Psycholinguistic Investigations: A Survey and Assessment

1
Psycholinguistic Investigations A Survey and
Assessment

• Anne Bezuidenhout
• Trondheim, September 18-22, 2006

2
BeginningsGibbs Moise (1997)
GM interested in gathering experimental evidence
that could be used in the debate about the notion
of what is said (a) Explicature (Relevance
Theory) (b) Minimal proposition (Griceans) Forced
choice tasks participants chose paraphrases that
best captured what is said Sentences were
presented either in isolation or in story
contexts, including contexts supporting the
minimal interpretation. Tapping peoples
intuitions testing Recanatis Availability
Principle.
3
Gibbs Moise (1997) Materials
GM classified their sentence materials into 5
categories (1) Cardinals Billy has two
balls. (2) Quantificational sentences Everyone
is a vegetarian. (3) Time-Distance sentences The
mountains are some distance away. (4) Possession
sentences Robert broke a finger. (5)
And-coordinations Judith insulted him and he
left. Note Not all these sentences are
associated with Gricean generalized
conversational implicatures (GCIs). Only types
(1), (4) and (5) are. And the only ones
associated with scalar implicatures (SIs) are
type (1). Forced choice task Billy has two
balls (i) Billy has at least two balls. (ii)
Billy has exactly two balls. Pick the one that
best represents what the (imagined) speaker said.
4
Gibbs Moise (1997) Results
People favored explicature paraphrases (which in
conditions (1), (4) and (5) meant favoring
GCIs). In minimal contexts people favored minimal
paraphrases. Even when given brief pre-test
training in which Gricean minimal sayings were
distinguished from enriched sayings, people
continued to favor explicature paraphrases. GM
took this to vindicate Recanatis Availability
Principle and to show that people have intuitions
about what is said. This intuitive notion
corresponds to the enriched one favored by
Relevance Theorists, Recanati and other
contextualists.
5
Nicolle Clarks (1999) Challenge
NC challenge the idea that GM have vindicated
Recanatis Availability Principle. Their
experiments used contextualized sentences from
GMs 5 categories. They used variants of GMs
forced-choice task, where participants had to
chose a paraphrase that best matched what the
speaker had said, but they added options. NCs
participants favored implicature paraphrases
rather than explicature paraphrases. Moreover,
people seemed to be rather insensitive to the
theoretical distinctions between what words mean,
what is said by the use of those words, and what
is communicated by their use. NC offer a
relevance theoretical explanation for their
pattern of results and then attempt to directly
test this explanation.
6
Bezuidenhout Cutting (2002)
Experiments 1 and 2 were off-line tasks similar
to those of GM (1997) and NC (1999), and were
attempts to replicate some of their
findings. Experiments 3 and 4 were attempts to
get on-line measures to address more directly the
question whether minimal propositions play a role
in processing. We used contextualized sentences
of the 5 sorts used by GM, but we also added a
6th category, namely (6) Perfectives Ive had
breakfast. Also, like NC (1999), we restricted
our possession sentences to ones expressing
inalienable possession.
7
BC (2002) Exps 1 and 2
Example of story context with 5 paraphrase
choices used in Experiment 2 Jane was planning
to spend a week in Memphis. She couldnt decide
which of her friends, Brian or Paul, to stay
with. Jill, who knew that Jane was allergic to
animals, said, Brian has three cats. (1)
Minimal Proposition Brian has at least three
cats. (2) Explicature Paraphrase Brian has
exactly three cats. (3) Strong Implicature Jane
should stay with Paul. (4) Weak Implicature
Brian likes cats. (5) Implicated Premise
Brians cats will cause Jane to have an allergic
reaction.
(How did we decide on these choices? Based on Exp
1, which was an open-ended task for each item,
participants were asked to take about one minute
and write down as many of the things that they
thought were conveyed by the target sentence in
its context as they could.)
8
BC (2002) Exp2 Results
Results of forced choice task used in Experiment
2 Type of Paraphrase Chosen Frequency with
which Paraphrase Type was Chosen (express
ed as of total number of choices
made) Strong Implicature 38.43 Explicature
Paraphrase 23.55 Implicated Premise 22.56 We
ak Implicature 9.17 Minimal Paraphrase 6.30
9
BC (2002) Exps. 3 and 4 Materials
Sample of a target sentence with its two story
contexts and the sentence used for the
match/mismatch judgment task Enriched
Context Roger was directing a musical. For one
scene he needed extras to play a group of
onlookers watching a street fight. But the stage
would already be pretty crowded with the
principal actors in this scene. He figured that
in order to prevent the scene from looking too
chaotic, he needed six people. Minimal
Context Roger was trying to arrange a rafting
trip for his scout troupe during the summer.
