Depth of lexical-semantic processing and sentential load
Introduction
There is now a considerable amount of evidence that the words encountered during reading are not processed to a uniform degree. For example, work in the tradition of the Moses illusion (Erickson & Matteson, 1981) has shown that words that fit well in a broad context, but that are actually anomalous in their local semantic setting, may not be perceived as anomalous during reading, as in:
- (1)
How many animals of each sort did Moses put on the Ark?
- (2)
Suppose there was an aircrash right on the border of France and Spain. Where should the survivors be buried?
Example (1) is the well-known Moses illusion (Erickson & Matteson, 1981): it was Noah, not Moses, who put the animals on the Ark. Example (2) was extensively studied by Barton and Sanford (1993): you don’t bury survivors. Many studies show just how easily a reader or listener can miss such anomalies (e.g., Hannon and Daneman, 2001, Hannon and Daneman, 2004, Reder and Kusbit, 1991, Van Oostendorp and de Mul, 1990, Wason and Reich, 1979); their prevalence suggests that shallow semantic processing may be ubiquitous, although much psycholinguistic work of recent years has tended to emphasize incremental interpretation, and assume uniform deep processing (see Ferreira, Ferraro, & Bailey, 2002 and Sanford & Sturt, 2002 for complementary overviews).
Our basic tenet is that the extent or depth of semantic (and syntactic) processing is of potential importance for any proper understanding of how discourse comprehension works. Anyone who is communicating is trying to induce a listener to attend to and process some things and not others. Processes that control the extent of semantic analyses of words in a discourse should play a role in determining which things are processed, and the degree to which they are processed. Devices of emphasis in writing must work by encouraging some kinds of processing at the expense of others, and complex or obfuscatory writing could well serve to discourage thorough processing where appropriate. The present paper is aimed at exploring one of these issues.
A key question is what factors determine whether semantic processing is deep (extensive) or shallow. There are many possibilities, including stylistic devices designed to draw a reader’s attention to some things and not others. In our own research, we have studied two main candidates for influences on depth of processing. First, linguistic devices of focus, putting attention on certain parts of a discourse rather than other parts, should enhance semantic processing. A second factor that may serve to modulate the quality of semantic processing is the complexity of syntactic processing, this time by reducing the extent or success of processing. The present paper explores this possibility.
In the remainder of the introduction, we describe existing evidence that shows how focus affects the extent of semantic processing that is afforded to words. We then describe our manipulations of processing load, relating this to empirical and theoretical work in the literature.
It has been demonstrated that focus can enhance semantic processing. Using the Moses illusion (Erickson & Matteson, 1981), Bredart and Modolo (1988) showed that people detected the anomaly much more easily with the it-cleft version (3) than with (4):
- (3)
It was Moses who put two of each kind of animal on the Ark.
- (4)
Moses put two of each kind of animal on the Ark.
Because (3) is effectively an answer to the question Who was it that put two kinds of animals onto the Ark?, Moses is afforded more thorough processing than is the case when it is not put into focus in this way (4). Similar results have been found for subordination. Baker and Wagner (1987) demonstrated that anomalous information in complex sentences was less likely to be detected if the clause with the anomaly was logically subordinate rather than being the main clause.
To address the focus question in a different way, Sturt, Sanford, Stewart, and Dawydiak (2004) devised a text-change task, in which participants saw passages twice, in immediate succession. On some trials, there was a change to one of the words on the second exposure and participants were asked to report which word changed and what it changed to. The change-detection task devised by Sturt et al. is analogous to the change-blindness paradigm used for detecting changes in complex visual scenes, where failures to detect changes have been taken as indicative of inattention (e.g., Hollingworth and Henderson, 2002, Hollingworth et al., 2001, Simons and Levin, 1997). Sturt et al. argued that failing to notice a change would index when a particular word had not been deeply processed. Although no other text-change detection experiments have been reported (except Sanford & Sturt, 2002), Raney and Rayner (1995) used a similar, repeat-reading paradigm to examine the effects of word frequency on repeated reading. Effects of change were found, but no measures of the conscious detection of changes were made.
