Grammatical and Coherence-Based Factors in Pronoun Interpretation

Grammatical and Coherence-Based Factors in Pronoun Interpretation Laura Kertz (kertz@ling.ucsd.edu) Andrew Kehler (kehler@ling.ucsd.edu) Department of Linguistics, 9500 Gilman Drive La Jolla, CA 92093-0108 USA Jeffrey L. Elman (elman@crl.ucsd.edu) Center for Research in Language, 9500 Gilman Drive La Jolla, CA 92093-0515 USA Abstract We describe a pronoun interpretation experiment in which a Coherence Hypothesis is tested against preference-based analyses of pronoun interpretation, including the Parallel Function Preference and the Subject Preference. We demonstrate that preferences can be systematically disrupted through the manipulation of coherence, and that only the Coherence Hypothesis can predict the full range of coreference patterns observed. We argue, following Kehler (2002), that apparent preferences follow from inferencing processes which support different types of coherence relation. Introduction Three decades of psycholinguistic research into pronoun interpretation have documented a broad range of biases and effects linked to syntax, semantics, pragmatics, and more. This has led some researchers to posit pronoun interpretation ‘heuristics’, that is, broadly-applicable strategies for matching pronouns to their antecedents. Of particular interest to this study are two proposals which gained prominence in the 1990’s and which continue to play a role in theories of pronoun interpretation today: the Parallel Function Preference (Smyth 1994, Chambers & Smyth 1998, inter alia) and the Subject Preference (Crawley & Stevenson 1990, inter alia). These preferences, however, are often in conflict, and each seems to surface only under certain conditions. Acknowledging this, proponents of preference-based models typically identify a basic antecedent selection preference, and then augment it with additional syntactic or semantic mechanisms. In this paper we offer an alternative analysis, following Kehler (2002), in which preference conflicts are best understood as side-effects of the establishment of different types of coherence. Parallel Function Preference The Parallel Function Preference holds that pronouns are preferentially resolved to antecedents that occupy a matching argument position, for example, subject pronouns prefer subject antecedents, and object pronouns prefer object antecedents, as demonstrated in (1-2) below. Samuel threatened Justin with a knife, and (1) … he blindfolded Erin with a scarf. [=Samuel] (2) … Erin blindfolded him with a scarf. [=Justin] However, the Parallel Function Preference fails in (3), where the object pronoun refers to a subject antecedent. (3) Samuel threatened Justin with a knife, and Erin stopped him. [=Samuel] In light of examples like (3), Smyth (1994) suggests a modification to the Preference, which requires a perfect match between the two clauses in both argument structure (number and type of arguments) and sentence structure (presence/absence of adjunct modifiers). In cases in which these conditions are not met, a subject preference kicks in, supporting resolution to a subject antecedent. This Qualified Parallel Preference makes the correct prediction in (3), because the first clause contains an adjunct modifier which is not matched in the second clause. It does not account for cases like (4), however, where the structures are fully parallel. The wrong prediction is made for (5) as well, where a lack of parallel structure does not trigger the subject preference. In both (4) and (5), the subject pronoun instead refers to an object antecedent. Samuel threatened Justin with a knife, and (4) he alerted security with a shout. [=Justin] (5) he alerted security. [=Justin] Note that neither the basic Parallel Function Preference nor its more restricted variant takes into account the semantic cues in (3-5) which support non-parallel reference. Subject Preference The Subject Preference, in its basic form, holds that pronouns are preferentially resolved to subject antecedents. The account assumes increased salience for arguments in subject position, which guides antecedent selection for both subject and object pronouns, as seen in (6-7). Samuel threatened Justin with a knife, and (6) … he blindfolded Erin with a scarf. [=Samuel] (7) … Erin stopped him with pepper spray. [=Samuel] The Subject Preference, as formulated by Crawley & Stevenson (1990), however, will not override explicit cues to antecedent reference, for example, pronoun gender. Nor is the Subject Preference predicted to override a resolution to a more plausible referent, as in (4-5) above, where the semantics of the sentence show a causal bias toward the non-subject antecedent. Note, however, that the Subject Preference does not account for cases like (2), where no such bias is present.