You are on page 1of 19

Context, Attention and Depth of Processing During

Interpretation
ANTHONY J. SANFORD
Abstract: The contribution that a word makes to the meaning and interpretation of
a sentence depends upon access to its (lexical) meaning, and to general knowledge
associated with the word. Evidence is presented to support the argument that accessing
lexical meaning, as with general knowledge, is a graded affair. We argue that the contri-
bution a word makes depends upon its relevance to the context, and to focus and related
variables. Extensions of the argument are made to other aspects of language processing.

1. Backgound
In this paper I present a case for variability in the degree to which the pro-
cessing of the meanings of words and utterances is carried out during language
understanding. I use meaning in the broad sense to include both semantic
and pragmatic influences on the establishment of mental representations. The
conventional view of the interpretation of language is that input is parsed
according to some internal grammar, and that the (local) meanings of words
in the input are recovered from memory and integrated, largely under syntactic
control, to produce a (sentence) meaning for that input. Interpretation against
knowledge then follows. The main claim in the present paper is that the extent
to which these basic processes are uniform and complete is dependent upon
two factors: how readily pragmatics overrides the local processing of word
meaning, and structural aspects of language strings that determine how much
attention, or processing effort, to invest. Indeed, various pointers to selective
processing are rife in language. Focussing devices, such as clefting, clearly put
processing emphasis on different aspects of sentences:

(1) Harry dropped the valuable Spode plate in the museum.


(2) It was Harry who dropped the valuable Spode plate in the museum.
[Emphasis on Harry]
(3) It was the valuable Spode plate that Harry dropped in the museum.
[Emphasis on the plate]

Clefting gives prominence to individuals mentioned in the sentence, and this

Address for correspondence: Department of Psychology, University of Glasgow G12,


Scotland, UK.
Email: tony얀psy.gla.ac.uk

Mind & Language, Vol. 17 Nos 1 and 2 February/April 2002, pp. 188–206.
 Blackwell Publishers Ltd. 2002, 108 Cowley Road, Oxford, OX4 1JF, UK and 350 Main Street, Malden, MA 02148, USA.
Context, Attention and Depth of Processing During Interpretation 189

prominence is reflected as an increased accessibility to anaphoric reference


(Gernsbacher and Shroyer, 1989), and ease of the subsequent integration of
material (Morris and Folk, 1998). Similarly, individuals mentioned in the main
clauses of complex sentences are more accessible than those mentioned in sub-
ordinate clauses (Cooreman and Sanford, 1996). Still other possibilities for
selective processing are widely accepted in the literature. For instance, in narra-
tology, it is widely accepted that narratives are understood from the point of
view of main characters (the Focalization Effect; Genette, 1980; Toolan, 1988).
Our claim in the present paper is that even the most fundamental semantic
aspects of sentence processing are influenced by foregrounding, focus, and
focalization.
Surprisingly, theories of focus effects have concentrated almost exclusively
on what guides pronominal reference (e.g., Centering Theory; Walker, Joshi
and Prince, 1998), while psycholinguistic studies of both inference-making and
semantic processing have largely ignored the possibility that these activities
might themselves relate directly to focus. We view accounts of foregrounding,
focus and focalization as providing a theory of what will be attended to in a
language stimulus, and claim that those things that are attended to receive
deeper, more extensive processing (a common enough idea in the psychology
of attention).
The bulk of this paper is concerned with what psychological evidence there
is to support our claims. We begin with some evidence concerning degrees
of semantic processing.

2. The Variable Impact of Words

There is evidence to show not only that the impact of word-meaning on


discourse comprehension forms a continuum (i.e., not all lexical-semantic
information is necessarily used), but that its impact varies in a way which is
consistent with two factors: the fit of a word to a situation, and linguistic
variables which modulate attention. In this section, we shall examine the idea
that the contribution of a word to the meaning of a sentence (or utterance)
is dependent on the fit of a word to the scenario the sentence is identifying.
It is usually the case that a clearly anomalous word will be detected auto-
matically during reading. For instance, using sentences like She drank her coffee
with dog, participants recognise the anomalous word dog without any difficulty,
and immediately upon reading it (Kutas and Hillyard, 1980). The usual view
of this is that it is because selection restrictions on what one has with coffee
rule out dog, and this is picked up automatically during the mapping of the
meaning of dog into the role of something that accompanies coffee. However,
a steady trickle of evidence has shown that such processes do not always occur,
or at least are not always successful at detecting anomalies. The best known
is the so-called Moses Illusion, which is a very powerful illustration of the fact
 Blackwell Publishers Ltd. 2002
190 A.J. Sanford

that lexical semantic processing is incomplete, and should be viewed as a matter


of degree. Erickson and Matteson (1981), asked people the following question:

(4) How many of each sort of animal did Moses put on the ark?

