example, in Stowe (1986), reading times on 'us' were slower in sentences like (1a), in which a gap may be expected in this position, compared to sentences like ...
Working memory in the processing of long-distance dependencies: Interference and filler maintenance Tal Ness1 & Aya Meltzer-Asscher1,2
1
Sagol School of Neuroscience 2
Linguistics Department Tel Aviv University
1
Abstract During the temporal delay between the filler and gap sites in long-distance dependencies, the "active filler" strategy can be implemented in two ways: the filler phrase can be actively maintained in working memory (“maintenance account”), or it can be retrieved only when the parser posits a gap (“retrieval account”). The current study tested whether filler content is maintained during the processing of dependencies. Using a self-paced reading paradigm, we compared reading times on a noun phrase (NP) between the filler and gap sites in object relative clauses, to reading times on an NP between the antecedent and ellipsis sites in ellipsis sentences. While in the former type of dependency a filler by hypothesis can be maintained, in the latter there is no indication for the existence of a dependency prior to the ellipsis site, and hence no maintenance. By varying the amount of similarity-based interference between the antecedent and integration sites, we tested the influence of holding an unresolved dependency on reading times. Significantly increased reading times due to interference were found only in the object relative condition, and not in the ellipsis condition, demonstrating filler maintenance costs. The fact that these costs were measured as an effect on similarity-based interference indicates that the maintained representation of the filler must include at least some of the features shared by the interfering NP.
Key words: Sentence processing; Working memory; Long-distance dependencies; Filler-gap dependencies; Parsing; Syntax.
2
1. Introduction Filler-gap dependencies are unbounded relations between a displaced element (the ‘filler’, italicized in (1)) and its canonical thematic position (the ‘gap’, underlined in (1)), which are used to form wh-questions, relative clauses, and other structures. The parsing of these structures has been thoroughly studied in the past few decades, with one of the central findings being that they are resolved using an “active filler" strategy (Frazier & Clifton, 1989; Frazier & Flores D’Arcais, 1989), namely, once a displaced element is identified, the parser posits a gap at the first potential gap position, before any direct evidence in the input (i.e. a missing element) indicates it. (1) (a) (b)
What book do you think John read __ last night? This is the book that John read __ last night.
Much evidence was provided over the years in support of the "active filler" strategy. One type of evidence is the “filled gap effect” (Crain & Fodor, 1985; Lee, 2004; Stowe, 1986). For example, in Stowe (1986), reading times on 'us' were slower in sentences like (1a), in which a gap may be expected in this position, compared to sentences like (1b), where no gap is expected. This effect indicates that a gap was the parser’s first choice, which then had to be changed. (1) (a) My brother wanted to know who Ruth will bring us home to __ at Christmas. (b) My brother wanted to know if Ruth will bring us home to Mom at Christmas. Another type of evidence for active gap filling was found in experiments containing sentences such as (2a), in which the comprehender encounters a verb for which the filler is an implausible argument. Elevated reading times (Traxler & Pickering, 1996) and a greater N400 amplitude (Garnsey et al., 1989) were found at the verb when the filler was an implausible argument (compared to sentences such as 2b). Since these effects precede any direct evidence
3
for a gap in the input, they support an active dependency resolution strategy (though they may also be explained merely by the predictability of the verb in the preceding context). (2) (a) That’s the garage with which the heartless killer shot the hapless man ___ yesterday afternoon. (b) That’s the pistol with which the heartless killer shot the hapless man ___ yesterday afternoon. Similar and additional effects of active filler-gap dependency construction were demonstrated in many languages using a variety of paradigms (e.g. Aoshima et al., 2004; Boland et al., 1995; Keshev & Meltzer-Asscher, in press; Love & Swinney, 1996; Sussman & Sedivy, 2003; see Phillips & Wagers, 2007 for review). A critical property of filler-gap dependencies is the fact that they are non-local and can, in principle, be established between indefinitely distant elements (but see Futrell, Mahowald, & Gibson, 2015, who demonstrate that dependency length in actual utterances is minimized in many languages). During the temporal delay between the encoding of the filler and its integration at the gap site, there are two architectural accounts suggested for the representation of the unresolved dependency (Wagers & Phillips, 2014): A. Maintenance. Under such accounts, the filler phrase is actively maintained in working memory (or focal attention), and can be directly integrated at the gap site (e.g. Wanner & Maratsos, 1978). The benefit of such representation is that the filler is continuously accessible for the parser. However, maintaining this representation may demand working memory resources. B. Retrieval. According to these accounts, although the open dependency affects parsing, the content of the filler is not purposely kept in a privileged state and therefore it must be retrieved and reactivated when the parser posits a gap (e.g. McElree et al., 2003;
4
