Introduction

Garden-path sentences like (1) have been the focus of psycholinguistic studies for many years thanks to their specific structure containing a local syntactic (or semantic) ambiguity that initially leads comprehenders to interpret the sentence in a way that is not consistent with the following input (e.g., ‘Anna dressed the baby’). This misanalysis only tends to be resolved and reinterpreted later in the sentence. The temporary nature of the ambiguity allows researchers to probe various aspects of sentence processing, including the extent to which comprehenders are able to form coherent interpretations faithful to the input after realizing their initial analysis was wrong.

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

The initial misinterpretation of the sentence (1) is caused by the ambiguity of the verb dressed, which can be temporarily interpreted either as transitive, with the baby as a patient, or as reflexive, with Anna being both the agent and the patient. Because the transitive interpretation is preferred, comprehenders initially analyze the baby as the direct object of the verb dressed. This can be influenced by several factors, including syntactic complexity [1, 2], the frequency of the given verb form [3, 4], or plausibility of the baby as an argument of the verb [5, 6]. However, this initial analysis turns out to be incorrect, because the verb played would lack a subject if it was maintained.

According to Fodor and Inoue [7], the parser uses an operation called Attach anyway to attach the verb played to the current structure once it is encountered even though the result is ungrammatical. Then, it tries to repair the ungrammaticality by finding a suitable subject, in this case the baby. The baby is detached from its initial object position and attached again as the subject of played. Because the transitive verb dressed now lacks a direct object, a new ungrammaticality is created, which is solved by finding a reflexive interpretation of the verb. In general, the parser can continue adjusting the structure until the result is grammatical or until it encounters an ungrammaticality that cannot be repaired, in which case the reanalysis fails and the whole sentence is considered ungrammatical. Some approaches, e.g. cue-based retrieval, also argue that the difficulty and success of reanalysis can depend on the amount of interference caused by elements that are syntactically and semantically similar to the one that needs to be retrieved by the verb played [8–10].

It had long been assumed that comprehenders either manage to arrive at a full and faithful interpretation after they perform reanalysis, or they fail to build a coherent representation of the sentence altogether. However, as Christianson et al. [11] showed in their study, this might not be the case. They argue that reanalysis can be carried out only partially in some cases, resulting in interpretations that are not perfect but only “good enough” for the given situation. For example, for sentences like (1), the reanalysis might stop after finding a suitable subject for the verb played, but the initial interpretation of the baby as an object of dressed might not be erased, i.e., the initial misinterpretation of the garden path might still linger even after the whole sentence was processed.

These results were replicated in a study by Christianson et al. [12] which focused mostly on differences in processing between younger and older speakers. The tendency of the initial misanalysis to linger was also shown in other studies with various experimental designs such as paraphrasing [13], syntactic priming [14], multiple-choice questions [15], or the picture matching task [6]. A similar effect was also shown for L2 speakers of English [16–19]; and also for L1 speakers of other languages [15, 20].

Slattery et al. [21] suggest a different explanation for the lingering misanalysis. Unlike Christianson et al. [11], who propose that the syntactic structure built during reanalysis might be inaccurate or incomplete, the authors of this study argue that reanalysis does, in fact, lead to the creation of a syntactic structure that is complete, globally coherent and faithful to the input. According to them, reanalysis is incomplete in a different aspect, namely in that comprehenders fail to delete a memory representation built on the basis of the initial misanalysis, and this memory trace later interferes with the new representation of the whole sentence. The authors used the Gender Mismatch Effect to test this hypothesis in an eye-tracking experiment. In sentences like (2), a full and faithful syntactic structure needs to be built for the parser to detect a gender mismatch between the reflexive pronoun (himself) and its antecedent (David’s mother); specifically, David’s mother needs to be interpreted as the subject of the matrix clause.

1. (2) After the bank manager telephoned David’s mother grew worried and gave himself approximately five days to reply.

The experiment has revealed that the Gender Mismatch Effect (in the form of longer first-pass reading times on the reflexive) is present in both the ambiguous and the unambiguous condition. The authors interpret this finding as an evidence that a full and faithful syntactic representation of the garden-path sentence was built. They also argue that the initial misanalysis remains active even after the reanalysis was performed, and that it interferes with new material. This can be seen in sentences like (3), where comprehenders usually slow down on the region himself.

1. (3) While Frank dried off the truck that was dark green was peed on by a stray dog. Frank quickly finished drying himself off then yelled out the window at the dog.

The authors explain this slow-down as an effect of interference caused by the initial misanalysis, which contains information that is contradictory to the information presented in later parts of the text. A similar effect was detected by Sturt [22].

It has been pointed out [20, 23] that Slattery et al. [21] did not use questions to test comprehension and that thus they could not be sure whether the sentences were really misinterpreted. The Gender Mismatch Effect thus could have been caused only by those trials in which the sentences were analyzed faithfully to the input.

The problem of the absence of comprehension questions was addressed by Qian et al. [24], who conducted a conceptual replication of the study by Slattery et al. and arrived at similar conclusions after examining the form of the final interpretation by combining self-paced reading and ERP measuring experiments with comprehension questions.

Another replication was reported by Huang and Ferreira [23], who carried out two experiments that combined online processing measures (eye-tracking, self-paced reading) with comprehension questions targeting the initial misanalysis, using the same stimuli as Slattery et al. [21]. Incorrectly answered trials were analyzed separately from those that were answered correctly. This allowed the authors to link the presence of the Gender Mismatch Effect (which indicates that a proper syntactic structure was built) with the response accuracy (which indicates whether the sentence was interpreted in accordance with the input). The Gender Mismatch Effect was present in both correctly and incorrectly answered trials, and its presence was independent of the grammatical features of other nouns in the sentence, suggesting that the parser does not ignore syntactic constraints while searching for an antecedent. These results are in accordance with the hypothesis of Slattery et al. [21].

Huang and Ferreira [23], however, point out that there is in fact no need for a full syntactic representation of the sentence to be built for the Gender Mismatch Effect to appear. As has been said above, it is crucial that the ambiguous noun (e.g. David’s mother) is analyzed as the subject of the matrix clause for it to serve as the antecedent for the reflexive. It does not, by itself, mean that the previous analysis (with the noun as the object) has to be deleted and that the whole sentence has to be reanalyzed faithfully. The Gender Mismatch Effect would still be present if the reanalysis was not performed until the end and instead stopped after a suitable noun for the subject position was identified. It is thus possible that the parser only builds several locally coherent structures (e.g., the bank manager telephoned David’s mother + David’s mother grew worried). These structures cannot be combined into a single global one without violating the rules of grammar, but each of them can generate its own interpretation for the given part of the sentence. Current data are in accordance with the explanations given by both Slattery et al. [21] and Huang and Ferreira [23].

It should be noted that most of the research on garden-path sentences within the good-enough approach is focused on studying a specific type of local ambiguity: one where the combination of a verb and a noun is initially interpreted as a transitive verb and its direct object, and later has to be reinterpreted as an intransitive or reflexive verb and the subject of the matrix clause, or, in the case of Sturt’s experiment [22], as a transitive verb with a subordinate clause in the position of the direct object.

A different type of garden-path structures has been studied by Christianson and Luke [25], who worked with sentences containing an ambiguous noun phrase coordination like (4).

1. (4) The publisher called up the editor and the author refused to change the book’s ending.

