Which Of The Following Would Be Considered A Garden-path Sentence?

Sentence Processing

E.-K. Lee , D.G. Watson , in Encyclopedia of Human Behavior (Second Edition), 2012

Garden-Path versus Constraint-Based Model

Under both the garden-path and constraint-based models, nonsyntactic sources of information influence syntactic processing. However, these theories differ with respect to when these sources of information are thought to come into play. Experimental evidence from the psycholinguistic literature has shown very early effects of these nonsyntactic sources of information during syntactic processing, suggesting that the processing system is interactive. Proponents of the garden-path model, however, interpret these early effects as reflecting the process of reanalysis, arguing that they do not necessarily rule out the garden-path model. For the garden-path model to account for these data, reanalysis must occur very rapidly, and thus the outstanding differences between the garden-path and constraint-based models are now fairly minimal. Thus, consistent with both models, we can conclude that a wide array of information sources are used very rapidly to determine the most likely structure of a sentence.

Read full chapter

URL:

https://www.sciencedirect.com/science/article/pii/B9780123750006003219

Effects of the Focus Particle Only and Intrinsic Contrast on Comprehension of Reduced Relative Clauses

Charles CliftonJr., ... Janina Radó , in Reading as a Perceptual Process, 2000

Referential theory

All psycholinguists know that sentences like the garden-path sentence of Bever (1970) in (1) are difficult to comprehend. A reader or listener takes the initial verb as the main verb of the sentence and has great difficulty arriving at the required relative clause analysis. Experimental demonstrations of this fact abound (e.g., Frazier and Rayner, 1982). On the other hand, some psycholinguists have pointed out that sentences superficially like (1), such as that in (2), are easy to comprehend as relative clauses when they occur in appropriate contexts.

(1): The horse raced past the bam fell.
(2): A person presented with the evidence examined by the lawyer may interpret … (Ferreira and Clifton, 1986, quoted by MacDonald, Pearlmutter and Seidenberg, 1994, p. 678).

Crain and Steedman (1985) proposed a 'referential theory of sentence processing' to account for some of the variability in difficulty of comprehending relative clause constructions, among a variety of other effects. In their model, the parser constructs all syntactic analyses of a fragment, choosing which analysis to pursue on semantic/pragmatic grounds, namely, whichever analysis violates the fewest presuppositions or requires the fewest additions to the evolving discourse model. They assumed that a restrictive relative clause felicitously modifies a definite noun phrase only if it is used to identify a subset from an already established context set. In the absence of a context set, only the main clause interpretation of, say, 'The horse raced past the bam …' will be felicitous. This is not because the main clause analysis is syntactically preferred, but because a relative clause can be used to modify a definite noun phrase only if the discourse context contains multiple possible referents of the head noun from which the relative clause permits selection. If this discourse presupposition is not met, the reader/listener must accommodate it by adding possibly contrasting entities to the discourse context, a process which is costly and therefore to be avoided.

There is good reason to doubt the validity of this assumption about the discourse requirements of relative clauses, and hence reason to doubt the validity of the Crain and Steedman analysis (cf. Clifton and Ferreira, 1989). It is not at all clear that relative clauses are predominantly, or even commonly, used in the way Crain and Steedman assume that they must be used. Fox and Thompson (1990) present a careful pragmatic analysis of the actual use of relative clauses in a corpus of spoken English. Their analysis of 414 relative clauses (all at least arguably restrictive) indicated that while all relative clauses were 'grounded', in the sense of being linked to given referents in the discourse, the linkage took a variety of forms. A common use was as a 'file-establishing' relative clause, which introduces a new entity into the discourse. In their example (4), … the uh heater thing we put in I think was a hundred … dollars, the relative clause we put in is linked by the word we to the conversational participants (who are necessarily represented in the discourse model). Another way of grounding is called 'proposition-linking'. An example is the relative clause who isn't a Catholic from Fox and Thompson's example (6): … The mother's sister is a real bigot. Y'know and she hates anyone who isn't a Catholic. Here, the relative clause is grounded through the frame provided by the word bigot.