Lewis & Vasishth, 2005). Such an architecture avoids maintenance costs, but may increase the possibility of retrieval errors or difficulties. Importantly, both explanations can be compatible with the "active filler" strategy, as the existence of search for a gap may be active independently of the maintenance of the filler’s content. Support for the retrieval account comes from the pattern of cross-modal lexical priming (CMLP) effects found for semantic associates of the filler (in a lexical decision task: Nicol et al., 1994; Nicol & Swinney, 1989; in a probe recognition task: McElree, 2001; and in a naming task: Love & Swinney, 1996). In these experiments, priming effects for words semantically related to the filler were observed at the filler and gap sites, but not in between the two positions. These findings suggest that the semantic content of the filler may not be actively maintained or that it is maintained in a manner that cannot influence the task (e.g. using a partial representation, as explained further below). Initial support for the maintenance account is provided by the preference for shorter dependencies over long ones (e.g. Crain & Fodor, 1993; Gibson, 1998, Futrell et. al., 2015) and the greater processing costs that longer dependencies incur (Gibson, 2000). However, such findings can also be explained by higher retrieval costs for longer dependencies. Stronger support for maintenance views comes from studies demonstrating increased processing load during the interval when an open dependency exists. Behaviorally, this was demonstrated by increased reading times on the words between the filler and gap sites in relative clauses compared to sentences with a sentential complement (which do not include a dependency) (Chen et al., 2005; Stepanov & Stateva, 2015). Relatedly, event-related potentials (ERP) studies found a sustained anterior negativity (SAN) on the words intervening between the filler and gap site in object-extracted dependencies (King & Kutas, 1995; Fiebach et al., 2002; Phillips et al., 2005), which was interpreted as reflecting increased working memory demands
5
incurred by maintaining a filler in an open dependency (see also Meyer, Obleser, & Friederici, 2013, who showed EEG oscillatory activity correlated with maintenance of arguments in German verb-final sentences). It is important to note, however, that these effects (both behavioral and ERP) do not contradict the retrieval account, since they do not provide any indication as to what causes the increased processing load. When a dependency is open, the “active filler" strategy in itself is likely to cause increased processing load due to the need to look for a gap site, regardless of what exactly is being maintained in working memory, namely a full (or partial) representation of the filler, or its mere existence. Another factor to consider in deciding between the two hypotheses is the motivation for the "active filler" strategy. Over the years, several proposals have been put forward to account for the preference to posit a gap as early as possible. This strategy was suggested to stem from the need to assign thematic roles (Pritchett, 1992) or from the parser's aim to form the maximal interpretation possible at any point (Altmann, 1999). Another explanation relies on frequencybased parsing (Mitchell & Cuetos, 1991), which states that the parser would always initially assume the most probable option in terms of its previous experience. All of these explanations do not hinge on working memory as the motivation for active gap filling. Alternatively, the preference for early dependency resolution can be due to demands on working memory resources, a limited capacity mechanism (Wanner & Maratsos, 1978; Just & Carpenter, 1992; Kane et al., 2001). This explanation aligns best with the maintenance account, as this account predicts elevated working memory costs while a dependency is unresolved. The maintenance and retrieval accounts are not necessarily mutually exclusive. A hybrid maintenance/retrieval account would suggest that maintenance applies to some but not all features of the filler, and retrieval may be needed for the unmaintained content (Fiebach et al., 2002; Wagers & Phillips, 2014). This means that the content maintained in working memory
6
for an unresolved dependency is a partial representation of the filler, and the full representation is reactivated at the gap site. In order to explore the hybrid maintenance/retrieval account, Wagers & Phillips (2014) used three different indices of dependency formation drawing upon different sources of information to test what aspects of a filler are available to the parser and at what stage, when forming short versus long dependencies. Using the filled-gap effect, their results showed that the existence of a displaced element (the filler) and its basic syntactic category (NP versus PP) have an immediate effect at the potential gap site, and are thus arguably actively maintained throughout the dependency. However, semantic anomaly detection (that requires specific semantic information associated with the displaced element), as well as the detection of a preposition mismatch between the filler the verb's requirement, showed this immediate effect only for short dependencies. For longer dependencies, the mismatch had a somewhat delayed effect. The fact that in long dependencies some information about the filler (i.e. its syntactic category) has more immediate influence than other information types (i.e. its semantic content and type of prepositional head) may indicate that only the former type of information is maintained available for the parser, while the unmaintained information has to be retrieved when needed, which leads to a delayed effect.1
1
The delayed effect of semantics seems to contradict with the results seen in the cross-modal priming paradigm
(e.g. Nicol & Swinney, 1989), which show priming for probes semantically related to the filler immediately at the offset of the verb. The results can be reconciled, however, when considering that (1) response time in a lexical decision task typically takes several hundred milliseconds, during which priming may have occurred at any point, namely, priming need not have occurred instantaneously at the verb offset; (2) whereas priming effects can arise as soon as the reactivation of the filler’s content started, semantic anomaly detection may require initial integration processes that may only occur after full retrieval of the filler’s content (a more time-consuming process) is completed.