In sentence (4) the whole phrase the editor and the author is interpreted as the direct object of the verb called up. However, as it later turns out, the conjunction and does not connect two noun phrases but two clauses. The participants were once again asked to answer a comprehension question targeting the initial misanalysis (Did the publisher call the editor and the author?). The response accuracy was much higher in this case, around 80%, suggesting that the tendency of the initial misanalysis to linger was less strong than for the structure used in the studies mentioned above, where the response accuracy reached only 50%.

An even lower response accuracy (around 20%) was produced by the participants in a study by Christianson et al. [26]. This study used sentences like (5) with yet another type of ambiguity where a reduced relative clause tossed the ball is at first interpreted as a continuation of the matrix clause (predicate and direct object).

1. (5) The player tossed the ball interfered with the other team.

Another type of temporarily ambiguous structure was examined by Fujita and Cunnings [19]. The authors examined whether the initial misanalysis persists in the case of filled-gap and non-filled-gap sentences, such as (6) and (7), when compared to their unambiguous counterparts, such as (8).

1. (6) Elisa noticed the truck which the policeman watched the car from earlier that morning.

1. (7) Elisa noticed the truck which the policeman watched very quietly from earlier that morning.

1. (8) Elisa noticed the truck from which the policeman watched the car earlier that morning.

Sentences like (6) and (7) contain long-distance dependencies where an argument, in this case the filler ‘the truck’, is dislocated from a post-verb region of a relative clause, the gap (which is located after the preposition from in our case). In order to interpret the sentences correctly, comprehenders need to link the filler and the gap. Since comprehenders tend to assign a filler to a gap as soon as possible, they usually tend to assume the policeman watched the truck. However, this representation is proven incorrect by the presence of the preposition, and in case of sentence (6) also by the presence of the car.

In a series of experiments working both with online and offline comprehension measures, the authors show that the initial misanalysis has a tendency to linger for this type of sentence. When asked about the initial misanalysis, e.g. Did the policeman watch the truck?, the participants tended to answer correctly in approximately 70% of the cases with filled-gap sentences like (6). In cases with non-filled-gap sentences, the response accuracy varied between 30% and 70% depending on the experimental design.

The results of these studies suggest that garden-path structures may vary in difficulty and by extension in the form of the final representation. For some structures, it even seems that a coherent syntactic representation is not formed at all. For example, in the case of certain rather demanding Czech garden-path structures, comprehenders might even give up on trying to form a coherent representation of any kind and instead rely on general inference when making sense of the sentence [20].

This difference can be caused by typological characteristics of the two languages (unlike English, Czech is a language with a relatively free word order and a rich morphology), but it can also be linked to linguistic properties of the given garden-path structures. As Chromý [20] argues, resulting representations of garden-path sentences may vary based on the difficulty of the structure. It may be the case that for some structures, a coherent representation is not formed at all; for some, reanalysis results in a full and coherent syntactic representation that is faithful to the input, but the initial misinterpretation lingers; for others, both the syntactic representation and interpretation may be successfully repaired.

There is one more issue that we would like to point out. The previous studies have typically approached the question of lingering misanalysis by trying to find a single explanation for the effects of garden-path sentences on answering comprehension questions. At least three possibilities were introduced: (1) Reanalysis leads to the creation of several partial syntactic structures that are locally coherent but that cannot be combined into a single grammatical structure. These partial structures generate interpretations for given parts of the sentence. (2) Reanalysis leads to the creation of a globally coherent syntactic structure that is grammatically correct and faithful to the input. At the same time, comprehenders fail to erase the interpretation built during the initial misanalysis. (3) Readers fail to generate a coherent representation of the sentence (at least in the case of more complicated garden-path sentences) and rely on the lexical content of the sentence retained in their memory, general inference and encyclopedic knowledge when answering the questions. The previous studies do not explicitly claim that the resulting representation always looks the same. For example, Christianson et al. [12] or Malyutina and den Ouden [6] explicitly examine individual differences between comprehenders. Nevertheless, the authors still focus on trying to find a prototypical, general, model explanation for what happens when people are garden-pathed (for example the explanations offered by Christianson et al. [11] and Slattery et al. [21] are tested against each other, suggesting they are mutually exclusive).

However, there might be an alternative approach to the issue at hand. Perhaps, a single explanation is too simplistic. The reasons why readers answer comprehension questions incorrectly may differ from participant to participant and also from item to item. Since the current data cannot reliably distinguish between such reasons, it is possible that all of the potential explanations mentioned above are in fact not mutually exclusive. They all can represent different processing outcomes that readers arrive to after encountering garden-path sentences. This cannot be addressed properly if we analyze only the grouped data and if we use solely yes–no questions, because these do not tell us much about the resulting representation. Our approach will overcome this by focusing on the use of a combination of quantitative and qualitative analyses, namely open-ended comprehension questions, to examine the resulting representations of garden-path sentences in their full variety, without limiting the comprehenders’ options by asking them whether they agree or disagree with a specific analysis of a given sentence. This would allow us to treat the resulting representations of garden-path sentences like a scale with a variety of options, ranging from full and faithful representations to incoherent ones, instead of focusing on finding a prototypical explanation for the phenomenon. We could thus see whether a single explanation may account for all (or most of) the incorrect answers that seem to suggest the presence of lingering misanalysis or whether the types of incorrect answers are rather diverse and cannot be explained unitarily.

In the current study, we thus introduce a set of experiments that focus on probing the final representation of Czech garden-path sentences. To do this, we use both yes–no and open-ended questioning and also test the difference between word-by-word self-paced reading (which does not allow for regressions, but informs us about the presence of the garden-path effect during reading) and whole-sentence presentation (which allows for regressions, but does not inform us about the process of reading).

Structure of Czech and the current study

Czech is a highly inflectional language. A single suffix (ending) thus expresses several grammatical categories. Czech nouns typically have one ending which specifies their case (nominative, genitive, dative, accusative, vocative, locative, instrumental), gender (masculine, feminine, or neuter), and number (singular or plural). This leads to rather complex paradigms comprising multiple forms. Importantly, some forms may be syncretic (homonymous). For example, the lemma obchodnice (‘female storekeeper’) can have the form obchodnici, which can be either dative, accusative, or locative, depending on the context. Morphological syncretism of this type may be easily used for creating locally ambiguous (garden-path) sentences. This can be illustrated with sentences such as (9), which we used in our experiments.

1. (9)

The verb prohledal (‘searched’) is transitive, it takes one direct object, and since the following noun, obchodnici (‘a storekeeper’), corresponds with the accusative case (or dative, or locative), it is analyzed as a patient (since this presents the easiest way of integrating the noun in the sentence at that point of processing). In other words, the initial interpretation is that the policeman searched the storekeeper. However, comprehenders later encounter the noun dodávku (‘a van’), which is also in the accusative, but non-syncretic (i.e. not homonymous with another form of the noun). As such, it has to be incorporated into the sentence as the direct object of the verb. But the verb prohledal already has the position of the direct object filled by the noun obchodnici and it cannot take another object. So, the initial analysis has no sufficient attachment site for the newly encountered noun and the sentence has to be reanalyzed, i.e., the noun obchodnici has to be reanalyzed as a dative form (it is not possible to analyze it as a locative form, since in Czech, the locative case is strictly prepositional, i.e., a locative form is never used without a preposition requiring a locative complement).