The most common answer was ‘two’. The correct answer is ‘none’: Moses
was not the person who built and used the ark. Many people simply miss this
fact, even if they know it. Subsequent work has shown that changing the
anomalous word to Adam results in much better detection (Van Oostendorp
and De Mul, 1990), so it’s not so much that people don’t read the critical
word, as that they don’t process it as fully as might be expected.
Attempts to determine why this effect should occur have included a test
of the idea that it is the similarity of meaning between Moses and Noah that
is the essential ingredient. Testing the rated semantic similarity of Moses and
Noah, and Adam and Noah, Van Oostendorp and De Mul found that Adam and
Noah came out as most different. (In this test, they simply asked respondents to
rate the similarities in meaning: it seems most likely that the similarity scores
stem from a function of the number of retrieved features in common and the
number of retrieved features not in common). Van Oostendorp and De Mul
went on to explain the original effect in terms of sampling the features associa-
ted with the critical word as it is read: a small sample of the features associated
with Moses would be much the same as those associated with Noah, so no
distinction is possible during comprehension. With Adam, the difference would
be greater, and so the anomaly would be noticed. This makes sense if rather
than pull out all of the features associated with a word, only some are used
during reading. It seems likely that it is the high availability of the information
that Adam was the first man that explains these findings.
The general view that processing can be very shallow is bolstered by many
other examples of anomalies that people simply fail to detect. Thus Sanford
and Young (unpublished) asked 20 participants to answer a series of questions
about marriage laws, and obtained results suggesting that accessing the encoded
semantics of a word can be overridden by pragmatic considerations. The ques-
tions used included:

(5) Can a woman marry her daughter’s boyfriend?


(6) Can a man marry his sister’s mother?
(7) Can a man marry his widow’s sister?

Of the 20 people tested, 16 answered positively to question (7). Post-test inter-


views revealed that the people who had answered the last question positively
believed that the woman of the couple had died, and that of course the man
could marry her sister. What they didn’t realise was that to have a widow, the
man himself must be dead. Just what is happening here with our non-
detectors?
 Blackwell Publishers Ltd. 2002
Context, Attention and Depth of Processing During Interpretation 191

We suggest that the question is interpreted as being about the possibility


of marriage within the context of socially acceptable norms for the living,
avoiding incestuous relations, and so on. So, a question about whether some-
body can marry presupposes that they are alive, and this presupposition over-
rides the local semantic interpretation of the noun phrase his widow. The global
picture, combined with the presupposition, constrains his widow to be the
only one who had died. Of course, all of this assumes that his widow is not
deeply semantically analysed, similar to the way that knowledge associated with
Moses was not deeply analysed in the original Moses Illusion. In brief, local
semantic processing seems to have been overshadowed by pragmatics, and per-
haps even replaced by it.
There are earlier and even more powerful examples of the same type of
phenomenon, the most notable of which are Peter Wason’s ‘Depth Charge’
sentences (Wason and Reich, 1979). The best-known example is taken from
a notice in a hospital casualty department:

(8) No head-injury is too trivial to be ignored.

This is typically read as meaning:

(9) However trivial a head injury might appear, it should not be ignored.

However, to see that it does not mean that, consider:

(10) No missile is too small to be banned.

Here the meaning is clear: however small, it should banned. By parallel structure,
(8) actually means however trivial, it should be ignored, which is plainly not what
was intended. We only understand the false but intended message in (8)
because we relate the words to the situation being depicted, where the actions
demanded are clear, and so interpret the message as a reminder of what should
be done. Again, as Wason and Reich show, pragmatics, in the form of situ-
ation-specific knowledge, overrides full local semantic interpretation.
The author recently encountered an example of the same problem in a
magazine article in which the efforts of an individual had had an impact on
a hobby. It was said of the individual that ‘. . .it is easy to overestimate the
importance of his work on the subsequent events in the hobby. . .’, hardly a
match between the intended tribute and the actual meaning of the utterance.

3. Situation-Specificity and the Detection of Anomalies


Barton and Sanford (1993) investigated the impact that situation-descriptions
had on anomaly detection by using a case study of a single anomaly. Like the
 Blackwell Publishers Ltd. 2002
192 A.J. Sanford

‘widow’ example, it is the actual meaning of a word that was at issue. Parti-
cipants were asked to read the following passage, and answer the question:

There was a tourist flight from Vienna to Barcelona. On the last leg of
the journey, it developed engine trouble. Over the Pyrenees, the pilot
started to lose control. The plane eventually crashed right on the border.
Wreckage was equally strewn in France and Spain. The authorities were
trying to decide where to bury the survivors.
What is your solution to the problem?

In the basic version of this problem, as presented above, only 60% of parti-
cipants recognised that you simply don’t bury survivors (we carefully screened
out people who knew the problem already). Participants who missed the ano-
maly didn’t believe that the survivors had later died, or anything of that nature.
They were frankly stunned that they had missed the anomaly.
We suspected that how well the word survivor goes with the situation (an
air-crash), and the core meaning of the word, jointly determine the chances of
it being detected. By fit to a situation, we mean how relevant the word is to
the situation; and by core meaning, we mean the accessible parts of its meaning
that would make up a dictionary entry, or the definition a person might give
for it.
Because the fit to the situation is good (survival is highly relevant to any
air-crash scenario), we surmised that the processor moves on quickly on
encountering it, and so misses its core meaning on some occasions (survive
means to be alive after a life-threatening event), leading to a failure to detect the
anomaly. We therefore compared other words that fit the situation well:
injured, wounded, and maimed. Now these words do not have ‘being alive’ as
part of their core meaning. For instance, injured was defined by a group of
participants as ‘having damage done to the body by something’, with no men-
tion of life-and-death. In the case of the injured set of words, however, we
established that the assumption of being alive is part of the presupposed state
that would go with the choice of the word (i.e., in our terminology, it is part
of what is presupposed).
Thus, if the degree of processing for lexical meaning afforded a word is
low when the word fits the situation well (producing a 40% miss rate for
survivor), we would expect the miss rate to be even higher for the injured set,
because the ‘alive’ feature is not part of the core meaning. The percentage
detections conformed to this expectation (from Barton and Sanford, 1993,
Experiment 1):