7
Although these results are compatible with the hypothesis that some features of the filler, e.g. its syntactic category, are being maintained rather than reactivated, it is also possible that these features are easily retrieved instantaneously, thus showing early effects, while e.g. the semantic contents of the filler take longer to fully reactivate (since these are much richer and more detailed). Since the effects in Wagers & Phillips (2014) are exhibited at or after the gap site, and not before it, they cannot provide direct evidence for maintenance, rather than fast retrieval. In addition, as mentioned above, although several reading time and ERP studies did show "maintenance" effects before the gap site (Chen et al., 2005; King & Kutas, 1995; Fiebach et al., 2002; Phillips et al., 2005; Stepanov & Stateva, 2015), these effects could be attributed simply to the processing cost of holding an open dependency, rather than to actual maintenance of some content associated with the filler. In the current experiment, we wanted to provide direct evidence for filler maintenance by looking at processing costs between the filler and gap sites, and testing whether these costs are modulated by interference. A prominent attribute of working memory is the vulnerability of the representations it holds (e.g. Baddeley, 1992). Among the suggested causes for this are time-based decay (Baddeley et al., 1975; Page & Norris, 1998) and interference between items held simultaneously (Nairne, 1990; Oberauer & Kliegl, 2006, Van Dyke & Johns, 2012). The degree of mutual interference between two representations in working memory was shown to depend greatly on their similarity. In language processing, this was shown for example by Gordon et al. (2002), who gave participants syntactically complex (object-extracted cleft) and simpler (subject-extracted cleft) sentences to read, while they had to remember a short list of words. Comprehension performance was lower for complex sentences, but this effect was greater when the nouns in the sentence matched the nouns in the list in type (common nouns or proper nouns). In a similar experiment, Fedorenko et al. (2005) showed that similarity also affects online processing. Reading times were longer for syntactically complex relative clauses
8
(object-extracted) than for simpler relative clauses (subject-extracted), and this effect was greater when participants had to remember a list of nouns that were similar to the nouns in the sentence than when they were dissimilar (see also Van Dyke & McElree, 2006). Effects of interference were also demonstrated when similarity held between items within the sentence. For example, Van Dyke (2007) found increased reading times when an interfering NP (an NP located in a clause embedded between the subject and the verb of the main clause) shared syntactic or semantic features with the subject of the main clause. Furthermore, it was shown that interference effects arise whether the similar NP is located between the constituents of a dependency (i.e. retroactive interference) or prior to both constituents (i.e. proactive interference) (Van Dyke & McElree, 2011). Another example of similarity being the cause of working memory interference comes from a comparison between the difficulty of object relative clauses (RCs) and subject RCs, when the animacy of the filler matches or does not match that of the noun phrase (NP) argument inside the RC (Mak et al., 2002, 2006). Object RCs were found to be more difficult than subject RCs (as reflected by increased reading times), but only when the animacy of the filler matched that of the NP inside the relative clause (see also Gordon et al., 2001; Traxler et al., 2002, 2004). Dissimilarity in gender and number of the filler and the noun in the relative clause was also shown to improve accuracy in children (Adani et al., 2009) as was dissimilarity in semantic restrictedness (Friedmann et al., 2009). Similarity-based interference is also a key feature in computational models of memory mechanisms in sentence processing (e.g. Oberauer & Kliegl, 2006; Lewis & Vasishth, 2005). In view of this, the current experiment employed an interference manipulation to measure filler maintenance effects on on-line reading times. Using self-paced reading, we compared object relative (OR) sentences, in which an unresolved filler-gap dependency is held between the filler and gap sites, with NP-ellipsis sentences, in which there is no indication for the
9
dependency prior to the retrieval site, hence no maintenance is needed. Interference was manipulated by the addition of an interfering NP (similar to the filler in grammatical and semantic features). The critical region for analysis was an NP between the antecedent and integration sites. Since in this position no retrieval is necessary, any observed effect can be attributed to maintenance only. Under these conditions, an interaction between interference and the type of dependency, namely an increased effect of interference in the filler-gap condition, would strongly support the maintenance account (or the hybrid maintenance/retrieval account). This is so since, as explained above, working memory interference is similarity-based, and thus, in order for it to be affected by processing an open dependency, the maintained representation of the filler must include at least some shared features with the intervening material.
2. Methods 2.1. participants Participants were 64 Tel-Aviv University students (23 male), all native Hebrew speakers, with an average age of 25.15 (range: 20-38). All participants had normal or corrected to normal vision. Participants volunteered for the study or were paid 20 NIS (~$5) for their participation.