Many of his scouts were on family vacations. He
asked the rafting company how many committed
participants he would need to be able to secure
the reservations for the trip. They told him
that he needed six people. HE NEEDED AT LEAST
SIX PEOPLE Exp 3 was used to select the most
effective 24 pairs of contexts from an initial
set of 36 pairs.
10
BC (2002) Predictions of 3 Models
(1) Literal-First Serial Model (2) Ranked
Parallel Model (equivalent to Enriched-First
Serial Model) (3) Local Pragmatic Processing Model
11
BC (2002) Results

12
Shift in Emphasis
One criticism of Gibbs Moise (1997) is that
they cast their net too wide, focusing on too
diverse a sample of context-sensitive
expressions (a) Scalar implicatures (b) Other
sorts of non-scalar generalized conversational
implicatures (GCIs) (c) Cases of quantifier
domain restriction (d) Other cases of what Bach
calls sentence non-literality (the some time
and some distance cases) Criticism applies also
to Bezuidenhout Cutting (2002), who raised this
as a worry for their own findings. Subsequent
work has tended to focus very narrowly on cases
involving scalar implicatures.
13
Noveck Posada (2003) ERP Study
Compares neo-Gricean (Levinson, 2000) and
Relevance Theoretical accounts of the scalar
implicature (SI) not all associated with
some. According to NP, RT says that SI
generation is a non-automatic, late-occurring,
effortful pragmatic inference that will only be
made if it is contextually needed to achieve the
expected level of relevance (p. 209). The
Neo-Gricean, on the other hand, says that SIs are
generated automatically and effortlessly. Only
when context clashes with the SI will there be a
need to expend effort to cancel the inappropriate
SI.
14
Noveck Posada (2003) Materials
Materials (i) Patently True Some animals have
stripes. (ii) Patently False Some dogs have
wings. (iii) Underinformative Some elephants
have trunks. In earlier work and in a pretest of
these new materials, Noveck found two sorts of
responses by adults to type-(iii) sentences.
(Children in the earlier study were notoriously
different from adults in this regard but that
is another story!). Logical responders judge
type-(iii) sentences to be true, whereas
pragmatic responders judge them false.
15
Noveck Posada (2003) Predictions
Reaction Times Relevance Theory Pragmatic
responses will be longer than logical ones,
because to derive the SI involves
effort. Neo-Gricean account Logical responses
will be longer than pragmatic ones, since SI
retrieval is automatic, whereas a logical
response requires costly cancellation of the
SI. ERPs The ERP data is supposed to provide
evidence for the time-course of SI recovery. NP
assume that RT predicts late recovery of SIs
(after the literal meaning has been decoded),
whereas the neo-Gricean predicts early recovery,
during the stage of semantic integration (p.
205). They claim that large N400s for type-(iii)
sentences would be evidence of on-line
detection of pragmatic anomalousness and
support the neo-Gricean idea that pragmatic
processes intrude into semantic ones.
16
Questionable Assumptions
(1) NPs understanding of the RT notion of
underspecificity is, in my view, incorrect. Both
the some and possibly all and the some but not
all interpretations are potentially effortful
according to RT, since both are enrichments of an
underspecified form. So RT does not necessarily
predict that pragmatic responses should be
longer than logical ones. It all depends on the
context. (See NPs own ex. Do all Italians like
ice cream? Answer Some do.) (2) NP seem to
attribute a literal-first serial model of
processing to RT. This is to confuse RT with
classical Griceanism. RTists have long argued for
pragmatic intrusion.
17
Noveck Posada (2003) Results
Reaction times (msec) Patently-True Patently-F
alse Underinformative Logical responders
647 633 655 (n7) Pragmatic responders
1064 856 1203 (n12) Totals 911 774 1014 T
here was a main effect for type of responder and
for statement type and a significant
interaction. NP conclude that they have support
for the RT view, since pragmatic responders are
slower overall and are undertaking deeper
processing, which is effortful (p. 209).
