Sample analysis scripts and data sets collected from graduate students are included in the supplementary materials. In the following, we present the results from sample01a.

For our regression, a 1 encodes an initial actor, while a 0 encodes an initial undergoer. Because of the mutual exclusivity of the actor-undergoer relation, we do not need to encode the other argument. We chose one to correspond to initial actor so that more prominence would correlate positively with more actorhood. Table 2 provides the results of the regression with sample01a.

| **Estimate** | **Std. Error** | **z value** | **Pr( > |z|)** |

(Intercept) | 0.2878 | 0.2064 | 1.4 | 0.163 |

animacy | 0.6269 | 0.167 | 3.8 | 0.0001743 |

case | 1.3480 | 0.2318 | 5.8 | 6.014e-09 |

number | −0.4678 | 0.1596 | −2.9 | 0.003381 |

index | 0.0005 | 0.001767 | 0.3 | 0.7614 |

Table 2 **AIC**: 215.6 **Deviance**: 205.6 **Null Deviance**: 263.7 Results of probit regression for actor-initial order

While the exact meaning of the estimates in probit regression is difficult, the relationship in the size of the estimates is straight forward. Case clearly has the strongest estimate, which fits well given that unambiguous case marking is known to work deterministically in German. Case also has the least amount of variance relative to its influence – this is reflected in the large *z*-score. Animacy has the next highest estimate and *z*-score, with both about half as large as for case. Number has the estimate and *z*-score with the smallest magnitude. Interestingly, the sign is also reversed for number. This could reflect the late disambiguating nature of number agreement in the verb final sentences. Index (i.e. trial number within the experiment) has a very small estimate and *z*-score, which indicates that the test subject was unable to develop and apply a strategy during the course of the experiment.

Figure 1 Actor identification by feature

This is also clear graphically. Figure 1 presents the likelihood of choosing an initial actor based on the difference in prominence. High initial prominence followed by low initial prominence – the high end of the scale – increases the odds of assigning actorhood to the first argument. (Shaded regions indicate the 95% confidence interval; the number of samples of each condition is shown as a rug plot.) The strength of a cue is reflected in the slope of the individual lines. The preference for an initial actor can also be clearly seen here. At 0, i.e. at a tie in prominence between the two arguments, all features show a preference for an actor-initial interpretation. Despite the low power from a short, unbalanced design, a clear ranking is visible.

Figure 2 Convergence of estimates

Figure 3 Effect of estimate precision on actor identification by feature

The emergence of such a clear ranking is also interesting for exploring the learnability of the actor strategy. Figure 2 shows the convergence of the parameter estimates by recomputing the model for the first *n* trials, starting with trial 25. Despite the low statistical power in such an experiment, the parameter estimates quickly sort themselves into a clear ranking. Figure 3 shows how this learning would look in terms of language processing, displaying the identification curve as in Figure 1, but after 50, 100, 150 and 200 trials. Again, after 50 trials, the strength of case is established, but the strength of the next strongest cue, animacy, is established after twice as many (100), and the third strongest cue settles down between 150 and 200 trials (1.5–2x as many as animacy), suggesting perhaps a rank-power law in cue strength (such as in Zipf’s Laws).

Figure 4 Individual actor space

Although we did not model separate interaction terms here, it is nonetheless interesting to consider how the model handles interaction. For this, we used the model to predict outputs for the grid of animacy × case, treating both as semi-continuous measures. Figure 4 shows the contour for this simulation holding constant index = 1 and number = 0 (plural). Darker tones indicate a higher probability of initial actor interpretation.
Using a physical metaphor, we can view the darker tones as being valleys and the lighter tones as being hills. Actorhood works as an attractor, with the ideal attractor basin being an initial, animate, nominative argument. However, even an inanimate initial nominative is more likely to be interpreted as an actor than an animate initial accusative – the basin slopes more sharply along case than along animacy.

Figures and models for four test subjects can be found in the Appendix.

** ** | **Estimate** | **Std. Error** | *t* value | **Pr( > |***t*|) |

(Intercept) | 863.2570 | 57.57 | 15 | 2.618e-34 |

animacy | −48.1050 | 44.68 | −1.1 | 0.283 |

case | −66.2133 | 56.59 | −1.2 | 0.2434 |

number | −16.8376 | 42.74 | −0.39 | 0.694 |

index | −1.8386 | 0.4956 | −3.7 | 0.0002706 |

Table 3 **AIC**: 2973.8 **Adjusted ***R*^{2}: 0.1 **Residual standard error**: 402.7 on 195 degrees of freedom, *F*(4, 195) = 3.96, *p* = 0.0041. Results of linear regression model for reaction time

## Comments (0)