2.2. Materials The materials for the experiment consisted of 24 Hebrew sentence sets, each with four experimental sentences, and 48 filler sentences. The experimental conditions were built in a 2x2 factorial design, with the factors Filler Maintenance (with / without) and Interference (with / without). Filler Maintenance was manipulated by using two sentence structures: object relative clauses (with filler maintenance) and NP-ellipsis sentences (without filler maintenance). Interference was manipulated by the addition of an interfering NP between the
10
filler/antecedent and the gap/retrieval site. The interfering NP was similar to the filler/antecedent in animacy, gender and number, and was also semantically related to it. An example set is presented in Table 1, with the filler/antecedent underlined, the critical NP marked in bold, and the intervening NP, where present, capitalized.2 Table 1: Example sentences Condition
Example sentence Itai zirez
ha-mitlamed šel ha-itonai
et
Itai hurried ACC the-intern
ha-xarif
me-aruc
eser
of the-journalist the-bright from-channel ten
_ šel ha-orex ha-kašuax
aval lo et
- Maintenance
but not ACC _ of the-editor the-strict
- Interference
'Itai hurried the intern of the bright journalist from channel ten but not _ of the strict editor.' Question: Did Itai hurry the intern of the editor? (Correct answer: No)
Itai zirez
et
ha-mitlamed šel ha-taxkiran
Itai hurried ACC the-intern ha-xarif
me-aruc
me-ha-toxnit
šel ha-itonai
of the-researcher from-the-program of the-journalist
eser aval lo et
_ šel ha-orex ha-kašuax
- Maintenance
the-bright from-channel ten but not ACC _ of the-editor the-strict
+ Interference
'Itai hurried the intern of THE RESEARCHER from the program of the bright journalist from channel ten but not _ of the strict editor.' Question: Did Itai hurry the journalist? (Correct answer: No)
Itai zirez
et
ha-mitlamed še ha-itonai
Itai hurried ACC the-intern edrix
ha-xarif me-aruc
eser
that the-journalist the-bright from-channel ten
_ bemaalax ha-kaic
+ Maintenance
mentored _ over
- Interference
'Itai hurried the intern that the bright journalist from channel ten mentored _ over the
the-summer
summer.' Question: Did the journalist mentor the intern? (Correct answer: Yes)
2
ha-mitlamed še ha-taxkiran
+ Maintenance
Itai zirez
+ Interference
Itai hurried ACC the-intern
et
me-ha-toxnit
šel
that the-researcher from-the program of
The full list of experimental materials is available upon request from the corresponding author.
11
ha-itonai
ha-xarif
me-aruc
eser edrix
_ bemaalax ha-kaic
the-journalist the-bright from-channel ten mentored _ over
the-summer
'Itai hurried the intern that the researcher from the program of the bright journalist from channel ten mentored _ over the summer.' Question: Did the researcher mentor the intern? (Correct answer: Yes) Note: The antecedent is underlined, the critical NP is marked in bold and the integration site is marked by an underscore (_).
Four lists were created in a Latin Square design such that each participant saw only one sentence from each set (6 items in each condition). The order of items was randomized for each participant. Filler sentences were mainly aimed at varying the syntactic structures employed in the experiment to avoid anticipation of the ellipsis site in the ellipsis conditions, and they consisted of eight different structures (subject relatives, VP-ellipsis sentences and others). Seventy-five percent of the experimental sentences (and overall 50% of all trials, including fillers) were followed by a yes/no question (randomly distributed) to make sure that participants read and understood the sentences.
2.3. Procedure The sentences were presented in a self-paced moving window paradigm using the Linger software (Rohde, 2003). Responses were made using a Cedrus RB840 response box. Participants responded with their preferred hand. Participants were instructed to read at their normal rate. Feedback was given for incorrect responses to the comprehension questions. Before the experiment began participants completed a practice block of 5 sentences.
2.4. Data Analysis
12
The data of one participant was excluded since his mean reading time (RT) exceeded 2.5 standard deviations from the group’s mean. Both RT and accuracy data were analyzed with mixed-effects models. Analyses were conducted using the lmerTest package (Kuznetsova et al., 2014) in the R software environment (R Development Core Team, 2011). Linear models were used for RT data. Both factors were coded for simple contrasts (one level of the factor coded as 0.5, and the other as −0.5). A logistic model was used for the accuracy data, including random effects for subject and item. RT
data
were
analyzed
using
a
two-step
approach
(see
http://www.hlplab.wordpress.com/2008/01/23/modeling-self-paced-reading-dataeffects-ofword-length-word-position-spill-over-etc/), similar to the analysis described by Hofmeister (2011). The log-transformed RTs from all stimuli (all words in the experimental items as well as in the filler items) were entered into a regression model which included: 1) the word’s length, 2) the restricted cubic spline of the word’s position in the sentence, 3) the log-transformed position of the trial in the experiment, and 4) random effects of subject and item. The residuals from this model, excluding residual log RTs exceeding 2.5 standard deviations from the mean, were then entered into a second model. The second model was fitted only for data from the critical words in the experimental sentences. The model included the experimental factors Maintenance and Interference, and their interaction. Further, to account for possible spill-over effects, the model also included the log RT of the three previous words (as three predictors). Additionally, it included a full random effects structure.