18
Noveck Posada (2003) ERP results
NP did not find increased N400s in pragmatic
responders for underinformative statements, as
allegedly predicted by the neo-Gricean
account. (In fact there was no difference in the
ERP record at the N400 time point between logical
and pragmatic responders, despite their obvious
behavioral differences.) The largest N400s were
for the Patently-False statements, which NP
finally decide has to do with the fact that these
statements paired semantically unrelated items
(e.g., dogs with wings), whereas the
underinformative ones paired related items (e.g.,
elephants with trunks). They conclude that RT is
better supported than the neo-Gricean view, since
they think the flat response at the N400 point
shows the that SI is not generated early.
19
Reactions to Noveck Posada (2003)
(1) If their behavioral data really does show
that SIs are only accessed late, after recovery
of the literal meaning, this would support a
classical Gricean literal-first serial model of
processing, not RT. (2) Their ERP data shows
nothing about the time course of implicature
generation. They cant say one way or the other
whether SIs occur early or late, the flat N400
notwithstanding. (3) Their logical responders
could have gone in for only shallow semantic
processing. (See Barton Sanford, 1993).
Elephants and trunks go together, and that
association is enough to yield an answer
true. (4) On the other hand, the pragmatic
responders could have accessed the SI
interpretation right away (as the neo-Gricean
claims), and the effortful part would then come
at the verification stage Can this
interpretation be allowed to pass through the
filter of my encyclopedic knowledge? No, because
I know that all (normal) elephants have trunks.
20
Bott Noveck (2004)
Materials 1. Some elephants are mammals (?) 2.
Some mammals are elephants (T) 3. Some elephants
are insects (F) 4. All elephants are
mammals (T) 5. All mammals are elephants (F) 6.
All elephants are insects (F) Exp 1 Participants
instructed to interpret some as either some
and possibly all (logically) or some but not
all (pragmatically) and to judge sentences
such as the above either true or false. Exp 3
Similar judgment task but without explicit
instructions as to the interpretation of
some. Exp 4 Same task as in Exps 1 and 3, but
time available for judgment was manipulated.
21
Bott Noveck (2004) Results
Exp 1 Pragmatic responses to type-1 sentences
were less accurate and slower than logical
responses (as well as than responses to
controls). That is, people seemed to have a hard
time sticking to the some but not all
instruction. Taken as support for RT over Default
view. Exp 3 Roughly 60 of responses were
pragmatic. Reaction times were significantly
slower than to logical responses and to control
responses. Again taken to indicate that pragmatic
responding is effortful and to count against the
Default view. Exp 4 When a short time was
available for responses, there were more
logical responses than when there was a longer
time available. This was taken as direct support
for RT and for the prediction that a reduction
in the cognitive resources available reduces the
likelihood that the scalar inference will be
made. (p. 454).
22
Reactions to Bott Noveck (2004)
Again, I think that BN are operating with a
faulty picture of Relevance Theory (RT), which
they assimilate too much to classical Gricean
views. RT does not subscribe to a literal-first
serial model of processing. On the contrary,
RTists think implicatures and explicatures are
derived in parallel, that there is pragmatic
intrusion into semantic interpretation, and that
both logical and pragmatic responses are
enrichments arrived at using the same sorts of
inferential mechanisms and calculations of
relevance. So on the face of it, Bott Novecks
results count against RT as much as against the
Default view. But these sorts of off-line
judgment tasks dont show us much about
underlying processes, so I think the jury is
still out.
23
Bezuidenhout Morris (Unpublished 1)
Materials 1. Some/some Some of the leaves are
green. (T) 2. Some/all Some of the leaves are
green. (?) 3. All/some All of the leaves are
green. (F) 4. All/all All of the leaves are
green. (T) 5. At least some/some At least some
of the leaves are green. (T) 6. At least
some/all At least some of the leaves are
green. (T) 7. Only some/some Only some of the
leaves are green. (T) 8. Only some/all Only some
of the leaves are green (F) Task After reading
a sentence, participants pressed a button and the
sentence was replaced by a visual array of 4
objects. The task was to judge whether or not the
sentence just read correctly described the array.
Displays were either Some or All displays.
