The potential application of both neural networks and parallel processing is most obvious in the areas of natural language understanding and speech recognition. To break the language barrier to knowledge, computer translation of scientific correspondence, papers and books would be a tremendous achievement. With the exponential proliferation of publications of all kinds, something which may be called the bulk barrier has to be overcome in all kinds of scientific work. The internal communication within the scientific community could be speeded up manyfold and the reinvention of the wheel occur less often.
Imagine the task of a scientific library which has the responsibility of translating, summarizing and storing a steady stream of papers in many foreign languages. To scan a paper, translate it into English, create an abstract, define keywords and store it in a database constitutes a well-defined undertaking for the area of artificial intelligence. A short step from this capability is the quality scanning of texts. The identification of grammatical and syntactical errors, stylistic shortcomings and a weak logical structure would utilize the whole power of AI. Another related task would be natural language interaction with databases. Today’s communication demands some kind of formally rigid, abbreviated query language in the retrieval phase.
Speech recognition is an area which ultimately poses the hardest challenge for AI. One aim is voice recognition and translation of a telephone conversation into a second language, and there are already some applications for the input of unconstrained continuous speech using a diversified vocabulary, although with several weaknesses. The real task is to make it possible for say a Japanese and a Swede to converse in their respective native languages. No one has yet been able to create a program with enough grammatical rules to understand every sentence in a single language, let alone two.
A closer look at the translating telephone reveals that three technologies are required. First a device for automatic speech recognition, then a language translator and finally a speech synthesizer. Furthermore, the set of devices has to be duplicated; one for the hypothetical Swedish and one for the hypothetical Japanese side (see Figure 7.4).
Figure 7.4 Main parts of a translating telephone.
The combination of the six modules should allow for a large, relatively unrestricted vocabulary, continuous speech input and speaker independence (little or no system training) in both directions. An interesting technical aspect relates to the intermediate link of text use. If an artificial language like Esperanto was used here instead of a natural language, the technical problems might appear less daunting. This is due to the fact that Esperanto is a very structured and simplified language, designed to be easy to handle although rich enough to facilitate good human communication.
Computers are however growing increasingly powerful and data storage is becoming cheaper. This fact was the impetus for an ongoing Japanese project to solve the translation problem by brute force. The JETS (Japanese-to-English Translation System) searches for entire sentences in a huge database consisting of all imaginable standard phrases with their variations. Because many similar sentence structures can be identified by a single string, the amount of memory needed is more restricted although still extensive. The main problem with this solution is however the prolonged search time.
Another application not to be forgotten is the listening typewriter which would be able to write down and edit the incoming speech on paper. A further use is pattern recognition which has applications in several areas: picture analysis, radar detection in military and civilian systems, cartography, etc.
Language understanding, translation and pattern recognition are extremely difficult areas for computer scientists and demand large computers capable of high-speed processing. The systems must be both self-learning and self-organizing in relation to the extremely comprehensive information to be worked upon. In spite of impressive funding and tremendous efforts in the last decade, any real breakthrough within AI has yet to come. One impediment must be the contemporary over-confidence in formal languages and the opinion that all phenomena can be expressed in logical concepts. Science seems to have been snared by expressions inherent to mathematics, statistics and logic. While the computer, working formally, that is, according to fixed models and algorithms fits this style, human beings act instead on the basis of direct communication and associations. Certain semiconscious processes in human thinking often emerge as concepts which might seem fuzzy but which are superior when reality has to be described.
The need for sophisticated AI is especially pronounced in the area of robotics where the equivalent of goal-directed behaviour found in biological organisms is to be implemented by mechanical machines. From a purely mental point of view, no robot yet constructed can compete with the capacity of an ant. Some sceptics contend that it never will be possible to reach the level of that marvellous insect. It should not, however, be forgotten that robots are often useful because they do not reflect the nature of human sensing and being.
In fact, robot technology does play an important role in scientific, economic and military affairs. The main development has however been within the area of practical industrial robots, where reliability and cost- effectiveness is more important than an advanced artificial intelligence. All over the world several manufacturing processes, both menial and unhealthy for human labour, for example, welding and spray painting, have been replaced by industrial robots.
The real need for AI becomes apparent in the special categories of robots used in bomb disarming, nuclear-environment repair operations and similar tasks. In contrast to symbolic problem-solving as in language understanding, these special robots interact with the physical surroundings much as living beings would. They therefore have to be sensory-interactive and managed by a hierarchical control system. Problems to be solved may be illustrated by a robot manipulating one arm while pursuing an uncooperative target. Here the calculation of the coordinate’s transformations have to be performed continuously, with delays of no more than milliseconds. A set of six equations with six unknowns has to be solved; this requires massive computational capability.
Source: Skyttner Lars (2006), General Systems Theory: Problems, Perspectives, Practice, Wspc, 2nd Edition.