3. Results Mean accuracy rate on the comprehension questions was 80.1% (OR – without interference: 81.7%, with interference: 77.0%, Ellipsis – without interference: 83.3%, with interference:
13
78.5%. There were no significant differences between the conditions, although the effect of interference was close to significance (Estimate = 0.289, SE = 0.154, z = 1.87, p = .062). Reading times for the critical NP in the different conditions are reported in Table 2 and Figure 1. A significant interaction between Filler Maintenance and Interference was found (Estimate = 0.086, SE = 0.019, t = 4.65, p < .001). Pairwise comparisons revealed a significant effect of Interference in OR sentences (t = 3.62, p < .001) but not in ellipsis sentences (t = 1.60, p = .119). Table 2: Mean reading times (SE) in milliseconds for the critical NP in the different conditions With
Without
interference
interference
659.8 (24.22)
593.7 (27.15)
629.6 (32.85)
617.1 (28.37)
With filler maintenance (OR) Without filler maintenance (Ellipsis)
Reading time (ms)
Figure 1: Reading times (log-transformed) for the critical NP 2.76 2.74 2.72 2.7 2.68 2.66 2.64 2.62 2.6 2.58
*
OR + Interference
Ellipsis - interference
Note: an asterisk (*) indicates p < .05
14
We also analyzed the data from the integration site. For the object-relative conditions, the integration site was at the verb. For the ellipsis conditions, the integration site was on the word 'of' (šel), but since this is a function word which may be read fast across conditions, we also analyzed the following noun. Reading times at the integration site in the different conditions are reported in Table 3. Since this region differed between the object relative sentences and the ellipsis sentences, they were analyzed separately. No significant effect of interference was found for the object-relative conditions (+maintenance) (Estimate = -0.009, SE = 0.010, t = 0.88, p = .382). The effect of interference for the ellipsis conditions (– maintenance) was nonsignificant on the word 'of' (šel, a function word), but it was significant on the following noun (on 'of,' Estimate = 0.038, SE = 0.010, t = 1.47, p = .089, on the following noun, Estimate = 0.027, SE = 0.008, t = -3.36, p < .001). Table 3: Mean reading times (SE) in millisecond at the integration site for the different conditions With
Without
interference
interference
605.2 (26.73)
613.4 (26.91)
“of” (šel)
416.2 (14.02)
414.0 (9.39)
noun
657.1 (16.56)
602.5 (18.91)
With filler maintenance (OR) Without filler maintenance (Ellipsis)
Reading times for non-critical as well as critical words are presented in Figure 2. The critical NP was word number 6 in the conditions without interference and word number 9 in the conditions with interference, due to the addition of the interfering NP (critical words from all conditions are displayed in position 9 in Figure 2). Each of the first five words (identical for all four conditions) was also analyzed with the same model as the critical word. No significant effects were found.
15
Figure 2: Word-by-word mean reading times (log-transformed) 2.85 2.8 2.75 2.7 2.65 2.6 2.55 2.5 1
2
3
4
5
6 7 Ellipsis_Int OR_Int
8
9
10 11 12 Ellipsis_NoInt OR_NoInt
13
14
15
16
17
18
Note: Since the conditions with interference have 3 extra words (words 6-8, the interfering NP), in the No Interference conditions words 6 and on are displayed as words 9 and on in order to align identical words. The critical NP is in position 9 in the graph. The integration site is in position 13 for the OR sentences and in positions 16-17 for the ellipsis sentences.
4. Discussion The aim of the current study was to test whether filler content is maintained in working memory, as predicted by the maintenance account (or the hybrid maintenance/retrieval account) for the processing of dependencies. In order to do that, we measured the influence of an open dependency on additional working memory load, as reflected in reading times, by varying the amount of similarity-based interference between the filler and gap sites. Our results showed longer reading times due to interference only inside object relative clauses, where the comprehender holds an unresolved filler-gap dependency, and not in NPellipsis sentences, in which there is no indication for the dependency prior to the retrieval site and hence no maintenance is needed. This effect was measured on an NP between the antecedent and integration sites and therefore cannot be explained by retrieval costs. The results support the maintenance (or hybrid) account for the processing of dependencies by behaviorally demonstrating filler maintenance costs. The fact that these costs were measured as an effect on similarity-based interference indicates that the maintained representation of the 16
filler must include at least some of the features shared by the interfering NP, rather than simply an expectation for resolution of the dependency. The demonstration of modulation of filler maintenance costs by an interfering NP is also in line with the suggestion that crossing over an intervening NP in object relative clauses is the source of difficulty in this structure (compared to subject relative clause) for children (Arnon, 2005) and aphasic individuals (Sheppard et al., 2015). In order to evaluate effects of interference on retrieval, we also looked at the integration site. Increased reading times due to interference were found only in the ellipsis condition, and not in the object-relative condition. These results are also in line with the maintenance (or hybrid) account, as they show that interference only affects reading times at the integration site when full retrieval is needed (ellipsis) and not when maintenance could apply (object-relative). This demonstrates the core purpose of “wasting” working memory resources on proactively maintaining a filler, i.e. it helps the parser avoid retrieval difficulties or costs (caused by interference or otherwise). An alternative account