(Fillers w/ other DPs variation in some
displays)
24
BM Advantages Predictions
Advantages Our materials seem more natural than
the Bott Noveck (2004) materials. The partitive
construction (some of the ) is more naturally
used with the scalar interpretation than is a
categorial statement of the Some newts are
reptiles variety. We controlled the contexts
that people used rather than relying on peoples
(possibly variable) categorial knowledge (of
langoustines, newts and the like!). Predictions I
f people automatically derive the some but not
all interpretation, then the Some-All condition
should present problems. Also, responses should
be similar to those in the Only Some-All
condition, where the upper-bound interpretation
is made explicit. The SI should be blocked when
the locution at least some is used, and this
should facilitate the judgment of true in the
At Least Some-All condition. If the SI is not
accessed in the Some-All condition, responses
under that condition should resemble those in the
At Least Some-All condition.
25
BM (Unpublished 1) Results
26
BM (Unpublished 1) Results cont.
27
BM (Unpublished 1) Assessment
Just as Noveck found with his participants, our
participants were either consistently logical
or consistently pragmatic in their responses in
the Some-All condition. So we analyzed the data
for these two groups of responders separately. We
assume that logical responders did not derive the
SI and so their profiles in the Some-All and At
Least Some-All conditions should be similar,
which they are (both in judgment times and in
accuracy rates). Similarly, pragmatic responders
are assumed to have derived the SI and so their
profiles in the Some-All and Only Some-All
conditions should be similar, and they are (both
in judgment times and accuracy rates).
28
BM (Unpub. 1) Assessment cont.
However, we did NOT, as Noveck did, find that
pragmatic responders were slower or less accurate
than logical responders in the crucial
conditions. In fact, in the crucial Some-All
condition, pragmatic responders were
faster! (Pragmatic responders were slower in all
three DET-some conditions, but not in the three
DET-All conditions, which are the crucial
comparison conditions). There is one way in which
pragmatic responders were remarkable, and this is
that they seemed to treat the Some-All and At
Least Some-All conditions similarly. They made
significantly more errors in this condition than
the logical responders did (62 as opposed to
13). So, these pragmatic responders apparently
derived the SI in this condition too! This
suggests that they are relatively insensitive to
the idea of a lower-bound, even when it is made
explicit.
29
Bezuidenhout Morris (2004)
This study used eye monitoring during reading to
assess the time course of scalar implicature
retrieval. It aimed to test the predictions of
two rival models about the generation of scalar
interpretations, namely the Default Model
(inspired by Steve Levinsons work) and the
Underspecified Model (inspired especially by
Robyn Carstons work). The focus was on the
interpretation of sentences such as Some of the
trees have blossoms. Horn scale some Q-Heuristic What is not said is not the
case GCIs Not all/many of the trees have
blossoms. According to the Underspecified view,
the but not all meaning is an explicature not
an implicature. Context determines whether
some, whose meaning is underspecified, is
interpreted as some but not all or some and
possibly all.
30
BM (2004) Materials
According to the Default View, GCIs are
automatically generated and cancelled if the
context doesnt support this meaning. Materials
30 items in three conditions, plus 63 fillers
random presentation order. (1) Some books had
color pictures. In fact all of them did, which is
why the teachers liked them. (2) Many books had
color pictures. In fact all of them did, which is
why the teachers liked them. (3) The books had
color pictures. In fact all of them did, which is
why the teachers liked them. Cancellation
clauses in fact, actually, as a matter of
fact, in all truth, etc. Cant tell until the
anaphor them is encountered that there is a
cancellation, because the following is a possible
continuation (1) Some books had color pictures.
In fact all of the pictures were highly colored,
which is why the children liked them.
31
BM (2004) Eye Monitoring
Text display on computer monitor 35inches in
front of participants eyes. Their right eye
movements were monitored during reading. Measures
collected (a) Gaze duration/First pass time
(early processing) (b) Second pass time
(re-reading) (c) Regressions in (allows
inferences about reanalysis and text integration
processes) (See Rayner et al. (1989) and Rayner
Morris (1990) for arguments supporting the use of
this methodology to draw inferences about the
time course of processing.) Text regions of
interest (in bold)
32
BM (2004) Predictions
Default Model Underspecified Model Increased
time on all? No Yes Increased time on them
were/did? Yes No Regressions/rereading of
Some N? Yes No Some should behave like
Many? Yes ?? Some should behave like
The? No No Many should behave like
The? No ??