According to the Attach Anyway principle [7], the parser attaches the noun dodávku to the position of the object of the verb prohledal, even though the result is ungrammatical, and tries to fix the mistake by adjusting the structure. To do this, it needs to detach the noun obchodnici from its current position and find its dative interpretation. In Czech, nouns in the dative form can have the function of an indirect object (e.g. recipient) or of an adjunct. In the case of our stimuli, the dative noun is a benefactive or an external possessor (see Fried [27] for more information on external possessivity in Czech). Thus, the parser can reattach the noun to the structure with a different function. The final, faithful and globally coherent interpretation should be that the policeman searched a van that belonged to the storekeeper. The sentence might be difficult to reanalyze because the disambiguating noun does not offer any information (positive symptoms, in terminology of Fodor [28]) about what needs to be done to fix the structure—there is no cue or symptom that would lead the parser to the dative interpretation of obchodnici, the only information is that dodávku needs an attachment site and that prohledal cannot have two direct objects.

Thus, comprehenders should perform three key steps to interpret the sentence faithfully to the input. They should:

1. Correctly attach the noun dodávku to the existing structure as the object of the verb prohledal.
2. Detach the ambiguous noun, obchodnici, from the object position and deactivate the initial interpretation.
3. Reanalyze the ambiguous noun and reattach it to the structure as a benefactive/possessor.

If the hypothesis of Slattery et al. [21] is correct and a coherent and faithful syntactic structure is built during the reanalysis, comprehenders should be able to perform steps 1 and 3, i.e., to correctly identify the patient and to discover the relationship between the patient and the benefactive/possessor. They should not, however, perform step 2, where they deactivate the initial misinterpretation. Thus, they should arrive at the conclusion that the policeman searched the van and the van belonged to the storekeeper, but the interpretation where the policeman searched the storekeeper should still be active.

We can use simple comprehension questions to test whether comprehenders managed to go through the specific steps during the reanalysis. Namely:

1. a) Did the policeman search the van? for step 1.
2. b) Did the policeman search the storekeeper? for step 2.
3. c) Did the storekeeper have a van? for step 3.

If a comprehender managed to successfully attach the object noun to the structure, their answer to the question a) should be “yes”. If they detached the ambiguous noun from its initial position as an object and deactivated the initial misanalysis, the answer to the question b) should be “no”. And if they managed to successfully reanalyze the ambiguous noun and to discover its dative interpretation, they should answer “yes” to the last question.

The yes–no version of comprehension questions can influence the way comprehenders respond in several ways. All of the questions inherently suggest a specific interpretation, questions a) and b) can reactivate specific analyses of the sentence through similarities in syntactic structure, and there is also the problem of acquiescence bias [29]—participants have a higher tendency to answer “yes” than “no” to comprehension questions. Since all of this can lead to a higher response accuracy for questions a) and c), and a lower response accuracy for question b) for reasons that are unrelated to sentence processing, we decided to also use open-ended (free recall) questions. Questions a) and b) can be substituted with the question What did the policeman do?, and question c) can be changed into Whose was the van?. The use of open-ended questions also allows us to perform a qualitative analysis (along with a quantitative one), which may shed more light on the resulting representations of garden-path sentences.

The type of ambiguous nouns that we chose to create the garden-path sentences also allowed us to create non-ambiguous conditions with a very similar meaning, since they have a masculine counterpart that is not ambiguous in its dative form (the dative form is homonymous with the locative form, but nouns in the locative case are preceded by a preposition). This allowed us to create a control condition that only differs from the experimental condition in the grammatical gender of the potentially ambiguous noun, as in sentence (10).

1. (10)

Since the noun obchodníkovi can only be interpreted as dative in the given context, participants should not be garden-pathed, and there should not be any problems with comprehension caused by temporary ambiguity. There is, however, one more factor that can complicate processing, namely semantics of the disambiguating noun. There are three nouns in every condition, and processing can get more difficult when these nouns are more similar to each other due to encoding interference [30]. Since the ambiguous noun and the object noun already interfere with each other syntactically in the garden-path condition (both can be interpreted as accusative and thus as the direct object of the verb), we decided to control the semantics of the object noun to differentiate between a low interference condition such as (9), where the noun refers to an inanimate object (and thus is semantically less similar to the ambiguous noun), and a high interference condition, where the object noun refers to a person, as in (11). The high interference condition should be more difficult to process, leading to a lower response accuracy in both the garden-path and the control condition.

1. (11)

As our reviewers pointed out, it is also important to note that interference can play a role in information recall. In order to answer the comprehension questions, participants need to reconstruct the previously read sentence and pick the correct response. Doing this might be more difficult when different parts of the sentence are interfering with each other due to semantic or syntactic similarity. This effect might be especially relevant when the initial misanalysis persists. The memory trace it leaves behind can interfere with the new representation, making it more difficult to respond correctly when these two representations are semantically similar to each other. It has also been shown that animate patients are generally considered less acceptable than inanimate ones [31], which might play a role in the decision making process if participants compare the two representations against each other. Since plausibility plays an important role in garden-path sentence processing [5, 6], we can assume that conditions where both representations are similarly plausible or acceptable will be more difficult to interpret correctly than conditions where the correct representation (i.e., the one containing an inanimate patient) is also more acceptable than the initial one.

Overall, this manipulation allows us to create a more diverse set of experimental sentences in terms of difficulty, and lets us explore whether the pattern of responses changes for more problematic sentences.

In sum, we used items consisting of four conditions as shown in Table 1. The same set of 48 items was used in all five experiments.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. The structure of items used in all five experiments with glosses.

https://doi.org/10.1371/journal.pone.0288817.t001

Since the number of verbs that can be plausibly combined with both animate and inanimate patients is limited, we decided to use every verb twice. We also repeated the animate nouns in the disambiguating region for the same reason. The combinations of the verb and the nouns never repeated.

Altogether, five experiments were conducted. Experiment 1 employed an acceptability judgment task and thus aimed to test whether there is indeed a difference between the ambiguous and unambiguous conditions, as well as a difference between the high-interference (animate object) and low-interference (inanimate object) conditions. The four remaining experiments used the self-paced reading task and differed in two important aspects, i.e. presentation mode and questioning type. Experiments 2 and 4 used whole-sentence-at-once presentation self-paced reading, which provided the participants with more natural reading conditions. On the other hand, Experiments 3 and 5 employed the word-by-word presentation mode in order to provide information about reaction times (RTs) on specific regions of the sentence (especially the disambiguating region, i.e. the object noun, where we expected elevated RTs as a result of the garden-path effect in ambiguous sentences). Experiments 2 and 3 used classic yes–no comprehension questions targeting various aspects of the resulting representation, whereas Experiments 4 and 5 employed two open-ended questions, which allowed for a more qualitative analysis of what the resulting representation looks like.

Data accessibility

Experimental items, data and R scripts used for their analysis are accessible in the following repository: https://osf.io/98g26/.

Ethics approval

Research ethics board approval for all five experiments was acquired from the Research Ethics Committee of the Faculty of Arts, Charles University (Ref. No.: UKFF/160211/2022–1-1). Participation in the experiments was voluntary, and all participants provided written informed consent. All data used in the current analyses were fully anonymized.