Fox and Thompson explicitly distinguish between two functions of the relative clauses they collected, namely, "provid[ing] a characterization of a New Head NP referent, not previously known to the hearer," and "help[ing] to identify a Given Head NP referent, previously known to the hearer" (pp. 301–302). Crain and Steedman's analysis claims that only (a subset of) the latter use of relative clauses is felicitous when the relative clause is used to modify a definite NP. Fox and Thompson's data suggest that this claim may be too strong. They provided explicit statistics on 382 relative clauses with nonhuman head nouns. They were able to categorize the communicative function of 228 of these: 141 (62%) were used as 'identifying' and 87 (38%) as 'characterizing'. Including the 154 relative clauses whose function could not be securely determined (but were presumably not used as 'identifying') only these 141 (37% of the total) met the criterion of selecting from a set of discourse entities. If all of these relative clauses modified definite NPs, only the 'identifying' usage should be possible according to Crain and Steedman. Unfortunately for the purposes of this analysis, Fox and Thompson did not provide a breakdown by NP definiteness. They did note that 85% of subject NPs modified by a relative clause were definite and that only 43% of all subject NP relative clauses were used as 'identifying' (and that 28% of object NPs modified by a relative clause were definite, with 25% of all object NP relative clauses being 'identifying'), suggesting that at least a substantial number of definite NPs had relative clauses that were not used to identify a specific referent.

Read full chapter

URL:

https://www.sciencedirect.com/science/article/pii/B9780080436425500280

Sentence Comprehension, Psychology of

C. CliftonJr., in International Encyclopedia of the Social & Behavioral Sciences, 2001

3.4 Lexical and Frequency Effects

Once the existence and basic nature of garden-paths were discovered, researchers realized that garden-paths did not always occur. The sentence presented earlier, 'The car driven into the garage caught fire' does not seem to lead to a garden-path. This may be due in part to the fact that the verb 'driven' is unambiguously a participle, not a main verb, and the normally preferred simpler analysis is grammatically blocked. However, other cases indicate that more subtle effects exist. Sentences with verbs that are obligatorily transitive are relatively easy (e.g., 'The dictator captured in the coup was hated' is easier than 'The dictator fought in the coup was hated'). Sentences with verbs like 'cook,' whose subject can take on the thematic role of theme/affected object, are particularly easy ('The soup cooked in the pot tasted good'). Sentences in which the subject is implausible as agent of the first verb but plausible as its theme are easier than sentences in which subject is a plausible agent (e.g., 'The evidence examined by the lawyer was unreliable' is easier than 'The defendant examined by the lawyer was unreliable,' although this difference may depend on a reader being able to see the start of the disambiguating 'by' phrase while reading the first verb). Consider a different grammatical structure, prepositional phrase attachment. It is easier to interpret a prepositional phrase as an argument of an action verb than as a modifier of a noun (e.g., 'The doctor examined the patient with a stethoscope' is easier than 'The doctor examined the patient with a broken wrist'), which is consistent with a grammatical analysis in which the former sentence is structurally simpler than the latter. However, this difference reverses when the verb is a verb of perception or a 'psychological' verb and the noun that may be modified is indefinite rather than definite (e.g., 'The salesman glanced at a customer with suspicion' is harder than 'The salesman glanced at a customer with ripped jeans') (Spivey-Knowlton and Sedivy 1995; see also Frazier and Clifton 1996, Mitchell 1994, Tanenhaus and Trueswell 1995, for further discussion and citation of experimental articles).

