Identifying Syntactically Relevant Units of Meaning

Identifying Syntactically
Relevant Units of Meaning
Hans C. Boas
The University of Texas at Austin
hcb@mail.utexas.edu
http://hcb.gmc.utexas.edu

Overview
• Introduction
• Using syntactic behavior for defining verb
classes
• Verb classes and limitations on
predictability
• A frame-semantic approach to verb
classification
• Conclusions and Outlook
2

Introduction
• Locative alternation
(1) a. Sophie sprayed water on the bushes.
(locative variant)
b. Sophie sprayed the bushes with water.
(with-variant)
(See Fillmore 1968, Anderson 1971, Salkoff 1983, Rappaport & Levin 1985,
Pinker 1989, Levin 1993, Goldberg 1995, Nemoto 1996, Dowty 2000, Boas
2003, Iwata 2005, Levin & Rappaport Hovav 2005, among many others)
3

Pustejovsky (1995: 8)
It is the goal of any lexical semantic theory
to adequately classify the lexical items of
a language into classes predictive of their
syntactic and semantic expression.
Furthermore, such a theory should not
merely map the meanings of lexical items
per sentence, on an individual basis.
Rather, it should capture the semantic
relations between words in such a way
which facilitates this mapping.
4

Using syntactic behavior for
defining verb classes
• Pinker’s (1989) lexical rule approach to
locative alternation
- Verbs of putting and covering
- Semantic criteria for identifying two verb classes
(force, dimensional geometry of solids, etc.)
5

Content-oriented into/onto verbs
(Pinker 1989: 126)
6

Content-oriented with-verbs
(Pinker 1989: 126-27)
7

Pinker’s lexical rule approach
The lexical rule takes a verb containing in
its semantic structure the core ‘X causes Y
to move into/onto Z,’ and converts it into a
new verb whose semantic structure
contains the core ‘X causes Z to change
state by means of moving Y into/onto it.
(Pinker 1989:79)
8

Lexical Rule: Derivation
• with-variant locative variant
• locative variant with-variant
• Directionality depends on ability of direct
argument to stand as the sole complement
9

Linking of underlying thematic roles
to syntax
The difference in argument structure follows from
the linking rules: in the old verb, the moving
thing was the theme and hence was linked to
direct object; in the new verb, the location is the
theme (of a state change) and hence is linked to
object. The argument not linked to object gets
linked to an oblique function or position by virtue
of other linking rules in combination with lexical
entries for specific prepositions.
(Pinker 1989:79)
10

Problem 1
• Sole complement diagnostic does not make the correct
predictions
(5) a. John heaped books on the shelf.
b. John heaped the shelf with books.
c. *?John heaped the books.
d. *John heaped the shelf. (Pinker 1989:38)
(6) a. John packed books into the box.
b. John packed the box with books.
c. John packed the books.
d. John packed the box. (Pinker 1989:39)
Possibility of standing as a sole complement does not
truly serve as a diagnostic for the derivational base
11

Problem 2
• Verbs such as drizzle, drip, and slop
exhibit with-variants, contrary to Pinker.
(7) I tried using a normal bit with a standard drill as I dripped it with oil.
(re.antiques.radio+phono)
(8) I made a fresh mashed-spider cake and drizzled it with raspberry sauce
just for you. (acadia.chat)
(9) I slopped it with Ketchup, smacked my lips, and took a bite.
(alt.religion.christian.roman-catholic)
12

Problem 3
• Which verb sense is the “basic” variant?
• E.g., stuff-verbs (cram, crowd, stuff, etc.)
Stuff-verbs can trigger the formation of
corresponding content-oriented or
into/onto forms” which “involve the notion
of a container’s intended capacity (Pinker
1989:234)
13

Stuff-verbs
• Verbs that fit the semantic characterization of stuff-verbs
don’t exhibit similar syntactic behavior, e.g. squeeze
(10) a. Dawn stuffed the turkey with breadcrumbs.
b. Dawn stuffed breadcrumbs into the turkey.
(11) a. *Christian squeezed the turkey with breadcrumbs.
b. Christian squeezed breadcrumbs into the turkey.
Why does lexical rule apply to stuff but not to squeeze?
14

Summary – Problems with Pinker (1989)
• Not clear which verb sense is “basic”
• Lexical rules do not always have access to
the proper types of inputs
• Semantic classification not fine-grained
enough
Does not make the correct generalizations
15

