Competitive queueing models of serial order memory in speech and language

Abstract

A recent focus in cognitive and brain sciences is to construct models for memory starting with assumptions that are neurophysiologically plausible. The behavior of the model should fit the empirical data. In this thesis, we model serial order in human memory. Serial order in memory refers to the memory of items whose features are ordered temporally. We concentrate on specific tasks in speech and language. The models we consider are connectionist, and the individual units in the model have a localist representation. The models are designed on the principle that during recall of the serial order, the items constituting the sequence are activated in parallel with dynamically varying strengths at each time step. The selection of a winner among these items takes place at each time step. This mechanism, called the Competitive Queueing principle, is the foundation on which the models considered in this thesis are built. We study sequencing of serial order in memory at two levels: at the level of the lexical items and at the sub-lexical level. Lexical level sequencing is of a list of words. Sub-lexical level sequencing is addressed by looking at the spelling of a word and the phonemes constituting a word. Addressing these issues takes us deeper into their domains. The sequence of graphemes/letters constituting a word (i.e., its spelling) is modeled. The model replicates the behavior resulting from a spelling disorder arising from Acquired Dysgraphia, upon being subject to “damage.” Speech signals, when analyzed, have anticipatory and left-over features influencing the current phoneme being produced. This phenomenon, called co-articulation in speech production, can also be explained by the Competitive Queueing principle. Recent evidence points to verbal short-term memory being influential in vocabulary acquisition. We study two models which address these issues. One model gives a detailed account of sub-level sequencing and addresses the key empirical results. It explains the presence of short-term memory capacity for lists without rehearsal. The second model satisfies a wide range of empirical results and includes visual word learning. It further replicates and predicts the behavior caused by lesions in the brain, as in Aphasia, by “damaging” the model.