These results indicate that grammatical structure is not the only factor that affects sentence comprehension. Subtle details of lexical structure do as well. Further, the sheer frequency with which different structures occur and the frequency with which particular words occur in different structures seems to affect sentence comprehension: For whatever reason, more frequent constructions are easier to comprehend. For example, if a verb is used relatively frequently as a participle compared to its use as a simple past tense, the difficulty of the 'horse raced' type garden-path is reduced. Similarly, the normal preference to take a noun phrase following a verb as its direct object (which leads to difficulty in sentences like 'I understood the article on word recognition was written by an expert') is reduced when the verb is more often used with sentence complements than with direct objects. MacDonald et al. (1994) review these findings, spell out one theoretical interpretation which will be considered shortly, and discuss the need to develop efficient ways of counting the frequency of structures and to decide on the proper level of detail for counting structures.

Read full chapter

URL:

https://www.sciencedirect.com/science/article/pii/B0080430767015497

Commas and Spaces: Effects of Punctuation on Eye Movements and Sentence Parsing

Robin L. Hill , Wayne S. Murray , in Reading as a Perceptual Process, 2000

Relative-clause ambiguities

Like closure ambiguities, reduced relatives are also generally held to promote very strong garden-paths. This is especially true in 'unforgiving' structures, such as The farmer sold the pigs fattened them up, where the second noun phrase (NP) is potentially a direct object. There have been a large number of studies examining a range of factors which might influence the processing of this ambiguity. But while some have claimed effects of factors such as context, verb structure, etc., on its resolution (e.g. MacDonald, 1994; Spivey-Knowlton and Tanenhaus, 1998; Trueswell, Tanenhaus and Garnsey, 1994), others have denied the existence of such effects (e.g. Clifton and Ferreira, 1989; Murray and Liversedge, 1994; Patterson, Liversedge and Underwood, 1999). Wherever the truth may eventually be found to lie, it again seems uncontroversial to claim that this, especially in sentences where the initial verb is followed by a potential direct object, presents a very robust garden-path. However, unlike early closure, it is not immediately intuitively obvious that punctuation would necessarily help. But, rather than resolving this ambiguity by expanding the relative clause into its unreduced form, a similar effect might be obtained by physically 'chopping' the sentence into its primary constituents. Delineating the clausal boundaries with commas is just such a possible (and non-radical) mechanism of textual segmentation. While contextual and other factors might influence the processing of this structure, it seems pretty clear that if manipulations of this sort are to have any hope of success, they must be robust (and probably compound). But if a couple of humble commas can alleviate this kind of processing difficulty, irrespective of other factors, then punctuation must surely have a rapid and dominant influence on parsing.

Read full chapter

URL:

https://www.sciencedirect.com/science/article/pii/B9780080436425500279

Contemporary Theories of Humor

Rod A. Martin , Thomas E. Ford , in The Psychology of Humor (Second Edition), 2018

Comprehension and Elaboration

Yu-Chen Chan at the National Taiwan Normal University, Taiwan, has collaborated with a number of colleagues ( Chan et al., 2012, 2013 Chan et al., 2012 Chan et al., 2013 ) to investigate whether the processes thought to underlie the comprehension and elaboration of verbal jokes are truly distinct, i.e., whether they involve different neural circuits in different regions of the brain. Chan et al. (2012) placed participants in an fMRI scanner and presented them with 60 verbal stimuli. The setup was shown for 20 seconds and the punch line (ending) for 9 seconds. Participants pressed a button on a keypad to indicate that the joke/statement was funny or not funny.

The three types of verbal stimuli were: funny jokes, unfunny statements, or garden path statements. Garden path statements are nonhumorous statements that contain an incongruity, a conclusion that does not make sense. Upon encountering the incongruity, the reader must reread the sentence to construct a new sensible reinterpretation. Accordingly, garden path statements involve the same incongruity-resolution processing associated with humor comprehension, but not the processing required for appreciating humor associated with the elaboration stage (p. 901). Chan et al. provided the following example of a funny joke and its conversion to an unfunny statement and to a garden path statement.

Funny Joke: Setup: One day after work, a mother buys some donuts from a store close to her office. When she gets home, she says to her eldest son, "Peter, Mom brought some sweets home. You can take one donut to share with your little brother. Don't eat it all yourself!" So, Peter takes the donut, thanks his mom, goes to his little brother and says,

Punch line: "Hey, we have a donut to share! I'll take the circle, and you can have the hole!"

