Effects of NP type on reading English clefts

Tessa Warren$^1$, Edward Gibson$^2$, Mariko Jameson$^2$ and Christopher Hirsch$^2$
tessa@psych.umass.edu, egibson@mit.edu, mariko@mit.edu, ckh@mit.edu
1. University of Massachusetts at Amherst
2. MIT

This paper reports results from an experiment testing Gibson's Dependency Locality Theory (DLT; Gibson, 1998/2000; Warren&Gibson, 2002) against Gordon, Hendrick & Johnson (2001; GHJ)'s similarity-based interference hypothesis. In the experiment providing the strongest evidence for GHJ's hypothesis over the DLT, GHJ manipulated the types of NPs in subject- and object-extracted clefts between definite descriptions and names. GHJ found that in object-extracted clefts, conditions with matching NPs (name-name or description-description) were read more slowly on the verb than conditions with non-matching NPs. Warren & Gibson's (2002:W&G) DLT distance metric predicts a different pattern. W&G predict that at the verb, reading times should be slower for the description-description and name-description conditions and faster for the description-name and name-name conditions. This is because W&G hypothesize it is easier to build dependencies crossing NP types that are more given, and names are more given than descriptions.

However, there is a concern about GHJ's experiment: While typical self- paced reading RTs are in the 250-450 ms/word range, GHJ's reported RTs averaged around 500-600 ms/word for simple items, and up to 800-900 ms/word for complex items. It is possible that GHJ's methods may have induced participants to memorize the items during the processing of the sentences, leading to longer RTs.

The current self-paced word-by-word-moving-window experiment extended GHJ's experiment by including pronouns, a strong test case for both theories. The materials were object-extracted cleft sentences in a 3x3 design crossing three NP-types (definite descriptions, names and pronouns) in two positions (matrix subject and embedded subject).

(1)

It was {you/Patricia/the lawyer} who {we/Dan/the businessman} avoided at the party.

In contrast to GHJ's results, RTs averaged under 400 ms/word, even on the most complex regions, possibly because of the increased ratio of fillers to targets in the current experiment. Reading times at the critical region (the verb) were as predicted by W&G. RTs were fastest after pronouns, slower after names and slowest after descriptions. Contrary to GHJ's similarity-interference hypothesis, the pronoun-pronoun condition was numerically the fastest condition at the critical region. However, the pattern for conditions without pronouns provided some support for both theories. There was evidence of GHJ's interaction as well as a main effect of the type of NP2. Interestingly, off- line measures such as question-answering reaction times and error rates over every condition, including those with pronouns supported similarity based- interference. The implications of these results for the timing of different memory effects during reading will be discussed.