Experiment 1

As Experiment 1, we ran a pilot study using the acceptability judgment task. Since the structure we worked with has not, to our knowledge, been used in any previous work, and the experimental items were created specifically for the purposes of our experiments, we needed to ensure that the unambiguous versions of the sentences will be considered unproblematic by the readers. If this was not the case, we could not confidently say whether any potential effects on processing and comprehension detected in the next four experiments could be attributed to the garden-path phenomenon, or whether they resulted from the complex nature of the stimuli. We also aimed to see whether we would find the expected difference in acceptability between garden-path and non-garden-path sentences and also between sentences with animate and inanimate objects.

Results

The average scores for each condition together with 95% confidence intervals are shown in Table 2.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Acceptability judgments for the experimental items used in this study together with raw mean reaction times in ms (with 95% confidence intervals) for the four conditions in Experiments 2 and 4.

https://doi.org/10.1371/journal.pone.0288817.t002

The linear-mixed effects model yielded three significant effects: sentence type (β = −0.733, SE = 0.063, t = −11.559, p < 0.001), object animacy (β = −0.695, SE = 0.059, t = −11.678, p < 0.001) and the interaction between sentence type and animacy (β = −0.245, SE = 0.073, t = −3.3357, p < 0.001).

Experiment 2

The aim of this experiment was to test whether comprehenders form a full and coherent global syntactic representation of garden-path sentences and whether the initial misinterpretation lingers even after reanalysis is performed due to comprehenders’ failure to deactivate it in their memory.

We used intraclausal garden-path sentences with patient/benefactive ambiguity, such as the one in (9). We used three yes–no comprehension questions to probe whether the participants performed all steps required to carry out a successful reanalysis of the sentence and to reach a faithful final interpretation, namely a) the correct attachment of the disambiguating noun as the patient, b) detaching the ambiguous noun from the object position and deactivating the interpretation based on this syntactic structure, and c) successful reanalysis of the ambiguous noun and reattaching it as a benefactive/external possessor.

The experiment used a self-paced reading paradigm. Sentences were presented at once as a whole to provide the participants with a more natural reading environment. Since rereading previously encountered material is a crucial part of reanalysis processes for many comprehenders, we considered it important not to limit the participants in their ability to perform regressions. Each sentence was followed by a single comprehension question and three options for answer (yes, no, I don’t know). The “I don’t know” responses were always counted as incorrect and the option was used mainly to allow the readers not to guess the right answer in case they did not know it.

The experiment was preregistered on the Open Science Framework: https://osf.io/vbcp2.

Method

Participants.

One hundred seventy four Charles University undergraduate students (147 female and 27 male; mean age = 22.19 years) participated in Experiment 2. All participants were native speakers of Czech and participated for course credit.

Materials.

Forty-eight experimental items were used in Experiment 2 (see the example of an item in Table 1).

The experiment had a 2x2x3 within-subject design. We manipulated the ambiguity of the sentences by changing the grammatical gender of the nouns in question (feminine vs. masculine). We also manipulated the difficulty of the sentences by changing the animacy of the disambiguating noun (animate vs. inanimate), with animate nouns being semantically more similar to the potentially ambiguous noun (which always referred to a person), thus leading to a stronger interference effect.

We used three yes–no comprehension questions to probe the final representation (see Table 3). The first one focused on the correct attachment of the disambiguating noun (qcor), the second one on the initial misinterpretation (qmis) and the third one on the correct interpretation of the ambiguous noun (qpos). The correct answer to qcor and qpos was always “yes”, the correct answer to qmis was always “no”. The number of positive and negative responses was balanced through the experiment.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Yes–no comprehension questions used in Experiments 2 and 3.

https://doi.org/10.1371/journal.pone.0288817.t003

Ninety-six filler items and three practice items were used. Altogether, each participant read 147 sentences. Twenty-eight of the filler sentences were ungrammatical. In eighteen of them, there was a missing or a redundant constituent to mirror the situation of an unfaithful garden-path analysis (if the reanalysis fails, there are two constituents with the same grammatical function in the representation of the sentence and there is no possessor). The remaining ungrammatical fillers were constructed in a different way (in five of them, there was a missing preposition, and in another five, there was an agreement violation). Each filler was followed by a comprehension question. None of the questions targeted a region influenced by the ungrammaticality. For forty of the filler sentences, the correct answer to the comprehension question was “yes”, and for the rest, the correct answer was “no” (to balance the overall number of positive/negative responses). For the ungrammatical sentences, the number was also balanced.

Procedure.

The experiment was web-based and conducted on the IbexFarm platform [32].

At the beginning, participants were informed about the general experimental procedure and they were asked to fill in their gender, age, native language, field of study, and whether they experienced any reading problems such as dyslexia. In the actual experiment, participants first saw a sentence as a series of underscores. Once they pressed the space bar, the sentence appeared. Their task was to read the sentence at their normal reading pace, and once they were finished, they had to press the space bar again, which caused the sentence to disappear and a yes–no comprehension question to appear. Participants responded by clicking the mouse. Once they responded, the next sentence appeared as a series of underscores.

Each participant saw only one of the conditions of each item and the conditions were distributed based on the Latin-square design. Altogether, each participant was exposed to four cases of each condition. The order of sentences was fully randomized for each participant. The experiment took about 20–25 minutes.

Data analysis.

First, participants’ accuracy for filler items was calculated. Only those participants with accuracy scores of 75% or higher were included in subsequent analyses. Fifteen participants scored below 75% and were excluded (these are already not included in the number of participants reported in Participants subsection).

Second, reaction times were trimmed. We excluded RTs lower than 1000 ms, the remaining RTs were log-transformed and the upper cut-off point was set as 2.5 standard deviations from the mean, i.e. 9.928 log(ms) which corresponded to 20 491.67 ms. This resulted in 2.48% of the RTs being excluded.

Third, we used linear-mixed effects regression with the lme4 package [33] to examine the differences in RTs between the conditions. The degrees of freedom and p-values were estimated using Satterthwaite’s approximations from the lmerTest package [34]. As the dependent variable, we used the log(RTs). As fixed effects, sentence type and object animacy were used (in interaction). Both effects were coded using sum contrasts. For sentence type, GP sentence was coded as 0.5 and non-GP sentence as −0.5. For object animacy, animate object was coded as 0.5 and inanimate object as −0.5. Following the approach by Bates et al. [35], we used the following random effects structure: the intercept for participants and items, the by-participants random slope for sentence type and a by-item random slope for sentence type. For this analysis, we report the beta estimates, standard errors (SEs), t-values, and p-values.

Fourth, response accuracy was analyzed using logit mixed-effects models [36] with question type as a fixed effect, sentence type and object animacy as nested effects within question type (in interaction), and intercept for participant and item as random effects (with question type as a random slope both for items and for participants) which again followed recommendations by Bates et al. [35]. This allowed us to directly compare general response accuracy for the three question types and also to examine the animacy and sentence type effects for each question type separately. Question type was coded using Helmert contrasts [37] and the contrast matrix is shown in Table 4. Contrast 1 targeted the difference between questions targeting the real object (qcor) and the question targeting the possessive relationship (qpos). Contrast 2 targeted the difference between the mean of the two questions (qcor and qpos) and the question targeting the initial misanalysis (qmis). Sentence type and object animacy were coded using sum contrasts (as in the RTs analysis). For this analysis, the beta estimates, standard errors (SEs), z-scores, and p-values are reported (only for significant results).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 4. Helmert contrasts matrix for the question type variable used in the analysis of response accuracy in Experiments 2 and 3.

https://doi.org/10.1371/journal.pone.0288817.t004

Fifth, an analysis of the incorrect response types was conducted. This analysis was supplementary to the analysis of response accuracy and due to the generally low number of incorrect responses, it is rather descriptive and qualitative.