Unfunny Statement: The punch line was replaced with, "Hey, we have a donut to share! I'll eat half and you can have the other half!"

Garden Statement: The punch line was replaced with, "I don't eat chocolate, donuts are more to my taste."

A comparison of brain images for participants in the unfunny statement and garden path statement conditions indicates the brain regions associated with the comprehension stage. This revealed that the left and right inferior frontal gyri (bilateral IFG) and the left superior frontal gyrus were associated with humor comprehension. By comparing the brain images for participants in the funny joke condition versus those in the garden path statement condition, Chan et al. identified the brain regions associated with the elaboration stage. They found that the ventromedial prefrontal gyrus (left vmPFC) in the cortical regions and bilateral amygdalae and bilateral parahippocampal gyri in the subcortical regions were uniquely associated with the processes involved in elaboration that result in the feeling of amusement.

Chan et al. (2013) further distinguished between the comprehension processes of incongruity detection and reinterpretation/incongruity resolution. Fig. 3.2 illustrates the three stages of neural circuitry underlying comprehension and elaboration.

Chan et al.'s (2012, 2013) Chan et al.'s (2012) Chan et al.'s (2013) findings support Wyer and Collins' (1992) model that comprehension and elaboration represent two distinct sets of cognitive processes underlying the amusement with verbal jokes.

Read full chapter

URL:

https://www.sciencedirect.com/science/article/pii/B9780128121436000035

Psycholinguistics: Overview

B.J. MacWhinney , in International Encyclopedia of the Social & Behavioral Sciences, 2001

3.2 Garden-pathing

There are times when the initial decisions that we have made take us down the garden path. A classic example of garden-path processing occurs with sentences such as 'The communist farmers hated died.' It often takes the listener awhile to realize that it was the 'communist' that died and that it was the 'farmers' who hated the 'communist.' Inclusion of a relativizer to produce the form, 'The communist that farmers hated died' might have helped the listener sort this out. A somewhat different example is the sentence, 'The horse raced past the barn fell.' Here, we need to understand 'raced past the barn' as a reduced relative clause with the meaning 'The horse who was raced past the barn.' If we do this, the appearance of the final verb 'fell' after 'barn' no longer comes as a surprise.

Garden paths arise when a word or suffix has two meanings, one of which is very common and one of which is comparatively rare (MacDonald et al. 1994). In a sentence like 'The horse raced past the barn fell' the use of the verb 'raced' as a standard transitive verb is much more common than its use as the past participle in a reduced passive. In such cases, the strong meaning quickly dominates over the weak meaning. By the time we realize our mistake, the weak meaning is fully suppressed by the strong meaning and we have to try to comprehend the sentence from scratch. A classic garden-path example from Lashley (1951) is the sentence 'Rapid righting with his uninjured left hand saved from destruction the contents of the capsized canoe.' When this sentence is read aloud, listeners find it extremely difficult to understand the second word as 'righting' rather than 'writing.'

Read full chapter

URL:

https://www.sciencedirect.com/science/article/pii/B0080430767030242

Mental Models and the Mind

Barbara Hemforth , Lars Konieczny , in Advances in Psychology, 2006

4 Eliminative mental model Theory: The role of parsimony

Since Bever (1970), principles governing human sentence processing have been investigated by looking at a very specific type of sentences, so-called garden-path sentences. These are sentences containing some kind of local ambiguity that is initially interpreted in a way incompatible with semantic or syntactic information showing up later in the sentence. A classic example of a garden-path sentence is (3).

(3): The horse raced past the barn fell.

Initially this sentence is rated as ungrammatical even by most native speakers. The problem is that there is a local ambiguity on the word "raced," which may be the main verb of the sentence, or a past participle starting a reduced relative clause (… that was raced past the barn …). There is a strong preference to interpret "raced" as the main verb of the matrix clause, but then the verb "fell" cannot be integrated, giving the impression of ungrammaticality.

