36 D.Reirter.I.D.Moore/Journal of Memory and Language 76 (2014)29-46 乙pgg case (PP):-0.35.F ffect correlations between all variab es was lower tha 025.ANOVA F-values shown. Covariate OR F p(> 2099 001 0.000 26 (FEQ) n(DisT):C 8 CP.Map Task PP.Map Task etitio they are 0.165.and 0.141.respectively.Th make To derive these -0.30-0.25-0.20-0.15-0.10-0.050.00 We of Cdactorsandverager vn in Table be with a preposition (such as below that bend there )The of this Results Refer to Table 3.The estimate for In(Disr)describes the rime- arget dis of 1 and 0.120 at 9s Fo se values would be 0.158and0.109 M DisT 图品光芳 Discussion -0.106 d.(Ber and PF Experiment 1 holds:there is syntactic priming in both strong in task lends support to our hypothesis:evi 006 dence for stronge priming whe speakers Fig.2 contrasts different effect sizes.that is.estimatesof in the task oriented Map Task co us than in the spontane ous conversations of Switchboard. intervals as well as the model suggest that priming Discussion:results and methods data mpled toprovide an overa on prob erances rules,and bet repetition probabilities are dshor the corpora of spoken dialogue that we invest on svntactic repetition probability thus providing p=219.and at distances of 8-10s.0.143.For Switch- evidence for a structural priming effect for arbitrary syn-Results Refer to Table 3. The estimate for lnðDistÞ describes the slope of repetition probability over time for the baseline condition, that is, in Switchboard. We find a main effect of lnðDistÞ (b ¼ 0:165; p < :0001). This indicates priming in Switchboard. lnðDistÞ interacts with MAPTASK (b ¼ 0:058; p < :001), indicating reliably stronger prim￾ing in Map Task. As before, lnðDistÞ also interacted with lnðFreqÞ (b ¼ 0:092; p < :0001), i.e., priming is stronger for less frequent rules. An interaction between lnðDistÞ, CP and MAPTASK (b ¼ 0:106; p < :005) documents that there is a larger gap between CP and PP priming in Map Task than in Switchboard. (Between-speaker priming is strong in task-oriented dialogue, but not in spontaneous conversation, first reported in Reitter, Moore, & Keller (2006).) Fig. 2 contrasts different effect sizes, that is, estimates of priming strengths (lnðDistÞ interactions) for the four factor combinations of CP and SOURCE. The post hoc confidence intervals as well as the model suggest that priming between speakers (comprehension-production) may be stronger than priming within a speaker (production-pro￾duction) for Map Task only. To illustrate the relative magnitude of the effects, we give conditional repetition probabilities in our data. Recall that these data were resampled to provide an overall higher proportion of repetitions to facilitate model fitting. The average repetition probability is 0.170 for all samples from Map Task, and 0.156 in Switchboard. In Map Task, in the first two seconds after a prime, over all syntactic rules, and between speakers, repetition probabilities are p ¼ :219, and at distances of 8–10 s, 0.143. For Switch￾board, they are 0.165, and 0.141, respectively. That is, rep￾etition is not only more common in Map Task, but, crucially, its drop-off is greater. The model itself can make predictions. To derive these, effects and interactions have to be combined in logit-space, taking into account centering and log-transformations. (We assume average random effects.) For illustration pur￾poses, we chose the rule S ? PP S, which licenses a clause beginning with a preposition (such as below that bend there is an abandoned cottage). The frequency of this rule is 63% of that of the mean frequency (which is still more common than 87% of all rules in Map Task). In the Map Task data, for priming between speakers, the model’s prediction for rep￾etition of that rule in regular probability space is p ¼ :173 at a prime-target distance of 1 s, and 0.120 at 9 s. For Switchboard, these values would be 0.158 and 0.109, respectively. (Many constructions most commonly exam￾ined in priming studies, such as passives, are very rare in our speech corpora.) Discussion The model based on temporal distance confirms the earlier model based on utterances. The basic result from Experiment 1 holds: there is syntactic priming in both corpora. The experiment lends support to our hypothesis: evi￾dence for stronger syntactic priming when speakers engage in purpose-driven conversation. Priming is stronger in the task-oriented Map Task corpus than in the spontane￾ous conversations of Switchboard. Discussion: results and methods In summary, reliable syntactic priming effects can be detected in natural dialogue for general syntactic rules instead of selected constructions. We model syntactic priming as the decay of repetition probability of syntactic rules, both in the course of linguistic activity (utterances), and over time. Both of the corpora of spoken dialogue that we investi￾gated showed an effect of distance between prime and tar￾get on syntactic repetition probability, thus providing evidence for a structural priming effect for arbitrary syn￾Table 3 The regression model for the joint dataset of Switchboard and Map Task (Experiment 2), prime-target distance DIST in seconds. This is the minimal model without unjustified covariates. All variables were centred. Random variables for intercept and lnðDistÞ, grouped by utterances. All continuous variables were centred; CP and lnðFreqÞ are residuals after regressing out effect of lnðDistÞ; resulting coding: MAPTASK: 0:51 vs. base case (Switchboard) 0:49, CP: 0:65 vs. base case (PP): 0:35. Fixed-effect correlations between all variables was lower than 0:25. ANOVA F-values shown. Covariate b OR SE F z pð> jzjÞ Intercept 2:096 0.12 0:010 200.6 < 0:0001 lnðDistÞ 0:195 0.84 0:011 82.7 17.3 < 0:0001 CP 0:263 0.83 0:014 304.5 18.8 < 0:0001 MAPTASK 0:054 1.06 0:015 2.6 3.49 < 0:001 lnðFreqÞ 0:759 2.14 0:007 10388.8 102.2 < 0:0001 lnðDistÞ: CP 0:033 1.02 0:019 5.4 1.77 < 0:10 lnðDistÞ: MAPTASK 0:058 0.94 0:017 12.6 3.35 < 0:001 CP: MAPTASK 0:166 0.85 0:028 35.1 5.93 < 0:0001 lnðDistÞ: lnðFreqÞ 0:092 1.10 0:009 113.3 10.65 < 0:0001 lnðDistÞ:CP:MAPTASK 0:106 0.90 0:037 8.2 2.87 < 0:005 Fig. 2. Relative Decay effect sizes in logits for lnðDistÞ with different combinations of CP and SOURCE factors and average residual frequency, based on model shown in Table 3. Longer bars indicate stronger decay and priming. Error bars show standard errors. 36 D. Reitter, J.D. Moore / Journal of Memory and Language 76 (2014) 29–46