33
BM (2004) Results
Mean initial reading times (in msec) for 2
critical regions in "cancellation"
sentences Gaze Duration on all First Pass
Time on them were/did Some
N 275 301 Many N 260 308 The N
256 329
34
BM (2004) Follow-up
A different control (a) Some books had color
pictures. In fact all of them did, which is why
the teachers liked them. (b) At least some books
had color pictures. In fact all of them did,
which is why the teachers liked
them. Results Mean reading times (in msec) for
2 critical regions in the "cancellation"
sentence Gaze Duration on all First Pass
Time on them were/did Some N 264
303 At least some N 249 325
35
BM (2004) Assessment
The most robust finding, as evidenced by the
regression data, is that readers did not engage
in re-analysis. This does suggest that there was
no need for cancellation of a some but not all
interpretation. But it is still very unclear what
the mechanisms involved in the interpretation of
sentences containing SI triggers are. Defeasible
reasoning of a weaker sort may very well be going
on. E.g., reasoning in conformity with Gioras
(1997) Graded Salience Hypothesis. Also, van
Rooij, cited by Sauerland (2005), has argued that
the in fact locution does not act to cancel an
implicature but to correct what was said and
shift relevance. So one might question whether
our materials were suitable for studying the
effects of cancellation.
36
Conjunction-Buttressing
Recent psycholinguistic work in this area tends
to focus on scalar implicatures, and even more
narrowly on a few paradigmatic types of scalars,
mainly those involving the following
scales 1, But scalars are only one sub-class of a more
general class that Levinson (2000) calls
Q-implicatures. And in addition to this class, he
posits two additional classes of GCIs
I-Implicatures and M-Implicatures. Examples of
I-Implicatures that have been much discussed are
those generated by the use of and in various
contexts (a) The old king died and a Republic
was declared. and then (b) A Republic was
declared and the old king died. and as a
result (c) I went to London and visited the
Queen. and while (d) Judith insulted him and he
left. and for that reason
37
Bezuidenhout Morris (Unpublished 2)
Levinson (2000) argues that GCIs are default
inferences. As far as conjunction buttressing
(CB) is concerned, he acknowledges that there are
potentially many understandings of and,
including temporal, causal and intentional ones,
but he suggests and then is the default, which
is derived and cancelled if inappropriate. Carston
(1991, 1993, 2002) on the other hand suggests
that and is semantically underspecified and
context determines the relevant
interpretation. BM have a series of
(unpublished) experiments focused on CB. Here
Ill mention just one simple one, that looked at
contrastive vs. temporal understandings of and.
38
BM (unpublished 2) Materials
Participants read sentences of the following sort
and reading times on the last conjunct were
measured Contrastive Mary was obese / and her
baby was underweight. Temporal Mary was ill
during her pregnancy / and her baby was
underweight. Reverse-temporal Mary was
convicted of child abuse / and her baby was
underweight. This was a self-paced reading task.
Conjuncts were displayed a line at a time on a
computer screen and participants pushed a button
when they were ready for the 2nd half to display,
at which time the 1st part disappeared. There
were 60 experimental items and 60 filler items,
presented in random order. Comprehension
questions were associated with some of the filler
items. Participants were asked simply to read for
comprehension and be prepared to answer such
questions when they appeared. Note that they were
not called on to make any specific judgments
about the experimental items.
39
More Examples of Materials
Contrastive The bathroom looked dirty / and the
kitchen looked clean. Temporal The maid scrubbed
hard / and the kitchen looked clean. Reverse-tempo
ral The homebuyers were impressed / and the
kitchen looked clean. Contrastive Coras poetry
was praised / and Joels novel won a
prize. Temporal The judges finally came to a
decision / and Joels novel won a
prize. Reverse-temporal The money came pouring
in / and Joels novel won a prize. Contrastive
Anne took the train to Toronto / and Milo flew to
Tokyo. Temporal All the travel arrangements were
finalized / and Milo flew to Tokyo. Reverse-tempor
al The Japanese hotel staff was welcoming / and
Milo flew to Tokyo.
40
BM (unpublished 2) Predictions
If the and then interpretation is retrieved by
default when readers encounter and, they should
have difficulties both with sentences in the
Contrastive and the Reverse-Temporal conditions.
Construing the order of the conjuncts as
representing the order of events gives an
interpretation that clashes with our world
knowledge about these sorts of events. This would
support the Default view. On the other hand, if
and remains unspecified until it is integrated
with any encyclopedic information that is
triggered by reading the sentence, then responses
in the Contrastive condition should be more
similar to those in the Temporal condition, where
interpreting and as and then can be
consistently integrated with world knowledge.
This would support the Relevance Theory view.