The task used by Sturt et al. (2004) enabled a theory of how focus influenced semantic processing to be tested. Two factors were manipulated, first whether the word in question was in focus, and second, whether the change made was to a semantically related or unrelated word. An example is:
- (5)
Everyone was wondering which man had been arrested.
The man with the hat [→ cap/dog] had been arrested that night.
(brackets indicate change options, cap being closely related, and dog being unrelated).
- (6)
Everyone was wondering what had happened.
The man with the hat [→ cap/dog] had been arrested that night.
In (5), the implicit question is which man, making the information with the hat salient (focused); in (6), it is the whole event that is focused, with the man with the hat not particularly relevant, just an associated property. In (5), the critical issue is to pick the relevant man out of the (hypothetical) set of alternative men (i.e., focus is contrastive; see, e.g., Rooth, 1995).
The experiment revealed a main effects of focus, with more changes being detected when the critical word was in focus, and of semantic distance, with the change to dog being more noticeable than the change to cap. There was also an interaction between these two, such that the distance effect was reduced when the word-change in question was in focus. Sturt et al. had predicted this interaction on the basis of the Granularity Theory of focus.
According to this theory, when a word is in focus, its meaning is represented in memory at a more detailed level than when it is not in focus. Consider the case where hat is not in focus. Under these circumstances, the meaning of the word may be represented in such a way that it may not be discriminated from cap, being perhaps represented simply as something corresponding to some unspecified form such as headgear. However, when it is in focus, the level of representation should be more fine-grained, carrying the distinction between hat and cap. The impact of this on change detection is straightforward. When a word is in focus, it should be possible to detect changes to a word with a similar meaning much more easily than when it is not in focus. On the other hand, if the change is to a word that is dissimilar in meaning, then it should be relatively easy to detect whether the word is in focus or not. Hence, focus should reduce the effect of semantic distance. This is exactly what Sturt et al. (2004) found, and it has been replicated with auditory presentation (Sanford, Molle, Sanford, & Healy, 2004).
In the present paper, we examine whether similar effects occur with the complexity of sentence processing. One possibility is that when a sentence is difficult to process for reasons of syntactic or other complexity, resources that might be spent on semantic analysis may well be compromised, leading to a reduced capacity to recognize a change. This is consistent with any pooled resource account, such as the Capacity Theory of comprehension (Just & Carpenter, 1992). A further question is whether a reduction in processing resources due to sentential complexity leads to a cruder granularity of representation, or whether it degrades processing in some other way.
There is some evidence that syntactic complexity influences semantic processing. Thus, Eastwick and Phillips (1999) examined the use of a semantic cue, animacy, as a source of information in disambiguating by-phrase sentences. An example is:
- (7)
The evidence/defendant recently examined by the lawyer was not very reliable.
- (8)
The judge remembered that the memo stating that the evidence/defendant recently examined by the lawyer was not very reliable had been stolen from the filing cabinet.
- (9)
The reporter who sent the photographer hoped for a story.
- (10)
The reporter who the photographer sent hoped for a story.
It is also recognized that not all sentences containing object-extracted relative clauses are equally difficult to process. Indeed, they can seem relatively easy to understand if they contain a first- or second-person pronoun (indexical pronouns) rather than a full noun-phrase (e.g., Bever, 1970, Bever, 1974, Kac, 1981). Examples from Warren and Gibson (2002) illustrate nicely the intuitive difference in processing difficulty when a full noun phrase and an indexical pronoun are compared within the same structure:
- (11)
The professor who the student had recently met at a party was famous, but no-one could work out why.
- (12)
The professor who I had recently met at a party was famous, but no-one could work out why.
Example (11) is intuitively much harder to read than (12). The important difference is that (11) contains a full definite NP (the student) while (12) contains a first-person pronoun (I). In controlled experiments, both ratings of difficulty and self-paced reading times confirm that (11) is harder to process than (12) (Warren, 2001, Warren and Gibson, 2002; see also Gordon et al., 2001). This manipulation of the form of the subject in the relative clause also meets two of our own requirements for exploring load with change detection. There is an empirically demonstrated and robust effect of referential complexity on reading time, and there is a reasonably well-understood sentential locus for where these complexity effects influence processing. Assuming this locus is where processing load is high, we may also assume that if there is a localized effect of load in our own studies, this is where it will occur.
