Pronoun Interpretation as a Side Effect of Discourse Coherence

Pronoun Interpretation as a Side Effect of Discourse Coherence Hannah Rohde (hannah@ling.ucsd.edu) Andrew Kehler (kehler@ling.ucsd.edu) Department of Linguistics; University of California, San Diego 9500 Gilman Drive, La Jolla, CA 92093-0108, USA Jeffrey L. Elman (jelman@ucsd.edu) Department of Cognitive Science; University of California, San Diego 9500 Gilman Drive, La Jolla, CA 92093-0515, USA Abstract over Sources) and an event-structure hypothesis (a bias to focus on the end state, where the Goal is assumed to be more salient than the Source in transfer passages). Rohde et al.’s aspect manipulation distinguished between these two explanations because the thematic roles remain the same in (1) and (2), but the perfective verb in (1) describes a completed event, which is compatible with end state focus, while the imperfective verb in (2) describes an ongoing event that lacks an end state. They found that perfective sentences yielded more Goal continuations than imperfective sentences, thereby supporting Stevenson et al.’s event-structure hypothesis. Rohde et al.’s work further supports a model which incorporates deeper discourse-level factors. Specifically, they found that the influence of event structure was observed only in passages in which certain relationships could be inferred to hold between the two clauses (see Arnold 2001 for similar results). These results predict that a shift in the distribution of such relationships (henceforth ‘coherence relations’) ought to induce a shift in the distribution of pronoun interpretations. If people mainly process ambiguous pronouns using surface-level biases, one might expect their continuations to expose these biases since they are free to write continuations that are consistent with the preferred interpretation. If, on the other hand, participants are sensitive to coherence-driven factors that allow them to generate predictions about where a discourse is going, they may use the pronoun differently depending on the context. In this paper, we present two experiments that manipulate the coherence distribution in story continuations by making only minimal changes to Rohde et al.’s original stimuli. Our results show that the pattern of pronoun interpretation corresponds directly to the distribution of coherence relations. No model of interpretation that relies entirely on surface-level cues can account for these results. From the story continuations, we derive estimates for the likelihood that a pronoun refers to a specific referent given a coherence relation and discuss how conditional probabilities (p(referent|coherence)) can lay the groundwork for the an expectation-driven model of pronoun interpretation. Recent story completion studies (Arnold 2001, Rohde et al. 2006) show that passages exhibiting different coherence relations yield different patterns of pronoun interpretation. These results predict that a shift in the distribution of coherence relations in participant responses ought to induce a shift in the distribution of pronoun interpretations. Experiment 1 manipulates the coherence distribution by varying the direct object in Rohde et al.’s original stimuli. We find that the likelihood of a pronoun referring to a specific referent varies by relation and that the conditional probabilities (p(referent|coherence)) remain consistent across conditions. In Experiment 2, we vary only the instructions, having participants write continuations that answer either the question ‘Why’ (Explanation relation) or ‘What happened next?’ (Occasion). As predicted, both pronoun interpretation and the coherence distribution differ significantly by instruction type, and the pattern of pronoun interpretation corresponds directly to the distribution of coherence relations. Models of pronoun interpretation that ignore discourse coherence relationships cannot account for results like these. Keywords: discourse processing, pronoun interpretation Interpreting Ambiguous Pronouns Whereas previous work on pronoun interpretation has appealed to surface-level cues like subjecthood, first- mention, recency, and parallelism, recent work suggests that interpretation should be analyzed in part as a byproduct of deeper discourse-level comprehension processes. Rohde, Kehler, & Elman (2006) mimicked the design of a story continuation study by Stevenson, Crawley, & Kleinman (1994) in order to show that pronoun interpretation differs in transfer-of-possession passages that vary by verbal aspect. (1) John SOURCE handed a book to Bob GOAL . He __________. (2) John SOURCE was handing a book to Bob GOAL . He ______. The context sentences in (1) and (2) contain two possible referents for the pronoun, one that appears in subject position and fills the Source thematic role, and one that appears as the object of a prepositional phrase and fills the Goal thematic role. Stevenson et al.’s work, based on stimuli like (1), showed that participants wrote just as many continuations that corresponded to a Goal interpretation for the pronoun as to a Source interpretation. Two explanations were considered: a thematic-role preference (favoring Goals