41
BM (unpublished 2) Results
42
Breheny et al. (2006) Experiment 1
Upper-bound context John was taking a university
course/ and working at the same time./ For the
exams/ he had to study/ from short and
comprehensive sources./ Depending on the course,/
he decided to read from/ the class notes or the
summary./ Lower Bound context John heard that/
the textbook for Geophysics/ was very advanced./
Nobody understood it properly./ He heard that/ if
he wanted to pass the course/ he should read/ the
class notes or the summary./ Condition Reading
Time Mean (StdDev) Upper-bound 1291
(352) Lower-Bound 1204 (292) Significant main
effect of condition. RT on the implicature
trigger in UB contexts was significantly longer
than in LB contexts.
43
Breheny et al. (2006) Experiment 2 Manipulating
Information Structure
In a conversation, there is typically a question
or issue being resolved. Call this the contextual
issue. Breheny et al. assume that this issue is
connected to the old information in an utterance.
(E.g., if the issue is Who did the consultants
meet? then The consultants met with . is old
information). They also assume that old
information is typically stated first. For
out-of-the-blue utterances, a reader is likely to
accommodate an appropriate contextual issue by
construing the first-mentioned NP to be related
to the issue. If SIs are context-driven, they
are more likely to be derived when connected to
the (assumed) contextual issue. If generated by
default, sentence position should make no
difference.
44
Experiment 2 Materials Results
Sentence-initial some Some of the consultants/
had a meeting/ with the director./ The rest/ did
not manage/ to attend./ Sentence-initial only
some Only some of the consultants/ had a
meeting/ with the director./ The rest/ did not
manage/ to attend./ Sentence-final some The
director/ had a meeting/ with some of the
consultants./ The rest/ did not manage/ to
attend./ Sentence-final only some The
director/ had a meeting/ with only some of the
consultants./ The rest/ did not manage/ to
attend./ Reading time on the anaphor The
rest (msec) Sentence-initial some 613 Sentenc
e-initial only some 611 Sentence-final
some 628 Sentence-final only
some 586 There was a significant interaction
between Position and Explicitness. Some is
slower than only some in the sentence-final
position but not in initial position.
45
Follow-up Completion task
Completion task Ruling out an explanation for
the effect found in Exp. 2 that relies on the
focusing effect of only. Participants were
given the first sentences used in Exp 2 and asked
to supply their own completions. NP picked
up in continuation NP1 NP2 Sentence-final
some 30 52.5 Sentence-final only
some 24.2 59.2 Interaction between
Explicitness and Continuation was not
significant, but there was a main effect of
Continuation. That is, people preferred to pick
up on NP2 in their continuations, whether they
saw some or only some.
46
Breheny et al. (2006) Experiment 3
Maybe there was just something about the sentence
final position that blocked generation of
SIs. Compare predictions of three models (1)
Revised Default (Default-with-blocking) (2)
Standard Default (2) Context-Driven Present SI
triggers in contexts that either support the SI
or are neutral. As before, Breheny et al. call
these upper-bound and lower-bound contexts
respectively.
47
Experiment 3 Materials
Upper-Bound Context with some Mary asked John/
whether he intended to host/ all his relatives/
in his tiny apartment./ John replied/ that he
intended to host/ some of his relatives/ in his
apartment./ The rest/ would stay/ in a nearby
hotel./ Upper-Bound Context with only
some Mary asked John/ whether he intended to
host/ all his relatives/ in his tiny apartment./
John replied/ that he intended to host/ only some
of his relatives/ in his apartment./ The rest/
would stay/ in a nearby hotel./ Lower-Bound
Context with some Mary was surprised/ to see
John/ cleaning his apartment/ and she asked/ the
reason why./ John told her/ that he intended to
host/ some of his relatives/ in his apartment./
The rest/ would stay/ in a nearby hotel./
48
Predictions Will SIs be generated?
Context-driven Revised Default Standard
Default UB-some Yes No Yes UB-only
some n/a n/a n/a LB-some No No Yes, bu
t cancelled If the SI is generated, this means
longer times on SI trigger and shorter times on
the anaphor. If SI is not generated, the opposite
pattern is expected. Yes Long then short No
Short then long
49
Breheny et al. (2006) Exp. 3 Results

• Reading times (in msec) on SI trigger and
  anaphor
• UB-some UB-only some LB-some
• SI trigger 1027 n/a 927
• Anaphor 604 612 655
• RT on trigger in UB-some significantly different
  from trigger in LB-some.