Warren and Gibson predicted the observed differences in difficulty on the basis of two arguments, an argument about the relative “givenness” of full NPs and indexical pronouns, and the role that new individuals play in the load that has to be carried while sentence parsing is occurring. According to the idea of a Givenness Hierarchy (Ariel, 1990), different types of nominal expressions vary with respect to accessibility of their referents (Ariel, 1990, Gundel et al., 1993). Warren and Gibson hypothesize that processing load at an NP is related to its referent’s expected accessibility, and given the assumption that the referents of first- and second-person pronouns are highly accessible in context, they predict that these pronouns will be very easy to process (Warren & Gibson, 2002). On the other hand, full noun phrases refer to individuals that are much less accessible, either in the prior discourse (Garrod and Sanford, 1994, Sanford and Garrod, 1981) or most obviously if there is no prior discourse.
Warren and Gibson’s second assumption comes from Gibson’s (1998) dependency locality theory of syntactic complexity. On this account, processing load effects occur at points where prior parts of the sentence have to be integrated, for instance, where the object of the verb has to be integrated with the verb [The professor with met in (11) and (12)]. The crucial determinant of processing difficulty with (11) is that a new entity is introduced as the subject (the student) prior to the integration; with (12), the assumption is that nothing new needs to be introduced. Warren and Gibson predicted greater difficulty of integration of met in (11) than in (12). They argued that the difficulty emerges only at the point of integration (met), and not at other points (such as the adverb, recently, or the following NP, at a party).
A similar conclusion regarding the locus of the effect can also be reached from other theoretical positions. For instance, a potential source of integration difficulty is interference between similar elements in memory during retrieval of what is to be integrated. (Gordon et al., 2001, Lewis, 1996, Van Dyke and Lewis, 2003). Thus, Gordon et al. (2001) proposed that one reason why (11) is more difficult than (12) is that in (12) there are different forms of NP (the professor, as well as the pronoun I), whereas in (11) there are two full definite NPs (the professor, the student). The similar forms of NPs lead to competition, which makes the item more difficult to read. This does not occur in (11). Our purpose is not to enter this debate (see, for instance, Warren & Gibson, in press), but to select a load manipulation that is well researched in terms of the locus of difficulty. Using self-paced reading time, Warren and Gibson showed that there was a slower reading time for the embedded verb when a full NP was used, as predicted (i.e., people were slower to read the verb met in 11 than in 12), and that there was no effect on the following NP (i.e., in example 11, the party).
Our basic question is whether we can obtain a difference in detection of word changes at critical points as a function of sentential complexity. We explore this question by contrasting performance with sentences containing subject- and object-extracted embedded clauses (Experiment 1), and by manipulating referential load (Experiment 2). The reading time experiments discussed earlier revealed longer dwell times on embedded verbs under high-load conditions, consistent with localized processing difficulty. However, it is unclear what happens during this extra time. For instance, if the extra time leads to the repair of any difficulties that emerge as a result of high load, then it may be the case that no load effect will occur at all using word change-detection. However, if the extra time is simply a correlate of high demand on processing resources, then load could have a variety of effects that are not overcome by the extra processing time, including effects on depth of semantic processing.
Do similar load effects occur with reading and with listening? Under reading conditions, participants have as long as they require (in principle) to compensate for sentence complexity. However, with speech, the time available for processing is determined simply by the producer, not by the listener. Under such conditions, it is possible that load effects on lexical processing might occur, even if they do not with reading, where by taking more time readers might compensate for added complexity. This is tested in Experiment 3.
A third question is whether any load effects obtained are localized to those points theoretically predicted, and supported by reading time data, or whether the load effects spread over other sites in the sentence. A diffuse effect of load on change detection is certainly possible, even if that effect emanates from difficulties at a particular sentential locus in the first place, and such an outcome would suggest that sentential complexity has a general effect on the short-term representation of a sentence. We investigate the localization question in Experiment 5. (Experiment 4 deals with a control issue.)