Levin’s (2003)
Projectionist Approach
Verb meaning
Root
(idiosyncratic part)
Constant
(Structural part, LCS, event structure)
- Core meaning - combinations of primitive predicates
- Not relevant for grammatical behavior - relevant for grammatical behavior
- Number is in principle open-ended - limited set of possible event structures
See also Rappaport Hovav & Levin (1998, 2001), Levin & Rappaport Hovav (2005)
16

Simple vs. complex event structures
(12) a. [x ACT ]
b. [[ x ACT ] CAUSE [ BECOME [y ]]]
(Rappaport Hovav and Levin 1998:108)
(13) The argument-per-subevent condition
There must be at least one argument XP in the syntax per
subevent in the event structure.
(Rappaport Hovav and Levin 2001: 779)
Syntactic differences between run (manner verb) vs. break
(result verb)
17

“Generating” verb meaning
(14) Template Augmentation
Event structure templates may be freely
augmented up to other possible templates
in the basic inventory of event structure
templates.
(Rappaport Hovav and Levin 1998: 111)
18

Locative Alternation (Levin 2003)
• Simple event structures are found in complex event
structures
• Each verb is characterized by a particular type of end
result
• Only means or manner roots may exhibit multiple
alternations
a. [x ACT ]
b. [[ x ACT ] CAUSE [ BECOME [y ]]]
19

Why can some means and
manner verbs alternate but not others?
• The root of sew is associated with a simple event structure, and
different end results are possible (template augmentation)
Creating an object
(15) a. Dale sewed the piece of silk into a ball gown.
b. Dale sewed a ball gown out of the piece of silk.
Covering a surface
(16) a. Dale sewed bows on the costume.
b. Dale sewed the costume with bows.
Attaching things/impressing an image
(17) a. Dale sewed the lining to the skirt.
b. Dale sewed the lining and skirt together.
(Levin 2003:10)
20

Differences between sew and
vacuum
• The root of vacuum describes a means/manner used
only to obtain a very specific result and can thus not be
associated with various types of results
(18) a. Avery vacuumed the dust off the rug.
b. Avery vacuumed the rug.
(19) a.*Avery vacuumed the dust onto the rug.
b.*Avery vacuumed the rug with the dust.
(20) a.*Avery vacuumed the dust into a pile.
b.*Avery vacuumed a pile from the dust.
(Levin 2003:10-11)
21

Event structures determines
alternating behavior
• Non-alternating verbs have roots associated with
complex “result state” event structures, not with simple
event structures
(21) a. Ashley filled the bucket with water.
b.*Ashley filled water into the bucket.
(22) a. Ashley poured water into the bucket.
b.*Ashley poured the bucket with water.
(Levin 2003:11)
22

Event structures determines
alternating behavior
Manner/Means Verbs
(smear, sew)
Associated with simple event
structure
Template augmentation
possible
Alternation possible
Result Verbs
(pour, fill)
Associated with complex
event structure
Template augmentation not
possible
Alternation not possible
23

Problem 1 with Projectionist Approach
• Manner/means verbs closely related in meaning do not
exhibit same alternation pattern (sew/stitch)
• Which semantic factors are responsible?
(23) a. *Joe stitched the piece of silk into a ball gown.
b. *Joe stitched a ball gown out of the piece of silk.
(24) a. Joe stitched bows on the costume.
b. *Joe stitched the costume with bows.
(25) a. Joe stitched the lining to the skirt.
b. Joe stitched the lining and skirt together.
24

Problem 2 with Projectionist Approach
• Manner/verbs should be able to alternate to
express different types of results, but they don’t
(26)a. Collin {wiped/swept/rubbed} the dust off the rug.
b. Collin {wiped/swept/rubbed} the rug.
(27)a. Russell {wiped/swept/rubbed} the dust onto the rug.
b.*Russell {wiped/swept/rubbed} the rug with the dust.
(28)a. Petra {wiped/swept/rubbed} the dust into a pile.
b.*Petra {wiped/swept/rubbed} a pile from the dust.
25

Levin’s event structure
classification not precise enough
• Despite our expectations, hang alternates
(29) a. They hung pictures of the president onto their walls.
b. They hung their walls with pictures of the president.
• Irregular behavior of tape-verbs (anchor, band, belt, bolt, bracket,
buckle, button, cement, chain, clamp, clasp, clip, epoxy, fetter, glue, gum,
etc.)
(30) a. Paul taped pictures onto the wall.
b.*Paul taped the wall with pictures.
27

Summary – Problems with Levin (2003)
• Event structure classification is too coarsegrained
• Template augmentation does not capture
full distribution of verb senses
• Similar problems as Goldberg (1995),
where verbs and constructions “fuse” to
generate new verb senses
28

