In Ik-Hwan Lee et al (2012) Issues in English Linguistics (Papers from the 1st World Congress of Scholars of English Linguistics, Hanyang University, Seoul, South Korea, June 30, 2012) Hankookmunhwasa, Seoul. pp. 119-131.

Resultatives and the Problem of Exceptions Stephen Wechsler University of Texas [email protected]

Abstract English resultatives are secondary predicates indicating the result of the action described by the primary predicate.

The class of adjectives used in such constructions is conditioned by

lexicosemantic factors: gradable adjectives with an inherent scalar maximum are strongly favored, leading to an understanding of resultatives as event-argument homomorphisms (Wechsler 2001; 2005a; 2005b). Counter-examples have been noted (Boas 2003; Broccias 2003; Iwata 2008a; 2008b).

But a statistical analysis of corpus data confirms that the correlation is highly significant.

The problem of exceptions should be addressed by means of a multi-factorial approach, which is argued to be a good way to analyze lexical syntax.

1. Introduction As syntacticians we attempt to determine the rules of grammar, but we are also accustomed to finding many exceptions to those rules.

In the famous words of Edward Sapir, ‘Were a

language ever completely “grammatical” it would be a perfect engine of conceptual expression. Unfortunately, or luckily, no language is tyrannically consistent. All grammars leak.’ (Sapir 1921, 38)

Perhaps nowhere is that leakiness more apparent than in the interaction of word meaning

with syntax.

The immense body of research on that interface, in particular in the study of

argument expression, has yielded many important insights.

But there is a tendency to rely upon

anecdotal observations in the absence of more rigorous methods for choosing between competing hypotheses.

Scholar X proposes a theory and supports it with his favorite language data, then

scholar Y shoots it down with a different set of examples, and so on in endless circles. How can we break out of those circles and advance the field? causal factors.

One way is to tease out the

Instead of seeking categorical rules that could be refuted with any arbitrarily

small set of counter-examples, we can look for statistically significant causal factors. approach is standard in every other field of science. multivariable world.

This

As a biostatistician notes, ‘We live in a

Most events, whether medical, political, social, or personal, have multiple

causes.’ (Katz 2006, 1)

Multivariable analyses of language have become more common; see for

example Bresnan et al (2005) for a such multivariable study of the dative alternation. This paper is a case study on the English resultative construction, such as Sue hammered the metal flat.

In earlier work I argued, based upon a corpus study, that such constructions

119

strongly favor a certain semantic type of adjective, namely maximal endpoint closed-scale adjectives (Wechsler 2005a; 2005b).

I tried to show that identifying the lexicosemantic

restrictions on the adjective could be the linchpin for a broader understanding of the construction as a whole, perhaps helping to unravel some of its mysteries by pointing to an analysis of such constructions as expressing event-argument homomorphisms.

But at least three scholars have

criticized this work, pointing to exceptions that they found (Boas 2003; Broccias 2003; Iwata 2008a; 2008b).

So the present paper bolsters those earlier results with a demonstration of strong

statistical significance.

Section 2 describes the facts and presents the earlier analysis. Section 3

summarizes the results of a corpus study. those results is demonstrated.

In Section 4 the strong statistical significance of

Section 5 constrasts resultatives with lexical causatives.

Section 6

discusses the problem of exceptions, and Section 7 concludes.

2. Resultatives and the semantics of scalar adjectives English resultatives are secondary predicates indicating the result of the action described by the primary predicate. 1

The predicate flat in sentence (1) is a resultative because the sentence

entails that the metal became flat as a result of the hammering (entailment is indicated by the double arrow ⇒).

(1) Resultative (predicate in italics; its subject underlined): John hammered the metal flat. ⇒‘John hammered the metal; as a result, the metal became flat.’

In contrast, The chairman came to the meeting drunk is a depictive, not a resultative, because drunkenness is not an entailed result of the action. Constraints on the adjective appearing in the English resultative construction have long been noted, as in these contrasts observed by Green (1972):

(2) He wiped it clean / dry / smooth / *damp / *dirty / *stained / *wet. (Green, 1972, ex. 6b/7b)

Is there any pattern to these constraints on the adjective?

If so, what is the pattern and how

can it be explained? In Wechsler (2001, 2005a,b) I proposed a solution that drew upon independently identifiable semantic subcategories of scalar adjectives (Kennedy 1999; Rotstein and Winter 2004; Kennedy 1

This section is based on Wechsler (2005b).

120

and McNally 2005).

Gradable adjectives can be modeled with scales, that is, ordered sets with

a measure function.

The adjective is interpreted relative to a standard degree on that scale:

Michael Jordan is tall means that Jordan’s height is greater than some contextually determined standard of tallness (Kennedy 1999, inter alia). the structure of the scale.