Results

Reaction times.

Raw reaction times for each condition are presented in Table 2. The linear mixed-effects model [33] showed significant effects for sentence type (β = 0.062, SE = 0.016, t = 5.345, p < 0.001), object animacy (β = 0.069, SE = 0.008, t = 8.469, p < 0.001) and also for the interaction between the fixed effects (β = 0.05, SE = 0.016, t = 3.032, p < 0.01).

Response accuracy.

The overall response accuracy for participants was relatively high. Participants’ accuracy scores ranged from 44.68% to 100%, the average accuracy being 88.32% (median 89.58%). The between-item variability was 73.41% to 96.51% (mean 88.39%, median 88.19%).

Table 5 and Fig 1 show the response accuracy for each condition. The logit mixed-effects model revealed several significant effects. For question type, there was a significant effect of Contrast 2 (β = −0.237, SE = 0.039, z = −6.019, p < 0.001), meaning that the response accuracy for questions targeting the initial misanalysis was significantly lower than the mean accuracy for questions targeting the true object and questions targeting the possessive relationship. For questions targeting the true object, the model yielded significant effects of sentence type (β = −0.572, SE = 0.144, z = −3.968, p < 0.001) and object animacy (β = −0.621, SE = 0.144, z = −4.315, p < 0.001). For questions targeting the initial misanalysis, we found only a significant effect of sentence type (β = −1.342, SE = 0.119, z = −11.25, p < 0.001). And for questions targeting the possessive relationship, we found significant effects of object animacy (β = −0.766, SE = 0.144, z = −5.329, p < 0.001) and the interaction between sentence type and object animacy (β = −0.795, SE = 0.287, z = −2.772, p < 0.01).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 5. Response accuracy in Experiments 2 and 3 which employed yes–no questions.

https://doi.org/10.1371/journal.pone.0288817.t005

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Response accuracy for the three question types used in Experiment 2 with 95% confidence intervals.

qcor = question targeting the real object; qmis = question targeting the initial misanalysis; qpos = question targeting the possessive relationship.

https://doi.org/10.1371/journal.pone.0288817.g001

Experiment 3

Experiment 3 was a replication of Experiment 2 with one important difference. The stimuli were not presented in a sentence-at-once mode, but using a classic word-by-word presentation. Such presentation provide us with important information about the course of processing, which will be crucial for the interpretation of our results. Measuring RTs on specific regions of the sentence can also give us information about additional difficulty in processing that relates to the changes of animacy of the disambiguating noun.

The experiment was preregistered on the Open Science Framework: https://osf.io/c5wby.

Results

Reaction times.

Fig 2 shows RTs in ms for each sentence region used in Experiment 3. The analysis targeted the disambiguating region 7 (i.e., the real object) and the two following regions (i.e., 8 and 9). For all three regions sentence type was used as a random slope for participants. For regions 7 and 9, object animacy was included as a random slope for items and for region 8, it was used as a random slope for participants together with sentence type (without interaction). The pattern of results was practically identical for these regions. For region 7, the model yielded separate significant effects for sentence type (β = 0.035, SE = 0.008, t = 4.256, p < 0.001) and object animacy (β = 0.028, SE = 0.012, t = 2.342, p < 0.05), but no interaction effect. The same results were documented for region 8: sentence type (β = 0.075, SE = 0.008, t = 9.817, p < 0.001), object animacy (β = 0.039, SE = 0.006, t = 6.048, p < 0.001), and also for region 9: sentence type (β = 0.047, SE = 0.007, t = 7.19, p < 0.001), object animacy (β = 0.022, SE = 0.007, t = 2.989, p < 0.01).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Reaction times in ms for experimental sentences used in Experiment 3 together with their 95% confidence intervals.

https://doi.org/10.1371/journal.pone.0288817.g002

The secondary analysis targeted the ambiguous region 4 and the two following regions (i.e. 5, and 6). For these regions, no significant effect of sentence type was found, meaning that the unambiguous dative and ambiguous dative/accusative nouns did not differ in their processing.

Response accuracy.

The overall response accuracy for participants was again relatively high. Participants’ accuracy scores ranged from 52.08% to 100%, the average accuracy being 86.36% (median 89.58%). The between-item variability was 77.22% to 92.36% (mean 86.37%, median 87.34%).

Table 5 and Fig 3 show the response accuracy for each condition. The logit mixed-effects model again revealed several significant effects. As in Experiment 2, there was a significant effect of Contrast 2 for question type (β = −0.293, SE = 0.041, z = −7.126, p < 0.001). In other words, response accuracy for questions targeting the initial misanalysis was significantly lower than the mean accuracy for questions targeting the true object and questions targeting the possessive relationship. For questions targeting the true object, the model yielded significant effects of sentence type (β = −0.622, SE = 0.147, z = −4.23, p < 0.001) and object animacy (β = −0.783, SE = 0.148, z = −5.308, p < 0.001). For questions targeting the initial misanalysis, we found significant effects of sentence type (β = −1.489, SE = 0.118, z = −12.603, p < 0.001) and of the interaction between sentence type and object animacy (β = −0.758, SE = 0.23, z = −3.291, p < 0.001). And for questions targeting the possessive relationship, we found a significant effect of object animacy (β = −0.447, SE = 0.137, z = −3.264, p < 0.01).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Response accuracy for the three question types used in Experiment 3 with 95% confidence intervals.

qcor = question targeting the real object; qmis = question targeting the initial misanalysis; qpos = question targeting the possessive relationship.

https://doi.org/10.1371/journal.pone.0288817.g003

Experiment 4

The aim of Experiment 4 was to use open-ended questions to probe the final representation of given garden-path sentences. Experiment 4 is a replication of Experiment 2. It employs the same experimental design (self-paced reading with sentences presented at once) and the same set of stimuli, but instead of three yes–no comprehension questions, we used two open-ended questions targeting the regions of interest in the sentence.

The first question, qwhat, tested which of the two nouns the comprehenders chose to attach to the verb as its object, e.g., What did the policeman do? Because the verb in the target sentence is transitive, the response needs to contain an object in the accusative to be grammatical and informative. The response can provide us with information about whether participants stayed committed to the initial interpretation, or whether they correctly attached the disambiguating noun into this position (or whether they maintained both analyses at the same time). The second question, qwhose, targeted the faithful (dative) interpretation of the ambiguous noun via its relation to the disambiguating noun, e.g., Whose was the van?. The responses were manually coded for the correctness (yes/no) and type of error.

The experiment was preregistered on the Open Science Framework: https://osf.io/km3ga.

Results

Reaction times.

Raw reaction times for each condition are presented in Table 2. The linear mixed-effects model [33] showed significant effects for sentence type (β = 0.049, SE = 0.012, t = 4.032, p < 0.001), object animacy (β = 0.069, SE = 0.014, t = 4.897, p < 0.001). In other words, an identical pattern of results has been found as in Experiment 2, except the interaction between the fixed effects has not reached significance this time.

Response accuracy.

The overall response accuracy for participants was lower than in both of the experiments using yes–no questions. Participants’ accuracy scores ranged from 22.92% to 100%, the average accuracy being 80.09% (median 83.33%). The between-item variability was 65.92% to 90.5% (mean 80.09%, median 79.61%).