• (ii) RT on anaphor in LB-some significantly
  different from anaphor in other two conditions,
  which in turn are not significantly different
  from each other.

50
Storto Tanenhaus (2005)
Method Visual-world eye tracking
paradigm Experimental question Is the scalar
implicature not both P and Q accessed early, as
soon as the lexical item or is encountered,
yielding the exclusive rather than inclusive
interpretation of or? Design While
participants were gazing at a visual display,
their eye movements were monitored.
Simultaneously, participants heard auditory
stimuli of the following sort The grapes or the
oranges are next to some locks. Please click on
those locks. What was of interest was the
pattern of looking prior to hearing the
instruction. That is, what are participants
looking at while hearing the sentence describing
the visual display?
51
3 Conditions for Or
The grapes or the oranges are next to some
locks. (a) Please click on those locks. (b)
Please click on some other locks.
(The shaded rows will be called the relevant
rows)
52
Participants strategies
Participants know that they will be asked to
perform some action on objects of the type
mentioned in the VP of the description
sentence. So they need to identify those
objects. As soon as they hear The grapes or the
oranges they can begin to narrow the search to
the relevant rows. Furthermore, if they derive
the SI The grapes or the oranges but not both,
they can rule out as potential targets any cells
adjacent to grapes and oranges that contain
similar objects. So in the OR EARLY condition,
they can narrow the search in the relevant
rows from 4 cells to 2 cells even before hearing
the target locks. They cannot do this in the OR
LATE condition.
53
Fixations to Target vs. AlternativeOR EARLY
54
Fixations to Target vs. AlternativeOR LATE
55
Assessing STs results
Is this support for the claim that SIs are
accessed automatically/ by default? Not clearly
so. Task design rewarded the strategy of deriving
the SI, since it simplified the search for the
target in the OR EARLY condition and didnt
obviously penalize people in the OR INCLUSIVE
condition. On the face of it, the OR INCLUSIVE
condition would penalize people who derived the
SI, since they would have a preference for
looking away from the target(s) in this
condition. Hence one would expect some delay
while they recalibrated once the target was
specified. But in fact ST found no differences
between the OR LATE and OR INCLUSIVE conditions.
56
Chierchia, Frazier Clifton (2006)
This is an attempt to show that in the absence of
any biasing context, peoples tendency to compute
SIs will be influenced by grammatical
structure. In particular, people will be less
likely to compute SIs in downward entailing (DE)
contexts than in non-downward entailing
ones. Chierchia et al. presented people with
disjunctions in either DE or non-DE contexts.
They hypothesized that the exclusive
interpretation of or would be more likely to be
derived in the non-DE contexts.
57
DE-contexts
E.g., the antecedents of conditionals and the
restriction clauses of universal quantifiers are
DE contexts, whereas the consequents of
conditionals and the nuclear scopes of universal
quantifiers are nonDE contexts. If set to subset
entailments are licensed in a linguistic
environment, it is downward entailing
(DE) Every shirt is pink entails Every
torn shirtis pink. BUT Every shirtis
pink doesnt entail Every shirtis a dirty
pink.
58
Materials Chierchia et al. Exp.1

• People saw one or other of the following
  variants
• Simple assertion Jeremy is a child or foreign.
  He must fill out a form.
• Consequent If someone must fill out a form, he
  is a child or foreign.
• Antecedent If Jeremy is a child or foreign, he
  must fill out a form.
• Universal restriction Everyone who is a child or
  foreign must fill out a form.
• People were then asked Is the writer talking
  about someone
• ___ who is either a child or foreign but not
  both?
• ___ who is a child or foreign or possibly both?

59
Results Chierchia et al. Exp 1
As predicted , people chose exclusive
interpretations more often in the non-DE
cases Condition Percentage exclusive-or 1.
Simple assertion 67 2. Consequent 68 3.
Antecedent 59 4. Universal restriction 42 The
difference between conditions 3 and 4 was a
little puzzling, but in a follow-up experiment
(Experiment 2), where the items in these
conditions were close paraphrases of each other,
the difference disappeared If a man reads or
writes, he may vote. (31 exclusive) Every man
who reads or writes may vote. (29 exclusive)
60
Assessing results of Chierchia et al.
