langbrain banner

Relational Networks

The relational network model grew out of an attempt to understand a person's linguistic system as one that is able to operate — for speaking and for understanding. The newcomer to linguistics might suppose that such concerns are shared by all those who try to construct theories of language, since after all the linguistic systems of our minds are able to operate. But the fact is that the majority of linguists just try to analyze and describe various properties of texts, without a concern for how our systems work in speaking and understanding. (See analytical linguistics.)

Relational network theory began its development without any formal notation and allowed notation to develop as the analysis unfolded. The result was a network notation — not as a design feature present from the outset but as the conclusion of the process of analysis.

This network model is justified and was arrived at by analyzing relationships among linguistic units. Starting with the traditional assumption that a lexical item (likewise morpheme, phoneme) is a unit of some kind, an object or symbol or combination of symbols, we analyze its relationships to other units to which it is related. For example, morphemes are somehow related to elements of phonological expression on the one hand and to elements of conceptual information on the other. They are usually represented as symbols — for example, 'boy', and these symbols represent their phonological (or graphic) realizations. Obviously, they are therefore related to the elements of their phonological realization and have to access them for production of speech. These relationships can be represented as connections, the minimum requirement for any kind of relationship. Of course, positing connections leads inevitably to the need to posit nodes — the points connected. A morpheme also 'has' higher-level properties — grammatical and lexical or semantic — and if we ponder what the meaning of 'has' is here, it is that the morpheme is connected to such properties. So we have further connections. But after the relationships of the morpheme to other elements — phonological, grammatical, etc. are thus plotted, the symbol that has been representing it can be removed from the resulting diagrams with no loss of information. Whatever information it can be considered to represent is now already represented in the depiction of its relationships (Lamb 1999: Chapter 4). This is especially clear in the case of the morpheme, since the symbol was just representing the components of its phonogical representation, but these are now directly represented by connections. Similarly, what do phonological symbols represent? They are just abbreviated notations for type and manner of articulation; so they have connections to those phonological properties, and after such connections are plotted, these symbols too are superfluous. A similar line of reasoning applies to all the other linguistic units for which symbols have been used. They turn out to be just abbreviations for sets of connections.

Following this procedure, every unit of phonology and lexicon can be seen to be what it is by virtue of what its relationships are, and so can be seen as just the point in the system which has those relationships. And it appears that semantic and conceptual relationships can likewise be handled as relationships, with no units (Pathways of the Brain, Chapter 9).

Upon removal of the symbols from the structure, the result is a network of relationships — not symbols and relationships, just relationships, represented as interconnections among nodes. For what do the usual symbols for morphemes, for example boy, consist of? They are (spoken or written) expressions. But in the network representation the actual expression is already provided by the downward connections to elements of the phonological system, so there is no need to have a symbol within the system in addition.

In any case, the presence of symbols within the system would raise a serious problem: Just how is the symbol to be detected by the system, or distinguished from other symbols? After all, the brain does not have little internal eyes to read symbols.

Thus the linguistic system, understood as a network, does not contain symbols but is a system which interprets and produces symbols — spoken or written. Symbols are inputs and outputs of the system; they are not inside it. If they were inside, the system would need a little symbol interpreter in there to interpret them, as well as a symbol producer, to write them. That is, somewhere within the brain there has to be a system capable of interpreting and producing symbols which does not contain symbols. Well, then, why not let this system be the linguistic system itself? What use has it of internal symbols?

A pathway in the network from the auditory or visual perception of a spoken or written word to the set of concepts or images which embody its meaning within the cognitive system is precisely what provides the system with its means of interpreting symbols presented to it from the outside.

Traveling Activation
The network operates for producing or understanding speech by means of activation traveling along its pathways. (Examples.)

Compact and Narrow Notation
Relational networks of the form usually seen in the literature are drawn at a relatively abstract level. They are like highway maps drawn to a large scale. In keeping with their aim of accounting for linguistic data, they leave out considerable detail that is of interest for other purposes. But their lines are really two-way lines, able to carry activation in two directions. But underlying this notation is a model which does not have two-way lines. Rather, each line of the ordinary (or COMPACT or CONDENSED) notation is considered to represent (in the typical case) a pair of oppositely directed lines of the actual model, as may be shown directly using the EXPANDED (a.k.a. NARROW) RELATIONAL NETWORK NOTATION. Similarly, a node of the condensed notation generally corresponds to two or more nodes of expanded notation (Lamb 1999: Chapter 5). (Illustrations). The "ordered AND" node of condensed notation is more complex, as it has to provide not only for a combination but also for the sequencing of the components.

For phonology, there are evidently two distinct (but of course interconnected) subsystems, one for production, the other for recognition, each made up of one-way lines and nodes. Lexical structure, on the other hand, may achieve bidirectionality without separate subsystems, by having both feed-forward and feed-backward output connections from their nodes (Lamb 1999:Chapter 8). The model has this property also in the conceptual and perceptual systems and is in general agreement with Antonio Damasio's hypothesis of representation of information in neural networks (Damasio 1989a, 1989b, 1989c).

Also, in the actual networks of the theory, connections differ in strength. Similarly, the activation received by a node has a continuous range of values, and the node responds according to a sigmoid threshold function such that greater input results in a higher degree of outgoing activation. A line of a given strength may therefore carry varying degrees of activation.


Go to Menu (for non-frame version.)

This page was last modified on 31 January 2002.

© 2000 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.