Recycling prosodic boundaries

Yuki Hirose
University of Electro-Communications, Tokyo

hirose@cs.uec.ac.jp

The idea of implicit prosody (Bader, 1998; Fodor, 1998) assumes that prosodic structures as well as syntactic structures are computed in silent reading.  Our self-paced reading experiments using Japanese relative clause (RC) constructions reveal: (a) unpreferred RC analyses can be facilitated and (b) boundary ambiguity resolution in reanalysis are influenced, by the prosodic structure created on the initial syntactic analysis.

The word string up to the first verb ("found") in (1) is temporarily ambiguous between the simplex clause (initially preferred) and RC analyses, until the head noun (HN) is encountered. In auditory input, the left RC boundary position ( [ ) is often marked by a major prosodic phrase (MajP) boundary, eliminating the ambiguity. For visual input, such information isn't available and a garden path occurs at HN ("researcher"), when the initial (simplex) analysis is falsified.

Experiment 1 tested pairs such as (2) and (3), where MajP phrasings are manipulated by the number of phrases and lexical accents (').  Our phonetic studies on spoken word strings up to the first verb confirmed a MajP boundary occurs following the subject (conjoined accented nouns) in (2), whereas, no MajP occurs within the sequence (where the subject is an unaccented noun) in (3).  Phrase-by-phrase self-paced (silent) reading times for the HN were significantly faster in (2) than (3) (despite the number of words to process and possible memory load increase), demonstrating the RC analysis was facilitated when the input provides an environment that elicits a MajP boundary during the initial parse.

Even at HN, there remains an ambiguity regarding the RC boundary position, as in (1) versus (4), although (1) is structurally preferred.  Experiment 1's results lead to a prediction that ambiguity resolution can be influenced by the MajP boundary position postulated during the initial parse; specifically, by the alignment (cf. misalignment) of MajP with the correct left RC edge.  Pairs such as (5) & (6) and (7) & (8) were subjected to further self-paced (silent) reading experiments, where the final disambiguating phrases (underlined) were of interest.  In (5)-(8), all items prior to the first verb are accented and MajP boundaries predicted to occur during the initial analysis according to our phonetic studies are marked as {MajP.  As predicted, in Experiment 2, reading times for NP-ACC were significantly faster for (5) than (6), with the initial-parse prosody reinforcing the preferred RC analysis.  In Experiment 3, there was a trend for the final verb in (8) to be read faster than (7), where the syntactic preference is presumably counteracted by the initial-parse prosody.  Implications of our results for the mechanism of "recycling" prosody will be discussed.

Examples