Introduction

Parsing refers to the process of assigning syntactic structure to language stimuli during on-line language comprehension in real time. Basically, each incoming word is integrated incrementally into a syntactic representation of a sentence, a parse tree [1–4]. For each word, the human sentence parser must decide how to integrate it into the current structure, and in the case of potential ambiguity, there are multiple theoretically possible options. A key question in the literature on sentence processing is how the parser makes these decisions; that is, which types of information are taken into account during on-line parsing [1,4–10]. Does the parser rely on a few (syntactic) heuristics, or does it employ all available information? A radical answer to this question is provided by Mitchell [11] and van Gompel and Pickering [12] who demonstrate that noun phrases are attached as direct objects of intransitive verbs (but see 13). For a parser to do this, it must rely on some basic syntactic heuristics, such as late closure, which states that the parser prefers to attach new words into the phrase currently being processed [1], and it must ignore (or have restricted access to) non-syntactic information and information about subcategorization frames (i.e. transitivity, requirements on the number and types of objects/complements a verb can or must take). At the other extreme, it has been argued that the human parser has access to all types of information, including syntax, semantics and pragmatics, simultaneously when making attachment decisions [3,14].

A classic method of determining which factors affect attachment decisions is the plausibility paradigm, where a filler (a wh-phrase, such as which rule, occurring at the left edge of the sentence) is associated with a gap at the end of the sentence, but where a possible gap site intervenes [14,15]. In (1a) below, which language is the filler, and (✓) indicates a possible and plausible attachment site; the intermediate attachment site is plausible (which language could be the object of learnt) though the wh-phrase is ultimately attached as the object of speak at the final attachment site.

1. (1)
2. a. [Which language] has the student learnt (✓) to speak (✓)?    (Plausible)
3. b. [Which drain] has the student learnt (*) to clean (✓)?     (Implausible)

In (1b), however, the extracted wh-phrase, which drain, is not a plausible object of the matrix verb, learnt, indicated with (*); only the final attachment site yields a plausible interpretation, indicated with (✓).

Using an incremental method of stimulus presentation (e.g. self-paced reading), it is possible to measure localized increases in response time (RT) due to increases in processing cost. If the parser avoids intermediate attachment when the filler is an implausible object, as in (1b), then the prediction is that there would be no anomaly effect giving rise to increased RT at the matrix verb learn. On the other hand, if the parser does attach the implausible filler at the intermediate position, an increase in RT at the matrix verb is predicted due to the anomaly effect.

Several studies have investigated whether implausibility blocks intermediate attachment or not. For example, Boland, Tanenhaus, Garnsey, and Carlson [16] found that plausibility affected attachment, whereas Traxler and Pickering [4,17] concluded that it did not. Likewise, for verbs with multiple subcategorization frames, Trueswell, Tanenhaus, and Kello [18] observed an effect of frequency on attachment, whereas Pickering and Traxler [4] found no such effect. In fact, there is currently little agreement as to which factors influence intermediate attachment decisions [13].

Implausible fillers have been found to increase RT at the first (intermediate) attachment site (i.e. when the parser encounters the matrix verb) compared to plausible fillers [10,17]. Arguably, however, this RT effect in itself does not really tell us much about whether attachment takes place or not. If the theoretical assumption is that structural attachment precedes semantic evaluation [19], then clearly the increase in RT could be taken to suggest early attachment. But on the other hand, if the assumption is that various types of information are used simultaneously, then it could be argued that the RT effect is the result of the decision process (i.e. choosing between competing alternatives); other things being equal, it is likely that it takes longer for the parser when the syntactic preference (‘fill the gap’) is in conflict with the semantic preference (‘avoid implausible objects’), compared to when they converge [14].

