NEUROLINGUISTIC MODELS

Neurolinguistic models investigate the links between any external speech activity of human beings and the corresponding electrical and humoral activities of nerves in their brain.

It would be interesting to investigate in detail what part of the brain is activated when a human is preparing and producing an utterance or is trying to understand an utterance just heard. Unfortunately, the problem to discover the way people think while speaking or understanding is tremendously difficult. Indeed, the unique objective way for a researcher to reveal the ways of the human thinking by neurophysiological methods is to synchronously investigate electrical and humoral activities in a multiplicity of places in the human brain.

The brain of any human consists of a tremendous number of neural cells, or neurons. The neurons are not elementary units by themselves, since each of them is connected with hundreds of other neurons by their dendrites and axons in a very complicated manner.

The dimensions of neurons, dendrites, and axons are so small that the techniques of modern neurophysiology provide no practical opportunity to observe each neuron separately. What is more, the intrusion of observing tools inside the brain evidently changes the activity under observation. To understand in detail the internal functioning of the brain, it would be necessary to observe vast groups of neurons activated in the multistage process and trace back a grandiose multiplicity of causal chains.

Then, after this tracing, the researcher would be able to express the detected traces as rules and algorithms presumably used by the human brain. This is what might be called a neurolinguistic model of language processing. However, the extremely large number of possible combinations of signals traced in such a way is well beyond the capability of any computer to handle.

Leaving aside the inadequacy of the modern neurophysical techniques, the mathematical tools for such a modeling are insufficient either. We cannot even hope to get close to the information activity of neuron groups in our brain with rather simple mathematical models. According to the modern point of view, the neurons are complicated elements of logical type. Thus, a modeling system should contain elements of the switching circuitry. However, neither deterministic nor stochastic theory of approximation by such a circuitry has been developed well nowadays.

About 30 years ago, neural networks were proposed as a tool of artificial intelligence research. They consist of elements referred to as formal neurons. These are standard logical computational units with rather limited abilities. There have been numerous studies attempting to apply neural networks to various tasks of artificial intelligence and particularly to language processing. If there are some statistical trends in phenomena under observation and modeling, these trends can be recognized and taken into account by this technique.

However, for computational linguistics the problem of how to choice the input and output signals for modeling of the brain activity remains unclear. If we propose some specific inner representations for these purposes, then the formal neurons will only model our inventions rather than real language processes.

For this reason, without revolutionary changes in the research techniques and new approaches for treatment of observable data, neurolinguistic models of natural language understanding are unlikely to give good results in the nearest future. At present, only very crude features of brain activity can be observed effectively, such as which areas within the brain show neuronal activity associated with human memory, general reasoning ability, and so forth.