Table 6 and Fig 4 show the response accuracy for each condition. The logit mixed-effects model revealed several significant effects. There was a significant effect of question type (β = 0.206, SE = 0.077, z = -2.695, p < 0.01). For questions targeting the initial misanalysis, the model yielded significant effects of sentence type (β = −0.702, SE = 0.087, z = −8.07, p < 0.001) and object animacy (β = −0.654, SE = 0.087, z = −7.518, p < 0.001). And for questions targeting the possessive relationship, we found significant effects of object animacy (β = −0.661, SE = 0.082, z = −8.081, p < 0.001) and the interaction between sentence type and object animacy (β = −0.393, SE = 0.163, z = −2.419, p < 0.05).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 6. Response accuracy in Experiments 4 and 5, which employed open-ended questions.

https://doi.org/10.1371/journal.pone.0288817.t006

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Response accuracy for the two question types used in Experiment 4 with 95% confidence intervals.

qwhat = question targeting the initial misanalysis; qwhose = question targeting the possessive relationship.

https://doi.org/10.1371/journal.pone.0288817.g004

Incorrect responses.

The frequency of the four different types of incorrect answers is shown in Table 7. It can be seen that the Different type (the case in which the participants seemingly answered a different question) occurred only rarely. The three other types were more frequent. For the question targeting the initial misanalysis, qwhat, the biggest number of incorrect responses belonged to the Substitution category and to the “Other” category. For the question targeting the possessive interpretation of the ambiguous noun, the most frequently represented category was “I don’t know”, followed by “Substitution”. This pattern was similar for all four conditions.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 7. Frequency of the four types of incorrect responses in Experiment 4.

https://doi.org/10.1371/journal.pone.0288817.t007

We also performed a more detailed analysis of the types of responses within the Substitution category. The number of responses that were consistent with the initial misanalysis (i.e., when the question What did the policeman do? elicited responses like he searched the storekeeper, he searched the storekeeper or the van, and he searched the storekeeper and the van) can be seen in Table 8.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 8. Frequency of the GP error type in response to qwhat in Experiments 4 and 5.

https://doi.org/10.1371/journal.pone.0288817.t008

As we can see, garden-path sentences clearly elicited more responses that were consistent with the potential initial misanalysis than non-garden-path sentences. The given type of error also formed more than 50% of the Substitution category for both the high- and low-interference garden-path sentences.

Experiment 5

Experiment 5 is a replication of Experiment 4 using the word-by-word presentation mode, which allows us to investigate RTs on specific regions of interest. The results can also be compared to Experiment 3, which used the same paradigm but with yes–no comprehension questions. The experiment was preregistered on OSF: https://osf.io/hcf28.

Results

Reaction times.

Fig 5 shows RTs in ms for each sentence region used in Experiment 5. The analysis targeted the disambiguating region 7 (i.e. the real object) and the two following regions (i.e. 8 and 9). For region 7, sentence type was used as a random slope for participants and object animacy was used as a random slope for items. For regions 8 and 9, sentence type was used as a random slope for participants together with object animacy and for items, sentence type was used as a sole random slope. For region 7, we found separate significant effects for sentence type (β = 0.053, SE = 0.01, t = 5.162, p < 0.001) and object animacy (β = 0.032, SE = 0.013, t = 2.54, p < 0.05), but no interaction effect. For region 8, the model yielded a similar pattern of results with two independent significant effects: sentence type (β = 0.077, SE = 0.011, t = 7.024, p < 0.001), object animacy (β = 0.065, SE = 0.008, t = 7.894, p < 0.001). For region 9, two independent significant effects were also found: sentence type (β = 0.043, SE = 0.008, t = 5.425, p < 0.001), object animacy (β = 0.028, SE = 0.007, t = 3.702, p < 0.001).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. Reaction times in ms for experimental sentences used in Experiment 5 together with their 95% confidence intervals.

https://doi.org/10.1371/journal.pone.0288817.g005

Response accuracy.

The overall response accuracy for participants was similar to Experiment 5. The participants’ accuracy scores ranged from 33.3% to 100%, the average accuracy being 76.26% (median 79.17%). The between-item variability was 48.57% to 87.86% (mean 76.36%, median 76.95%).

Table 6 and Fig 6 show the response accuracy for each condition. The logit mixed-effects model revealed several significant effects. There was a significant effect of question type (β = 0.353, SE = 0.117, z = 3.015, p < 0.01). For questions targeting the initial misanalysis, the model yielded significant effects of sentence type (β = −0.796, SE = 0.096, z = −8.311, p < 0.001) and object animacy (β = −0.738, SE = 0.096, z = −7.713, p < 0.001). And for questions targeting the possessive relationship, we found a sole significant effect of object animacy (β = −0.754, SE = 0.088, z = −8.607, p < 0.001).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 6. Response accuracy for the two question types used in Experiment 5 with 95% confidence intervals.

qmis = question targeting the initial misanalysis; qpos = question targeting the possessive relationship.

https://doi.org/10.1371/journal.pone.0288817.g006

Incorrect responses.

The frequency of the four different types of incorrect answers is shown in Table 9. The distribution of the incorrect responses was similar to the previous experiment. The “Different” category occurred only twice. The rest of the categories occurred more frequently. For the question targeting the initial misanalysis, most responses belonged to the “Substitution” and “Other” category. For the question targeting the possessive interpretation of the ambiguous noun, the “I don’t know” and “Substitution” categories were the most represented. This pattern was similar for all four conditions.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 9. Frequency of the four types of incorrect responses in Experiment 5.

https://doi.org/10.1371/journal.pone.0288817.t009

Further analysis of the Substitution category showed similar results as in Experiment 4. The number of responses that were consistent with the initial misanalysis of garden-path sentences is captured in Table 8.

Once again, garden-path sentences led to visibly more responses consistent with the potential initial misanalysis than non-garden-path sentences. Just as in Experiment 4, the GP type of error was responsible for more than 50% of the responses within the Substitution category for both the high- and low-interference garden-path sentences.

General discussion

In this paper, we aimed to examine various aspects of the resulting representation of garden-path sentences in detail. We tested whether comprehenders often create full and faithful syntactic representations of garden-path sentences, even though they fail to delete the initial misanalysis from memory, as assumed by Slattery et el. [21]; or whether resulting representations tend to be more sparse. For example, speakers may form several locally licit syntactic representations [23] or not form a coherent representation of the sentence at all and rely on general inferences instead [20].

To do that, we examined Czech garden-path sentences which contained a local dative-accusative ambiguity of a feminine noun following a transitive verb (e.g., Ostražitý policista prohledal obchodnici před prodejnou dodávku/klientku a odjel na služebnu.). We conducted four experiments which differed (a) in their mode of presentation (two used the classic moving-window, word-by-word self-paced reading presentation, two presented the sentences as a whole at once, which allowed the participants to read naturally), and (b) in the type of comprehension questions asked (two experiments used yes–no questions, and two used open-ended, free recall questions). By doing this, we were able to probe whether participants managed to perform all steps that were necessary for creating a faithful and coherent representation of the sentence. If what Slattery et al. [21] assume is correct, participants should be able to correctly attach the disambiguating noun to the syntactic structure in most cases, and they should also be able to interpret the ambiguous noun as dative, i.e., benefactive or external possessor. They should not, however, inhibit the initial misanalysis in their memory.