The crucial question, rather, is whether there is an effect, not at the actual point of early (intermediate) attachment (i.e. at the matrix verb learnt in (1) above), but at the subsequent point of disambiguation (i.e. at the infinitive marker to in (1)), where reanalysis must occur in case of attachment. This means that a model that assumes that implausibility can hinder attachment predicts a higher RT for the plausible condition at to in (1a) because there is attachment and reanalysis. In contrast, a parser that is immune to implausibility and operates solely on syntactic information is predicted to show no difference in RT at to since attachment takes place regardless, and there is ultimately reanalysis in both (1a) and (1b). The plausibility paradigm can thus be used to distinguish between what we will refer to as a syntactic parser (i.e. one that relies mostly or solely on syntax in initial attachment) and a semantic parser (one that relies mostly on semantic cues, or at least ranks semantic cues higher than syntactic ones). Note that the question of which sources of information the parser uses to make decisions is largely orthogonal to the question of serial vs. parallel parsing, because both ranked parallel parsers and serial parsers could base their structural preferences on either syntactic or semantic cues [20–23].

A third kind of parser, which we shall refer to as a blind parser, is a fully autonomous syntactic parser that also ignores (or has no access to) subcategorization information [11]. The prediction for the blind parser is the same as for the syntactic parser, namely that early attachment occurs regardless of whether the filler is plausible or not. In order to distinguish between the blind parser and the syntactic parser, we need to look at verbs that show a contrast in their subcategorization frames, as exemplified in (2) below. The matrix verb in (2a), noticed, can take either a clausal complement, i.e. a complementizer phrase (CP) ((that) the pig … was missing), or a noun phrase (or, rather, a determiner phrase, henceforth, DP) object (the pig). The matrix verb in (2b), presumed, on the other hand, is only compatible with a CP complement (*John presumed the pig is ungrammatical).

1. (2)
2. a. John noticed the pig in the pen was missing.    (CP or DP object)
3. b. John presumed the pig in the pen was missing.    (CP object only)

In (2a), the pig is attached as a DP object by all three types of parser, and reanalysis must take place at the disambiguating verb was; until this point, the pig in the pen is a possible and plausible object. In (2b), however, only the blind parser attaches the pig as a direct object. The syntactic and the semantic parsers do not have this option available since information about matrix verb subcategorization is accessible and blocks erroneous early attachment; instead they have to make a more complicated analysis and attach the pig as the subject of an embedded clause. Crucially, no reanalysis is needed at was for the syntactic and the semantic parsers because there is no attachment at the intermediate potential gap. Hence, with a syntactic or semantic parser, a higher RT is predicted at the disambiguating was in (2a) (the CP/DP condition, where there is reanalysis) compared to (2b) (where there is no reanalysis).

In our study we used the plausibility contrast in (1) and the subcategorization contrast in (2) to determine which of the three parsers was most successful in accounting for the data: the syntactic, the semantic or the blind parser. In other words, we aimed to determine which sources of information affect the attachment decisions and, more specifically, whether plausibility and subcategorization requirements affected the parser’s initial attachment decision.

Below we report the results of one off-line acceptability judgment task and two experiments using G-Maze [24]. Based on these results, we shall argue in favor of a syntactic parser.

Experiment 1

Before we attempted to locate the plausibility effect precisely using self-paced reading, we needed to make sure that the effect really was there in Danish. Experiment 1 was designed to test whether there is an acceptability contrast between fully grammatical sentences with implausible intermediate attachments and sentences without. In an earlier study using acceptability judgments on wh-extraction in Danish, we found an effect of plausible vs. implausible intermediate, temporary attachment on overall acceptability such that an implausible intermediate attachment gave rise to lower acceptability of an otherwise fully grammatical sentence [25], see also 26. A similar result is reported by Pickering and Traxler [4] using self-paced reading and eye-tracking. Based on these prior results we expect to find reduced acceptability in the implausible condition, and we assume that the lower acceptability reflects the additional processing load caused by the implausibility of the temporary attachment cf [26].

Materials and Methods

Materials and Methods

The stimulus set used in the actual experimental acceptability survey consisted of the 40 sentences (20 pairs) remaining after norming plus 24 fillers (12 fillers were judged unacceptable, and 12 were judged to be completely acceptable in previous studies; see examples in (7) below); 64 sentences in total. The task consisted of providing acceptability judgments on a 7-point Likert-scale (1 = helt uacceptabel, ‘completely unacceptable’, 7 = helt acceptable, ‘completely acceptable’). Using an internet-survey created using Google Drive, we obtained answers from forty-five participants.