Since Frazier & Fodor (1978), or even since Kimball (1973), garden-path effects have been explained on the basis of the syntactic structure of the respective sentences. In (3), the main verb reading is syntactically less complex than the reduced relative clause so that any principle involving a minimal amount of structure (e.g. minimal attachment, Frazier 1987, or simplicity, Gorrel 1995) can predict the preferences.

A principle of parsimony in syntactic structure building can also explain preferences in constructions like (4i, ii).

(4)

(i)

The psychologist told the woman that he had problems with

a.: her husband
b.: to leave her husband.

(ii)

The cop watched the spy with the binoculars.

In (4i) the (a)-version is syntactically less complex and easier to understand than the (b)-version. The ambiguity lies in the interpretation of the "that"-clause either as a complement clause attached to the main verb (He told her that …) or as a relative clause modifying the object noun phrase (… the woman that …). (4ii) is fully ambiguous (the prepositional phrase "with the binoculars" can be interpreted as an instrument of watching or as an attribute of the spy) but there is a preference to interpret the phrase "with the binoculars" as an instrument of the verb. This attachment is often assumed to be less complex (e.g. Frazier 1987; but see Hemforth 1993 for a critical analysis of this assumption).

Crain & Steedman (1985) proposed that the preferences established for these ambiguities are not at all due to a preference for a simple syntactic structure but to a simple and parsimonious mental model. During parsing, syntactic analyses are pursued in parallel. However, in cases of ambiguity, only those are kept which are compatible with the most parsimonious mental model. What does parsimony mean for a mental model? According to Crain and Steedman there are at least two factors determining more or less complex models. One is the number of presuppositions that have to be adjusted when integrating a new piece of information. For example, a definite noun phrase like "the woman" in (4i) implies that there exists a uniquely identifiable woman in the discourse universe. "The woman" in a context where there is more than one woman is infelicitous. A second factor concerning complexity is the number of entities to be represented in a mental model. Increasing the number of entities increases the complexity of the model. A preference for parsimonious models guarantees the construction of mental models with the minimal number of entities necessary to represent the current discourse.

How can such an approach explain garden-path phenomena? Consider a discourse as in (5). In the target sentence the "that"-clause can be a complement clause ("told … that") or a relative clause ("the woman that"). In the context given here, there is one married couple, presumably a man and a woman. Hence, the definite noun phrase "the woman" finds a unique referent without any further need of modification. Since there is no need for a modifying relative clause, the that clause is interpreted as a complement clause.

(5)

Context:

A psychologist was counseling a married couple. One of them was nice to him, but the other one was fighting with him.

Target:

The psychologist told the woman that he was having trouble with

a.: her husband.
b.: to leave her husband.

What if there were two married couples and consequently two women in the context as in (6)? Then the definite noun phrase "the woman" would not on its own allow identification of a unique referent. There would be need for more information. In this case a modifying relative clause would be felicitous. In their experiments, Crain and Steedman found a preference for the complement reading in one-referent contexts like (5) and a preference for the relative clause reading in contexts like (6).

(6)

Context:

A psychologist was counselling two married couples. One of them was nice to him, but the other one was fighting with him.

Out of any context, readers seem to prefer the complement clause reading (Frazier 1987). Why should that be so? Crain and Steedman assume that readers construct the minimal mental model compatible with the linguistic input. When they read "the woman," they assume that there is only a single woman in the current universe of discourse. Hence, there is no need for further information, no need for a modifying relative clause.

This line of research inspired an enormous amount of follow-up studies (e.g. Altmann & Steedman 1988), some of them pointing out empirical problems with the approach. It has been shown, for example that preferences for sentences like (4i) or (4ii) often depend on lexical biases (Britt et al. 1992), mostly from the verb, and also on syntactic biases like a preference for simple structures (Desmet et al. 2002, Konieczny & Voelker 2000). In particular, lexical and syntactic effects show up as early processes, as they can be established in experiments registering eye movements while participants are reading texts. A more or less complex context, on the other side, only affects interpretation at later stages.

