Collocations and lexical functions. In get go crazy, turn red, get naked
Cowie 1998
Comment:
Mel'cuk, Igor' (1995), The future of the lexicon in linguistic The unit rid of seems to
description and the explanatory combinatorial dictionary. X naked occur only with a SV.
Linguistics in the Morning Calm 3. 181-270. Hanshin: Seoul
consideration
© 2003 Charles Fillmore 59 / 92 © 2003 Charles Fillmore 60 / 92
15
Support Verbs with PP
V
Verb > P > N is the lexical unit,
N is semantic head, V is
support verb. With
X prep possession there are two
alignments of the arguments:
N Possessor - Possessed
come
I came into possession of
these documents. Transparent Nouns
Possessed - Possessor
X into
These documents have come
into my possession.
possession
16
Types of transparent nouns "Transparency"
1. Aggregates
bunch, group, collection, herd, school, flock • The term, I think, is from Naomi Sager.
2. Quantities • The idea is that the dependent noun - the second noun in a N+of+N
flood, number, scores, storm phrase - has certain selectional relations to the external context of the
phrase; the contextual pattern can be noticed by "seeing through" the
3. Types
transparent noun.
breed, class, ilk, kind, type, sort
• The first example on the next page does not involve an MWE, but we
4. Portions and Parts can at least notice that the locative prepositions in and on go with
half, segment, top, bottom, part nouns of particular semantic types (in with a volume and on with a
5. Unitizers surface, for example), and this is independent of whether there's an
glass, bottle, box, serving expression like part of between the preposition and its preferred
6. Evaluations collocate.
gem, idiot, prince
17
N > N Compounds
N N
N > N is the lexical unit; listed
compounds have the dependent
in red; the syntactic head is the
frame evoker, the dependent is
N N either a frame element or a
"quale". The order is Modifier +
Head.
Compounds risk knife
health fish
18
A-N Compounds "Pertinative" adjectives
N N
N > A is the lexical unit; listed Pertinatives are adjectives whose senses are
compounds have the dependent
in red; the syntactic head is the defined in (some) dictionaries with the phrase "of or
frame evoker, the dependent is pertaining to". Traditional term: relational adjectives.
A A either a frame element or a WordNet term: pertainyms.
"quale".
Ready-made A+P compounds: They are not used predicatively in the same meaning.
hot news, friendly fire, blind They aren't scalar, e.g., they don't get modified with
police news alley, dead end
very.
federal hot
19
3. Minigrammars Personal Names
Some MWEs are generated by simple Reverend Dr T. Allen Hampton-Smith III
generative structures, usually finite state Components: titles, honorifics, given names,
automata, for which dependency – or patronymics, family names, extensions, ...
constituency – representations are not always It would be best if names could be "sealed":
relevant. handed over to the grammar with a category
– Names name and nothing else
– Numbers
Sometimes personal names are penetrable.
– Locations (addresses, coordinates)
(Croatian)
– Time Expressions
– Marko je Ivir zubar.
– Kinterms
– *Patrick is Hanks lexicographer.
– Miscellaneous constructions
© 2003 Charles Fillmore 77 / 92 © 2003 Charles Fillmore 78 / 92
20
cousins second cousins
X X
A B A B
C D C D
E F E F
G H G H
A B A B
C D C D
E F E F
G H G H
21
Digression Self-indulgence
Ordinary techniques of computational • This is just a bit of self-indulgence. In the full write-up of
this paper I want to give a lot of attention to special
linguistics/corpus linguistics won't be grammatical constructions, and this unusual pattern with
able to recognize the constructional an indefinite determiner and a plural noun - with the
nature of some expressions. requirement that there be both some kind of qualifier and a
quantity (usually number) seemed strange, in the first
Test case place, and frequently hidden from view by the way we
express numbered units of currency (e.g., by putting the
another $600 "$" before the number while pronouncing it after the
number).
Indefinite article Qualifier Number Plural Noun Indefinite article Qualifier Number Plural Noun
22
Continuity Hypothesis Why this?
I assume the continuity of the lexicon and the • Here the point is that in dealing with lexically headed
constructicon. special grammatical constructions, Kay and I have often
found that they can be described in pretty much the same
Reference: Paul Kay & Charles J. Fillmore way as complex lexical items with contextual constraints
(1999), "Grammatical constructions and that are pretty much like what you get with ordinary
linguistic generalizations: the What's X Doing words. The full force of the "continuity hypothesis" is that
Y? construction", Language 75 1-33. from highly specific lexical items to maximally general
grammatical patterns, there are always simple head-to-
Claim: many lexically-headed constructions dependency relations of the kind mentioned in this paper.
can be analyzed as dependency subtrees.
23