Gradable adjectives are classified according to

Maximal endpoint closed-scale gradable adjectives like full, empty,

straight, and dry have an inherent maximum that serves as a default standard which applies when it is not overridden by context (Kennedy 1999).

Such adjectives are often paired with

antonymous minimal endpoint closed-scale gradable adjectives:

dry/wet, clean/dirty, and so on.

Rotstein and Winter 2004, following Yoon 1996, calls the members of such pairs total and partial adjectives, e.g. dry is total while wet is partial.

The inherent standard for the total

adjective (e.g. clean) is always at the minimal point of the scale associated with the antonymous partial adjective (e.g. dirty), but not vice versa.

So on the dirtiness scale, clean means roughly

‘no degree of dirtiness’ while dirty means ‘some degree of dirtiness’.

But on the cleanliness

scale, dirty does not mean ‘no degree of cleanliness’, nor does clean mean ‘some degree of cleanliness’. Unlike the maximal endpoint (or total) and minimal endpoint (or partial) adjectives, openscale gradable adjectives like tall, long, wide, short, and tall lack endpoints altogether, and hence must always rely on context for their standards (Kennedy and McNally 1999, Hay et al. 1999). One test for closed- versus open-scale is the appropriateness of modifiers like totally or completely:

(3) a. completely full/ empty/ straight/ dry

(closed-scale)

b. ?? completely long/ wide/ short/ tall

(open-scale)

Kennedy and McNally (1999, fn. 1) note that the completely-test is complicated by the fact that completely sometimes means not ‘maximal on the scale’ but rather ‘very’, and in that latter meaning it appears with open-scale adjectives.

(Similarly, Sapir (1944:114) commented that the

use of the superlative in locutions like a most pleasing personality is ‘logically unreasonable but psychologically somehow inevitable.’)

They point out that entailments differ, making (4a) but

not (4b) contradictory:

(4) a. #The line is completely straight, but it could be straighter. b. I’m completely uninterested in finances, but Bob is even less interested.

By coupling the completely test with a check of such entailments, we get a fairly reliable test of open- versus closed-scale adjectives.

121

As a second test, total adjectives systematically differ from both other types in the acceptability and interpretation of modifiers like almost (Rotstein and Winter 2004):

(5) a. It is almost dry.

(total)

b. #It is almost wet.

(partial)

c. #It is almost long.

(open-scale)

In certain rich contexts, partial and open-scale adjectives also allow modification by almost, but the interpretation differs systematically from total adjectives: almost wet, if acceptable, entails ‘not dry’, while almost dry does not entail ‘not wet’ (Rotstein and Winter 2004:267):

(6) a. The towel is wet but it is almost dry. b. #The towel is dry but it is almost wet.

A consequence of that asymmetrical entailment is that (6a) is coherent but (6b) is not.

Other

phenomena that have been claimed to distinguish these adjective types include the (a)telicity of cognate ‘degree achievement’ verbs like straighten, flatten, and coolV

(Hay, Kennedy, and Levin

1999) and differences involving plurals and ‘donkey anaphora’ (Yoon 1996). With these semantic categories in hand, we can solve the resultative puzzle.

When the verb

in a resultative construction is durative, a maximal endpoint closed-scale gradable adjective (i.e. a total adjective) is strongly preferred.

Referring to example (2) above, the acceptable adjectives

clean, dry, and smooth are all maximal endpoint adjectives (MaxEndpt), while the unacceptable damp, dirty, stained, and wet are minimal endpoint adjectives (MinEndpt). The next section reviews this hypothesis in more detail, and then presents the results of a corpus study supporting this correlation.

3. A corpus study supporting the Maximal Endpoint Hypothesis The essential findings reported in my two short papers are as follows.

Boas’ (2000) exhaustive

study of a 100 million word corpus, primarily from the British National Corpus, turned up 869 resultative constructions containing the following nine adjectives:

clean, dry, flat, full, open, red,

shut, smooth, solid (out of 89,206 total tokens of these adjectives).

Meanwhile, the same sample

turned up only a single token resultative construction containing any of the following eight adjectives: famous, fat, ill, sleepy, sore, tired, dirty, wet (out of 26,909 tokens).

The sole

instance of a resultative from the second group was a single example using the word sore. Every adjective in the first list above can be independently classified as a ‘maximal endpoint

122

closed-scale adjective’ (MaxEndpt) based upon a battery of semantic tests that are independent of resultatives.

No adjective in the second list belongs to this type.

So out of a controlled sample

of these 17 adjectives, MaxEndpt adjectives are strongly favored over non-MaxEndpt adjectives, when it comes to appearing in resultative constructions: 869 of the former but only 1 of the latter showed up in the sample (although there were more of the former type of adjective overall; see Section 5 below).