Our results are largely consistent with the accounts offered both by Slattery and colleagues [21] and by Christianson and colleagues [11]. We found strong evidence suggesting that the initial misanalysis was still active even after participants finished reading the sentence. Thus, we replicated the results of Christianson et al. [11] and various further studies targeting the lingering initial misanalysis. Our results also showed that participants often managed to interpret the rest of the sentence faithfully to the input (they successfully attached the disambiguating noun to the structure and they also reanalyzed the ambiguous noun). However, the evidence of the participants forming a complete and coherent representation was mixed.

In Experiments 2 and 3, which employed yes–no questions, we documented a clear difference in response accuracy for questions targeting the real object between garden-path and non-garden-path sentences. Participants thus often did not form a full and faithful representation after reading garden-path sentences. Needless to say, response accuracy for questions targeting the correct interpretation of the real object was significantly higher than for questions targeting the initial misanalysis. This points out that the problem with forming a full and faithful representation (in the sense of attaching the real object to the transitive verb) occurs to a considerably lower degree than the problem with erasing the initial misanalysis from memory.

Experiments 3 and 4, which employed open-ended questions targeting (a) the grammatically correct and faithful interpretation of the patient and (b) the possessive relationship between the ambiguous noun and the object noun, allowed us a deeper look into the aspects of the resulting representation of garden-path sentences. From a quantitative perspective, we found a clear difference for questions targeting the correct interpretation of the sentence. Participants made more errors while answering these questions after reading garden-path sentences than after reading control sentences. Moreover, the majority of the incorrect responses for garden-path sentences pointed towards lingering misanalysis. However, the qualitative analysis showed that part of the incorrect responses for garden-path sentences occurred due to a failure in forming a complete and coherent analysis of the garden-path sentences.

We may thus say that the general pattern of results somehow conforms to the claims by Slattery et al. [21]. However, our results also bring evidence that the processing of garden-path sentences may often lead to different ends than to forming a full and faithful interpretation of the sentence input.

The findings from the first two experiments show that when participants reach the disambiguating noun, they may not incorporate it into the structure they have built so far and instead choose to stick with the initial misanalysis of the first part of the sentence, while not building a coherent representation of the rest of the clause. This does not seem to happen too often (in comparison to the lingering misanalysis), but it is still well attested in the data.

In other cases, comprehenders may also form a “merged” representation (where both the ambiguous and the disambiguating noun are interpreted as a patient) or they may remain unsure about the meaning of the sentence and choose not to commit to a specific interpretation. In other words, when deciding between the two possible interpretations (the policeman searched the storekeeper vs. the policeman searched the van), participants sometimes either accept both, or neither. We found multiple responses where the participants directly stated both the ambiguous and the disambiguating noun as the object of the verb (e.g., prohledal obchodnici a dodávku ‘he searched the storekeeper and the van’) and where they directly stated they were not sure about the correct interpretation (e.g., prohledal obchodnici nebo dodávku ‘he searched the storekeeper or the van’).

Slattery and colleagues [21] do not claim that reanalysis of GP sentences is always successful. However, they decide to explore the cases where the correct syntactic structure is built during reanalysis. As was mentioned previously, our results are largely consistent with Slattery’s account (in many cases, the disambiguating region and the ambiguous region are interpreted correctly, while the initial misanalysis of the ambiguous region still persists). Nevertheless, the number of trials, where the reanalysis failed and a proper syntactic structure was not built, is significant in our results, and these cases occur reliably and repeatedly across all experiments, regardless of their design. The results from the yes–no comprehension questions, and especially from the open-ended questions, show that the pattern of responses is very diverse. Representations built during processing of the sentences or when answering comprehension questions can be incoherent or partially disrupted in multiple different ways. The accounts offered by Christianson et al. [11], Slattery et al. [21] or Huang and Ferreira [23] are all consistent with some portion of our data. However, the big picture seems to be much more diverse than that. We thus want to advocate for an approach that would focus on reanalysis processes and their results in all their diversity.

Another interesting finding of our study is related to the semantics of the disambiguating region. Sentences with animate patients generated more incorrect responses regardless of the presence of temporary ambiguity. We also found an interaction of object animacy with ambiguity, showing that sentences with animate objects further complicated the processing of garden path sentences and lowered the chance of correctly answering the comprehension question targeting the initial misanalysis.

We assume that this is mostly an effect of encoding [30] and retrieval [9] interference. When the two nouns are both animate and thus semantically more similar, it makes the sentence more difficult to process and it also makes it more difficult to reconstruct correctly when answering the comprehension questions. The effects of encoding interference are observable in the on-line measures, in the form of elevated reading times on the second noun and the following regions. The effect could also be potentially caused by the fact that animate patients are less acceptable or more difficult to process in general [31], making the processing of the whole sentence more difficult for the participants, and making them less willing to accept animate nouns as patients. However, we saw that unambiguous sentences with animate patients were still judged as acceptable by the participants (even though they were less acceptable than sentences with inanimate patients), so we can assume that the animacy should not be as problematic as to make the participants unwilling to interpret the noun as a plausible patient (i.e., assume that a policeman was searching a client). The cause of the low response accuracy should not lie in the animacy of the object itself.

It is also possible that the potential cause for the decrease in response accuracy lies in the nature of the possessive relationship, since the possessive relationship between two animate entities (the storekeeper’s client) is qualitatively different from the relationship between an animate and inanimate entity (the storekeeper’s van), e.g. in terms of alienability [27]. The fact that the nature of the possession might be crucial for answering the questions is supported by the detected significant effects of animacy for questions targeting the possessive relationship. The most likely explanation for the animacy effect on response accuracy is similarity-based interference, however, retrieval of the correct answer to this question should not be influenced by potential interference of the two nouns, since one of the nouns is not involved in the search at all (the possessum is already mentioned in the question and the participants should thus only search for the possessor). It is possible that it is simply more difficult or less plausible to construe a possessive relationship between two animate entities. However, it is important to mention that unambiguous sentences with animate patients still received relatively good acceptability ratings in our study. We can thus assume that the possessive relationship itself was probably not considered too problematic by the participants, and the detected effect of the object animacy was at least to some extent caused by interference, as assumed.

The fact that participants struggle more with finding a coherent and faithful interpretation of garden-path sentences when the two potential object nouns are semantically closer can be explained at least in two ways. First, it is possible that the similarity between the ambiguous noun and the disambiguating noun leads to encoding interference [30]. And since the higher level of interference makes the processing of the sentence more difficult (as documented by higher RTs and lower acceptability ratings), the cognitive load the comprehenders have to face in such a case is higher, and in order to compensate for it, they have to rely more strongly on heuristics while processing the sentence [40]. It has been pointed out multiple times that one such outcome of heuristic processing is the tendency to not inhibit the initial misanalysis of garden-path sentences (and, potentially, not performing a complete reanalysis of these sentences in general). Thus, higher level of interference might cause the initial misanalysis to linger more often, exactly as we have seen.