Survivors: 60%
Injured: 5%
Wounded: 25%
Maimed: 25%
 Blackwell Publishers Ltd. 2002
Context, Attention and Depth of Processing During Interpretation 193

Our argument was, then, that words in discourse are first checked in some
way for their fit (or relevance) to the situation depicted (the base domain of
reference): if the fit is high, then (other things being equal) the processor does
a less detailed analysis of the meaning of the word than if it is low. If processing
proceeds beyond recognising that the word is relevant to the situation, then
information that is next most available is likely to enter processing, and we
take this to be the core meaning. So, survivors is more likely to be detected as
anomalous than the injured group of words. Finally, the background presuppo-
sition, that to predicate ‘injured’ or ‘maimed’ of a person means they are alive
at that time, is the least accessible information.
The other major prediction was that if a word does not fit the situation
well, then greater use would be made of lexical-semantic information (core
meaning). Barton and Sanford (1993, Experiment 3) tested this by comparing
materials in which the scenario depicted was an air crash with one in which it
was a bicycle crash. A pre-test showed participants to believe that the probability
of someone dying was very much lower in the case of the bicycle accident.
The materials used in the experiment were:

(11) When (an aircraft crashes / a bicycle accident occurs), where should
the survivors be buried? [The / denotes alternatives].

With this version of the problem, detection rate averaged 33% for the air crash,
but went up to 80% for the bicycle accident.
So, the picture we have is one in which, when a word is directly relevant
to a situation, less use is made of lexical-semantic information than when it is
not so relevant, and the information used does not always include the core
meaning of the word in question.
The evidence of all of these examples suggests that when one reads the
examples, there is access to the basic situation in which the examples are
grounded. This requires two things: a situation-centred set of representations
in long term memory, and speedy access to it, not dependent upon fully
establishing sentence meaning before hand. The idea of situation-centred
knowledge representations has a long history, as schemata (Rumelhart and
Ortony, 1977), scripts (Schank and Abelson, 1977), frames (Minsky, 1975),
and in discourse processing, as scenarios (Garrod and Sanford, 1988a; Sanford
and Garrod, 1981, 1998; Sanford and Moxey, 1995a, 1995b). Sanford and
Garrod (1981, 1998) have argued extensively that the primary basis of under-
standing requires a mapping of a sentence onto a recognised situation, and that
this does not require going first through a complete local semantic interpret-
ation. Of course, the idea of early pragmatic intervention in the process of
local semantic analysis is completely at odds with the view that local analysis
precedes subsequent interpretation. It is even somewhat at odds with the idea
that a literal meaning is assigned to utterances. For an argument for the invo-
cation of schema-driven interpretations for metaphorical language, rather than
 Blackwell Publishers Ltd. 2002
194 A.J. Sanford

an invocation of literal meaning, see Recanati (1995), and Sperber and


Wilson (1998).

4. Linguistic Structure and Degree of Analysis


Having illustrated how the recruitment of lexical meaning is not all-or-none,
we now consider what might modulate the extent to which semantic infor-
mation is used, and also the extent to which other sources of information
are utilised. A good starting point turns out to be cues regarding focus and
foregrounding. In the present paper, I am not concerned with particular
theories of focus and foregrounding, though the details are obviously of impor-
tance. The interest here is that very general notions and illustrations of focus
and foreground constitute the basis for an account of how attentional resources
might be allocated during language comprehension. For instance, the experi-
mental literatures on focus and foregrounding have been very much concerned
with the ease with which pronominal anaphora takes place. Entities that are
focussed are generally considered to lead to easy pronominal reference, and
ease of pronominal reference has become treated as an index of focus. We
might, therefore, conjecture that cues to focus determine resource allocation
and so are good candidates for controllers of the extent to which various sorts
of processing take place. In fact, focus and foreground seem to also influence
the extent to which lexical-semantic and encyclopaedic information is used
when constructing an interpretation for a sentence. One of the earliest illus-
trations of this was with two versions of the original Moses Illusion (Bredart
and Modolo, 1988), in which participants had to evaluate the statements as
true or false:

(12) Moses put two of each kind of animal on the ark.


(13) It was Moses who put two of each kind of animal on the ark.

Detection rate was very much higher for the clefted version which focalises
Moses. So, putting a word in focus through clefting brings about a fuller analysis
of that word.
A second structural effect has been found by Sanford and Young
(unpublished data). They compared the following versions of the ‘Widow’
task:

(14) If a man wants to re-marry, can he marry his widow’s sister? (30%
detection)
(15) If a man’s wife becomes a widow, can the man marry her sister?
(65% detection)

In example (14), the first part of the sentence is about the man wanting to
remarry, whereas in (15), it is about the man’s wife becoming a widow. We pre-
 Blackwell Publishers Ltd. 2002
Context, Attention and Depth of Processing During Interpretation 195

dicted that the analysis of the state of the wife would be greater when the first
clause was about the wife. In (14), the woman being a widow is backgrounded
information in the second clause, and so should receive relatively shallow
analysis. The detection rates for participants were 30% for (14), rising to 65%
for (15), in line with this prediction.
An additional area where differential processing has been demonstrated as
taking place is in sentences with subordinate clauses. Thus Baker and Wagner
(1987) presented participants in two experiments with complex sentences that
contained false information in either the main or subordinate clause, for
example:

(16) The liver, which is an organ found only in humans, is often damaged
by heavy drinking. [Subordinate clause position]
(17) The liver, which is often damaged by heavy drinking, is an organ
found only in humans. [Main clause position]

