Some remarks on the psycholinguistic relevance of LTAGs

Alexandra Kinyon (TALANA, University of Paris 7)

The psycholinguistic relevance of LTAGs has been addressed in the
past: e.g. (Frank 92) shows that the adjunction operation is
psycholinguistically relevant in the context of Children Language
Acquisition, (Joshi 90) discusses psycholinguistic results on crossed
and serial dependencies, (Kim et al. 97) discuss the psycholinguistic
relevance of lexicalization to capture preferences due to frequency.

In this paper we will discuss the psycholinguistic relevance of a "surfastic" TAG-based theory of syntax (Abeillé, Abeillé et al. 99) with respect to 3 points:

  1. Garden path (GP) phenomena: there is a continuum between sentences that are not not GPs and sentences that constitute strong GPs. LTAGs allow to predict accurately the degree of processing difficulty that will be encountered when parsing a sentence from left to right: while building a derivation tree, the more ancestors and descendants have to be modified, the more the sentence will be perceived as a GP. This vision of GP phenomena also allows to shed a new light on psycholinguistic results that have aimed at proving the existence of Wh-traces , but that can be explained in terms of GP effects, which we will discuss in the final version of this paper.
  2. Parsing preferences expressed in terms of derivation trees (i.e. dependancy information): These structural principles allow to account for 2 widely accepted types of preferences: argument vs. adjunct, idiomatic vs. literal interpretation and have given good results in parse ranking applications (Kinyon 99a) . Linguistic motivations can be found in (Kinyon 99b).
  3. Notion of treelets : Most psycholinguistic studies have been done in relation with a GB theory of syntax. (Fodor 98a,b], though, develops the notion of "treelet" to explain how we learn to parse and how we parse to learn. A treelet is "a small piece of a syntactic tree" that strikingly resembles the notion of "elementary tree", thus indicating that researchers from varied backgrounds tend to converge towards a LTAG-like conception of syntax.