This correlation is described and supported in more detail in Section

5 below. Let us review the argument, starting with the empirical observation that formed the basis for my analysis:

(7) A controlled sample of resultative constructions (870 tokens) exhibited a very strong preference for MaxEndpt adjectives (869 token) over non-MaxEndpt adjectives (1 token).

Based on this strong correlation, and in the absence of any confounding factors that I could think of, I hypothesized that:

(8) The Maximal Endpoint Hypothesis (MEH). For English adjectives, there is a causal link between the MaxEndpt property and the ability of the adjective to appear in a resultative construction.

Then I proposed a particular explanation for this causal connection, namely:

(9) The compositional semantics that combines the verb and adjective meanings in a resultative construction involves a homomorphic mapping between the adjective’s property scale and the temporal structure of the verb-denoted event.

Assuming the

event-argument homomorphism theory of telicity (Krifka 1998; Jackendoff 1996), we predict that resultatives, being telic, would favor adjectives with scales bounded by an inherent maximum, hence MaxEndpt adjectives.

This parallel between the resultative adjective’s property scale and the event’s temporal scale fits in with a theory of telicity that is independently supported in various other domains within the work of Manfred Krifka and others. The explanation provided above crucially relies upon the stipulation that resultative constructions are telic. the construcion.

I posited simply that telicity was a grammatically specified property of

But in recent unpublished work, Richard Larson (personal communication) put

forth an interesting proposal that does away with that stipulation.

123

Simplifying (and perhaps

reinterpreting and modifying) slightly, Larson’s proposal is based on the notion that the grammar must allow the scalar adjective some means for fixing its standard.

Recall that the standard for

an attributive or primary predicate scalar adjective can normally come from the context, and for an open-scale adjective it must come from context: Michael Jordan is tall means that Jordan’s height exceeds some standard of tallness defined by the context, perhaps relative to other basketball players, or to people in general.

I have assumed that in a resultative construction, the

adjective and verb compose semantically so that the scalar structure of the adjective fuses with the temporal structure of the verb.

Suppose that this fusing renders the verb-adjective semantic

complex opaque to contextual valuation of the scalar standard. to set the standard, an inherent standard must be exploited.

Since the context is not available

Thus, only those adjectives with an

inherent standard, that is, only maximal endpoint closed scale adjectives, are possible. The last sentence may be an overstatement of the strength of the MaxEndpt requirement, since the verb-adjective combination may not be completely impervious to contextual standardsetting.

If it ‘leaks’ and a contextual standard penetrates this semantic complex, then an open

scale adjective becomes possible.

Uegaki (2009) analyzes Japanese resultatives as differing from

English in allowing such contextual setting of the standard.

Perhaps under certain conditions

English allows the Japanese-style contextual valuation of the standard.

If so then the conditions

for such exceptional cases are not understood, but this may eventually explain some of the exceptions discussed in Section 6 below. The MEH has been somewhat influential in research on resultatives (Beavers 2002; 2008; Uegaki 2009).

But at least three scholars have responded to this work critically, pointing

out counter-examples to the MEH and rejecting the hypothesis on the basis of those counterexamples (Boas 2003; Broccias 2003; Iwata 2008a; 2008b). Before examining their attempts to refute the MEH (Section 6), I will present some quantitative data supporting it.

4. Statistical significance of the Maximal Endpoint Hypothesis Wechsler (2001, 2005a, 2005b) did not include a statistical analysis, but I will add one now.

I

have also rechecked the data and filled out the analysis by adding all the other adjectives from Boas’ original study.

For the record, further below I also analyze the original data reported in

Wechsler (2001, 2005a, 2005b), since it is that earlier work that some scholars have critiqued. Tables 1, 2, and 3 display data from Boas’ (2000) corpus study, for three types of adjective, respectively: Maximum endpoint closed scale adjectives (MaxEndpt), Minimal endpoint closed scale adjectives (MinEndpt), and open scale adjectives.

124

Table 1.

Maximum endpoint closed scale adjectives (MaxEndpt’s)

MaxEndpt

all

Resultatives

ECM

make/etc.

awake

1379

39

2

0

clean

6092

102

0

6

dry

6236

77

1

8

empty

5409

1

1

0

flat

5899

34

0

2

full

26545

35

0

1

open

20457

395

0

1

red

11609

11

0

4

shut

5333

207

0

0

smooth

3045

5

0

12

solid

3990

3

0

2

sober

605

0

1

2

unconscious

1386

29

1

0

TOTAL

97985

938

6

38

causative verb

1 make; 1 render

Table 2. Minimal endpoint closed scale adjectives (MinEndpt’s) MinEndpt

all

Resultatives

ECM

make/etc.

causative verb

dirty

2591

0

0

41

8 make; 33 get

wet

3787

0

0

12

TOTAL

6378

0

0

53

The all column shows the total number of tokens of this adjective in the corpus.

