On-line Reference Assignment for Anaphoric and Non-Anaphoric Nouns: A Unified, Memory-Based Model in ACT-R

The computational model in present paper confirms that memory-based accounts are sufficient to account for a high rate of success at first-pass referent retrieval for anaphoric (and non-anaphoric) nouns. Because even definite noun phrases can often be non-anaphoric (e.g., Poesio & Vieira, 1998), an adequate model must account for how a reader makes an explicit or implicit decision about the anaphoric status of a noun (herein: The Anaphoric Classification Problem). We explain why we are inclined to reject the conventional intuition that: the failure to find/retrieve a referent within the discourse then, serially, leads to treating a (possibly anaphoric) noun as a new referent. Instead, we extend the memory-based account to address this classification problem. We suggest that LTM contains both generic referent types and specific referent tokens, which simultaneously compete for retrieval via resonance. The nature of what is retrieved (token vs. type) determines whether the reader effectively treats a noun as anaphoric or not. Our model predicts whether an anaphor in a given text will be misinterpreted as a new referent during first-pass processing. The influence of anaphor word choice is explained, and encompasses metaphoric anaphors.