First, note that the constraint of grammar seems
to be a relatively soft one. There is not
anything close to unanimity, with at most 68 of
people choosing enriched interpretations in
non-DE contexts (only 53 in Exp 2). Note also
this is a forced choice task. So it is not clear
what people would do if they were simply reading
and comprehending passages of this sort. It would
also be interesting to see how soft a constraint
this is. Does the non-DE context trigger the SI
and the DE context suppress the SI? Or is the SI
triggered automatically and then cancelled in the
DE context? What happens when biasing context
goes against the grammatical constraint?
61
Katsos, Breheny Williams (2006)
Katsos et al. have actually begun to address the
questions asked on the previous slide, in a
series of studies, using both off-line judgment
tasks and on-line reading tasks. (This work is
not yet published, but it has been presented at
several conferences over the past 2 years,
including IPrA, GLOW, CogSci, Lisbon.) I will
discuss one of their on-line studies, meant to
assess the effects of grammatical context vs.
extra-linguistic context. To assess the effect of
grammar, Katsos et al. focus on upward-entailing
(UE) vs. downward-entailing (DE) contexts. To
asses the effect of extra-linguistic context,
they focus on upper-bound (UB) vs. lower-bound
(LB) contexts, using van Rooij Schultzs (2006)
distinction between mention-all and
mention-some questions.
62
Katsos et al. (2006) Study 3 Materials
UE-UB context The director/ asked his
consultant/ Who is representing our company/ at
the court hearing?/ His consultant replied/
Turner or Morris/ from the Legal
Department. UE-LB context The director/ asked
his consultant/ Who is available to represent
our company/ at the court hearing?/ His
consultant replied/ Turner or Morris/ from the
Legal Department. DE-UB context The director/
asked his consultant/ Who is representing our
company/ at the court hearing?/ His consultant
replied/ I believe that if/ Turner or Morris/
from the Legal Department/ do so,/ we need not
worry too much.
63
Katsos et al. (2006) Predictions of Rival Models
Default Model grammar-first, context-later
order of processing. Context-Driven Model
grammar and context operate in tandem from the
start Will the SI be generated? Default
Context-Driven SI trigger in UE-UB context
Yes Yes SI trigger in DE-UB context No No
(!?) SI trigger in UE-LB context Yes, but
cancelled No Note these are the predictions
that Katsos et al. themselves make. In
particular, they seem to be suggesting that
grammatical processes will overwhelm
extra-linguistic ones in DE-UB contexts! Also,
they are saying that in UE-LB contexts, grammar
will be overwhelmed by extra-linguistic context.
So somehow DE contexts are more powerful than UE
ones. (Prohibition stronger than permission?)
64
Katsos et al. (2006) Study 3 Results
Reading Times (msec) on SI trigger (Turner and
Morris) and spill-over region (from the Legal
Department) SI trigger Spill-over UE-UB
condition 785 815 DE-UB condition 760 765 UE
-LB condition 780 775 Warning These RTs are
ones that I have inferred from a very
course-grained line graph from a PowerPoint slide
created by Katsos for a conference presentation
in Lisbon. Katsos et al. conclude that they have
support for the Context-Driven approach. In my
opinion results are at best equivocal (i)
Apparent strength of DE contexts is anomalous for
Context-Driven approach. (ii) But no evidence of
SI cancellation in UB-LB context counts against
the Default approach.
65
Overall Assessment Future Directions
The almost exclusive current focus on SIs means
that the original issue as to what role (if any)
minimal propositions play in utterance
understanding has become submerged. Even those
who take themselves to be supporting Relevance
Theory no longer talk about underspecification
and enrichment. Scalars are only one sort of case
in which relevance theorists have argued for the
primacy of enriched over literal
interpretations. It would be interesting to
experimentally investigate the pragmatic
mechanisms underlying the interpretation
of comparative adjectives like tall and
red quantifier domain restriction possessive
constructions like Johns book adjectives and
adverbs like enough and ready etc.
66
More Future Directions
One of Recanatis persistent themes has been that
literal sentence meaning may not be recovered
during processing, since pragmatic processes of
enrichment operate locally to modulate meanings
before they are composed. An example he gives
is (1) There is a stone lion in the
courtyard. What is input to the compositional
mechanism are not the literal meanings of stone
and lion but their modulated meanings made of
stone and lion representation. There is
experimental work by McElree et al. (2001, 2006)
on Jackendoffs notion of enriched composition
that is potentially relevant to anyone wanting to
test Recanatis claims about modulation and
compositionality. There is also relevant
experimental work by Frisson Pickering (1999)
and Pickering Frisson (2001) on the processing
on metonymies.