for our results, that does not require that the filler’s content be maintained, is offered by Gibson’s Dependency Locality Theory (DLT, Gibson, 1998, 2000). Gibson’s DLT model includes two components that consume working-memory resources: storage, the cost of holding expectations for syntactic heads required to complete the current input as a grammatical sentence, and integration, the cost of connecting a word into the previously built structure. Initially, in Gibson’s SPLT model (Gibson, 1998), both components incurred a distance-sensitive cost. For the storage component this meant that the cost of maintaining an expectation (e.g. the need for a predicate and/or a gap) increases as additional input is processed. If this is the case, Gibson’s model would predict our results even without assuming maintenance of filler information, since the expectation for a verb (and a gap) in the OR conditions would be more costly when the critical NP (on which reading times were
17
analyzed) is preceded by an additional NP, regardless of similarity between the filler and the NP. However, while much evidence supports the length-sensitivity aspect of the retrieval component (see Gibson, 1998), no direct evidence for length-sensitivity of the storage component was provided to date. Gibson therefore suggested a variation to his theory that does not assume that storage cost is length-sensitive, the DLT model (Gibson, 2000). The current results, showing increased maintenance costs with the addition of an intervening NP, are therefore hard to reconcile with the view that only a prediction for a verb is held in memory. Rather, they are compatible with a model in which actual features of the filler are being maintained. Additionally, the results at the integration site in our experiment are inconsistent with what Gibson’s model would predict. Since the integration component in Gibson’s models (both the SPLT and the DLT) is length-sensitive and the dependency’s length is longer in the interference conditions in both the OR sentences and the ellipsis sentences, Gibson’s models would predict a similar “interference” effect (due to length) in both types of sentences, as opposed to the current results in which no indication of an interference effect was found in the OR sentences. Since the current results are compatible with both the maintenance account and the hybrid maintenance/retrieval account, a question remains regarding the specific content maintained in working memory for an unresolved dependency – does maintenance apply to the full representation of the filler or only to some of its features (and retrieval is then needed for the unmaintained content)? When considering this question in light of previous studies, the hybrid maintenance/retrieval account seems more plausible, as evidence indicates that at least some of the filler’s content is not kept active between the filler and gap sites. The most robust finding regarding this issue is that lexical priming effects for words semantically related to the filler are observed at the filler and gap sites, but not in between the two positions (Nicol et al., 1994;
18
Nicol & Swinney, 1989; McElree, 2001; Love & Swinney, 1996). This leads to the conclusion that between the filler and gap sites some of the filler’s features must be maintained (in order to affect interference, as observed in the current study) but the full semantic representation is not maintained active (at least not to the degree that it can prime related words). This semantic content is later retrieved at the gap site, when the priming effect occurs again. This adds up with Wagers & Phillips’s (2014) results that show a delayed semantic anomaly detection for long dependencies, which was interpreted as caused by the need to retrieve the filler's unmaintained semantic content, which decayed over the extended dependency. In fact, accumulating evidence suggests a three-way distinction between different features of the filler. While, as detailed above, some features are maintained and others decay and then retrieved at the gap, some features may only decay, without being retrieved later. This seems to be the case for the filler’s phonological form. The decay of the phonological form of a word was demonstrated by Hudson & Tanenhaus (1985), who showed rhyme priming occurring at a four-word distance from the prime, but not at a seven-word distance. Love and Swinney (1996) further demonstrated that in sentences where the filler is lexically ambiguous but placed in a biasing context towards a single meaning, probes related to both meanings are primed at the filler site, but only probes related to the context-relevant meaning are primed at the gap site. This indicates that the filler’s phonological form (which has both meanings) is likely not reactivated at the gap site. In the present experiment, the filler and the interfering NP shared syntactic category (NP) and animacy, as well as gender and number agreement features. Additionally, the two nouns were taken from the same semantic field. Based on this experiment it is thus not possible to pinpoint the specific type/s of maintained information, which engendered interference. Manipulating the number and type of similar features of the filler and the interfering NP and testing whether the interaction found in the current experiment still occurs may provide insights
19
regarding the specific features that are maintained, as only features that remain available in working memory can affect interference. Based on Wagers & Phillips’s (2014) results, manipulating the distance between the filler and the interfering NP may also be interesting in order to examine the changes to the representation that may occur over time.
Acknowledgements The authors would like to thank Maayan Keshev for extremely valuable discussions during work on the paper. Funding This study was funded by the EU Marie Curie Career Integration Grant No. 631512 (Aya Meltzer-Asscher).