Apart from these fundamental issues, there is the question of whether processing load influences the granularity of the representation of words during reading. If granularity is coarsened under high-load conditions, then, all other things being equal, one would expect high load to have the greatest impact on close semantic changes.
Section snippets
Experiment 1: Subject- and object extraction
In this experiment, we manipulate processing load through a purely syntactic manipulation, comparing subject- and object-extracted constructions. Critical changes were made at the embedded verb, where load effects are predicted to occur. The aim is to establish whether syntactic complexity influences the likelihood of detecting changes.
Design and materials
A two-way repeated measures design was used to explore the effects of referential load and the semantic relatedness of the words that changed across presentations. Referential load was manipulated through the use of full noun phrases to induce a high load and use of first-and second-person nominal references to induce a low load. Semantic relatedness was manipulated by changing the critical word (the embedded verb) to either a semantically related word or a semantically unrelated word. The
Experiment 3: Auditory change detection
The difficulty associated with referential load should not be modality specific. The pronoun condition should be easier than the full NP condition to understand when being listened to, or when being read. The present experiment used an adaptation of the change procedure in which the Experiment 2 materials were voice-recorded and presented for listening.
Experiment 4: Comprehension under change detection
This experiment was carried out as a control for the materials used in Experiments 2 and 3. No conventional comprehension questions were used in studies 1, 2, and 3, and it is possible that the high load sentences were sufficiently complex to cause participants to give up on trying to comprehend them. The sentences which we used were based on, and hence very close to, those used by Warren and Gibson (2002) in their reading time study. Warren and Gibson asked simple questions to test basic
Experiment 5: Control positions
The final experiment was designed to check whether the clear load effect is localized or whether it is diffuse, due simply to a more general difficulty associated with the high referential complexity sentences. According to Warren and Gibson (2002), the referential complexity effect originates at the embedded verb, and has no major influence on adjacent loci, for instance at the adverb or the NP following the verb. In practice, they found a very small but nevertheless reliable effect of load on
General discussion
Text-change detection is a new technique for investigating the fineness, or detail, of the representation of various aspects of discourse. The technique is sensitive to a number of things that in global terms should influence the amount of attention being paid to certain parts of a text, including focus (Sturt et al., 2004), and subordination (Sanford & Sturt, 2002). The main aim of the present studies was to investigate the possibility that a high sentential complexity would reduce the amount
Acknowledgments
This research was supported by ESRC Grant R000239888 to Anthony Sanford, and by study leave to Alison Sanford from the University of Strathclyde. Alison Sanford thanks the Department of Psychology, University of Glasgow for providing facilities for carrying out parts of this research. Experiment 3 was carried out by Nichola Healy. The authors are grateful to Sandy Pollatsek and Jerry Myers for statistical advice on the independent processes account, and to Ted Gibson, Lyn Frazier and Chuck
References (43)
- et al.
Controlling the intelligibility of referring expressions in dialogue
Journal of Memory and Language
(2000) - et al.
Moses strikes again: Focalization effects on a semantic illusion
Acta Psychologica
(1988) - et al.
The independence of syntactic processing
Journal of Memory and Language
(1986) Linguistic complexity: Locality of syntactic dependencies
Cognition
(1998)- et al.
Eye fixation patterns during the reading of relative clause sentences
Journal of Verbal Learning and Verbal Behavior
(1981) - et al.
Individual differences in syntactic processing: The role of working memory
Journal of Memory and Language
(1991) - et al.
The influence of animacy on relative clause processing
Journal of Memory and Language
(2002) - et al.
Locus of the Moses Illusion: Imperfect encoding, retrieval, or match?
Journal of Memory and Language
(1991) - et al.
Depth of processing in language comprehension: Not noticing the evidence
Trends in Cognitive Sciences
(2002) - et al.
