Setting Syntactic Parameters with Implicit Negative Evidence: The Case of Zero-derived Causatives in English

PACLIC 28 !39 Setting Syntactic Parameters with Implicit Negative Evidence: The Case of Zero-derived Causatives in English [name] [address1] [address2] [address3] [email] Isaac Gould Department of Linguistics and Philosophy Massachusetts Institute of Technology igould@mit.edu [name] [address1] [address2] [address3] [email] Abstract In this paper, I introduce a learning challenge for various models of parameter setting in generative syntax, namely a scenario where all input to the learner underdetermines the target parameter setting. This scenario is exempli- fied by the case of zero-derived causatives in English, as discussed in Pylkkänen (2008). I then propose a model for parameter setting that uses a simple Bayesian learning proce- dure to learn from implicit negative evidence and arrive at the target parameter setting. 1 Introduction An important question in language learnability is how to converge on a target grammar when all rel- evant grammars are compatible with the input. In- deed, this is a general challenge for various prominent models of parameter setting in genera- tive syntax (e.g. Gibson and Wexler, 1994; Sakas and Fodor, 2001; and Yang, 2002). Consider a bi- nary Parameter P concerning the complement of a head X0: the complement could simply be YP (1a) or the more complex ZP containing YP (1b). (1) a. [XP X [ YP ] ] b. [XP X [ZP Z [YP Y ] ] ] Further, suppose that the target setting for a learner is the simpler structure in (1a), but that all the input the learner receives is ambiguous as to the para- metric choice in (1). In such a case, we can ask how the learner can be sure to arrive at the adult grammar of (1a). In this paper, I present a simple case study that illustrates the learning challenge in (1) with zero-derived causatives (ZDCs) in English under Pylkkänen’s (2008) theory of causatives. I propose a Bayesian model for parameter setting that learns the target setting from implicit negative evidence: given repeated instances of ambiguous input, the structure in (1a) has a greater likelihood of being the correct analysis. This result is a con- sequence of the learning process itself, and there is thus no need to invoke some principle such as the Subset Principle (Berwick, 1986), or to resort to default values for parameter setting. 2 The Learning Challenge with Zero- derived Causatives Pylkkänen observes that examples like (2a) are not ambiguous: only the causer John can be character- ized by gumpiness, not the causee Bill. This con- trasts with (2b), in which Bill’s action can be characterized by grumpiness. (2) a. John awoke Bill in a state of grumpiness. !John is grumpy (high reading) !Bill is grumpy (low reading) b. Bill awoke in a state of grumpiness. The question Pylkkänen asks is: if we follow Par- sons (1990) in assuming that causatives involve a causing and caused eventuality, why do the PPs in (2a) unambiguously modify the causing event (and thus the state of the causer) and not the caused eventuality?1 I call the possible adverbial interpre- tation in …

Content truncated to 500 words