5. References Adani, F., Van der Lely, H. K., Forgiarini, M., & Guasti, M. T. (2010). Grammatical feature dissimilarities make relative clauses easier: A comprehension study with Italian children. Lingua, 120(9), 2148-2166. Altmann, G. T. (1999). Thematic role assignment in context. Journal of Memory and Language, 41(1), 124-145. Aoshima, S., Phillips, C., & Weinberg, A. (2004). Processing filler-gap dependencies in a head-final language. Journal of memory and language, 51(1), 23-54. Arnon, I. (2005). Relative clause acquisition in Hebrew: Towards a processing-oriented account. In Proceedings of the 29th Boston University conference on language development (Vol. 29, pp. 37-48). Somerville, MA: Cascadilla Press. Baddeley, A. (1992). Working memory. Science, 255(5044), 556. Baddeley, A. D., Thomson, N., & Buchanan, M. (1975). Word length and the structure of short-term memory. Journal of verbal learning and verbal behavior, 14(6), 575-589. 20
Boland, J. E., Tanenhaus, M. K., Garnsey, S. M., & Carlson, G. N. (1995). Verb argument structure in parsing and interpretation: Evidence from wh-questions. Journal of Memory and Language, 34(6), 774. Chen, E., Gibson, E., & Wolf, F. (2005). Online syntactic storage costs in sentence comprehension. Journal of Memory and Language, 52(1), 144-169. Crain, S., & Fodor, J. D. (1985). How can grammars help parsers? In D. Dowty, L. Kartunnen, & A. Zwicky (Eds.), Natural language parsing (pp. 94–128). Cambridge: Cambridge University Press. Crain, S., & Fodor, J. D. (1993). Competence and performance in child language. in: E. Dromi (Ed.), Language and Cognition: A Developmental Perspective, Ablex, Norwood, N.J., 1993, pp. 141– 171. Fedorenko, E., Gibson, E., & Rohde, D. (2006). The nature of working memory capacity in sentence comprehension: Evidence against domain-specific working memory resources. Journal of Memory and Language, 54(4), 541-553. Fiebach, C. J., Schlesewsky, M., & Friederici, A. D. (2002). Separating syntactic memory costs and syntactic integration costs during parsing: The processing of German WHquestions. Journal of Memory and Language, 47, 250–272. Fodor, J. D. (1978). Parsing strategies and constraints on transformations. Linguistic Inquiry, 9(3), 427-473. Frazier, L., & Clifton Jr, C. (1989). Successive cyclicity in the grammar and the parser. Language and cognitive processes, 4(2), 93-126. Frazier, L., & Flores D’Arcais, G. B. (1989). Filler–driven parsing: A study of gap filling in Dutch. Journal of Memory of Language, 28, 331–344. Friedmann, N., Belletti, A., & Rizzi, L. (2009). Relativized relatives: Types of intervention in the acquisition of A-bar dependencies. Lingua, 119, 67-88.
21
Futrell, R., Mahowald, K., & Gibson, E. (2015). Large-scale evidence of dependency length minimization in 37 languages. Proceedings of the National Academy of Sciences, 112(33), 10336-10341. Garnsey, S. M., Tanenhaus, M. K., & Chapman, R. M. (1989). Evoked potentials and the study of sentence comprehension. Journal of Psycholinguistic Research, 18, 51–60. Gibson, E. (1998). Syntactic complexity: Locality of syntactic dependencies. Cognition, 68, 1–76. Gibson, E. (2000). The dependency locality theory: A distance-based theory of linguistic complexity. In A. Marantz, Y. Miyashita, & W. O’Neil (Eds.), Image, language, brain (pp. 95–126). Cambridge, MA: MIT Press. Gordon, P.C., Hendrick, R., & Johnson, M. (2001). Memory interference during language processing. Journal of Experimental Psychology: Learning, Memory and Cognition, 27, 1411-1423. Gordon, P. C., Hendrick, R., & Levine, W. H. (2002). Memory-load interference in syntactic processing. Psychological science, 13(5), 425-430. Hofmeister, P. (2011). Representational complexity and memory retrieval in language comprehension. Language and cognitive processes, 26(3), 376-405. Jackendoff, R. S., & Culicover, P. (1971). A reconsideration of dative movements. Foundations of Language, 397-412. Just, M. A., & Carpenter, P. A. (1992). A capacity theory of comprehension: individual differences in working memory. Psychological review, 99(1), 122. Kane, M. J., Bleckley, M. K., Conway, A. R., & Engle, R. W. (2001). A controlled-attention view of working-memory capacity. Journal of Experimental Psychology: General, 130(2), 169.
22
Keshev, M. & Meltzer-Asscher, A. (in press). Active dependency formation in islands: How grammatical resumption affects sentence processing. Language. King, J. W., & Kutas, M. (1995). Who did what and when? Using word-and clause-level ERPs to monitor working memory usage in reading. Journal of Cognitive Neuroscience, 7, 376–395. Kuznetsova, A., Brockhoff, P. B., & Christensen, R. H. B. (2014). lmerTest: Tests for Random and Fixed Effects for Linear Mixed Effect Models. R Package Version 2.0– 11. Available at: http://CRAN.R-project.org/package = lmerTest Lee, M.-W. (2004). Another look at the role of empty categories in sentence processing (and grammar). Journal of Psycholinguistic Research, 33, 51–73. Lewis, R. L., & Vasishth, S. (2005). An activation‐based model of sentence processing as skilled memory retrieval. Cognitive science, 29(3), 375-419. Love, T., & Swinney, D. (1996). Coreference processing and levels of analysis in objectrelative constructions; demonstration of antecedent reactivation with the cross-modal priming paradigm. Journal of Psycholinguistic Research, 25(1), 5-24. Mak, W. M., Vonk, W., & Schriefers, H. (2002). The influence of animacy on relative clause processing. Journal of Memory and Language, 47(1), 50-68. Mak, W. M., Vonk, W., & Schriefers, H. (2006). Animacy in processing relative clauses: The hikers that rocks crush. Journal of Memory and Language,54(4), 466-490. McElree, B. (2001). Working memory and focal attention. Journal of Experimental Psychology: Learning, Memory and Cognition, 27, 817–835. McElree, B., Foraker, S., & Dyer, L. (2003). Memory structures that subserve sentence comprehension. Journal of Memory and Language, 48(1), 67-91. Meyer, L., Obleser, J., & Friederici, A. D. (2013). Left parietal alpha enhancement during working memory-intensive sentence processing. Cortex, 49(3), 711-721.