Change blindness
Trends in Cognitive Sciences
(1997)
Processing subject and object relative clauses: Evidence from eye movements
Journal of Memory and Language
Semantic influences on parsing: Use of thematic role information in syntactic ambiguity resolution
Journal of Memory and Language
Distinguishing effects of structure and decay in attachment and repair: A cue-based parsing account of recovery from misanalyzed ambiguities
Journal of Memory and Language
Moses beats Adam: A semantic relatedness effect on a semantic illusion
Acta Psychologica
The influence of referential processing on sentence complexity
Cognition
Accessing noun-phrase antecedents
Evaluating information for truthfulness: The effects of logical subordination
Memory & Cognition
A case study of anomaly detection: Shallow semantic processing and cohesion establishment
Memory & Cognition
The cognitive basis of linguistic structures
The ascent of the spurious, or there’s a lot more we don’t know about mirrors
Cited by (25)
Separating the Activation, Integration, and Validation Components of Reading
2016, Psychology of Learning and Motivation - Advances in Research and TheoryCitation Excerpt :The key to the illusion lies in the underlying strength of the relation between the erroneous term (Moses) and the correct term (Noah) in long-term memory. Studies on semantic anomalies have typically utilized question answering paradigms (Erickson & Mattson, 1981; Kamas et al., 1996; Reder & Cleeremans, 1990; Reder & Kusbit, 1991), change detection tasks (Sanford, Sanford, Filik, & Molle, 2005; Sanford, Sanford, Molle, & Emmott, 2006; Sturt, Sanford, Stewart, & Dawydiak, 2004; Ward & Sturt, 2007), or short texts (Barton & Sanford, 1993; Bohan & Sanford, 2008; Daneman, Hannon, & Burton, 2006; Hannon & Daneman, 2004; Sanford, Leuthold, Bohan, & Sanford, 2011) to demonstrate that readers often fail to detect anomalies (ie, errors) unless the anomalies represent gross distortions of meaning and/or are highlighted via some distinguishing characteristic (eg, capitalization, syntactic focus, etc.). Furthermore, the tasks used in these studies typically require participants to consciously detect an anomaly (ie, provide a dichotomous response); the measure is discrete and does not allow for the possibility that information leading to detection could continue to accrue after a response has been made.
Conceptualist semantics: Explanatory power, scope and uniqueness
2013, Language SciencesCitation Excerpt :Accumulating evidence suggests that not all words in discourse are processed to the same degree (Glenberg et al., 1982; van Oostendorp, 1994; Sanford et al., 1995, 2005; Sanford and Sturt, 2002; Daneman et al., 2007). Instead, the granularity and extent of semantic processing are modulated by cognitive variables such as attention, task goals and context (Parkin, 1979; Smith, 1979; Smith et al., 1983; Henik et al., 1983, 1994; Kaye and Brown, 1985; Chiappe et al., 1996), as well as by more local, grammatical ones such as focus and embedding (Baker and Wagner, 1987; Bredart and Modolo, 1988; Sturt et al., 2004; Sanford et al., 2005). This is not the place to review this evidence.
The pistol that injured the cowboy: Difficulty with inanimate subject-verb integration is reduced by structural separation
2012, Journal of Memory and LanguageCitation Excerpt :Specifically, work by Ferreira and colleagues (Christianson, Hollingworth, Halliwell, & Ferreira, 2001; Ferreira, 2003; Ferreira, Christianson, & Hollingworth, 2001; see Ferreira & Patson, 2007, for a review) has demonstrated that participants who are presented with garden-path sentences or passive sentences that contain noncanonical agent-patient roles frequently misinterpret the meaning of the sentence. Similarly, Sanford and colleagues (Sanford, Sanford, Filik, & Molle, 2005; Sturt, Sanford, Stewart, & Dawydiak, 2004) have used a change-detection paradigm to demonstrate that readers are more likely to notice that a target word has changed from one presentation of the text to the next if the word is linguistically focused by the sentence structure or is highlighted by a prior discourse context. Although this previous work provides compelling evidence that linguistic representations are often inaccurate or incomplete, the measures used combine the influences of both online linguistic processing and offline memory-based retrieval of the linguistic information.
A Review of Studies on the Selective Attention Strategy During Language Comprehension: The Present and the Future
2023, Journal of Language Teaching and Research