3. Representing Objects of Different Levels.

Association net structure provides means for simultaneous representation of objects of various text processing stages. Let's consider representation of objects of three levels conventionally defined as syntactic, semantic and pragmatic levels.

The syntactic level serves for representing text words and syntactic relationships among them. Each text word is represented by a node; there is a link of "master" type represented by an arc leading from a subordinate word to its syntactic master. If in accordance with its syntactic model a word has a "slave" of a specific syntactic type then another arc may be drawn leading from the master to the slave and bearing the name of that relationship. This level of text structure is characterized by underlying hierarchy represented by a tree-like structure of syntactic links and lack of explicit links between different sentences. The links indicating the relationship of substitution also can be fixed between words. These links can be represented by a special type of arcs leading from the substitute to the antecedent. These links do not obey the above-mentioned hierarchical restriction but are also attached to the objects of the syntactic level.

To the semantic level we assign "referents" - objects, actions and alike participating in the real or possible situation presented by text. A distinctive feature for this kind of objects is its ability to participate in a substitution relationship as a new referent of the substituting word. It is known that such referent need not be directly named by a substituted word; sometimes a substituted word may even not exist in the text and then the referent is conveyed by the situation introduced by some other word. We let the semantic level be populated both by objects named directly and by those conveyed implicitly; but in order to establish relations with the syntactical level we distinguish between "direct referents" and implied "indirect referents" (an indirect referent of a word may coincide with a direct referent of another word).

The extraction of referents is not a phenomenal attribute of a domain; it is mainly of linguistic nature, depends on language peculiarities and may differ from one language to another. What is more, it is referents rather than words which are the bearers of, say, some linguistic features, say, gender or animation showing the ability to propagate them in the substitution process. Different words may refer to one and the same referent; for instance, direct referents of an antecedent and of a substituting word coincide in case of a coreferential substitution.

One of the permanent characteristic features of referent is its semantic class, one of the finite list of semantic classes assigned to the domain. Situations included in the domain description assign specific semantic classes to participating objects. A classification like this may seem too rough a means and eventually specific rules may require more subtle specification of the objects. But in many cases it is exactly such a crude classification that is needed for orientation and application of some basic rules.

Referents may be linked by specific relations indicating, say, a relationship between a situation and a participating object or between an object and one of its attributes (characteristic features). These relations can compose a rather complex graph extending beyond the sentence boundaries.

To the pragmatic level we assign objects immediately present in the real situation where we expect the computer to act (or on such a level of computer data representation after which all the processing is done by conventional methods). These may be references to certain data in a computer memory or a specific processing routine, etc. The extraction of pragmatic objects thus depends on the domain being studied and on the target task. In a system acting in a real situation only the objects belonging to this actual (rather than supposed) situation will be regarded as pragmatic ones. On the contrary, dealing with arithmetical problems where often a situation marked as real is juxtaposed with another one marked as unreal the two situations are treated equally on the pragmatic level due to the fact that that according to the "rules of the game" in these problems the motion laws are followed equally both in real and unreal situations. For a domain focused on the work with a database the objects having their database counterparts in the form of database records or fields of these records could possibly be regarded as pragmatic objects.

Pragmatic objects do not have individual representation in the association net (distinct from that of semantic objects) but they possess some extra attributes referring to appropriate structures representing the current situation in a computer memory. It's easy to see that not every object represented as a referent participates in processing on the pragmatic level.