More importantly, Konieczny and Voelker showed that some of the predictions to be derived from Crain and Steedman's model don't really hold in all cases. They presented participants with short texts like (7).

(7)

Yesterday, a girl / two girls was / were sitting on a bench.

Darja admired the girl a. with the pink dress. / b. with big eyes.

The interesting point here is that a fairly late preference for a noun modifying prepositional phrase ("the girl with the pink dress") could be established for the two-referent context. ² However, there was no penalty for a noun phrase modifier in a one-referent context. Participants did not really care whether or not a definite noun phrase like "the girl" that was uniquely identifiable without further information was then modified by a prepositional phrase. This additional information is apparently easily acceptable even when it does not serve any referential purpose. But if it is not the case that readers have to update their mental model to a more complex one in cases of modified noun phrases, this approach cannot really explain the basic preferences established for isolated sentences.

A more recent version of the eliminative approach can be found in research applying the visual world paradigm (e.g. Tanenhaus et al. 1995, Kamide et al. 2003, Trueswell et al. 1999). In this paradigm, participants are presented with visual scenes either on the screen or directly as a layout of objects on a table. They hear, for example, sentences like (8) including objects while looking at a scene that either has one baby or two babies. It has been shown that adults do not really consider the PP "in the cradle" as a directional object when there are two babies in the visual scene (whereas children apparently do, see Trueswell et al. 1999).

(8): Put the baby in the cradle on the highchair.

Moreover, given a visually presented scene with only a few objects, the given context appears to show an immediate and dominating effect on disambiguation in general, but also on the anticipation of verbal arguments (Kamide et al. 2003). Visual scenes obviously present a very strong cue for a mental model of a situation described in a sentence or text. Within this paradigm it is often argued that a strong enough mental model is sufficiënt to eliminate any effect of syntactic processes. On the other hand, one may argue that the presentation of a visual scene is not only a very strong cue for a mental model, but also a very specific situation for language processing. It is a situation where listeners construct a strongly reduced discourse universe with only a few clearly defined objects. Most of the time, participants are allowed to scan the scene for a few seconds before the linguistic input starts. It is probably true that these experiments show that syntactic factors in ambiguity resolution can be eliminated in principle, but it is far from clear in how far this extends to all sorts of everyday language processing, where the discourse universe is far less constrained.

All in all, we can say that some kind of enriched discourse model certainly affects sentence processing, though sometimes only fairly late. However, lower level representations, i.e. syntactic and semantic representations, seem to play an important role as well, and maybe even more so in earlier stages of processing.

There is another more theoretical problem with the approach presented in this section. Research on referential parsimony, since Crain and Steedman, assumes that the complexity of a mental model depends on the number of entities it contains. In the following section, we will discuss whether this is a viable assumption.

Read full chapter

URL:

https://www.sciencedirect.com/science/article/pii/S016641150680035X

Resolving Lexical Ambiguity Computationally with Spreading Activation and Polaroid Words

Graeme Hirst , in Lexical Ambiguity Resolution, 1988

5.3 Marker Passing, Path Strength, and Magic Numbers

One of the more vexed problems in using association cues for disambiguation, is knowing when an association is strong enough to be considered conclusive evidence. We know from the existence of semantic garden-path sentences that associations alone should sometimes cause immediate jumping to a conclusion; we also know that this isn't true of all associations, for we are not garden-pathed by sentences like (35):

(35): The lawyer stopped at the bar for a drink. (bar is not taken in any of its legal senses)

We therefore need some measure of the strength of an association, so that PWs will be able to jump to conclusions (rightly or wrongly) in the same situations that people do. ²³ Although frequency of the use of the concepts should be a factor in determining the strength of an association (see also [Anderson, 1983]), ²⁴ I shall limit my remarks below to a discussion of the semantic distance between two concepts.