The

resultatives column shows how many of those tokens occur in resultative constructions where the adjective heads the resultative secondary predicate.

The ECM (Exceptional Case Marking)

column is for special resultatives in which the object NP is not normally selected by the verb, such as We laughed ourselves sick (cf. *We laughed ourselves). The column labeled make/etc. lists the number of occurrences of this adjective in construction with a lexical causative like make, get, render, drive, or send (discussed in Section 5 below).

By far the most common lexical causative verb overall is make.

The causative verb

column indicates the breakdown of the lexical causative tokens enumerated in the make/etc. column between different lexical causative verbs: a blank cell in this column indicates that all the make/etc. examples were formed with the verb make, and any of them using other lexical causative verbs are noted.

125

Table 3. Open scale gradable adjectives open scale

all

Resultatives

ECM

make/etc.

crooked

314

2

0

1

famous

6344

0

0

37

fat

4137

0

0

5

ill

4808

0

0

65

sleepy

401

0

0

19

sore

852

1

0

11

tired

3989

0

0

18

insane

384

0

0

25

safe

8069

0

0

67

mad

2847

0

0

148

hoarse

236

0

9

1

sick

4243

1

12

136

soft

1107

1

0

0

stupid

3083

3

5

6

tender

1632

0

0

3

thin

5081

12

0

10

TOTAL

47527

20

26

552

causative verb

23 drive; 1 send; 1 make

108 drive; 5 send; 35 make

5 make; 1 drive

For example, the second row of Table I indicates that the adjective flat occurred 5899 times overall (including attributive uses as in a flat screen, predicative uses as in The screen is flat, etc.), of which 34 were resultative secondary predicates as in All the grass grinds their little molars flat; there were no ECM resultatives with flat; and there were two lexical causatives, one with make (…evaporation will have made the drop even flatter) and one with render (…they also render life experiences flat). The results indicated by the TOTALs in the bottom rows of Tables 1-3 are summarized in Table 4.

Non-MaxEndpt figures indicate the sums of figures for the open scale and

MinEndpt types.

As shown in the right-most column, the number of resultatives as a percentage

of total occurrences of the adjectives is much higher for MaxEndpt (almost 1%) than for nonMaxEndpt (less than .04%).

Table 4. Frequency of adjective by type in resultative versus other functions All

Resultative

non-Resultative

% resultative

MaxEndpt:

97,985

938

97,047

0.957%

non-MaxEndpt:

53,905

20

53,885

0.037%

126

Analysis:

The chi-square value (Yates chi-square, corrected for continuity) for these data is χ2 =

468.34 (df = 1); p < .0001 .

In prose, this means that the estimated probability of this

distribution being the result of chance is less than 1 in 10,000. for statistical significance is:

A common standard threshold

p < .05, or 1 in 20 chances of being the result of chance.

So

this is a very strong demonstration of significance of the correlation between MaxEndpt and resultatives. Since the critiques discussed below concerned Wechsler (2001, 2005a,b), for the sake of completeness I will analyze those data as well.

Tables 5 and 6 show the figures.

Table 5. Maximum Endpoint closed-scale adjectives (from Wechsler 2001) all

Resultatives

make/etc.

clean

6092

102

6

dry

6236

77

8

flat

5899

34

1

full

26545

35

1

open

20457

395

1

red

11609

11

4

shut

5333

207

0

smooth

3045

5

12

solid

3990

3

2

TOTAL

89206

869

35

Table 6. Open-scale and minimal endpoint adjectives (from Wechsler 2001) all

Resultatives

make/etc.

famous

6344

0

37

fat

4137

0

5

ill

4808

0

65

sleepy

401

0

19

sore

852

1

11

tired

3989

0

18

dirty*

2591

0

41

wet*

3787

0

12

total

26909

1

208

*Minimal endpoint adjectives

127

The frequencies are shown in Table 7.

Note that the correlation here is even greater.

In fact

there was only a single non-MaxEndpt resultative.

Table 7. Frequency of MaxEndpt and non-MaxEndpt adjectives in resultative versus other functions, from the adjective list in Wechsler 2005a,b All

Resultative

non-Resultative

% resultative

MaxEndpt:

89206

869

88337

0.974%

non-MaxEndpt:

26909

1

26908

0.0037%

The chi-square value (Yates chi-square, corrected for continuity) for these data is χ2 =

Analysis:

260.5 (df = 1);

p < .0001 .