Participants were required to evaluate a set of statements for being ‘true’ or


‘false’, with some of the statements not including an error. There was a small
but reliable effect of main versus subordinate clause. The probability of cor-
rectly detecting an error was 69% for the subordinate clause, and 80% for the
main clause. There are many possible explanations for this effect, although it
is consistent with the idea that processing is less detailed for subordinate clauses.
More recently, in our laboratory we have carried out two investigations that
explore the possibility more closely.
Data recently collected by Bohan and Sanford (in preparation) investigated
the detection of simple anomalies under two conditions of memory load. The
example below shows the high-load and low-load positions used in the two
conditions:

The future of the NHS has been a major electoral issue. There is increas-
ing concern from nursing unions that their members are under-paid.
UNISON has threatened strike action if the government does not
improve the present situation. However, critics argue that strike action
could dangerously affect the people in their care.
Would you support a national strike, possibly quite lengthy and disruptive,
that demanded a reasonable pay settlement for all patients in NHS hospi-
tals?
[High-load position]
Would you support a national strike that is demanding a reasonable pay
settlement for all patients, even though it may be quite lengthy and disrup-
tive to NHS hospitals?
[Low-load position]
 Blackwell Publishers Ltd. 2002
196 A.J. Sanford

In the high-load condition, supplementary information qualifying strike


(possibly quite lengthy and disruptive) appears between the start of the question
and the individuals about whom the question is being asked (the patients, an
anomalous reference). This serves to highlight the strike. In the low load con-
dition, the supplementary information comes after the question is completed,
and is presented as a modifying clause. Participants saw a number of passages
of these types, and were told that there would be errors in some of them such
that the passages did not make sense. The results showed that 97% of the errors
were detected in the low-load condition, while only 52% of the errors were
detected in the high-load condition.
It is difficult to find experimentally useful anomalies on a large scale, and
it might be argued that anomalies are in some way unusual, and if used over
and over again, may result in special strategies being adopted. To overcome
some of these problems, Sanford, Sturt, Stewart and Archambault (submitted)
used a somewhat different task, borrowed from visual perception. The task is
known as the ‘Change Blindness’ technique. With this technique, a display is
shown to a participant, then, after a short blank, the display is shown again
with some modification (e.g., Simons, 2000; Simons and Levin, 1997). The
question is whether the participant notices the change. Changes are noted
more readily in things that are attended to than things that are not. The change
blindness procedure provides a general method for establishing what kind of
processing has not taken place, just like anomaly detection. By changing a
word or a phrase in a text, and seeing whether readers detect the change, it
should be possible to map out the point at which different linguistic inputs
give rise to indistinguishable mental representations.
Sanford et al. predicted that changing a verb in a main clause or a subordi-
nate clause should have two constraints: first, changes should be more likely
to be noticed in main clauses than in subordinate clauses. Secondly, changes
to single words should be most likely to be noticed if the change in meaning
was great rather than small. A contrasting pair of materials typical of those used
is given below. In the first case, the change is made to ‘finished’ in the main
clause position, while in the second case, the change was in the subordinate
clause:

(18) The newsagent had just hired a new paperboy to cover the down-
town area.
The paperboy finished his rounds after he ate his breakfast.
There were a lot of deliveries to be made.
(19) The newsagent had just hired a new paperboy to cover the down-
town area.
After the paperboy finished his rounds, he ate his breakfast.
There were a lot of deliveries to be made.
 Blackwell Publishers Ltd. 2002
Context, Attention and Depth of Processing During Interpretation 197

Table 1 Percentage changes detected on the first change.

Semantic distance Clause Type

MAIN SUBORDINATE

Small 26% 17%


Large 48% 40%

In each case, the crucial change was from finished -⬎ completed for the close
semantic change, and finished -⬎ started for the distant change.
Participants saw each passage in its original state, followed by a gap, fol-
lowed by the passage again but with a change in the verb. After a gap, the
original passage was presented again, and so on, for a total of 6 cycles. Some
other (filler) materials did not have a change at all. Participants pressed a button
as soon as they noticed a change, and said what the change was. There are
several ways of analysing these data, all with the same outcome. In Table 1,
we show the percentage of changes correctly detected on the first change. It
is clear that a small semantic change is much harder to detect than a large
one, and that either change is easier to detect in a main clause than in a
subordinate clause.
In sum, those parts of a text that are in focus (through clefting), or in a
main clause position, are more deeply processed than those in unfocussed or
subordinate positions. These patterns directly match prominence patterns for
ease of pronominal reference (e.g., Gernsbacher and Shroyer, 1989, for cleft-
ing; Cooreman and Sanford, 1997, for complex sentences).

5. Communicative Relevance and Degree of Analysis


Quite apart from structural aspects of texts and sentences, task relevance also
influences degree of semantic processing. Using the Survivors task, Barton and
Sanford (1993, Experiment 4) investigated the effects of providing more or
less information that is relevant to answering a question. The ‘basic’ version
used was:

(20) Suppose there was an aircrash with many survivors.


Where should they be buried?

In a second version, the ‘augmented relevant’ version, additional information


was provided, which pretests had shown to be relevant to giving an answer:

(21) Suppose there was an aircrash with many survivors, most of whom were
Europeans/of no fixed abode.
Where should they be buried?
 Blackwell Publishers Ltd. 2002
198 A.J. Sanford

Table 2 Percentage anomaly detections for the augmented survivors task. Adapted from Barton and
Sanford (1993), with permission of the authors, and thanks to the Psychonomic Society.