I mentioned in the previous section that most theories of spreading activation assume that different links have different strengths, though Frail does not attempt to model this. It is generally assumed that link strength is correlated with semantic distance—that a link between two concepts is strong exactly when they are very closely associated. Cases when this occurs may include one concept being a salient property of the other (edibility, food), or, possibly, a particularly good exemplar of the other (robin, bird); ²⁵ a high frequency of use also strengthens a link [Collins and Loftus, 1975; Anderson, 1983] and hence the association between the concepts. On the other hand, de Groot [1983] has found that activation does not spread to associates of associates of a node—for example, bull and cow are linked and so are cow and milk, but activation from bull does not reach milk. Thus, PWs need a way to take an MP path and determine its strength, i.e., the semantic distance between its end-points, by looking at the links and nodes that it includes.

The present, inadequate method of measuring path strength is a function of the length of the path, the nodes it passes through, and the links it uses. I use the following heuristics:

•: The shorter the path, the stronger it is.
•: The more arcs that leave a node, the weaker the connections through that node are (see also the anti-promiscuity rule, Section 2.3).

These methods, though I use them, are unsatisfactory because, like the marker passing constraints (Section 2.3), they rely heavily on magic numbers. For example, the second suggests that any node will not be vague if it has only N arcs, but N + 1 arcs will invariably tip the scale. This seems unlikely. And even if there were a neat threshold like that, how do we know that N is it?—it is merely a number that seems to work in the present implementation, but there is no principled reason for it. There is, of course, well-known evidence for the psychological reality of magic numbers in certain perceptual and short term memory processes [Miller, 1956], but it is hard to believe that this carries over to marker passing in long term memory, where activation seems to be a continuous, not discrete, variable.

It is hoped that future versions of MP will be able to include such features as path strength and the weakening of activation as it gets further from the origin, so that we won't have to worry about posthoc measurements of path strength. This would be a first step in approximating the continuous nature of spreading activation.

Read full chapter

URL:

https://www.sciencedirect.com/science/article/pii/B9780080510132500070

Unrestricted Race: A New Model of Syntactic Ambiguity Resolution

Roger P.G. van Gompel , ... Matthew J. Traxler , in Reading as a Perceptual Process, 2000

Introduction: Two-stage versus constraint-based theories

Current sentence processing theories can roughly be divided into two fundamentally different classes which make very different claims about the architecture of the language processor. One class of theories are what we will call the two-stage theories. The garden-path theory, initially proposed by Frazier (1979; see also Ferreira and Clifton, 1986; Frazier, 1987; Rayner, Carlson and Frazier, 1983), has been the most influential of these theories, but there are other theories which are very similar in their basic assumptions, although they differ in many details (e.g., Abney, 1989; Crocker, 1995; Inoue and Fodor, 1995; Pickering, 1994; Pritchett, 1992). Many of the basic ideas of these theories date back from Bever (1970) and Kimball (1973), and they are strongly influenced by the modularity hypothesis (Fodor, 1983; Forster, 1979). One of the basic assumptions that two-stage theories make is that some potentially useful sources of information are initially ignored in sentence processing, because the sentence processor is informationally encapsulated. In the garden-path theory, for example, only information about syntactic tree structures can be used initially, and all other sources of information, such as thematic roles, discourse context, semantic plausibility and lexical frequency, are ignored during this initial stage. Other theories assume that information such as thematic roles (e.g., Pritchett, 1992; Abney, 1989) is used initially. By virtue of the fact that some sources of information are used later than others, all these models share the assumption that the processor works in two stages. In the first stage, the processor adopts one analysis on the basis of a restricted range of information. In the second stage, other sources of information are employed. If the initial analysis of an ambiguous syntactic structure is incorrect, another analysis is pursued. This reanalysis process causes processing disruption (a garden-path effect).