23
Mitchell, D. C., & Cuetos, F. (1991). The origins of parsing strategies. In C. Smith (Ed.),Current issues in natural language processing. Austin: Center for Cognitive Science, University of Texas. Nairne, J. S. (1990). A feature model of immediate memory. Memory & Cognition, 18(3), 251-269. Nicol, J. L., Fodor, J. D., & Swinney, D. (1994). Using cross–modal lexical decision tasks to investigate sentence processing. Journal of Experimental Psychology: Learning, Memory & Cognition, 20, 1229–1238. Nicol, J., & Swinney, D. (1989). The role of structure in coreference assignment during sentence comprehension. Journal of Psycholinguistic Research, 18, 5–19. Oberauer, K., & Kliegl, R. (2006). A formal model of capacity limits in working memory. Journal of Memory and Language, 55(4), 601-626. Page, M., & Norris, D. (1998). The primacy model: a new model of immediate serial recall. Psychological review, 105(4), 761. Phillips, C., Kazanina, N., & Abada, S. H. (2005). ERP effects of the processing of syntactic long–distance dependencies. Cognitive Brain Research, 22, 407–428. Phillips, C., & Wagers, M. W. (2007). Relating structure and time in linguistics and psycholinguistics. In G. Gaskell (Ed.), Oxford handbook of psycholinguistics (pp. 739–756). Oxford: Oxford University Press. Pritchett, B. L. (1992). Parsing with grammar: Islands, heads, and garden paths. In Island Constraints (pp. 321-349). Springer Netherlands. R Development Core Team. (2011). R: A language and environment for statistical computing. Vienna, Austria: R Foundation for Statistical Computing. Retrieved from http://www.R-project.org/ Rohde, D. (2003). Linger: a flexible platform for language processing experiments.
24
Ruchkin, D. S., Johnson, R. Jr., Canoune, H., & Ritter, W. (1990). Short-term memory storage and retention: An event-related brain potential study. Electroencephalography and Clinical Neurophysiology, 76, 419–439. Sheppard, S. M., Walenski, M., Love, T., & Shapiro, L. P. (2015). The auditory comprehension of Wh-questions in aphasia: support for the intervener hypothesis. Journal of Speech, Language, and Hearing Research, 58(3), 781-797. Stepanov, A., & Stateva, P. (2015). Cross-linguistic evidence for memory storage costs in filler-gap dependencies with wh-adjuncts. Frontiers in psychology, 6. Stowe, L. (1986). Parsing wh–constructions: Evidence for on–line gap location. Language and Cognitive Processes, 1, 227–246. Sussman, R. S., & Sedivy, J. (2003). The time-course of processing syntactic dependencies: Evidence from eye movements. Language and Cognitive Processes, 18(2), 143-163. Traxler, M.J., Morris, R.K., & Seely, R.E. (2002). Processing subject and object relative clauses: Evidence from eye movements. Journal of Memory and Language, 47, 69-90. Traxler, M. J.,& Pickering, M. J. (1996). Plausibility and the processing of unbounded dependencies. Journal of Memory and Language, 35, 454–475. Traxler, M.J., Williams, R.S., Blozis, S.A., & Morris, R.K. (2004). Working memory, animacy, and verb class in the processing of relative clauses. Journal of Memory and Language, 53, 204-224. Van Dyke, J. A. (2007). Interference effects from grammatically unavailable constituents during sentence processing. Journal of Experimental Psychology: Learning, Memory, and Cognition, 33(2), 407. Van Dyke, J. A., & Johns, C. L. (2012). Memory interference as a determinant of language comprehension. Language and linguistics compass, 6(4), 193-211.
25
Van Dyke, J. A., & McElree, B. (2006). Retrieval interference in sentence comprehension. Journal of Memory and Language, 55(2), 157-166. Van Dyke, J. A., & McElree, B. (2011). Cue-dependent interference in comprehension. Journal of memory and language, 65(3), 247-263. Wagers, M. W., & Phillips, C. (2014). Going the distance: Memory and control processes in active dependency construction. The Quarterly Journal of Experimental Psychology, 67(7), 1274-1304. Wanner, E., & Maratsos, M. (1978). An ATN approach to comprehension. In M. Halle, J. Bresnan, & G. A. Miller (Eds.), Linguistic theory and psychological reality (pp. 119– 161). Cambridge, MA: MIT Press.
26
