langbrain banner

Phases of the Top-Down Modeling Procedure

These phases may be identified in the top-down modeling process which moves from structural linguistics to an understanding of the linguistic system of the brain.

Phase 0: Analytical Linguistics

Analytical linguistics, which analyzes spoken and written discourse, is a necessary precursor to the development of neurocognitive linguistics. From analytical linguistics we get various traditional notions such as phonology, syntax, morphology, along with a wealth of knowledge about phonetics, grammar, and semantics.

Phase 1: Person-Oriented Linguistic Structure

The first step in the neurocognitive direction is to adopt a different point of view of what the object of investigation is. Rather than the elusive abstract concept 'language' a system supposedly shared by members of a speech community, or the equally elusive 'ideal speaker-hearer' that has been proposed, we seek to understand something more tangible: the linguistic system of the typical individual, a neurocognitive system; and we recognize that each person's linguistic system differs from that of every other person, to a greater or lesser extent.

By examining things that people say we can form hypotheses about recurrent units in their speech and we can identify relationships among these units as well as patterns of combination. We find, among other things, different layers of structure, each presumably subserved by a subsystem having its own structural properties. These subsystems include the phonological, the lexico-grammatical, and the semological. Each of these strata has distinctive patterns of arrangement of the basis units. We recognize phonotactics (the structure of syllables and phonological words), morphotactics (the structure of words and grammatical phrases), lexotactics (the structures of clauses and sentences), and semotactics (the structure of thoughts, ideas, procedures, rituals, etc.)

Phase 2: Linguistic Structure as a Relational Network

Further study of the relationships among the structural units identified in Phase 1 leads to the conclusion that each such unit, when its relationships are fully analyzed, reduces to just a point in a network of relationships. The entire linguistic structure (a cognitive structure in the typical person) is thus seen as a relational network.

In this phase we no longer recognize phoneme, morpheme, lexeme, etc. as objects or symbols. In the network, a morpheme (for example) is represented as a distributed representation consisting of branching (tree-like) structures leading to phonological components in one direction and to semantic properties in the other.

The nodes of the networks differ from one another according to three dimensions of contrast: (1) 'and' (syntagmatic) vs. 'or' (paradigmatic); (2) upward (branching upward to or toward semantic values) vs. downard (branching downward to or toward phonetic values), (3) ordered vs. unordered. For an 'and' node, ordering is sequential — the under of understand precedes the stand); for the 'or' relation, the ordering is a precedence ordering: one line takes precedence, the other is the default.

At a more advanced subphase of Phase 2 we recognize the 'threshold node', a node with an integral threshold representing the number of incoming lines that must be active for the nodes threshold to be satisfied. The 'or' node may be then be reanalyzed as a threshold node with a threshold of one; and the 'and' node may then be reanalyzed as a threshold node with a threshold equal to the number of incoming lines. But a node with three or more incoming lines  may also have a threshold of intermediate value; for example, a node with three incoming lines and a threshold of 2 will be satisfied if any two of the incoming lines are active. Conceptual structures evidently make extensive use of threshold nodes with multiple connecting lines.

We can then analyze the structure of the relational network. We find that it consists of lines and nodes, organized into NECTIONS. The nections can be seen as the fundamental modules of network structure. Arrays of nections (e.g., for morphemes) are connected with supporting structures to form a subnetwork for each stratal system; and these subnetworks are interconnected to form the linguistic system as a whole, a vast network of millions of nections. But this linguistic system, if the various semological subsystems are included, has no boundaries to set if off from the rest of the cognitive system, since the semological systems (of which there are several) cover information of all the kinds we can be aware of and talk about. They are evidently distributed throughout the cerebral cortex.

Phase 3: Linguistic Structure as an Expanded Relational Network

The network of operates by the spread of activation from node to node. The networks of Phase 2 consist of bidirectional lines and nodes, capable of passing activation in both directions. In Phase 3 we analyze each bidirectional line as a pair of one-way lines, and do a corresponding analysis of the nodes of Phase 2 into their internal structures. The resulting expanded relational networks, drawn in narrow notation, have one-way lines and two kinds of nodes: junction nodes, where lines converge, and branching nodes, where lines diverge. Connections to junction nodes can be excitatory or inhibitory; and there is a special type of junction node called the 'blocking element' which connects directly to another line and inhibits the passage of activation on that line.

Phase 3 also takes a more refined look at connection and activation strengths. While a connecting line of Phase 2 is either present or absent, the lines of Phase 3 have varying degrees of strength, from very weak to very strong. Other things being equal, a line of greater strength carries more activation than a line of weaker strength, when given the same degree of activation. Likewise, strengths of activation may also differ in that a line of a given strength may be given a greater or lesser degree of activation by its source node. Thus another variable is degree of threshold satisfaction. Whereas in Phase 2 a node is either satisfied or not by incoming activation (hence Phase 2 thresholds have integral values), in Phase 3, a node may be satisfied to varying degrees, and each node has an input-output function (that can be plotted as an S-shaped curve) such that a greater amount of incoming activation leads to a greater amount of outgoing activation. In summing the incoming activation from various connecting lines, excitatory connections contribute positively while activation coming from inhibitory connections must be subtracted.

We can also define a 'nection' for Phase 3, on a similar basis to that of Phase 2; but the (bidirectional) nection of Phase 2 often corresponds to a pair of (directed) nections of Phase 3. Also, for nections of Phase 2 we make no attempt to locate precisely the boundaries between nections, just identifying 'external lines' as the loci of boundaries. But the nections of Phase 3 have a clear basis for distinguishing boundaries: a line belongs to the same nection as its source, and the boundary is the point at which it connects to its destination node, marked in the notation by an arrowhead.

Also assignable to Phase 3 is the division of the phonological system of Phase 2 into two distinct phonological systems: phonological production and phonological recognition. It remains an open question whether the grammatical system should be similarly divided.

Finally, Phase 3 includes a learning theory, which includes the 'abundance hypothesis' (genetically provided proliferation of weak connections) and the 'proximity hypothesis' (PATHWAYS, Chapters 10, 12).

Phase 4: Linguistic Structure as a Neurological Network

In Phase 4 we consider the structure of the cerebral cortex and of cortical neurons and their interconnections, and we test the various hypotheses of the theory of Phase 3 for neurological plausibility. The result is a theory of how the linguistic system is represented in the cerebral cortex.

According to the cortical column hypothesis. the nections of the theory (Phase 3) are implemented physically as cortical columns, each consisting of about 80-110 neurons.

The various hypotheses of types of interconnection required among nections by the theory (local and distal excitatory connections, local inhibitory connections, inhibitory connections on both nodes and lines) are all consistent with properties of cortical connectivity of neurons. Also confirmed are the hypotheses of differential activation levels resulting from differing degrees of stimulation, differing connections strengths, and the abundance hypothesis.

From the proximity hypotheses we derive a set of hypotheses about the likely cortical locations of various lingistic subsystems. These hypotheses are confirmed and refined by the neurological evidence.

Quantitative estimates of the learning capacity of the brain according to the theory, as compared with what we observe in humans, provide additional confirmation (PATHWAYS 341-343).


Go to Menu (for non-frame version.)

This page was last modified on 10 April 2007.

© 2000-07 Rice University. This document, or any portion hereof, may be used for non-commercial informational purposes only. Any copy of this document, or portion hereof, must include the copyright notice ( in its entirety.