1. (7)
2. a. Chaufføren røg en flot cykel    (Filler type A: implausible).
3.    Driver-the smoked a nice bicycle
4.    ‘The driver smoked a nice bicycle’

1. b. Kvinden elskede mørk chokolade.    (Filler type B: plausible)
2.    Woman-the loved dark chocolate
3.    ‘The woman loved dark chocolate’

Results

The mean acceptability rating was 6.3 for the plausible condition (corresponding to (4a)), and 5.8 for the implausible condition (as in (4b)). We used a linear mixed-effects regression model to analyze the data with the free software R [27] and the lme4 package for R [28]. The dependent variable was the acceptability rating and the independent variable was the four-level sentence type variable (Filler type A, Filler type B, plausible, and implausible). The model included random intercepts for participants and items. We used contrastive coding to get the pair-wise comparisons of neighboring levels in the sentence type variable [29], and p-values were subsequently Holm-corrected for multiple comparisons [30]. The final model is summarized in table 1 (excluding the intercept, which is not crucial here), and the results are illustrated in Figure 1.

Download:

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

Figure 1. Mean acceptability ratings across sentence types in experiment 2.

Error bars ±1 SE. Significant differences: *p<0.05, **p<0.001.

https://doi.org/10.1371/journal.pone.0076326.g001

Discussion

Experiment 1 confirms that a temporary implausible attachment of a wh-filler to a matrix verb has a significant negative effect on the overall acceptability of a sentence. This is consistent with the results reported by Fanselow & Frisch [26] and Christensen et al. [25] where a reduction in acceptability is argued to reflect an increase in processing cost. This causal link between processing cost and acceptability also supports the idea that an early attachment that violates the semantic selectional restrictions of the main verb increases processing cost. Having established that there is a plausibility effect in Danish, the next question was when it occurs in the on-line parsing. Experiment 2 was designed to this end.

Materials and Methods

We used a subset of the sentences that appeared in experiment 1 (all sentences were normed as described above), namely the 16 sentences (8 pairs) where the plausible and implausible versions of the sentences differed from one another in one respect only: the wh-phrase, as in (4) above. In this way we could be certain that any differences in reaction time (RT) could not be attributed to differences in word length or frequency (since all words except the wh-phrase were identical in the two conditions). The stimuli were divided into two lists, ensuring that participants saw only one version of each sentence (plausible or implausible). The non-targets (see below) were the same in the two lists, so the plausible and implausible version of each sentence appeared with identical non-targets. The same set of fillers appeared on both lists (14 sentences from an unrelated experiment, and 12 distractor items). Sixty native speakers of Danish participated in this experiment (mean age 23.7, range 17-39). The stimuli were presented in randomized orders (a unique random order for each participant) on a PC running DMDX [31]. Prior to the actual session, a training session was run to familiarize participants with the task. The entire session lasted approximately 6 minutes.

The G-Maze method was developed by Forster, Guerrera and Elliot [24] as a low-tech tool to localize processing effects giving rise to increases in RT, and Witzel, Witzel and Forster [32] report results demonstrating that the G-Maze method is comparable to eye tracking and superior to moving window self-paced reading in precision. Sentences are presented one word at the time, and each word is paired with a word that cannot possibly be a continuation of the sentence (a non-target). The task is to choose, using left and right arrow on the keyboard, the target word that continues the sentence in a grammatically possible way (see Figure 2). G-Maze enforces word-by-word incremental processing, and RT is recorded for each word pair presented to the participants, making it possible to localize any RT effects due to processing cost.

Download:

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

Figure 2. The G-Maze experimental design.

Words are presented incrementally in pairs and participants have to select the one that fits the context best. RT is logged for each selection.

https://doi.org/10.1371/journal.pone.0076326.g002

Feedback was displayed on the screen immediately after an incorrect choice and at the end of the sentence in the case of only correct choices. After feedback, participants pressed the spacebar to continue the experiment.

Results

The data was analyzed using a linear mixed-effects regression model (same software as in experiment 1). The dependent variable was RT, and intercepts for participants and items were included as random effects. The fixed effects were added one by one in following order (reflecting relative theoretical importance, from least to most important): sex, age, previous RT, trial, word position (word 1 to 6 in the sentence) and plausibility. Only significant effects were kept in the model. The final model is summarized in table 2 below (leaving out contrasts between the intercept and each word as they are irrelevant in this context). The results are illustrated in Figure 3.