Baseline Irrelevant Augmented Relevant Augmented

% detections 76% 80% 50%

The additional information, italicised here for convenience, can be (and


was) used in formulating answers.
The third version was the ‘augmented irrelevant’ condition, in which the
additional information was, for instance, about the time of the crash, and was
judged as conveying nothing which was relevant to where they should be
buried:

(22) Suppose there was an aircrash with many survivors, which happened
last week. Where should the survivors be buried?

A comparison of detection performance on items of these three types yielded


the pattern of results shown in Table 2. The ‘augmented relevant’ condition
produced a depressed detection rate compared to the ‘basic’ and ‘augmented
irrelevant’, which did not differ from each other. Our explanation is as follows:
on reading the sentences, processing of the word survivors is shallow. In the
‘augmented relevant’ case, no further analysis takes place, because information
relevant to answering the question (were it posed in good faith) is provided
by the text, and is simply used to answer the question. If that information is
not available, further analysis takes place on the material in order to answer
the question, elevating the detection rate for the anomaly. Our conclusion,
then, is that depth of semantic processing of a word is dependent upon the
degree of communicative relevance of other linguistically given material.

6. Other Evidence of Variable Depth of Processing

6.1 Garden Path Recovery


There is other evidence of variable depth of processing. Christianson, Holling-
worth, Halliwell and Ferreira (in press) studied garden-path sentences (Frazier,
1979), such as:

(23) While Anna dressed the baby played in the crib.

Readers spent a great deal of time processing the word played, and often re-
read the material, because of an initial misinterpretation of the sentence. The
initial reading closes as though Anna dressed the baby is a clause; this possibility
 Blackwell Publishers Ltd. 2002
Context, Attention and Depth of Processing During Interpretation 199

is then ruled out on encountering the word played. Because syntactic reanalysis
seems to take place, it has normally been assumed that the baby comes to be
taken as the subject of the verb played. However, when Christianson et al.
asked participants to answer questions like

(24) Did Anna play in the crib?


(25) Did Anna dress the baby?

they found that all participants correctly checked the baby played in the crib, but
that they were much less accurate at answering (25). For non-garden path
control sentences, participants could answer the questions correctly. This shows
that even though some syntactic re-assignment occurred, the meaning was
reanalysed at a shallower level, if at all. Like us, Ferreira and her colleagues
(Ferreira, Bailey and Ferraro, in press) are convinced that representations are
incomplete, and depend upon incomplete processing.

6.2 Syntactic Processing


The garden-path effects do raise the question of whether syntactic processing
is a matter of degree also. This is not an issue that is at the centre of the
present paper, but it is clearly of great interest, and, in terms of the present
concerns, there is some evidence to suggest that when parsing becomes diffi-
cult, processing may be more shallow, or incomplete. Early evidence from
Schlesinger (1968) suggested that centre-embedded sentences, notoriously dif-
ficult to comprehend, tend to be interpreted by using pragmatic (world)
knowledge. For instance, consider the sentence:

(26) This is the hole [that the rat (which our cat {whom the dog bit}
made) caught].
[Brackets added for reader convenience].

Interpretations were reported based on stereotyped behaviour (rats dig holes,


cats catch rats, dogs bite cats), although the added parse brackets show that
the actual sentence is anomalous. So, at least in extremis, syntactic processing
can be overridden by pragmatic information.

6.3 Narrative Focalization


In psycholinguistics, a distinction is drawn between necessary inferences (those
which must be made to maintain coherence), and elaborative inferences (those
which simply ‘elaborate’ on a theme) (e.g., McKoon and Ratcliff, 1992). An
example of a necessary inference is the bridge between the following sentences:

(27) Harry threw the delicate vase at the wall. It cost over a hundred
dollars to replace.
 Blackwell Publishers Ltd. 2002
200 A.J. Sanford

The inference The vase broke is necessary to make the two sentences cohere.
On the other hand, if one heard just the first sentence, to infer (however
likely) that the vase broke would be an elaborative inference. The point is
that elaborative inferences can go off in all directions. For instance, another
elaborative inference would be Harry is angry. It is an empirical question which
inferences are made, and when, and under what circumstances. Furthermore,
it seems likely that systematic variability in degrees of processing could be
established for these, possibly following the patterns discussed earlier for lexical-
semantic processing. In general, however, the question has been whether
certain plausible inferences are or are not made, and not whether there is a
different likelihood of an inference being made, or a different time course of
inference-making, as a function of attention variables. However, some experi-
ments by Sanford, Clegg and Majid (1998) examined a closely related issue,
and demonstrated a case of selective elaborative processing.
Their work was based on the distinction between main and secondary
characters in narrative discourse. It is widely believed in narratology that tracts
of narrative are typically understood from the perspective of particular individ-
uals, so-called focalizers (Duchan, Bruder and Hewitt, 1995; Genette, 1980;
Toolan, 1988). With simple multiperson texts, the focalizer will typically be
the main character. Sanford, Clegg and Majid (1998; see also Garrod and San-
ford, 1988b) examined this idea by studying how sentences like The air was
hot and sticky are processed during reading. Consider the following:

(28) Simon went to the cinema. The usherette showed him to a seat.
The air was hot
and sticky.

