Quantitative Neural Network Model of the Tip-of-the-Tongue Phenomenon Based on Synthesized Memory-Psycholinguistic-Metacognitive Approach P.M. Gopych Ukraine, Kharkiv, Kharkiv National University, 4 Svoboda Sq., 61077 Tel: +38(0572)493489, Fax: +38(0572)353977, E-mail: pmg@kharkov.com A new three-stage computer artificial neural network model of the tip-of-the-tongue phenomenon is proposed. Each word's node is build from some interconnected learned auto-associative two-layer neural networks each of which represents separate word's semantic, lexical, or phonological components. The model synthesizes memory, psycholinguistic, and metamemory approaches, bridges speech errors and naming chronometry research traditions, and can explain quantitatively many tip-of-the-tongue effects. Current models of spoken word production are developing within the scope of two research traditions based on speech errors analysis or picture naming chronometry [1]. Models of the tip-of-the-tongue (TOT) phenomenon emerged from speech error tradition, and they are developing within psycholinguistic, memory, or metacognitive approaches [2]. In present work a computational memory-psycholinguistic-metacognitive model of TOTs is proposed on the base of two-layer artificial neural network (ANN) memory mechanism from ref.3. Existing computer-implemented models of word production are network models and their nodes represent whole linguistic units [1]. Proposed network is build for each word's node from some interconnected learned auto-associative two-layer ANNs each of which represents separate word's semantic, lexical, or phonological components (such multidimensional memory representation consists with language processing brain imaging data [4]). ANNs operate with specific patterns of signals encoded specific linguistic information (they are sets of positive and negative units simulated sets of spikes affected simultaneously on excitatory and inhibitory synapses [3]). TOTs can arise due to many kinds of problems in mechanism of sought-for word free or cued recall. I name three main stages in currently inaccessible target word recall process (or resolving of TOT). 1. Word node selection when on the base of provided initial semantic information a learned group of neurons (target memories in the form of above-mentioned ANNs) containing information pertaining to a target word is selected and activated in part. Because TOTs are induced as usual by definitions of difficult or low-frequency words I suggest that appropriate learned ANNs (stored information) may be damaged (for example, due to the insufficient memory consolidation) and/or incompletely selected (for example, due to the insufficient or irrelevant input semantic information). The extent of selected ANNs (stored information) damage (and/or incompleteness) determines mainly the TOT power (strong or weak, imminent or nonimminent, etc). 2. Word (word's component) retrieval. Free recall retrieval process initiates a series of random sets of positive and negative units (i.e. random sets of spikes) enters to input neuron layer of learned ANN. Cued recall retrieval process is the same but spike patterns to be entered to the same ANN are not random in full and contain a fix part of true information on the recalled pattern (that is intensity of cue). The result of each attempt of retrieval is an output set of positive and negative units (or a spike pattern) from exit layer of ANN [3]. 3. Final stage of recall is comparison of pattern emerged from exit layer of ANN with reference pattern from metamemory and decision-making to stop (or not) the process. If emerged and reference patterns are the same then recall is finished. In other case stage 2 (see above) will be repeated, other random (or random in part) set of spikes enters to the same selected ANN, and so on until reference pattern will be detected (or the process will be stopped for independent external reasons). Comparison and decision-making are metacognitive (meta-level) processes and separable from the retrieval (object-level cognitive process). Proposed model can explain quantitatively many TOT effects: semantic priming; existence of negative, illusory TOTs, or TOTs for incorrect items; strong or weak TOTs; immediate, delay, or eventually full TOTs resolution; recollection of target word partial information (first letter or gender for example); age dependence in TOTs; TOTs in patient populations; etc. The number of different sets of spikes used for recall of missing word (the number of attempts of memory retrieval), duration of time intervals between successive sets of spikes, and duration of separate neuron spike determine retrieval chronometry. In that way proposed model bridges speech errors and chronometry research traditions. Also it supports Tulving's challenge to the doctrine of concordance. 1. Levelt, W.J.M. (1999) Models of word production. Trends in Cognitive Sciences, 3, 223-232. 2. Schwartz, B.L. (1999) Sparking at the end of the tongue: The etiology of tip-of-the-tongue phenomenology. Psychonomic Bulletin & Review, 6, 379-393. 3. Gopych, P.M. (1999) Determination of memory performance. JINR Rapid Communications, No.4[96]-99, 61-68. 4. Vigliocco, G. (2000) Language processing: The anatomy of meaning and syntax. Current Biology, 10, R78-R80.