Ambiguity, parsing, and the evaluation measure

ABSTRACT

An evaluation measure (EM) guides a learner’s choice of grammar when more than one is compatible with available input. EM must be universal, so children receiving comparable input acquire comparable grammars. It must favor the choices children actually make. The theoretical shift from rule-based grammars to principles-and-parameter-based grammars demanded a new conception of EM. I propose that it reflects biases of the sentence processing mechanism.

Notes

¹ Available at http://www.colag.cs.hunter.cuny.edu/.

² In this and other examples that follow, the syntactic properties of treelets are described in a very simple format for expository purposes, but I believe the general concept of parametric treelets is compatible with more sophisticated theories of syntactic tree structure.

³ Gagliardi, Mease & Lidz (2016) observe that early learners may understand some sentences without constructing a legitimate tree structure for them, relying instead on heuristics based on what they know of verb meanings and thematic roles. Presumably, at some point these heuristics prove inadequate, and tree construction must be undertaken.

⁴ In Sakas & Fodor (2012) we grappled with the fact that, due to uneven availability of triggers, for some parameters in the CoLAG domain, movement had to be regarded as the default setting, contrary to theoretical expectations. This does not bode well for the general claim about the EM that I wish to defend in the present article. However, it is quite possibly attributable to some quirks in the artificial domain that we constructed.

⁵ Han, Lidz & Musolino (2007) report an interesting case in which a parametric ambiguity in Korean (V-raising versus I-lowering) is not resolved by readily accessible input. Han, Lidz & Musolino show that Korean speakers, both adults and children, have more or less evenly divided opinions on which is correct. The lack of a preference (a between-parameter default) in this case is compatible with the fact that the two alternatives are roughly equal in derivational complexity.

⁶ For morphology and the lexicon, checking the weight of the evidence may be the only way to proceed. Yang (2016) presents language learning as “a search for productive generalizations,” governed by a Tolerance Principle that balances the benefit of the rule against the cost of memorizing exceptions to it.

⁷ A different association with semantics is suggested by Snyder (2016) who has proposed that “we count a parse as successful only when it allows a compositional interpretation that is compatible with the discourse context.” I think this is an excellent contribution to the learning-by-parsing approach and might mitigate some of the previous garden path worries.

⁸ Thanks to Dianne Bradley for pointing out the relevance of the syntactic priming findings to the activation-based treelet learning model I am proposing here. I should also acknowledge its affinity with the SRN model designed, implemented, and tested by Chang, Dell & Bock (2006), which is also founded on the assumption that “language learning and adult processing are part of the same mechanism” and also aims to reconcile lexically specific and structurally general syntactic learning.

⁹ See Marcus (2013), and references there, for evidence that some (much?) of the time it is not done accurately.