Suppose one now asks, Did Simon find the atmosphere hot and sticky? The answer,
intuitively, is yes. What about Did the usherette find the atmosphere hot and sticky?
Intuitively, the answer is less clear: its as though the question is not quite
legitimate. This intuition gave Sanford et al. the idea that the air being hot
and sticky would be interpreted as relevant to Simon, but not to The usherette.
Basically, the idea is that the use of a proper name causes Simon to be construed
as the main character (see Sanford, Moar and Garrod, 1988), and that psycho-
logical atmosphere statements are interpreted with respect to the main charac-
ter. To test this, Sanford et al. used materials like that shown in (29), where
two characters are introduced, one as main (named) character, and one as
secondary (role description) character. There is then a choice of two back-
ground atmosphere statements, followed by a choice of target sentences in
which one or other of the characters acts in such a way as to be consistent or
inconsistent with the atmosphere statement. For instance, fainting is consistent
with airlessness, but not with airy and refreshing. We expected that the difference
in speed of integrating the target sentence with the rest of the text would be
greater if the atmosphere statement supported the action, rather than did not
 Blackwell Publishers Ltd. 2002
Context, Attention and Depth of Processing During Interpretation 201

Table 3 Mean reading times for target sentences referring to main and secondary characters with
consistent and inconsistent atmosphere sentence information. Reading times in seconds. Adapted from
Sanford, Clegg and Majid (1998), with permission of the authors and thanks to the Psychonomic Society.

Fit of atmosphere statement Main Secondary


to target behaviour character character

CONSISTENT 1.695 1.810


INCONSISTENT 2.053 1.990

support it. However, we also predicted that this difference would be greatest
when the action was carried out by the main (named) character, because the
background information would be considered most relevant to that character
by the processor. This prediction was borne out.

(29) The Bank


The teller/Alistair called the next customer to his window.
{CHARACTER 1}
Margaret/A woman stood at the head of the queue.
{CHARACTER 2}
The atmosphere was airy and refreshing {INCONSISTENT
BACKGROUND}/
The atmosphere was airless and oppressive {CONSISTENT
BACKGROUND}
He/She fainted suddenly against a marble pillar. {TARGET}

As shown in Table 3, for the consistent backgrounds, main character targets


were read more quickly than secondary character targets, while for inconsistent
backgrounds, main character targets were read more slowly than secondary
character targets. Thus consistent information facilitated integration and incon-
sistent information adversely affected integration for main characters relative
to secondary ones. This experiment supports the view that the high textual
relevance of main characters causes background information that is not
explicitly assigned to any character to be interpreted as relevant to the behav-
iour of main characters. As with various other examples of depth of processing,
it has been shown that there is much faster pronominal reference to main
characters, defined for short texts as the character introduced by a proper name,
(Anderson, Garrod and Sanford, 1983; Sanford, Moar and Garrod, 1988), again
supporting the idea of a common basis for focus and depth of processing effects.

7. Discussion: Depth of Processing


The findings with respect to semantic processing suggest that a primary deter-
minant of the degree to which the semantics of a word is recruited during
 Blackwell Publishers Ltd. 2002
202 A.J. Sanford

comprehension is the fit of the word to the basic situation under discussion.
If the word has a low relevance to the situation under discussion, then it will
receive greater processing: if it has a high relevance to the situation, then it
will receive shallower processing. This idea assumes that it is possible to check
rapidly for the fit of a word to a situation. Such a mechanism could be realised
if part of semantic memory was structured in terms of words (lemmas) that
tend to be used together. It is not unreasonable to posit such a structure, since
this would be tantamount to preserving in memory some of the regularities
which attend thinking about and communicating about situations. Indeed,
there have been several attempts to capture at least some aspects of meaning
in these terms, in particular Latent Semantic Analysis (Landauer, Foltz and
Laham, 1998). This is a technique that extracts and represents the contextual
meaning of words by statistical computations applied to a large corpus of text.
Although it cannot capture all aspects of meaning (Kaschak and Glenberg,
2000), the idea is that ‘similarity of meaning’ can be indexed by aggregating
all of the word contexts in which given words do and do not appear. Data
about individual words obtained in this way has been shown to mimic a variety
of human judgements, including prediction of query-document similarity
judgements, matching reviewers of papers with papers reviewed based on
samples of reviewers’ own papers, and a variety of other tests which related
directly to the fit of words to contexts. If human semantic memory consisted
in part of a similar set of stored statistical constraints, then it would represent
a clear basis for fast word-situation matching.
Clearly more work is needed to establish precisely how situation-specific
control of degree of processing works. One possibility is that meanings are
‘fully’ recovered (at least to the level of core meaning), but that other con-
straints somehow over-ride or inhibit the fully-recovered word meaning. At
a theoretical level, this seems to us unlikely, since for some words (names of
people, for instance), it is difficult to see just what would constitute an encapsu-
lated meaning. An alternative point of view is that information about word
meaning is not recovered as a unit, but rather unfolds from memory over
time, and may draw upon broad, encyclopaedic information. This contention
requires examining in more detail.
The second major claim is that the extent to which lexical-semantic infor-
mation is used depends upon focus (either in terms of clefting, or in terms of
the main/subordinate clause distinction). This is unsurprising if we assume that
words in privileged positions (in focus, in main clauses, etc) are always more
important, and so the comprehension system is set to analyse them more
deeply. In the main, ideas of focus have tended to be interpreted within psych-
olinguistics as being concerned with ease of anaphoric reference resolution,
and the kind of anaphors that may be most effective. This is exemplified by
Centering Theory (Walker, Joshi, and Prince, 1998), for instance. Our con-
siderations of degrees of semantic processing suggest that a proper account of
focus and related phenomena should take into account other aspects of pro-
 Blackwell Publishers Ltd. 2002
Context, Attention and Depth of Processing During Interpretation 203