Another explanation relates to decision making processes during reanalysis or when answering the comprehension questions. As the responses to open-ended questions showed us, there are cases where comprehenders decide between the two possible interpretations of GP sentences (‘the policeman searched the storekeeper’ vs. ‘the policeman searched the van’) and sometimes remain unsure even after reading the whole sentence. The plausibility, frequency and general acceptability of these two options might play a role in their decision making. When these two local interpretations are more similar, it is more difficult for the comprehenders to choose one of them over the other, because they are both equally supported. And thus, the number of instances where comprehenders remain undecided, create merged representations or simply give up on trying to interpret the sentence might be higher. Since we failed to detect a significant interaction of ambiguity and animacy on the disambiguating region or on the following regions, we may assume that animacy of the patient does not directly influence the reanalysis and rather comes into play as a post-processing effect, most likely when answering comprehension questions.

The results of the present study also raise another very interesting question, namely the individual differences in garden-path processing. It has already been shown that comprehenders differ tremendously in what (if any) re-reading strategies they use during reanalysis of GP sentences [41, 42]. In each of the four self-paced reading experiments, we observed participants with a response accuracy of 100% or nearly perfect, but also those who scored much worse. Moreover, participants differed considerably in the magnitude of garden-path effects measured in the difference in reaction times. Unfortunately, the present study was not designed to examine the individual differences in detail. Each participant received only several examples of each condition which prohibits any meaningful analysis of the individual differences. Nevertheless, individual differences in garden-path processing is a potentially fruitful research topic since it seems that native speakers of a language may differ largely in how they process garden-path sentences.

Before we conclude, we should address three more, rather methodological aspects of the study. The first one is concerned with the different presentation modes. In Experiments 2 and 4, we used the whole-sentence-at-once presentation, whereas in Experiments 3 and 5, the traditional moving-window word-by-word self-paced reading mode was used. We found a slight decrease in response accuracy for the word-by-word presentation mode, which can be attributed to the fact that the participants were not able to re-read the material. The decrease mirrors a similar finding by Huang and Ferreira [23], who compared the results from a word-by-word SPR task with the results from an eye-tracking experiment (which also allows the participants to perform regressions). Nevertheless, despite this decrease in accuracy, the general pattern of results stayed the same independently of the presentation mode. In general, this suggests that the traditional moving-window paradigm is well suited for such recall experiments and can also provide researchers with information about the reading process.

The second methodological aspect which needs to be discussed is the use of comprehension questions in order to examine resulting representations. When answering the comprehension questions, participants are forced to recall information, reconstruct the sentences and settle on a single, explicit interpretation. In some cases, such responses might, to a certain extent, differ from the representation created directly during reading. Comprehension questions cannot decisively show whether the comprehenders create a full and faithful representation when processing sentences. Nevertheless, such questions (immediately presented after reading a sentence) offer us an important means to unveil the resulting representations readers create while reading sentences. They provide us with information which cannot be obtained otherwise. As Ferreira and Yang [43] have recently shown, the study of language comprehension should benefit both from the online measures (such as reaction times) and offline measures (such as comprehension questions).

We have seen that, in general, open-ended questions yielded a much higher rate of incorrect answers. This may be caused by two reasons. First, while answering yes–no questions, participants may guess the answer with a 50% probability of success. Second, yes–no questions are typically somehow leading, because they often have to contain cues to the sentence content. For example, we used questions such as Did the policemen search the client?, which practically reactivate the subject, the object and the verb of the sentence. Similarly, the classical study by Christianson et al. [11] and many others targeted the resulting representation of sentences such as While Anna dressed the baby played in the crib using questions such as Did Anna dress the baby?, which also activated three parts of the targeted sentence. This is especially clear in comparison to open-ended questions such as What did the policeman do?, which we used in the current study, or even more general questions such as What happened? [44].

The influence of question format is especially clear if we compare the results of yes–no and open-ended questions targeting the possessive relationship. In Experiments 2 and 3, which used yes–no questions such as Did the storekeeper have a van?, which contained the true possessor and the true object, and the correct answer was thus “yes”, the general response accuracy was relatively high. However, in Experiments 4 and 5, which used general questions such as Whose was the van?, the response accuracy was rather low and in some conditions even very similar to the response accuracy for the questions targeting the initial misanalysis. This clearly shows that the presence of cues and a high probability to correctly guess the answer may skew response accuracy to a considerable extent. This is problematic, especially in relation to questions like Did the storekeeper have a van? which are often very trivial and easy to answer correctly without any knowledge of the previous sentence. Such questions are in fact not very informative when used to probe sentence representations, because they allow comprehenders to rely on encyclopedic knowledge and inferences.

One possible problem related to the use of open-ended questions relates to the coding and interpretation of the results. We developed simple coding rules to decide which responses were correct and which were not, and for the latter, what type of error they contain. However, this might be done in different ways. Moreover, the problem is that responses to open-ended questions do not give us all information needed to assess the resulting representation of the given participant, since they are rather concise. For example, if someone responds to the question What did the policeman do? by saying He searched the storekeeper, it is a clear case of a lingering misanalysis, but it does not inform us about the rest of the representation (it might be the case that the given participant formed an otherwise full and complete representation, but also, this could be a case of an incomplete representation). We believe that using open-ended questions is generally more telling than employing traditional yes–no questions, especially since it enables a deeper, qualitative analysis of the data. However, it also has its disadvantages.

As one of the anonymous reviewers pointed out, another methodological problem of the present research is the relatively small number of examples of each condition for each participant (four in Experiments 2 and 3, and six in Experiments 4 and 5). We are fully aware of possible power problems this might present to the analysis. Nevertheless, we believe this does not present a serious problem for the present findings. Crucially, all four self-paced reading experiments used substantial sample sizes of more than 140 participants each and yielded the same pattern of results. Moreover, the creation of experimental items was already highly constrained due to grammatical factors. As we mention in the section Structure of Czech and the Current Study, the number of verbs that can be plausibly combined with both animate and inanimate patients is limited in Czech, and thus, we had to use each verb and disambiguating noun twice to get the current number of experimental items. To get even more items and thus raise the statistical power, we would also need to repeat the ambiguous nouns since they come from a feminine declension class which contains a relatively closed set of nouns denoting people (i.e., the růže-paradigm). Finally and importantly, using a substantially higher number of examples of each condition per participant might lead to unfortunate adaptation effects [45, 46]—in a way, the experiment would then train the participants to process the garden-path sentences used in the course of the experiment, which could skew both the reaction times and response accuracy.

Conclusion

Overall, the results of our experiments suggest that the resulting representations of garden-path sentences may be relatively diverse. The evidence from open-ended questions shows that the form of the final representation varies considerably. In the majority of the trials, comprehenders seem to be able to interpret the garden-path sentences faithfully to the input and delete the initial misinterpretation from memory. In approximately one third of the trials, they succeed in reanalyzing the sentence and manage to attach the disambiguating noun into the structure and reinterpret the ambiguous noun, but they fail to inhibit the initial misinterpretation. But there is also a significant, albeit smaller, number of trials where comprehenders only construct several local interpretations of the sentence. Sometimes they attempt to merge these local structures together even though the result is not consistent with the input, sometimes they remain unsure about the final interpretation and never build a coherent global representation of the sentence at all, deciding to maintain both of these unconnected local interpretations instead. And sometimes the reanalysis simply fails, and comprehenders only maintain the initial misinterpretation (with the ambiguous noun interpreted as an object) while giving up on the rest of the sentence, or they give up on trying to interpret the sentence altogether.

Supporting information

Acknowledgments

We would like to express our sincere gratitude to Jakub Sláma and James Brand for their invaluable help in proofreading and style cleaning our paper.