However, the χ2 is known to be unreliable for samples with very

small numbers in any of the cells and for highly unbalanced distributions, so other significance tests should probably be applied.

In any case, the correlation is clearly even stronger than in

the fuller data set discussed above. In conclusion, this corpus study shows that the probability that an adjective functions in a resultative construction, as opposed to some other grammatical function, is much greater for MaxEndpoint adjectives than for other types.

5. Lexical causatives The comparison between resultatives and lexical causatives provides another way to look at these data.

Resultatives express the result of an action:

Mary hammered the metal flat means

that Mary hammered the metal, which caused the metal to become flat.

So instead of

comparing frequency of resultative adjectives to overall frequency of the adjectives, we could compare frequency of resultatives with a given adjective to the overall frequency of sentences expressing the causation of the adjective-denoted state.

Such sentences include at least

resultatives like Mary wiped the table dry and lexical causatives like Mary got her hair dry. Fortunately we have data on those two types.

In addition, the corpus may contain various other

paraphrases, for which we lack data. But the resultative and lexical causative data allow us to compare MaxEndpt adjectives with other adjectives, across these two construction types. Totaling the resultative and lexical causative tokens, for each adjective X, gives us a rough measure of the number of times ‘cause to become X’ was expressed. shown in the All eventive column of Table 8.

128

The totals are

Table 8. Frequency of MaxEndpt and non-MaxEndpt adjectives in resultatives versus lexical causatives All eventive

Resultative

Lexical causative

% resultative

MaxEndpt:

976

938

38

96.1%

non-MaxEndpt:

625

20

605

3.2%

The percentages in the rightmost column show that for MaxEndpt adjectives X, resultatives make up a very large percentage (96.1%) of the total number of expressions of ‘causing to become X’, while for non-MaxEndpt adjectives, resultatives make up a very small percentage (3.2%). Analysis:

The chi-square value (Yates chi-square, corrected for continuity) for these data is χ2 =

1364.61 (df = 1); p <.0001 .

So the χ2 value in this test is even higher.

This suggests that

when a speaker of English chooses to express the notion ‘cause to become X’, they are much more likely to use a resultative to do so if the adjective X is a MaxEndpt adjective than if it is not one.

6. The data game

The correlation found in this controlled study seem to support the MEH so strongly that I would have thought a statistical analysis unnecessary. resultatives found the earlier work convincing.

Nonetheless, not all researchers working on At least three scholars responded to this work

critically, pointing out counter-examples to the MEH and rejecting the hypothesis on the basis of those counter-examples (Boas 2003; Broccias 2003; Iwata 2008a; 2008b). Let us take a look at their critiques. Boas (2003, 136-7) found sentences in some web Newsgroups containing resultatives that are formed with minimal endpoint adjectives: he cites three with damp, two with dirty, and two with wet.

For example, one of the sentences is Every day I wipe it wet with WD-40 before I ride

and then wipe it dry after my ride.

Broccias’ (2003, 144) counter-examples also included

sentences containing wet, in his case two constructed examples, one of them I sprinkled them wet with a garden hose. Iwata (2008a, 1097) repeated Broccias’ constructed examples containing wet, and Iwata (2008b) found some examples using wet on the web. Other apparent counterexamples, involving words that were not included in my study, were mentioned by each of these three authors, some discussed below.

But since my study includes wet/dry and dirty/clean, let us

focus on those first. Table 9 shows the figures for these four adjectives.

Note the very strong correlation: for

dry and clean, resultatives are the favored construction, making up over 90% of the instances where ‘causing to become X’ was expressed by either a lexical causative or resultative.

129

But for

wet and dirty, the percentage of resultatives drops to zero.

Table 9. Frequency of selected adjectives in resultatives versus lexical causatives

MaxEndpt:

non-MaxEndpt:

All eventive

Resultative

Lexical causative

% resultative

dry

85

77

8

90.6

clean

108

102

6

94.4

wet

12

0

12

0

dirty

41

0

41

0

Let us consider the responses of these three authors. Iwata’s (2008b) reaction strikes me as the most surprising. significance to the observed correlation.

Iwata seemed to deny any

Focusing specifically on the verb-adjective combination

wipe…wet, Iwata noted that I had reported that no resultatives using that verb-adjective combination were found in the BNC sample. wet with any verb.)

(Indeed the sample contained no resultatives using

Regarding this observation, Iwata makes the following comment (from

Iwata’s 2008b handout, p. 2):

The fact that no instance of wipe-wet is found in the BNC does not prove that wipe-wet is impossible.

It may well be the case that the BNC happens not to contain any wipe-wet

data.

With the statistical analysis above, we can now put a numerical estimate to the probability that the corpus sample ‘happens not to contain’ non-MaxEndpt adjectives: according to a standard statistical measure, the estimated probability that the overall correlation is the result of chance is less than one in 10,000.

