Glossary

activation: The value of a unit or set of units in a neural or semantic network. The term is often vaguely defined and used to refer to different concepts in different theoretical environments. For example, activation can refer to the base or resting activation of a unit, the continuously changing value of a unit, average value of a population of units, the triggering of some process, etc.

alignment model: Refers to models such as Cohort Theory which assumes the beginnings of words are mapped in sequence onto lexical representations. This type of model can be contrasted with continuous mapping models, such as TRACE and MERGE, which assume the mapping process is distributed over the course of a stimulus word and it's lexical representation.

algorithmic: algorithm (n): a finite step by step problem solving procedure. Algorithmic refers to models that are implemented as a computer simulation (i.e., TRACE, PARSYN, etc.)

canonical: An invariant structure defined by necessary and sufficient features. Canonical representations are often described in relation to linguistic features.

cognitive model: A model is a said to be cognitive if it describes any cognitive process (e.g., memory or speech processing) in relation to representational states and processes (i.e., algorithms) that operate on these states. Cognitive models describe and predict how information flows through the system. A cognitive model explains what the system is doing at a functional level. Cognitive models can be instantiated as verbal descriptions, processes (mathematical), connectionist, or as computer simulations (algorithms).

Cohort: Initial subset of words activated from the lexicon (superset) based on the initial 100-150 milliseconds of the speech waveform. See Cohort model.

feed-forward: (1) Refers to the direction of information flow in information processing models, including connectionist models. Strict feed-forward models (e.g., MERGE, FLMP, Cohort II) preclude any top-down contextual effects from changing the nature of the stimulus representation. Models that are not strictly feedforward and allow two-way bi-directional connectivity between representational levels (e.g., TRACE) are called interactive. (2) Feed-forward models are also called autonomous, whereas feedback models are called interactive.

FLMP: Fuzzy Logical Model of Perception

Frequency Effect: Ubiquitous empirical finding in psycholinguistics reflecting a negative correlation between word frequency and some performance variable (e.g., reaction time). As word frequency in the language increases, reaction time decreases. The relative frequency of words can also be conceptualized as a neighborhood effect. In neighborhood effects, words can be partitioned into two classes based on their own frequency and frequencies of their neighbors: 'Hard' words are characterized by a low frequency word with high frequency neighbors; 'Easy' word is characterized by high frequency word with low frequency neighbors. Neighbors are defined phonetically. (See NAM).

graded: Continuous. The activation level of a unit will vary with the degree of overlap between features in the speech input and features in the lexical representation. The degree of overlap is defined in different ways depending on the model. In some models (i.e., FLMP) overall goodness of fit between features in the input and features in the representation are defined as "fuzzy truth" values (ratio scale between 0-1.) In other contexts refers to the number of matching and mismatching features between input and representation.

Inhibition: In a connectionist network, when a set of units at one level inhibit each other. For instance, bottom up activation of features in the initial phoneme /k/ in /kat/ would simultaneously increase activation levels of /k/ and inhibit or reduce activation levels of all other phonemes in this initial position.

Luce Choice Rule: The Luce Choice Rule is an influential decision rule often employed in cognitive models in psychological science. The rule describes the probability of a response in some task given some stimulus evidence i as a function of the ratio of the degree of similarity between the stimulus evidence i and a given response alternative J over the sum of the similarities of stimulus evidence i to all relevant response alternatives in the set K:
As a simple binary example, consider a politically neutral voter deciding between republican and democratic presidential candidates from the set of all candidates. The likelihood this voter will choose one candidate over the other is related to the amount of positive evidence or support for one candidate relative to the total evidence or support for both candidates. As such, assume this voter rates the republican candidate a .7 and therefore the democratic candidate a .3 on a 0 to 1 scale (i.e., for the sake of simplicity this rating will constitute the degree of similarity between the stimulus and response alternative). Thus the probability of voting republican will equal .7 / .7 + .3 or .7 (i.e. 70% confidence in the republican candidate). Now, using the same formula plugging the degree of support for the democratic candidate into the numerator one obtains .3 / .7 + .3 or .3 (30% confidence in the democratic candidate). The Luce Choice Rule requires a number of assumptions to be met to be applicable, but the aforementioned description provides a brief summary of the use of this rule in psychological science.

Mapping Process: Process by which acoustic-phonetic input is mapped onto lexical representations. Essentially, given the representational assumption of canonical lexical forms (see above) refers to some algorithm (i.e., function) that maps a continuous, variable input domain onto discrete, invariant range.

Uniqueness Point (UP): Segmental point in a word where a given word diverges from all other words in the lexicon.