Reanalysis in discourse comprehension: Evidence from reading times

Abstract

This paper explores the evidence for ‘discourse reanalysis,’ the hypothetical case where an initial interpretation of a sequence of discourse units must be revised due to further context. If discourse reanalysis does occur, our representations may need to permit selective non-monotonic update of discourse-pragmatic meaning like coreference and coherence (Las carides and Asher 1993; Haug 2014). But a simpler alternative looms: fully-underspecified representations, which could maintain all grammatical discourse interpretations, without any intermediate selection or ranking to later be reanalyzed. To see which approach best models the representations used in actual incremental comprehension, we turn to psycholinguistic evidence. Discourse reanalysis should be associated with momentary difficulty akin to ‘gar den path’ effects in syntactic processing. But previous studies looking for this difficulty in discourse processing have been equivocal. Here, we investigate a case missing from those previous experiments, a joint ambiguity of coreference and coherence, for which we show robust initial preferences. Then, in a self-paced reading experiment, we observe evidence for difficulty when a preferred interpretation must be abandoned. We take this to solidify the empirical basis for discourse reanalysis as a consequence of incremental composition at the discourse level, and discuss two ways this might be accomplished, depending on one’s approach to the interface between logical representation and real-time comprehension.