If Iwata means that the sample ‘happens not to contain’ non-MaxEndpt

adjectives, then this is an unusual and highly dubious position to take, given the strength of the correlation.

Rarely is data this well-behaved in any field of science, let alone data involving

human behavior, as in social science or linguistics.

If Iwata means that the sample ‘happens not

to contain’ the specific verb-adjective pairing wipe-wet, then he may be right since the probability of this gap being due to chance is unknown.

But that probability has no bearing on

the MEH, as far as I can tell. Broccias (2003, 144) cited his constructed examples of resultatives using wet and argued against the MEH and in favor of an alternative hypothesis.

In the following quote he referred

to example (2) above from Green (1972), which he reproduces as his example (167) (Broccias 2003, 143):

130

I would like to argue that the exclusion of the adjectives damp, dirty, stained, and wet from (167) [= (2) above] does not stem from their contextual interpretation as a minimal endpoint vs. maximal endpoint adjective.

Rather, the notion of expected

consequence seems to be the determining factor.

Broccias (2003, 144) explains that wiping normally involves an intention to make something clean or dry, so in using this verb with damp, dirty, etc., ‘we are resorting to a more complex operation.’

This seems very reasonable.

Sentences describing more expected events sound better

to informants than those describing unexpected ones. earlier paper (Wechsler 1997).

In fact I said the same thing myself in an

Supposing that this is indeed a determining factor, it surely does

not follow that it is ‘the’ determining factor, excluding any relevance of the scalar semantics of the adjective.

Why would there be just one determining factor?

Most events in the world have

multiple causes and multiple causative factors, and Broccias gives us no reason to think that this particular phenomenon should be different.

Crucially, he does not address the evidence for MEH,

namely the correlation observed in a large controlled sample. Broccias’ ‘expected consequence’ hypothesis cannot replace the MEH unless it accounts for that correlation better than the MEH does alone, and at least as well as a combination of both factors does.

We will not know the answer definitively until someone devises a clear measure

of how ‘expected’ a consequence is, and then carries out a controlled study.

But it is unlikely

that the ‘expected consequence’ factor alone can predict the data as well as MEH.

For example,

as shown by the data in Table 9, speakers avoid the use of resultatives with wet or dirty, preferring the use of lexical causatives, while they had the opposite preference for dry and clean. I know of no reason why things becoming wet or dirty would be systematically less expected than things becoming dry or clean.

If anything the opposite seems more plausible: just expose

something to water or dirt and you can be virtually assured that it will become wet or dirty, while getting it dry or clean is harder and less predictable. It has been said that one researcher’s counter-evidence is another’s issue for further research. Broccias (2003, 146-7) cited the following as counter-examples to the MEH, noting that thin and wide are not MaxEndpt adjectives (Broccias’ examples 173 and 174):

(10) a. John cut the bread thin. b. He opened the door wide.

Considering 10a first, note that in the corpus study reported here, the adjective thin alone accounts for 12 out of the 20 resultatives with open-scale adjectives (see Table 3).

So the

correlation supporting the MEH would be even stronger if some alternative account of examples

131

like (10) could be found.

As it happens, in later work Broccias (2008) reanalyzed such

sentences as ‘pseudo-resultatives’, noting that thin ‘was originally an adverb’ in the history of English, and that forms such as thin in (10) ‘behave like adverbs in that they evoke notions like manner and they cannot be predicated of the constructional object (e.g. bread) but, rather, of the resultant object (e.g. slices of bread).’ (Broccias 2008, 33).

Similarly, (10b) does not entail that

the door became wide, but rather that the opening did; nor is there any overt predication subject in the standard locution by a doctor wanting to examine the patient’s throat: Open wide! In a similar vein, Broccias contrasted these two (constructed) examples, citing Verspoor (1997):

(11) a. b.

*John hammered the metal red. John painted the fence red.

This contrast is also problematic for the MEH, according to Broccias.

He claims that ‘the only

difference’ between these sentences is that ‘red describes a (relatively) transient property’ in the first but not the second (Broccias 2003, 145).

This is argued to be a problem for my analysis

because ‘Wechsler’s formal model is insensitive to such a criterion.’ (Broccias 2003, 145)

(I do

not understand what example (11a) would mean, even if it were grammatical; hammering metal does not make it red.)

Anyway, relative transience of the redness property is not the only

difference between these sentences.

Maybe the color adjective is an (optional) argument of the

verb paint but not the verb hammer.

Whatever the explanation for this contrast, it does not

necessarily count against the MEH, as far as I can tell. applying to certain words.

