False recognition through semantic amplification

False Recognition through Semantic Amplification Brendan T. Johns (johns4@indiana.edu) Michael N. Jones (jonesmn@indiana.edu) Department of Psychological and Brain Sciences, Indiana University 1101 E. Tenth St., Bloomington, In 47405 USA Abstract This paper describes a computational model to explain a variety of results in false recognition. The processing mechanism in the model is built around a co-occurrence representation of lexical semantics, affording an account of both structure and process. We show that this model can naturally account for levels of false recognition that are seen in studies using the DRM paradigm, including item-level effects, reaction times, and event-related brain potentials. Keywords: False recognition; co-occurrence representations; memory models; recognition memory; semantics Introduction False recognition is one of the most empirically studied phenomena in recent times, however very little formal modeling of this effect has been conducted. False recognition has been most studied using the Deese/Roediger-McDermott (DRM) paradigm (Deese, 1959; Roediger & McDermott, 1995). In this task, lists of words that are associated with specific critical words are presented to subjects, and on subsequent memory tests the unpresented critical items are falsely recognized at almost the same level as studied items (Roediger & McDermott, 1995). For example, given nurse, hospital, sick, and cure to remember, subjects are likely to subsequently produce a false alarm to doctor. Work within the DRM paradigm has provided fundamental evidence about the organization of human memory. The paradigm demonstrates that humans use semantic information to both store and retrieve items, and the use of this information can lead to profound memory errors. However, the exact mechanisms that underlie the false recognition of associates have evaded a formal explanation. Rather, theorists have focused more on general conceptual frameworks of cognition, such as Fuzzy Trace Theory (FTT; Brainerd & Reyna, 2002), the source- monitoring framework (Johnson, Hashtroudi, & Lindsay, 1993), and the discrepancy-attribution hypothesis (Whittlesea, 2002), to explain false memories. There are now many different computational models of recognition memory. However, a principled problem with these models is that the representations they use do not contain semantic information about specific words, due to the fact that their representations are typically constructed with random number generators. This practice is natural because the models are not typically used to simulate semantic effects in recognition memory. Further, it is still the subject of much debate what are the correct features to represent word meaning. However, we believe that in order to model semantic behaviors, such as is seen with false recognition, one must use a representation of words that contains semantic information. One promising avenue for these semantic representations are those created by co- occurrence learning models. In a co-occurrence model, a word’s semantic representation is constructed by observing statistical regularities in a large corpus of text. These models can account for a variety of different semantic behaviors (for a review see Jones & Mewhort, 2007). Due to the success of co-occurrence models in other domains, it seems natural to assume that they could also be used to account for semantic effects in recognition memory. The models are particularly promising given the observation that associative variables are important predictors of false recognition and false recall, particularly backward association strength (Deese, 1959; Gallo & Roediger, 2002). Because co-occurrence representations correlate with backward association strength (Johns & Jones, 2008), their representations are appealing to be used in a processing model of false recognition. By using a representation of words that is built up through exposure to the environment, we are not simply assuming a particular semantic organization. Instead we are both explaining how a certain memory structure is created, and how this structure interacts with the processing mechanism. The Recognition through Semantic Amplification (RSA) Model Our false recognition model is based on the Iterative Resonance Model (IRM) of recognition memory (Mewhort & Johns, E., 2005). The motivation for IRM comes from a series of experiments demonstrating that Old responses and New responses are based upon different types of information (Mewhort & E. Johns, 2000). Note that a subject responds Old if the probe item was in the encoded list, and New if it was not. In particular, the authors found that the amount of contradictory information contained within a probe predicted New responses, whereas Old responses were based on the similarity of the probe to the memory items. The original IRM used this dual-criterion decision process. If a decision is not made on a particular information sample, then successive iterations are employed to sharpen the evidence. The number of iterations required for the model to make a decision is taken as a proxy of response latency. Our Recognition through Semantic Amplification (RSA) model is kept within the same formal framework as IRM,