Université de Perpignan

When we examine examples of semantic networks from the phenomenological point of view-that is to say in trying to categorize what Charles S. Peirce called the "modes of being" of the networkconcept-nodes as well as relations, we are struck by the enormous diversity and confusion which holds good as much for the nodes as for the relations.

Thus Winograd [1] (see Figure 1) mixes together singular beings (Kazuo, Fido) and intentional concepts such as person, dog, animal, cow, meat, which are of very different kinds since it is clear that "Fido" is an example of the concept "dog"-but "dog" is a sub-concept of the super-concept "animal." "Meat" does not have the extension which we would give it if it were not linked with "cow" by the relation "gives" which implicitly totally restricts its extension to "meat of a cow."

There are many such distinctions which a simple node-and-arc notation fails to make. The attribution of qualities or properties by the relation "is" (Bagheera is black) introduces another kind

Figure 1. A fragment of Winograd's semantic network in which many node and link types are phenomenologically confused.

of node and relation. At least five 1 different meanings of the term "concept" can be distinguished in the fields of Artificial Intelligence and Computer Science.

D. J. Israel (5) is so conscious of the problems which arise from the diversity of elements in semantic networks that he considers it necessary to distinguish between the natures and qualities of things :

"There is to be one tree for kinds of things and another for qualities of things. Kinds must be distinguished from qualities: being a cat must be distinguished (in kind, no doubt) from being red."

For those who are familiar with the phenomenology of Charles Sanders Peirce (which we call "phaneroscopy" 2 ) these terms-"concepts," "kinds of things," "singular beings" and "qualities" immediately bring to mind the three universes which are distinguished by him as three "modes of being" of which the subjects are respectively:

Ideas or Possibles (potential qualities such as redness), which can be described using only monadic predicates ( _is red, _ is a cat). Existents and the Facts concerning these existents, which can be described using only monadic and dyadic predicates (for instance, the fact: "the cat eats the mouse" is described by the combination of the monadic predicates "_ is a cat" and "_ is a mouse" and of the dyadic predicate "_ eats _"). Necessitants (laws, habits and concepts), defined using only monadic, dyadic and triadic predicates (for instance, the concept of "sale" instanciated in the fact John sells a book to Mary can be described by the predicates "_ is a seller," "_ is a book," "_ is a buyer," "_ makes a transaction with _," "_ sells _ to _").
These three Universes are determined by the three fundamental categories of phenomenological elements (Phanerons in Peircean terminology) called respectively: Firstness, Secondness and

Thirdness. As we shall see, there is no corresponding need for any further "universes" such as Fourthness, Fifthness, etc. On this foundation a whole philosophy is built; see, for instance, [6]. Without going any deeper into the Peircean philosophy, we suggest that "phaneroscopy," the categorization of concepts and relations into the Peircean classes, would introduce an important taxonomic principle into the theory of semantic networks. This could be used to formalize distinctions among the present competing representations of the things we know, and could hasten the solution of problems which arise from the phenomenological disorder which seems to reign in the networks. Semantic networks, forming part of an "artificial phenomenology" (the computer thus being considered as a quasi-mind in the Peircean sense), would have everything to gain from integrating Peircean phaneroscopy

We are not going to try to compare different families of semantic networks from the point of view of Peirce's phaneroscopy or the Peircean semiotics which we have set out in [7]. Such a project would seem premature. However, the algebraic nature of our formalisations makes it easier to establish a connection through generalisations between Peirce's phaneroscopy and semiotics on the one hand, and Rudolf Wine's theory of lattice-concepts on the other, which provides a formal basis for categorization, subsumption and inference in semantic networks (reading the articles by Burch and by Wille in this volume, before reading this one, is highly recommended). This work combines subjects which are apparently very disparate: perception theory, Peircean phenomenology and relational algebra, Category Theory, Wille's lattice-concept theory, and semantic networks. Thus, virtually every reader will have to master some alien terminology.

1 See [ 2] . Moreover, J. P. Descles [3] when analysing the use of the relation "is" in Indo-European languages, brings to our attention the values of localisation, identification, ingredience and possession as well as the attributive values of belonging and inclusion. The formal algebraic definition of the "concept" taken from R. Wille's notion of "context" [4] which we will take into account later, should be added to the five meanings enumerated by Rustier. This one seems to be completely isolated from the other five-yet another illustration of the gulf which exists between formal and empirical sciences.

2 We use Peirce's own term out of respect for his concern for "moral terminology," meaning that he who invents something has the right to name it, and others should use that name. Generally, because this network theory had its origins outside the standard ones in Artificial Intelligence, the terminology used may be unfamiliar, although in most cases there are parallel terms in formalisms such as, for example, John Sowa's Conceptual Graphs.