It just involves a different factor

In fact examples with paint-red (11b), like the cut-thin combinations,

were also reanalyzed as ‘pseudo-resultatives’ by Broccias (2008, 28), who cited Halliday (1967) and Rapoport (1999). As noted above, Boas and Iwata found some resultatives with the adjective wet on the web. The practice of ‘cherry-picking’ examples from the web to support a theoretical claim is highly controversial (Lüdeling, Evert, and Baroni 2007).

In my opinion, it can sometimes be useful for

gathering anecdotal evidence for qualitative studies, that is, to find examples in order to examine them.

But use of the web raises many issues related to the unknown source the text (non-native

speakers, multilingual interference, children, dyslexics, aphasics, intoxication, jokes, typos, semiliteracy, prescriptivism, and so on). Moreover, the sheer size of the web and power of search engines like Google creates a serious problem in the absence of a control group. Further undercutting the value of such critiques is the opportunistic combining of cherrypicking the web to undercut negative predictions, and constructing unacceptable examples to undercut positive ones.

For example, Boas (2003, 137) saw a ‘more serious problem with

132

Wechsler’s analysis’, namely that it fails to rule out punctual verbs with path PP resultatives like to pieces. (PP resultatives were part of the earlier studies but are not discussed in this paper.) The exact reasons are not important here; basically, according to my analysis, path PPs are in principle allowable with punctual or durative events, since paths can be short or long. 2

The

point is just that Boas demonstrated the ‘serious problem’ for my analysis with these examples, where the judgments shown here are his:

(12) a. b.

The terrorist exploded the bomb (??to pieces). The soldier detotated the mine (??to pieces).

The problem, according to Boas, is that my theory wrongly predicts that such examples should be acceptable.

But on the previous page of his book, Boas (2003, 136) had presented some

web examples of resultatives with wet, etc., as problematic for a different aspect of my theory. Let us apply the same standard to his examples in (12).

A quick Google search for the string

exploded it to pieces returned 296 hits, of which the first ten appear to be genuine examples of the sentence type in (12a) that Boas deems unacceptable.

A couple of examples should suffice:

(13) The stupid thing had kept screaming until George finally went outside and exploded it to pieces. http://m.fanfiction.net/s/5697277/14/

(14) How is that a good hunting weapon? How could you eat what you hunted if you just exploded it to pieces? http://www.youtube.com/all_comments?v=DiSQAcyrF9M&page=1

In conclusion, there do not seem to be clear standards for hypothesis comparison in the critiques by Boas, Broccias, and Iwata.

None of them has actually addressed the correlation that I

presented as evidence for the MEH.

Neither cherry-picking from the web nor the use of

isolated constructed examples is sufficient to undercut a claim that has been established through a controlled quantitative study.

7. Conclusion

The Maximal Endpoint Hypothesis does not impose an absolute contraint on which adjectives can appear in the resultative construction, but rather a predicts that scalar structure of the adjective 2

Regarding the corpus data, I noted that punctual events showed an apparent preference for dead over to death, and suggested that might be for markedness reasons: to death is compatible with both, but being non-gradable, dead favors a short path so it is more specific. Boas argued that since there is no nongradable adjective analogous to dead that can block to pieces, therefore according to (his reading of) my reasoning, to pieces should be acceptable with punctual resultatives.

133

should be a factor.

It does not exclude other factors such as the influence of pragmatic context.

As discussed above, the explanation for language data is often multifactorial.

For example,

coercion of word meaning through pragmatic context is a well-established phenomenon from nearly all domains of lexical semantics (Pustejovsky 1993).

Count nouns can be coerced to

mass readings and vice versa; scalar readings can be forced onto normally non-scalar adjectives; and so on for virtually every phenomenon found to depend on word meaning.

Even reference

itself can be transferred, so that under the right conditions the phrase the scrambled eggs can be used to refer to a person (Nunberg 1995).

Hence wet, while a non-MaxEndpt adjective, might

be found to appear in a resultative construction if the pragmatic conditions are right for forcing a MaxEndpt reading onto this adjective.

This does not mean that we should abandon the

Maximal Endpoint Hypothesis and start over.

It rather indicates that additional factors may be

operative in determining whether an adjective can appear in resultative constructions.

