Events

A defining feature of human language is the ability to interpret syntactic material in different positions from where they appear. For instance, "Who" is interpreted as the object of "kissed", even though it appears in a different clause:

(1) Who did Jonny tell Denny that Lisa kissed _?

It has been shown that resolution of long distance dependencies in sentence processing is an active, predictive process (Crain & Fodor 1985, Stowe 1986). However, it is unclear what guides the dependency resolution in ambiguous questions, when there are multiple grammatically acceptable parses available. For example, in (2), "where" may modify the main verb "say" or the embedded verb "hurt":

(2) Where did Emily say _ that she hurt herself _ ?

There might be a bias to resolve the wh-dependency in the same clause (a "structural locality bias"), or with the first verb in the sentence (a "linear locality bias"). English data do not distinguish these hypotheses, because the first verb and the main verb are the same. Studies in Japanese help disentangle these hypotheses, since the main verb is not the first verb in this language (Nakano et al 2002, Aoshima et al 2004, Omaki et al submitted). This work suggests that the "linear locality bias" is the correct characterization, since ambiguous wh-dependencies are resolved with the first verb. However, these data do not rule out that structural locality is irrelevant in selecting between potential parses, since such a bias might be eclipsed by a (stronger) linear locality bias.

This work presents novel data from Bengali/Bangla, in which the main verb may either precede or follow the embedded verb (Bayer 1996, 2001, Simpson and Bhattacharya 2000), shown in (3).

(3) a. Jonny ka-ke Denny-ke boleche [je Lisa _ cumu kheeche?]
Jonny who Denny told [that Lisa _ kiss ate]?
b. Jonny ka-ke [Lisa _ cumu kheeche bole] Denny-ke boleche?
Jonny who [ Lisa _ kiss ate that] Denny told?
'Who did Jonny tell Denny that Lisa kissed
?'

Our results confirm that, in offline judgments, there is a strong bias to resolve a long distance dependency with the first available verb, regardless of whether it is the main verb. However, in online results, the results are less clear, which suggests the strength of the "linear locality bias" may have been overstated.