When does prosody influence parsing?

Duane G. Watson & Edward Gibson
Massachusetts Institute of Technology

dgwatson@psyche.mit.edu 

In this presentation, we investigate whether prosody is used in first pass parsing or in reanalysis in language comprehension.  A number of researchers have found that prosody reduces difficulty associated with dispreferred continuations of locally ambiguous structures, suggesting that prosody plays an early role in parsing (Speer & Kjelgaard, 1996; Marslen-Wilson et al., 1996; see Cutler et al., 1997, for a review).  However, it is not clear whether this reduction in difficulty is due to prosody's influence on first pass parsing or prosody's influence on subsequent reanalysis (Frazier & Clifton, 1996).

We address this question by testing unambiguous items using a cross-modal lexical decision task.  Participants listened to sentences like (1) up to the end of the direct object "picture".  Then they performed a visual lexical decision task for either a word which was unambiguously an argument of the immediately preceding noun ("of", an argument of "picture" in (1)) or a word which was unambiguously an argument of the verb ("to", an argument of "gave" in (1)).  The tone on the word "picture" was manipulated so that it signaled either the presence of an intonational phrase boundary (H* L-L%) or the absence of an intonational boundary (H*).  In verb-NP-PP sequences such as in (1), an intonational boundary typically occurs before the preposition in the VP attachment continuation, whereas a boundary usually does not occur before the preposition in the NP attachment continuation (Cooper & Paccia-Cooper, 1980; Price et al.  1991; Pynte & Prieur, 1996; Schafer et al., 2001).

If prosodic information is used in first-pass linguistic structure-building, lexical decisions will be faster for the locally attached preposition "of" than for the non-locally attached preposition "to" when there is no boundary tone on "picture".  The reverse pattern is predicted when a prosodic boundary is present: RTs to "to" should be faster than RTs to "of".  If prosody is accessed only after a syntax-first parser makes an error, no interaction is predicted since the items are unambiguous and at no point does reanalysis occur.

Analyses of the data revealed an interaction between the boundary and attachment site conditions (F1(1,27)=6.56, p<.05; F2(1,19) = 3.29, p=.08).  In particular, in the prosodic boundary conditions, RTs were slower for the locally attached preposition "of" than for the non-locally attached preposition "to" (601 ms vs. 576 ms), but in the no boundary conditions, RTs were slower for the non-locally attached preposition than for the locally attached preposition (613 ms vs. 598ms).  This result supports the hypothesis that prosodic information is used in first-pass linguistic structure-building, and is difficult to reconcile with the hypothesis that prosodic information is used only in later stages of linguistic structuring.

Example