cessing. Does variation in the ease of pronominal processing correlate directly


with those entities about which deeper semantic processing might take place?
There is substantial evidence that in continuation tasks, references are more
likely to be made to characters in main clauses than in subordinate clauses,
regardless of order of mention, and also that reading times for sentences that
refer to characters in main clauses are handled more rapidly during reading
(Cooreman and Sanford, 1997). The common link between ease of pronomi-
nal reference resolution and degree of semantic processing might well be the
amount of processing resource allocated to parts of an utterance.
The work showing focalizer effects on the processing of background infor-
mation also reveals that main characters give rise to deeper processing in that
background information is more obviously integrated with the behaviour of
main characters than secondary ones. Again, the integration effects appear to
parallel ease of pronominal reference effects: sentences making anaphoric refer-
ence back to main characters are processed more rapidly than the same ones
making reference back to secondary characters (Anderson, Garrod and Sanford,
1993; Sanford, Moar and Garrod, 1988).
Flexibility in the utilisation of processing resources is of paramount impor-
tance in resource-limited systems like the human mind. It is taken for granted
that in relation to the visual world, we only fully perceive those things to
which we attend (e.g., Simons, 2000). The present claim is that the same
applies to language, and that comprehension is based on elaborating certain
perspectives and aspects of language input, while ignoring other, less crucial,
material. This can be seen to be inevitable if one considers the magnitude of
the problem of coherence-checking a mental representation of what has been
said, both in terms of internal consistency, and in terms of how it relates to
the world. The amount of inferential work would be enormous. Rather, the
problem is to determine what to analyse in detail, and what to treat in a shallow
fashion, to preserve economy of effort. The fortunate thing about language as
input is that we know which devices serve as cues: focussing and foreground
effects. Speakers and writers choose to use devices of focus (e.g., Halliday and
Hasan, 1976) and foregrounding (e.g., Chafe, 1974; Tomlin, 1985) because
of the attention controlling properties they have.
Some, if not all, of these observations can be related to developments within
the relevance-theoretic framework, which explicitly seeks to develop accounts
of how inferential and other processing activity is managed during comprehen-
sion. For instance, Sperber (2000) has argued that encyclopaedic knowledge
associated with content words, may be retrieved over time (rather than all
together), and that retrieval stops when enough information relevant to the
comprehender’s purpose has been recruited. This kind of continuous retrieval
with optional stopping is precisely the kind of thing I am suggesting in the
present paper. In much the same way, from a message perspective, material
cued as foregrounded, and as focussed, is likely to be just that which the writer
intends to be most important, and will hence have the most processing effort
 Blackwell Publishers Ltd. 2002
204 A.J. Sanford

devoted to it. These connections remain to be explored, but offer an interest-


ing prospect for development.
In summary, we have presented a case for the following:

쐌 The recruitment of word meaning in determining the interpretation


of utterances is not all or none, but is a matter of degree, and is often
shallow. This appears consistent with arguments made by Sperber
(2000).
쐌 Situation-specific knowledge can intervene in local semantic analysis
and force interpretations that eclipse local semantic analysis. This con-
trasts with the conventional view that local semantic analysis necessar-
ily precedes global interpretation.
쐌 The degree of semantic analysis afforded words is greater when they
are syntactically prominent.
쐌 Focalization in narratives determines the processing of background
information in narratives.

Department of Psychology
University of Glasgow

References

Anderson, A., Garrod, S.C. and Sanford, A.J. 1983: The accessibility of pronominal
antecedents as a function of episode shifts in narrative texts. Quarterly Journal of
Experimental Psychology, 35A, 427–440.
Baker, L. and Wagner, J.L. 1987: Evaluating information for truthfulness: The effects
of logical subordination. Memory and Cognition, 15, 247–255.
Barton, S.B. and Sanford, A.J. 1993: A case-study of anomaly detection: Shallow sem-
antic processing and cohesion establishment. Memory and Cognition, 21, 477–487.
Bredart, S. and Modolo, K. 1988: Moses strikes again: Focalization effects on a seman-
tic illusion. Acta Psychologica, 67, 135–144.
Bohan, J. and Sanford, A.J. (in preparation) Main and subordinate clauses and the
control of degrees of semantic analysis.
Chafe, W.L. 1974: Language and consciousness. Language, 50, 111–133.
Christianson, K., Hollingworth, A., Halliwell, J. and Ferreira, F. (in press). Thematic
roles assigned along the garden path linger. Cognitive Psychology.
Cooreman, A. and Sanford, A.J. 1996: Focus and Syntactic Subordination in Discourse.
Research Paper RP-79, Human Communication Research Center, University
of Edinburgh.
Duchan, J.F., Bruder, G.A. and Hewitt, L.E. (eds.) 1995: Deixis in Narrative: A Cogni-
tive Science Perspective. Hillsdale, NJ: Erlbaum.
Erickson, T.A. and Matteson, M.E. 1981: From words to meaning: A semantic illusion.
Journal of Verbal Learning and Verbal Behavior, 20, 540–552.
 Blackwell Publishers Ltd. 2002
Context, Attention and Depth of Processing During Interpretation 205