Levin (1993)
• Classifications based on argument syntax
• 193 verb classes covering 3024 verbs, or
4,186 senses (Verbs of putting, verbs of removing,
verbs of communication, etc.)
• 79 alternations (Locative, Benefactive, Creation and
Transformation, etc.)
29

Multiple class membership and
limitations on predictability
• Levin (1993): verb classes based on syntactic behavior
of verbs
• E.g., classification of transitive verbs
30

“Other verbs show same pattern of behavior”
(31) a. Break Verbs: break, crack, rip, shatter, snap, …
b. Cut Verbs: cut, hack, saw, scratch, clash, …
c. Touch Verbs: pat, stroke, tickle, touch, …
d. Hit Verbs: bash, hit, kick, pound, tap, whack, …
(Levin 1993:7)
What are the relevant meaning components common
among the members of each class?
Do all members of a verb class exhibit identical syntactic
behavior?
31

Verbs from same class show varied
behavior in different alternations
• Cut verbs participate in middle alternation and
instrument subject alternation
• Not all cut verbs participate in body-part-possessor
ascension alternation
(31) a. Carol cut herself on the thumb.
b. Carol cut her thumb.
(32) a. *Carol hewed herself on the thumb.
b. Carol hewed her thumb.
(see also Boas 2003 on resultative constructions)
32

Syntactic frames not always
heterogeneous
(33)a. I ended up boxing with him.
b. Tyson will box Lewis.
(34) a. My son played/met with your son.
b. My son played/met your son.
33

World knowledge impacts syntactic distribution (Taylor 1996)
34

Syntax = Semantics?
• Not all verbal syntax is indicative of a verb’s
lexical semantic properties
• Syntactic classification gives only a limited
picture of underlying semantics
• Levin’s (1993) verb classes do not exhibit
uniform meaning, cf. break vs. split verbs:
• “It is possible that additional verbs from these
other classes might qualify as split verbs.” (1993:
167)
35

More verb class problems
• For each syntactic phenomenon there may be
a different verb class definition
• Verb classes are epiphenomenal
• Lexicalist paradox: two contradictory goals for
constructing lexical entries:
(1) Factoring predictable information out of lexical
entries (as minimal as possible)
(2) Explain a verb’s varied syntactic behavior (as
maximal as possible) – argument realization
36

Alternative: A frame-semantic approach
to verb classification
• Frame Semantics (Fillmore 1982)
• Lexical Unit (LU)
• Semantic Frame, Frame Elements (FE)
• Primacy of semantic information
•FrameNet (Fillmore et al. 2003; http://framenet.icsi.berkeley.edu)
37

Structure of Lexical Entry
• Frame description
• LUs evoking the frame: load, spray, etc.
• Exhaustive inventory of how FEs are
realized syntactically
• Annotated example sentences
38

Lexical Entry of load in the Filling Frame :
Frame description
39

Lexical Entry of load in the Filling Frame :
Summary of Valencies
40

Lexical Entry of load in the Filling Frame :
Annotated example sentences
41

The Placing Frame: Frame description
• Frame description: An AGENT places a
THEME at a location (the GOAL), which is
profiled (Langacker 1987). The THEME is under
the control of the AGENT at the time of its
arrival at the GOAL.
• LUs evoking the frame: bag, brush, drape,
load, pack, spray, etc.
• Exhaustive inventory of how FEs are
realized syntactically
42

Lexical Entry of load in the Placing Frame :
Summary of Valencies
43

Lexical Entry of load in the Placing Frame :
Annotated Example Sentences
44

Baker & Ruppenhofer (2002)
45

Advantages of frame-semantic
classification
• LUs are described with respect to the semantic frames
they evoke (splitting approach, coherence)
• More fine-grained classification
• Full syntactic information is available, but is not taken as
a necessary indicator for class membership
• See also Faber and Mairal Usón (1999:2): “Semantic
considerations are all-important because syntactic
distinctions in themselves are not sufficient as a basis for
the establishment of an inventory of semantic
categories.”
46

What is the relationship between the
syntactic behavior of a verb and its
meaning?
• Syntactic alternations offer limited insights
• Refine Frame Semantics with the notion of “verb
descriptivity” (Snell-Hornby 1983)
• More fine-grained analysis of verb meaning
• Analysis of participants and circumstances
• Additional semantic elements implicitly associated
with a verb’s meaning
50

Verb descriptivity
strut
ANu (act nucleus) Mod (modificants)
- walk - distinct physical characteristics (stiff, erect)
- value judgments passed on the agent and
his manner of walking (proud, pompous,
self-satisfied, etc.)
51