Download:

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

Figure 3. Mean RT, word for word, experiment 2.

Error bars ± 1 SE, *p<0.05.

https://doi.org/10.1371/journal.pone.0076326.g003

We found a small familiarization or training effect such that participants respond faster later in the experiment (see Trial row in table 2). We also found an effect of Previous RT, meaning that there was a significant correlation between a slow RT and a likewise slow previous RT, and vice versa for fast responses. This effect simply shows that fast responders were consistently fast and slow responders were consistently slow, even when the random intercepts for participants are taken into consideration.

The only significant effect of plausibility was located at the fourth word, the matrix verb (see Figure 3), where participants responded faster in the plausible condition than in the implausible condition (p=0.0294). There was no significant difference at the disambiguating word (word 5, p=0.3817) or at the embedded verb (word 6, p=0.4986).

Discussion

The result is in accordance with the predictions of both the blind parser and the syntactic parser; there is no difference at the disambiguating word, i.e. word 5 in Figure 3 (nor at the embedded verb, word 6). The lack of a difference at the disambiguating word is expected if reanalysis occurs in both cases, and the result thus supports this assumption.

The effect of plausibility at the matrix verb (word 4 in Figure 3) is consistent with both the view that semantic evaluation occurs after syntactic attachment (as argued in [19]), and with the view that the simultaneous consideration of syntactic and semantic cues prolongs processing in case of conflict [14]. It is, however, as argued in the introduction, the null-effect at the disambiguating word that reveals that attachment occurs in both conditions. The result suggests that plausible and implausible wh-fillers alike are attached as soon as possible. Similarly, in a self-paced reading experiment investigating whether plausibility affects attachment in English, Stowe, Tanenhaus and Carlson [10] found no RT difference at the disambiguating word.

The question is to which degree the parser ignores semantic information: Does the parser also attach impossible elements, or does it only attach implausible elements? In other words, is the parser syntactic or is it blind? To answer this, we designed experiment 3.

Materials and Methods

We constructed 10 sentence-pairs, where each pair only differed with regard to the matrix verb. Either the matrix verb subcategorized for a DP or a CP complement, such as hørte ‘heard’ in (8a), or it subcategorized for a CP complement only, tænkte ‘thought’ in (8b):

1. (8)
2. a. Filosoffen hørte forelæsningen om etik vakte begejstring.    (CP/DP)
3.    Philosopher-the heard lecture-the about ethics evoked enthusiasm
4.    ‘The philosopher heard the lecture on ethics evoked enthusiasm.’

1. b. Filosoffen tænkte forelæsningen om etik vakte begejstring.    (CP only)
2.    Philosopher-the thought lecture-the about ethics evoked enthusiasm
3.    ‘The philosopher thought the lecture on ethics evoked enthusiasm.’

In some cases verbs that usually cannot take a DP object, may, however, take a so-called cognate object. For example, you can think clever thoughts, but you cannot think clever people; likewise, you can die a certain death or sleep the sleep of the righteous, but you cannot die a certain fate or sleep a nap. Given that these cases are very limited, and to a large extend idiomatic, the possibility of a cognate object is probably not a factor in online parsing (unless it is clearly facilitated by the context). At any rate, the (im)possibility of a cognate object is substantially different from normal subcategorization for DP objects and is not relevant here.

Thirty-two people participated in the experiment with ages ranging between 21 and 49 (mean age 26.6 years).

Results

We fitted a linear mixed-effects regression model to RT using trial, verb type (CP/DP or CP-only) and word position (words 1 to 6) as independent variables (using the same software as in experiment 1). The final model is summarized in Table 3 and the results are illustrated in Figure 4.

Download:

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

Figure 4. Mean RT, word for word, experiment 3.

Error bars ± 1 SE, *p<0.05, **p<0.001.

https://doi.org/10.1371/journal.pone.0076326.g004

Again, we found an effect of trial, suggesting a small, but significant training effect. More interestingly, there was an effect of verb type at both the subject of the embedded clause (word 2 in Figure 4) and at the embedded verb (word 5). At the embedded subject, RT is higher for the CP-only condition (p=0.0394), whereas RT is higher at the embedded verb for the CP/DP condition (p=0.0006).

