Context effects in coercion

Abstract:

One self-paced reading and one eye-movement-monitoring experiment investigated processing of sentences containing coercing expressions.  Verbs such as began, finished, and enjoyed can be used felicitously only if one of their arguments denotes an event or an activity (e.g., Jackendoff, 1997; Pustoevsky, 1997).  However, these verbs are commonly used in combination with nouns whose default interpretations are simple entities.  It has been previously shown that entity nouns, such as book, are difficult to process when they follow verbs like started (McElree et al., 2001; Traxler et al., in press).  This difficulty may occur because expressions like began the book involve semantic misanalysis and reanalysis, because they require selection between alternative interpretations, or because they require readers to interpolate semantic structure that is not explicitly signalled by the text.  Previous work on such expressions has concentrated on the processing of sentences in isolation.  In the current study, we examined processing of coercing expressions in contexts that instantiated their preferred interpretation.  In the eye-movement experiment, 18 readers read twenty-eight two-sentence passages (like 1a - d).  The first sentence either mentioned an activity (like writing) that was the preferred interpretation of the coercing expression (like the author started the article) or an unrelated but plausible activity.  The second sentence either contained a coercing expression (like started the article) or an unambiguous expression of the same meaning (like wrote the article).

This led to a 2 (context: related vs. unrelated) x 2 (target: coercing vs. control) design, with all factors within subjects and items.  If coercing expressions require readers to interpolate additional semantic structure, then coercing expressions should be difficult (relative to controls) regardless of the context that precedes them.  This account would predict a main effect of target, with coercing expressions taking longer to process than controls.  If coercing expressions are difficult because readers must infer the activity represented by the ambiguous target expression, then explicitly mentioning the activity in the context sentence should reduce or eliminate the difficulty associated with coercing expressions.  This account would predict an interaction of context and target types.  The eye-movement data showed that coercing expressions were more difficult to process than controls, regardless of context.  When total reading time data from the target noun (e.g., article) were analyzed, a main effect of target was obtained in the by-subjects analysis [coerced: 406 ms; control: 372 ms; F1 (1, 17) = 4.66, p<.05, MSe = 4505], which was not significant in the by-items analysis [F2 (1, 27) = 2.58, p=.12, MSe = 11526].  No other main effects or interactions were obtained.  When the entire verb-phrase (e.g., started the article/wrote the article) was analyzed, total reading time data produced a robust effect of target [coerced: 742 ms; control: 688 ms; F1 (1, 17) = 11.9, p < .01, MSe = 15857; F2 (1, 27) = 9.49, p < .01, MSe = 27714].  No other main effects or interactions were significant (all F<1.5).

The self-paced reading experiment had the same design as the eye-movement experiment.  The two differed in that the self-paced reading materials had an additional sentence inserted between the context and target sentences.  The self-paced reading experiment did not produce any effects reliable by both subjects and items at the target word (e.g., article).  However, analyses on the self-paced reading data from the word following the target word produced an effect of target type [coerced: 457 ms; control: 430 ms; F1 (1, 35) = 7.72, p < .01, MSe = 3416; F2 (1, 26)* = 3.97, p = .06, MSe = 4190].  As in the previous experiment, coercing expressions (e.g., started the article) were more difficult to process than control expressions (e.g., wrote the article), and this difficulty was not moderated by context (both F<1, NS).

These data therefore favor accounts under which coercing expressions are interpreted via a structural reconfiguration, which requires interpolating semantic structure that is not explicitly represented in the text.  The data are less compatible with an account under which a general-purpose inferencing mechanism assigns an interpretation to the ambiguous string.

* 1 item was eliminated from the analyses due to an error in the display.