Verb descriptivity
• strut: high level of verb-descriptivity
• “descriptive verbs” (DV):
DV = ANu + Mod (+x)
x is an optional element without evaluative properties and not
expressible in terms of adjectives or manner adverbs
52

Two types of verb descriptivity
• Direct descriptivity: modificant refers
directly to the activity described by verb
(e.g. shout)
• Indirect descriptivity: modificant refers to a
participant or circumstance behind the
action or a combination of these (e.g.
strut)
53

Degrees of descriptivity
• The more semantic weight is taken by the
modificant as against the act-nucleus, the higher
the degree of descriptivity
• BUSTLE: high degree of descriptivity
– Modificant of bustle: excitedly, energetically, often
with apparent purpose, but usually noisily or
inefficiently, …
– Act-nucleus of bustle: not clearly definable (activity
such as behave, move about, etc.)
54

Low degree of descriptivity
• Modificant of shout: loudly
• Act-nucleus of shout: say, speak
55

Importance of level of descriptivity
• Level of descriptivity appears to influence
syntactic distribution (range of argument
realization) (see Hunston & Francis 2000)
• Light verbs: come, go, make, take, get, …
• Verbs with low level of descriptivity: shout, …
• Verbs with high level of descriptivity: grovel,
bustle, etc.
56

Integrating “verb descriptivity” into
Frame Semantics
• How do different lexical units evoking the
same frame differ in their degree of
descriptivity?
• Is there a correlation between a verb’s
level of descriptivity and the range of
syntactic patterns in which it occurs?
57

Case study: Self_motion frame
• Situations in which the SELF_MOVER, a
living being, moves under its own power in
a directed fashion, i.e. along a path what
could be described as a PATH, with no
separate vehicle. The presence of the FE
PATH presupposes the presence of other
FEs such as SOURCE, GOAL,
DIRECTION, and AREA.
58

First approximation of descriptivity
59

More motion verbs
60

Comparison of motion verbs
• They all evoke the Self_motion frame
• Semantics of FEs are drastically different:
–AGENT of walk, parade, totter, stagger
– PATH of walk and parade vs. totter and
stagger
• Profiling of Frame Elements differs
61

Differences in descriptivity
62

Correlation between degree of
descriptivity and range of argument
realization?
• Syntactic alternations and grammatical
constructions to test for range of argument
realization
63

Range of argument realization
Inclusion of locative PP
(33) a. Gerry {walked/paraded/staggered/tottered}.
b. Gerry {walked/paraded/staggered/tottered} down the street.
Zero-related nominal
(34) a walk, a parade, a stagger, a totter
64

Range of argument realization
Resultative construction
(35)a. Cathy {walked/?paraded/*staggered/*tottered} herself to exhaustion.
b. Cathy {walked/*paraded/*staggered/*tottered} Pat off the street.
Locative preposition drop alternation
(36) a. Julia {walked/paraded/staggered/tottered} across the town.
b. Julia {walked/paraded/*staggered/*tottered} the town.
65

Range of argument realization
Induced action alternation
(37) a. Claire {walked/paraded/*staggered/*tottered} the dog down the street.
b. The dog {walked/paraded/staggered/tottered} down the street.
Adjectival passive participles
(38) the {walked/??paraded/*staggered/*tottered} dog
66

Degree of descriptivity correlates
with range of argument realization
•LUssuch as walk: low level of descriptivity
and broad range of argument realization
•LUssuch as totter: high level of
descriptivity and narrow range of argument
realization
68

Combining Frame Semantics, Verb
Descriptivity, and Componential Analysis
71

Next steps
• Analyze each semantic frame in terms of an act
nucleus (abstract representation)
• Develop fine-grained analysis of Frame
Elements and their descriptive properties
(componential analysis/verb descriptivity)
• Compile set of test criteria that allow for a
stream-lined catalogue of meaning elements
• Investigate profiling properties and relate them
to syntactic properties
72

Next steps – locative alternation
• Profiling properties of FEs AGENT, THEME, GOAL
• Inherent meaning properties of verbs (components)
archive
load
wallpaper
Placing Frame
YES
YES
NO
Filling Frame
NO
YES
YES
73

Next steps – constructicon
• Extension of FrameNet methodology to
capture semantic and syntactic distribution
of syntactic constructions
• Identifying syntactically relevant units of
meaning in syntactic constructions
74

Conclusions
• Reliance on syntactic information for
defining verb classes is problematic
75