P600 Dominance Predicts Comprehension of Garden-Path Sentences

1. Introduction.   Research using the event-related potential (ERP) technique has provided many important insights into the neural mechanisms associated with language comprehension. The integration of lexico-semantic information is associated with an increased centro-parietal negativity between 300-500 ms known as the N400 (Kutas & Federmeier, 2011; Kutas & Hillyard 1980). Morphosyntactic integration is associated with an early left anterior negativity (LAN) maximal around 200-500 ms, followed by a late posterior positivity (P600) maximal between 500 and 800 ms (see Kutas, Van Petten & Kluender (2005) for review). The P600, in absence of early negativity, is also elicited by well-formed sentences that present increased difficulty due to temporary ambiguity (i.e. garden-paths; Gouvea, Phillips, Kazanina & Poeppel, 2010; Osterhout, Holcomb & Swinney, 1994). The use of ERP as a means of indexing the different neural mechanisms associated with language processing is contingent on the assumption that all neurologically normal, native speakers show consistent responses to sentence stimuli such that the grand averaged ERPs reflect effects that are manifest uniformly across individuals. This notion was recently challenged by Tanner and Van Hell (2014). In their innovative study, they showed that, although in the grand mean syntactic violations elicited a classic biphasic LAN/P600 response, most participants either showed an N400 or a P600 rather than a biphasic response. Given the topographical distribution of the effects for each group, they concluded that the LAN often found for syntactic violations in grand mean analyses is the result of the distributed negativity in some subjects being neutralized or minimized by the right lateralized positivity in the others such that only the left anterior negativity remains. Response dominance did not, however, predict acceptability judgment accuracy, nor did it correlate with measures of working memory (WM) and executive control.The individual differences in N400/P600 response dominance observed by Tanner and Van Hell (2014) lead to interesting questions regarding other contexts which tend to elicit these potentials. Garden-path sentences are known to elicit P600 effects in absence of early effects but there is some variability (Friederici, Mecklinger, Spencer, Steinhauer & Donchin, 2001; Gouvea et al., 2010; Horberg, Koptjevskaja-Tamm & Kallionen, 2013; Matzke, Mai, Nager, Russeler & Munte, 2002; Vos, Gunter, Schriefers & Friederici, 2001).  This variablity could suggest the possibility of individual differences in response profiles, as Tanner and Van Hell (2014) found for syntactic violations.One known source of variability in garden-path effects for both P600s and comprehension accuracy is working memory capacity (WMC). High WMC individuals show greater P600 effects for garden-path sentences compared to low WMC individuals (Friederici, Steinhauer, Mecklinger, & Meyer, 1998). High WMC individuals also show reduced garden-path effects in comprehension accuracy such that they have better comprehension accuracy for garden-paths (Just & Carpenter, 1992). Lower comprehension accuracy in low WMC individuals indicates they are more likely to arrive at “Good Enough” interpretations (Ferreira, Bailey & Ferraro, 2002) in which the faithful interpretation of the sentence is not adopted.In the current study we applied the RDI analysis to the ERPs associated with garden-path sentences in order to determine (1) if participants’ N400/P600 dominance for garden-path sentences will fall into a continuum such that there will be a continuous distribution of N400 and P600 effect magnitudes with negative correlations between them, (2) if response dominance will predict comprehension accuracy, and (3), if so, is that effect reducible to individual differences in WMC.    2. Methods.2.1. Participants.   Data were collected from 62 right handed participants, 25 of which were excluded due to eligibility issues, technical issues, noncompliance, or excessive artifacts. As a result, 37 participants (20 female) between the ages of 18 and 35 ( M = 21.6, SD = 3.21) were included in the analysis. All participants were right-handed, neurologically normal, native speakers of English with normal or corrected-to-normal vision, and none had had started learning a second language before age 12. 2.2. Sentence Stimuli This experiment used the same control and garden-path sentences as O’Rourke & Colflesh (2014) (based on Gouvea et al., 2010). See sentences (1) and (2) for examples of garden-path and control sentences, respectively.The patient met the doctor and the nurse with the white dress showed the chart during the meeting. The patient met the doctor while the nurse with the white dress showed the chart during the meeting.There were 36 sentences per condition and an additional 288 sentences including fillers and conditions not presented herein.  Fifty percent of the sentences were followed by a yes/no comprehension question. 2.3. Complex Span Tasks. As indices of WMC, three complex span tasks were used in the current study: reading span (Daneman & Carpenter, 1980; Unsworth, Heitz, Schrock & Engle, 2005), operation span (Unsworth, et al., 2005), and symmetry span (Unsworth, Redick, Heitz, Broadway, & Engle, 2009). In the reading span task participants were presented with a series of sentences and asked to indicate, via button press, if the sentence they read made sense. After each sentence they were then presented with a letter that they were to remember for later recall. At the end of the sequence, they had to recall the letters in serial order. Their score reflects the total number of letters recalled in the correct serial position out of a total of 75 items. Operation span was identical to reading span as described above except instead of making sense judgments on sentences, participants had to read math problems involving two operations, one addition/subtraction and one multiplication/division, and verify if the solution provided was correct.  