Ferreira, F., Bailey, K.G.D. and Ferraro, V. (in press). Good enough representations
in language. Journal of Psycholinguistic Research.
Frazier, L. 1979: On Comprehending Sentences: Syntactic Parsing Strategies. Ph.D. Disser-
tation, University of Connecticut. West Bend, IN: Indiana University Linguis-
tics Club.
Garrod, S.C. and Sanford, A.J. 1988a: Thematic subjecthood and cognitive constraints
on discourse structure. Journal of Pragmatics, 12, 357–372.
Garrod, S.C. and Sanford, A.J. 1988b: Discourse models as interfaces between language
and the spatial world. Journal of Semantics, 6, 147–160.
Genette, G. 1980: Narrative Discourse: An Essay on Method. (J.E. Lewin, trans.). Ithaca,
NY: Cornell University Press.
Gernsbacher, M.A. and Shroyer, S. 1989: The cataphoric use of the indefinite this in
spoken narratives. Memory and Cognition, 17, 536–540.
Glenberg, A.M. and Robertson, D.A. 1999: Indexical understanding of instructions.
Discourse Processes, 28, 1–26.
Halliday, M.A.K. and Hasan, R. 1976: Cohesion in English. London: Longman.
Kaschak, M.P. and Glenberg, A.M. 2000: Constructing meaning: The role of afford-
ances and grammatical constructions in sentence comprehension. Journal of Memory
and Language, 43, 508–529.
Kutas, M. and Hillyard, S.A. 1980: Reading senseless sentences: Brain potentials reflect
semantic incongruity. Science, 207, 203–205.
Landauer, T.K., Foltz, P.W. and Laham, D. 1998: An introduction to latent semantic
analysis. Discourse Processes, 25, 259–284.
Levine, W.H., Guzman, A.E. and Klin, C.M. 2000: When anaphor resolution fails.
Journal of Memory and Language, 43, 594–617.
McKoon, G. and Ratcliff, R. 1992: Inferences during reading. Psychological Review,
99, 440–466.
Minsky, M. 1975: A framework for representing knowledge. In P.H. Winston (ed.)
The Psychology of Computer Vision. New York: McGraw-Hill.
Morris, R.K. and Folk, J.R. 1998: Focus as a contextual priming mechanism. Mem-
ory & Cognition, 26, 1313–1322.
Recanati, F. 1995: The alleged priority of literal interpretation. Cognitive Science, 19,
297–232.
Rumelhart, D.E. and Ortony, A. 1977: The representation of knowledge in memory.
In R.C. Anderson, R.J. Spiro and W.E. Montague (eds.) Schooling and the Acqui-
sition of Knowledge. Hillsdale, NJ: Lawrence Erlbaum Associates.
Sanford, A.J. and Garrod, S.C. 1981: Understanding Written Language: Explorations
Beyond the Sentence. Chichester: John Wiley & Sons.
Sanford, A.J. and Garrod, S.C. 1998: The role of scenario mapping in text comprehen-
sion, Discourse Processes, 26, 159–190.
Sanford, A.J. and Moxey, L.M. 1995a: Notes on plural reference and the scenario-
mapping principle in comprehension. In C. Habel and G. Rickheit (eds.) Focus
and Cohesion in Discourse. Berlin: de Gruyter.
Sanford, A.J. and Moxey, L.M. 1995b: Aspects of coherence in written communi-
 Blackwell Publishers Ltd. 2002
206 A.J. Sanford

cation: a psychological perspective. In M.A. Gernsbacher and T. Givon (eds.)


Coherence in Spontaneous text. Amsterdam: John Benjamins.
Sanford, A.J. and Young, K. 2001: Unpublished data.
Sanford, A.J., Clegg, M. and Majid, A. 1998: The influence of character type on
processing background information in narrative discourse. Memory & Cognition,
26, 1323–1329.
Sanford, A.J., Moar, K. and Garrod, S.C. 1988: Proper names and the control of focus.
Language and Speech, 31, 43–56.
Sanford, A.J., Sturt, P., Stewart, A.J. and Archambault, A. (submitted) Text change
blindness.
Schank, R. and Abelson, R. 1977: Scripts, Plans, Goals and Understanding. Hillsdale,
NJ: Lawrence Erlbaum Associates.
Schlesinger, I.M. 1968: Sentence Structure and the Reading Process. The Hague: Mouton.
Simons, D.J. 2000: Current approaches to change blindness. Visual Cognition, 7, 1–15.
Simons, D.J. and Levin, D.T. 1997: Change blindness. Trends in Cognitive Sciences, 1,
261–267.
Sperber, D. 2000: Metarepresentations in an evolutionary perspective. In D. Sperber
(ed.) Metarepresentations: A Multidisciplinary Perspective. Oxford: Oxford Univer-
sity Press.
Sperber, D. and Wilson, D. 1998: The mapping between the mental and the public
lexicon. In P. Carruthers and J. Boucher (eds.) Language and Thought. Cambridge:
Cambridge University Press.
Tomlin, R.S. 1985: Foreground-background information and the syntax of subordi-
nation. In T. Givón (ed.), Quantified studies in discourse, Text (Special Issue), 5,
85–123.
Toolan, M.J. 1988: Narrative: A Critical Linguistic Introduction. London: Routledge.
Van Oostendorp, H. and De Mul, S. 1990: Moses beats Adam: A semantic relatedness
effect on a semantic illusion. Acta Psychologica, 74, 35–46.
Walker, M.A., Joshi, A.K. and Prince, E.F. 1998: Centering Theory in Discourse. Oxford:
Clarendon Press.
Wason, P. and Reich, S.S. 1979: A verbal illusion. Quarterly Journal of Experimental
Psychology, 31, 591–597.

 Blackwell Publishers Ltd. 2002

You might also like