Discussion

The syntactic parser predicts that reanalysis should only be necessary in the CP/DP condition, where the subject of the embedded clause (word 2) is first attached as an object of the matrix verb, and then later (upon reading word 5) reanalyzed to be the subject of the matrix clause. The blind parser, on the other hand, predicts attachment in spite of a potentially conflicting subcategorization frame, and hence, reanalysis in both conditions. The clear slowdown at the embedded verb (word 5, the disambiguating word) in the CP/DP condition is easily explained as reflecting the extra cost of reanalysis. This would mean that incorrect attachment does not take place in the CP-only condition, in accordance with the predictions of the syntactic parser, the one that is informed by syntactic information including subcategorization frames.

The other significant difference is at the embedded subject (word 2) in the CP-only condition. One could argue that the CP-only condition increases RT due to an anomaly effect from the evaluation of the alternative structures (i.e. ruling out the impossible DP object analysis) [14]; however, we take it to be a reflex of the fact that constructing a clause probably involves more processing than attaching an object, as demonstrated by Staub [13]. (Further support comes from a neuroimaging study by Shetreet, Friedmann, and Hadar [33] that showed increased brain activation for clauses compared to corresponding and matching nominal strings.) The number of syntactic nodes (branchings in the tree structure) has been suggested to be a measure of syntactic complexity [34–37]. Considering the syntactic complexity involved in constructing a CP vs. a DP, it is clear that a CP involves more structure than a DP, see Figure 5. A CP has (at least) an additional three maximal projections (XPs), namely, CP, TP and VP. Constructing a CP is in this respect a much more complex process, and the fact that the RT is higher at word 2 could be a reflection of precisely this.

Download:

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

Figure 5. Partial syntactic trees.

(A) The structure of a verb phrase headed by a verb with a CP complement, (B) a verb with a DP object. Notice that the verb notice appears in both (A) and (B). CP=Complementizer Phrase, TP=Tense Phrase, VP=Verb Phrase, DP=Determiner Phrase. In English, in examples like the ones used here, C can be realized by the optional complementizer that, T as a finite auxiliary or tense inflection, and V is where the lexical verb is inserted.

https://doi.org/10.1371/journal.pone.0076326.g005

Hawkins [35] argues that the parser projects (or constructs) a CP if and only if it encounters a word that uniquely requires the presence of a CP-node, such as a complementizer (e.g. that or if), a wh-word (e.g. what or where) or a nominative pronoun (such as he or she which must be an embedded subject, in contrast to him/her which must be the object). In our stimuli, neither is present (by deliberate design in order to ensure the possibility of ambiguity). The parser has to assume a CP at the second word (word 2 in Figure 4) in the CP-only conditions, despite the lack of a clear, unambiguous input, and this could contribute to the higher RT. Related to this issue is the observation that CP-only verbs are most frequently followed by a complementizer (at ‘that’) or a personal pronoun with nominative case (e.g. han ‘he’) ( [36], pp. 49-61). To test whether this observation also holds for Danish, we searched for the CP-only verb håbe ‘hope’ in the Danish on-line corpus KorpusDK and examined fifty random examples where the verb was used with a CP complement. In 44 cases (88%), the first word in the CP was indeed either a complementizer or a personal pronoun. That is, the parser predicts that a clausal complement is introduced by a complementizer, and when that expectation or preference is not met, there is an error signal and the integration cost goes up with a concurrent increase in RT [38].

Note also that Danish orthography has two alternative comma systems; the ‘grammatical comma’, with obligatory commas before and after embedded clauses (including complement and relative clauses), and the ‘new comma’, where there are only commas after embedded clauses (except parenthetical relative clauses where the initial comma is obligatory). This may potentially affect RT at the embedded subject if (at least some) participants preferred the ‘grammatical comma’ and therefore detected a missing comma, giving rise to an additional error signal [13].

Acknowledgments

The authors wish to thank Per Byrgren at the AU Arts IT department for help establishing the computer network for the project, Scott Jackson and Michael Krabbe Borregaard for valuable comments on the statistical models, the reviewer, Arild Hestvik, for insightful comments and finally Sten Vikner and Hanne Kizach for help with creating the stimuli.