Symmetry Span (Engle, 2005) is a complex span task like the aforementioned tasks, but it uses visuospatial stimuli. Participants were presented with a series of 8x8 black and white grids and asked to indicate, via button press, whether the design was vertically symmetrical. After each symmetry judgment they were presented with a 4x4 grid with a square filled in red that they were asked to remember for later recall. At the end of the sequence, participants had to recall the position of the red squares, in the order in which they appeared. The maximum score was 42. 2.5. Procedure. Electroencephalographic (EEG) data was recorded using the Electrical Geodesics Inc. (EGI) Hydrocel 256 channel system while participants performed the sentence processing task.  Sentences were presented word-by-word and participants responded to the comprehension questions with a button press.  Data was collected over two sessions.  Upon completion of the sentence processing task, participants performed the working memory assessments. 2.6. Data Analysis. Upon completion of pre-processing and averaging, ERPs were computed for each individual for each experimental condition for a 1500 ms interval time-locked to the presentation of the critical verb (“showed” in the examples above) relative to a 200 ms pre-stimulus baseline. The following time windows were considered in the analysis of P600 effects: 300-500, 500-700 and 700-900 ms. The analyses were performed on midline and dorsal and electrodes. The midline electrodes were divided into anterior (FPZ, AFZ, FZ, FCZ, CZ) and posterior (CPZ, 90, PZ, POZ, OZ) sections. The dorsal electrodes were grouped by anterior-posterior (AP) location and hemisphere:  Left anterior (FP1, AF3, F1, F3, FC3, C3), right anterior (FP2, AF4, F2, F4, FC4, C3), left posterior (CP3, CP1, P1, P3, P1, PO3, O1) and right posterior (CP4, CP1, P4, P2, PO4, O2). Sentence type effects in the ERP data were assessed in the dorsal regions with multiple three-way analysis of variance (ANOVA) (Sentence Type x AP x Hemisphere) and in the midline electrodes with a two-way ANOVA (Sentence Type x AP).  RDI was then calculated using Tanner & Van Hell (2014)’s formula, using the same centro-parietal region of interest and time windows. Participants were divided into groups according to response dominance (N400 or P600).  An analysis of covariance (ANCOVA) was run with garden-path comprehension accuracy as the dependent variable, RDI group as the independent variable and average, standardized WM score (average z-score for the three measures) as the covariate.3. Results.3.1. Sentence Type Effects. Accuracy for garden-path sentences ( M = 68.3%, SD = 14.4) was significantly lower than control sentences ( M = 73.6%, SD = 11.4; F (1,36) = 6.13, p < .05, η p 2 = .15).  In the ERP data, garden-path sentences (compared to controls) showed a significant interaction of Type and AP over midline sites in the 500-700 and 700-900 ms time windows ( F (1,36) = 4.18, p < .05, η p 2 = .10 and F (1,36) = 6.02, p < .05, η p 2 = .14, respectively) such that garden-paths elicited greater positivity than control sentences over posterior sites. Simple comparisons showed significant effects of type in posterior areas in both the 500-700 ( F (1,36) = 4.14, p < .05, η p 2 = .10) and 700-900 time windows ( F (1,36) = 4.93, p < .05, η p 2 = .12). There were no effects in the anterior sites.3.2. RDI Analysis. Analysis of N400 and P600 effect magnitudes for garden path sentences showed a strong negative correlation ( r (32) = -.90; p < .001). The data suggest a continuum between strong N400 and P600 dominance. Participants were divided into groups based on RDI values (negative values indicating N400 dominance and positive indicating P600 dominance). A total of 18 participants were N400 dominant and 19 were P600 dominant.Prior to running the ANCOVA, it was necessary to determine that the covariates affected the dependent variable equally across the two groups. In the entire sample, there was a significant correlation between average complex span score and garden-path comprehension accuracy ( r (32) = .52, p < .01). Each RDI group showed positive correlations (N400 dominant, r (14) = .63; P600 dominant, r (16) = .48). Using a Fisher transformation (Fisher, 1915), the difference between the group correlations was not significant ( z = .6, p > .50).The ANCOVA showed that there was a significant effect of Response Dominance on GP Comprehension Accuracy after controlling for WMC, F(1,31) = 4.45, p < .05, η p 2 =.13) such that P600 dominant individuals had greater accuracy.  Complex span performance accounted for a significant amount of variance (F(1,31) = 13.8, p < .005, η p 2 = .31).4. Discussion. The current study found evidence of distinct neural response profiles which were not apparent in the grand averaged data in neurologically normal, native English speakers during the processing of garden-path sentences and this individual differences measure predicted comprehension performance. The key finding of the current study is the effect of response dominance on behavioral performance. Response dominance emerged as an effective predictor of comprehension accuracy such that P600 dominant participants had better comprehension accuracy for garden-path sentences. The results of the ANCOVA show that response dominance is not a proxy for WMC but rather a distinct individual difference measure. This suggests that cognitive capacity alone does not limit the individual’s ability to resolve garden-paths. Response dominance may, instead, indicate the engagement of specific parsing strategies.  The results of the current study extend the utility of Tanner and Van Hell (2014)’s RDI as an individual difference to the processing of the garden-path sentences showing that individuals in the sample exhibited distinct response profiles (either N400 or P600 dominant). While future research will reveal the neurocognitive underpinnings of response dominance, the findings of the current study establish this individual difference measure as a means of predicting behavior from neural activity.