The goal of this research is to develop a model giving a rather deep representation of semantics of language including interaction with the real world and suitable for building practical systems with natural language control.
Inclusion of the level of actions of the text processing system is something that dissociates the view of the problem from traditional strategies where semantic studies mostly come to interpreting some words and utterances by means of some other words and utterances. On the other hand, studying system responses triggered by input texts seems too narrow to the researcher's semantic mind as compared to linking text to a system of concepts describing a domain. But ultimately any system of concepts is a result of an abstraction, more or less arbitrary and language-dependent. The concepts themselves can be expressed only through language, therefore any attempt to describe meaning in terms of these concepts runs the risk of going no further than plain periphrasing. And at last the approach to semantics based on meaning-preserving transformations rather than on attempts to make meaning emerge in some different form cannot account for the semantic equivalence resting on the knowledge of the real world; to discriminate this kind of equivalence would be confining. Besides, in order to build a system of practical use it's not enough to study equivalence only on linguistic level.
It is a well known fact that without recourse to the extra-linguistic reality it is often impossible to obtain even an unambiguous morphological or syntactical structure of a text. Therefore we do not intend to build an autonomous language processor following only linguistic rules and postponing all non-linguistic job to the subsequent stages. Linguistic and non-linguistic knowledge will interact throughout all stages.
There is also an inherent draw-back in studying meaning by constructing text-response models. Being realistic, today we can count on intelligent reaction to text only within a rather limited domain populated by real or imaginary automata. Accordingly, the ocean of natural language will be reduced to domain-specific brook of pertinent utterances that the system can perceive adequately. This is not just a restriction imposed on the subject matter or the vocabulary of a text but a very rigid limitation undermining researcher's interest in the most of potential automatic systems.
Still worse, in the prevailing approaches to constructing automatic systems building a practical automatic system is preceded by forming a rigid list of its functions which results in a rigidity of an interface language. In the situation like this natural language interface even if technically feasible becomes a bore for the user. Nevertheless,there is some justification for studying mechanisms of the natural language (or of its sub-languages) even under such circumstances. First, the experience showed that even for very simple domains (like the ones being discussed here) the transfer from text to a system functioning in the real world leads to focusing on new linguistic problems escaping traditional approaches. Second, the practice of building automatic systems starting from a rigid statement of the problem is becoming obsolete; we face the need of creating systems capable to enhance their functions and of combining different independently designed systems; in this case rigid formal languages are loosing their efficiency, especially for very complex systems. Third, there is a certain evidence [1] that language is an amalgamation of many relatively independent subsystems for particular domains and situations, and therefore the study of language behavior in these specific domains can contribute to understanding some general principles of language on the semantic level.
Taken as a whole, the present research which can be identified as "Reality -> Text" Model or, to be exact, "Text -> Response" Model (analogous to the well known "Meaning <-> Text" Model) gives no evidence of deep semantic structure being an indispensable stage of the transfer between reality and text. This follows from the fact that transformations necessary for solving any particular problem in the real world depend not only on linguistic but also on extra-linguistic information and it is impossible to fix a border between the phase governed by linguistic knowledge only and the phase governed by domain specific knowledge along with the description of the current situation.
The choice of transformation direction is determined by a specific goal which depends on nature of the problem being solved. In case of a narrow domain the domain itself specifies the goal (a kind of a response action such as triggering a certain routine, executing some action, solving an arithmetical problem, document retrieval by pattern etc.; of course, the goal may vary within one domain. The specific goal not only guides the transformations but also selects those features of the object that are essential for the task and decides which linguistic complexes are to be treated as single units while others should be processed separately.
Thus it is impossible to define a universal process of normalization of semantic representation not depending on the specific problem following the extraction of deep syntactic structure (and in the model proposed herein the syntactic parser doesn't operate separately either). On the other hand, all the transformations leading from text to the response action can be can be represented in a "semantic" language and be treated as semantic transformations. However, within the scope of natural language the transformations would never be recognized as meaning-preserving because of constant use of extra-linguistic knowledge. An attempt to define the "main" semantic representation of a text - its semantic equivalent - is unlikely to succeed.
In the research presented here several specific domains have been studied though not all the efforts resulted in complete computer implementation. One of the domains developed as far as a text understanding program was a "rigid" problem of a small application package control with an extra restriction that only one routine be triggered by one natural language command. Nevertheless, the actual application package being not ready for practical use at the time of the study, some specific requirements had to be invented in the course of the work.
Some domains were not subject to detailed linguistic study, only the correspondence between domain-specific concepts (terms) and computer output was examined. One such domain, building motions of an ideal robot, is considered in [2].
A rather profound linguistic study can be carried out working within the domain of solving arithmetical problems treating various situations around moving objects. It can't be regarded as a rigidly defined domain because of a wide variety of problems' casts and plots and use of implicit knowledge of real world essential for solving the problems. The investigation of the domain was carried out in two directions. First, not trying to attain a complete model of a domain one can "conceive" the domain-specific environment for each particular problem and thus get a field for linguistic observations and suggestions for model structure. Second, one can impose restrictions on the lexicon and decrease the number of plots essential for understanding the problem.
After all, the system discussed here must find a path from problem's wording to its solution. There is also an intermediate stage where the problem is transformed into a system of interrelated computational models each describing a typical situation and interrelations between the values concerned.
These particular models were used:
- uniform motion model (the relation between velocity, time, and distance);
- time interval (the relation between initial moment, final moment, and time);
- portion of a path (the relation between end points and the distance); - compound motion, i.e. two successive motions of the same object (the relation between characteristics of a whole motion and those of its components).
The transfer from the system of computational models to the solution of the arithmetic problem is well known [3].
The above-mentioned domain shows that automatic system response to a text may manifest itself not only in physical actions in external world. This may be data transformations in computer memory (e.g., modifying data in a conventional database). This might resemble traditional linguistic text transformations; however, here only such forms of representation are chosen as the final result for which the transition to practical applications is clear in principle.