The second class of theories are the constraint-based theories, which are descended from older interactive theories (e.g., Marslen-Wilson, 1975; Taraban and McClelland, 1988; Tyler and Marslen-Wilson, 1977). Since the early 1990s, constraint-based theories have become the dominant type of account of sentence processing. Recently, these models have been worked out in more detail, so that it is possible to derive some predictions from them (e.g., MacDonald, Pearlmutter and Seidenberg, 1994; McRae, Spivey-Knowlton and Tanenhaus, 1998; Tabor, Juliano and Tanenhaus, 1997). All constraint-based theories stipulate that the processor works in a single stage and that it can employ all sources of information immediately. Because all relevant information can be employed immediately, there should be no reason to predict differences between on-line preferences while people are reading a sentence, and off-line preferences such as measured by sentence fragment completion studies. Constraint-based theories assume a parallel processor: different possible analyses of an ambiguous syntactic structure are activated simultaneously by the various sources of information and compete with each other until a certain threshold level is reached (e.g., McRae et al., 1998). When one analysis receives more support than the others, little competition occurs between the analyses, resulting in little processing difficulty. But when one or more alternative analyses are about equally activated, there is a strong competition between the analyses, which should lead to processing difficulty.

Constraint-based theories usually claim that this process of competition is exactly the same as in lexical processing (e.g., MacDonald et al., 1994). A number of studies on lexical ambiguity resolution observed that ambiguous words for which both meanings were equally high in frequency (balanced words) were read more slowly than their unambiguous controls, whereas ambiguous words that had a frequency bias toward one meaning (biased words) were read as fast as unambiguous control words (e.g., Duffy, Morris and Rayner, 1988; Rayner and Duffy, 1986; Rayner and Frazier, 1989). This can be interpreted as evidence for a process of competition between two meanings of a balanced ambiguous word. Similarly, when biased words were put in preceding contexts favouring the low frequency meaning, reading times were slower than when they were put in a context favouring the high frequency meaning (e.g., Dopkins, Morris and Rayner, 1992; Duffy et al., 1988; Rayner and Frazier, 1989; Rayner, Pacht and Duffy, 1994; Sereno, Pacht and Rayner, 1992). This is also consistent with a competition process between two alternative meanings. Therefore, if syntactic ambiguity resolution is indeed similar to lexical ambiguity resolution, competition effects should obtain in syntactic ambiguities as well.

On the other hand, some proponents of two-stage sentence processing theories claim that syntactic and lexical ambiguity resolution are fundamentally different (e.g., Frazier, 1989; Rayner and Morris, 1991; Traxler et al., 1998; Van Gompel et al., 1999). Competition may arise in lexical ambiguity resolution, but by virtue of its serial architecture, it cannot arise in syntactic ambiguity resolution.

Read full chapter

URL:

https://www.sciencedirect.com/science/article/pii/B9780080436425500292

Syntactic Aspects of Language, Neural Basis of

A.D. Friederici , in International Encyclopedia of the Social & Behavioral Sciences, 2001

3.1.1 The syntax-related late positivity

The late positivity is a centroparietally distributed positivity usually at around 600 msec and beyond, called P600. P600 components are observed both in correlation with outright syntactic violation and with violations of syntactic preferences. Violations of the latter kind are present in so-called 'garden path' sentences, that is sentences for which it becomes clear at a particular word that the perceiver was misguided structurally by following a simple instead of a more complex syntactic structure (for example: 'The broker hoped to sell the stock was sent to jail' vs. 'The broker persuaded to sell the stock was sent to jail,' Osterhout and Holcomb 1992). At critical points in the sentence (at the word 'to' and the word 'was') the reader/listener has to revise and reanalyze the structure initially followed. The P600 is also present in sentences containing an outright syntactic violation, be it a phrase-structure violation realized by a word category error (Neville et al. 1991, Friederici et al. 1993) or be it a morphosyntactic violation realized by an incorrect inflection (Gunter et al. 1997, Münte et al. 1997). Functionally, the P600 is taken to reflect costs of reprocessing (Osterhout et al. 1994), syntactically guided processes of reanalysis and repair (Friederici 1995), or syntactic integration difficulties in general (Kaan et al. 2000).

Read full chapter

URL:

https://www.sciencedirect.com/science/article/pii/B008043076703494X