The model of analysis does not and cannot prescribe any predetermined sequence of actions. The sequence is extracted from the constituent words, or more exactly, from dictionary entries comprising computing procedures. Only the initial input of a sentence and dictionary search combined with morphological analysis are executed in a "centralized" way. After the morphological analysis is complete ambiguities can be left both on the level of dictionary entry choice and on the level of morphological features. In the former case several dictionary entries may go through the subsequent processing independently (but while one of them is being processed all the rest are blocked). In the latter case the ambiguous morphological features will not be ascribed to a word at all but later on a possible variant of these features may be ascribed to the word from a syntactic construct and then this variant will be compared with the morphological data).
The principal part of text processing is controlled by hypotheses (or predictions) which are included into dictionary entries of the words. A hypothesis comprises the following attributes:
- the base, i.e., the word referring to the dictionary entry it emerged from;
- the goal, i.e., the predicted syntactical or semantical object along with its predicted attributes - the object will be subject to several attempts to merge with an actual syntactical object or some semantical objects;
- the processing procedure ("attention" procedure) defining a certain order of the search for an object to be merged with the goal of the hypothesis as well as some necessary extra verification;
- the rank, i.e., a number determining the initial order of processing.
The procedures can use the failure and restoring mechanism. The mechanism allows to make provisions for a failure of a procedure before starting it (in particular, a failure may result from an attempt to merge two objects with imcopatible attributes or else from an explicit failure under certain circumstances). In case of such a failure an execution of the procedure stops and the state of the whole data net is restored back to the moment of entering the procedure. A procedure attempting to set a link (by merging the goal of a hypothesis with some other object) first sets the link and then performs a number of tests; if some test fails the link is cancelled.
The tests performed during the execution of a link setting procedure apart from compatibility checks on objects being merged can comprise other actions. They may attempt to set some additional links confirming the initial one and then possibly pass them through semantic and pragmatic interpretation which can fail as well. One must obviously avoid irreversible external actions before the result of the analysis is definitely confirmed.
The requirement that additional links be set before the initial link receives an ultimate confirmation reflects the existing interdependence of links in language. The system provides for two principal ways of seeking additional links, a search for a particular link needed to complete a syntactic structure, in particular a search for the syntactic master of a word (the dictionary entry must directly refer to the confirming link), and search for all links filling the gap between the words being linked (to secure projectivity of the syntactic structure being formed). Setting additional links is performed by a call to the list of all hypotheses including the hypotheses that will generate the links wanted (when the whole list has been processed one must make sure of all necessary links having been set).
The call to the list of hypotheses is a selective action: only those hypotheses are run which may lead to a "useful" link. The result is that the initially defined order of starting hypotheses according to their ranks can be altered when some hypothesis demands for that.
There are several predefined "attention" procedures differing in a way and order of searching for a prospective merging partner for the goal of the hypothesis. For the hypotheses with a syntactical object as the goal they are:
- a search for a specific lexical unit;
- a search by a specified syntactical class;
- a search by the semantic class of the direct referent;
- a search guided by word position (to the left or to the right from the base of the hypothesis or restricted only to the neighbors of the base).
Hypotheses seeking for a semantical object must employ procedures depending on specific features of the respective semantic classes.
Employing predefined procedures does not prevent from creating individual ones for the words with syntactic or semantic peculiarities.
For brevity we name a hypothesis syntactical (or semantical) if it seeks for a syntactical (or a semantical) object.
To show an example of interaction of hypotheses of different nature consider prepositional groups subordinate to a verb.
In a strongly governed prepositional group the verb predicts a preposition in the immediate surroundings (i.e. has the corresponding syntactic hypothesis) and also has a semantic hypothesis for a referent of the preposition governed substantive (i.e., predicts its semantic class and its place in the semantic structure). In a weekly governed prepositional group the preposition along with the subordinate substantive perform the task of predicting the verb and also interpreting the resulting link.
In the sentence "Поезд прибыл в три часа в город" ("The train arrived at three o'clock at the station", in the English version we use for convenience "station" instead of "town") the verb "прибыл" ("arrived") governs two prepositions "в" ("at"). One of them is strongly governed ("прибыл в город", "arrived at the station"), the other is weekly governed ("прибыл в три часа", "arrived at three o'clock"). The word "прибыл" ("arrived") has a syntactic hypothesis for the strongly governed preposition "в" ("at") and at testing the link set by the hypothesis we'll start another hypothesis predicting the preposition governed substantive from the dictionary entry of the preposition. Besides, the verb defines also the semantic class of the substantive direct referent (geographic point) and its semantic links (the motion termination point). If the verb hypothesis predicting the preposition erroneously chooses the first preposition "в" as a candidate to the hypothesis goal then the variant will be cancelled due to incompatibility of semantic classes. Another erroneous choice of the word "город" ("station") as a preposition governed substantive of the first preposition "в" ("at") will not succeed either due to projectivity violation, viz., the intermediary wordgroup "три часа в" ("three o'clock at") will not parse. Thus the correct link "прибыл в город" ("arrived at the station") will be set and the station will be marked as the motion termination point.
Projectivity testing of the link will trigger all the syntactic procedures from the gap between the verb and the preposition. However the strategy will be different here. The link between the preposition "в" ("at") and the prepositional group "три часа" ("three o'clock") will be set by the hypothesis introduced by the verb as before. The hypothesis predicting the syntactic master from the preposition will find the only candidate meeting projectivity conditions, viz., the verb "прибыл" ("arrived"). The link interpretation procedure will demand merging of the prepositional group direct referent with one of the referents within the limits of the verb group; which of the objects will be chosen depends on the semantic class (time moment). Thus we conclude that "три часа" ("three o'clock") is the motion termination time point.
An interaction of procedures coming from dictionary according to the above procedure if successful must result in building syntactic and semantic structures of the sentence including some pragmatic interpretation elements. This may be followed by a pragmatic response to a sentence-stimulus (performing some action, asking for more information etc.). The results of the analysis of a single sentence may affect the subsequent text processing through semantic objects inferred and the related hypotheses. Semantic objects may act in the subsequent text as semantic merging partners. The hypotheses remaining from the previous text may influence analysis of the new sentence if included in the general hypotheses list and make their own choice of referents from the new sentence to be merged with previously introduced ones (this can be regarded as a model of a "biased" perception of a new sentence affected by a prediction coming from the preceding text).
Thus the model proposed herein can handle textual links extending beyond the limits of a single sentence.