References Beavers, John. 2002. Aspect and Distribution of Prepositional Resultative Phrases in English. LinGO Working Paper #2002-7. Stanford CA: CSLI, Stanford University. ———. 2008. “Scalar Complexity and the Structure of Events.” In Event Structures in Linguistic Form and Interpretation, ed. J. Dölling, T. Heyde-Zybatow, and M. Schäfer, 245–265. Berlin: Mouton de Gruyter. Boas, Hans. 2000. “Resultative Constructions in English and German”. PhD Dissertation, University of North Carolina. ———. 2003. A Constructional Approach to Resultatives. CSLI Publications. Bresnan, Joan, Anna Cueni, Tatiana Nikitina, and R. Harald Baayen. 2005. “Predicting the Dative Alternation.” Proceedings of the Royal Netherlands Academy of Science Workshop on Foundations of Interpretation. Broccias, Cristiano. 2003. The English Change Network: Forcing Changes into Schemas. Vol. 22. Walter de Gruyter. ———. 2008. “Towards a History of English Resultative Constructions: The Case of Adjectival Resultative Constructions.” English Language and Linguistics 12 (01): 27–54. Green, Georgia. 1972. “Some Observations on the Syntax and Semantics of Instrumental Verbs.” In Chicago Linguistic Society, 8:83–97. Chicago, Illinois. Halliday, M. A.K. 1967. “Notes on Transitivity and Theme in English: Part 1.” Journal of Linguistics 3 (1): 37–81. Hay, J., C. Kennedy, and B. Levin. 1999. “Scalar Structure Underlies Telicity in ‘degree Achievements’.” In

134

Proceedings of SALT, 9:127–144. Iwata, Seizi. 2008a. “A Door That Swings Noiselessly Open May Creak Shut: Internal Motion and Concurrent Changes of State.” Linguistics 46 (6): 1049–1108. ———. 2008b. Another Look at the Maximal End-point Constraint on Resultatives. Presentation at ICCG5, Austin, Texas, Sept. 27, 2008. Jackendoff, Ray. 1996. “The Proper Treatment of Measuring Out, Telicity, and Perhaps Even Quantification in English.” Natural Language and Linguistic Theory 14.2: 305–354. Katz, Mitchell. 2006. Multivariable Analysis: a Practical Guide for Clinicians. New York: Cambridge University Press. Kennedy, Christopher. 1999. Projecting the Adjective—the Syntax and Semantics of Gradability and Comparison. Outstanding Dissertations in Linguistics. New York: Garland. Kennedy, Christopher, and Louise McNally. 2005. “Scale Structure, Degree Modification, and the Semantics of Gradable Predicates.” Language: 345–381. Krifka, Manfred. 1998. “The Origins of Telicity.” In Events and Grammar, ed. Susan Rothstein, 197–236. Dordrecht: Kluwer. Lüdeling, Anke, Stefan Evert, and Marco Baroni. 2007. “Using Web Data for Linguistic Purposes.” In Corpus Linguistics and the Web, ed. Marianne Hundt, Nadja Nesselhauf, and Carolin Biewer, 7–24. Amsterdam/New York: Rodopi. Nunberg, Geoffrey. 1995. “Transfers of Meaning.” Journal of Semantics 12 (2): 109. Pustejovsky, J. 1993. “Linguistic Constraints on Type Coercion.” In Computational Lexical Semantics, ed. P. Saint-Dizier and E. Viegas, 71–97. Cambridge: Cambridge University Press. Rapoport, T. R. 1999. “Structure, Aspect, and the Predicate.” Language 75.4: 653–677. Rotstein, Carmen, and Yoad Winter. 2004. “Total Adjectives Vs. Partial Adjectives: Scale Structure and Higher-order Modifiers.” Natural Language Semantics 12 (3): 259–288. Sapir, Edward. 1921. Language: An Introduction to the Study of Speech. New York: Harcourt, Brace and Company. Uegaki, Wataru. 2009. “Japanese Resultative Phrases as Verbal Degree Modifiers.” Chronos 9, University of Paris. Verspoor, Cornelia Maria. 1997. “Contextually-Dependent Lexical Semantics”. PhD Dissertation, The University of Edinburgh. Wechsler, Stephen. 1997. “Resultative Predicates and Control.” In Texas Linguistic Forum 38: The Syntax and Semantics of Predication, ed. Ralph Blight and Michelle Moosally, 307–321. Austin, Texas: University of Texas Department of Linguistics. ———. 2001. “An Analysis of English Resultatives Under the Event-Argument Homomorphism Model of Telicity.” In Proceedings of the 3rd Workshop on Text Structure, 1–15. Austin, Texas: Department of Linguistics, University of Texas.

135

———. 2005a. “Resultatives Under the Event-Argument Homomorphism Model of Telicity.” In The Syntax of Aspect—Deriving Thematic and Aspectual Interpretation, ed. Nomi Erteschik-Shir and Tova Rapoport, 255–273. Oxford: Oxford University Press. ———. 2005b. “Weighing in on Scales: a Reply to Goldberg and Jackendoff.” Language 81.2. Yoon, Youngeun. 1996. “Total and Partial Predicates and the Weak and Strong Interpretations.” Natural Language Semantics 4 (3): 217–236.

136