Author: Dr. Jacekas Antulis, Associated Partner, Head of the Patent Division at METIDA
- Artificial intelligence (AI) today: implementation and use
- AI problems related to automatic translation programmes
- Perfection is very close
The concept of artificial intelligence has been used for a long time already, while the first complex models of neural networks were created in the 60s. The first actual applications in practice began only in the last decade because only now our technological advance and capabilities (information density and processing speed) are gradually approaching the right level. It would take more than one decade to develop actual artificial intelligence that could approximate the activity of the human brain, however, the first prototypes which are now being developed do not require such a high level of functionality.
Very often, artificial intelligence is understood as a programme or algorithm that can perform the same thing differently, depending on past or prior experiences. However, the programme or algorithm itself is just a virtual object, on computers it may be an operating system that requires a specific implementation. This requires an appropriate structure – a matrix. In humans, this corresponds to the brain, on the computer it would be the hard drive. In this way, the artificial intelligence serves as the ‘soul’, while the neural network performs the function of the initiation of operations or the ‘initial body’.
Using a specially designed neural network and the system of corresponding algorithms, the automatic translation does not seem to be a big problem; the developers of automatic translation algorithms seem to be dealing with it quite well. Over the past few years, automatic translation programmes have improved to such an extent that after completing corresponding tests, they made us think about what to do next and how to use their new potential in our everyday practice.
The first translation problem is words and grammar. Twenty years ago, this was a big problem because it required a large database, matrix, and many rules. However, ten years ago, this problem was solved and the first automatic translation programmes appeared, which at that time were rather amateurish and formed the public opinion that automatic translation would never be perfect. In each language, the amount of words is quite precise, the rules and exceptions are clearly indicated, so from the mathematical point of view, this problem in automatic translation programmes has practically been solved.
The second problem is word order. In each language, the sequence of words is precisely defined by the rules of word order. With a fairly large field of analysis, this is an easy task. This problem still exists only because there are many languages and not all the inaccuracies in certain languages have been resolved yet, but speaking of the most popular languages, this stage has been completed.
The third problem is formulas and chemical compounds. The difficulties arise from the need to change the format of the output text – to change the editor (to install the compiler additionally). Currently, the simple format is most commonly used that does not allow us to work with structured or multi-level formulas. This problem could be realized in HTML format, but it would not be suitable for everyone, and programming in a special format would take extra time. So far, nobody wants to do it simply because of the lack of resources.
The fourth problem is the terminology of a specific field. Currently, this problem is actively dealt with because it exists in all major languages. Algorithms have managed to solve the problem of word order when only one sentence needs to be analysed instead of the whole text. However, in order to solve the problem of terminology at least a few sentences need to be analysed. When the context needs to be analysed and the relations between the terms in a particular text that is being translated need to be determined in the scope of several sentences. When analysing the texts that are being translated, we see that this problem can be solved well enough, but we still need to check the terms. When the Uniform Patent enters into force, in accordance with legal regulations, the human translation will not be needed – automatic translation will be enough. At the moment, when statistically analysing real translations, it has been noticed that translations by automatic programmes are better than the human translations.
The fifth problem is the mixing of different texts in fiction. In fiction, depending on the situation and the context, the terminology and topics from different fields are mixed, therefore, from the mathematical point of view, it is less ‘stable’.
Because the language is defined by the rules and their exceptions, an ideal automatic translation is possible. In the future, it will be as good or even better than a human translation, because humans make mistakes and machines do not. The use of programmes of automatic translation and storage in the future will help to preserve the original language, as well as compare it with what was many years ago. It will be possible to detect the changes that have taken place and to determine when and why they occurred. Only the machines will be able to preserve the authenticity of the languages of the people in the future because people tend to simplify, change and forget their language.
After a long process of improving automatic translation, it will be easier to achieve the next goal – automatic synchronic translation, when the programmes of automatic translation will provide a real-time translation of the language used into another language. This technology exists already and is being improved, but little is known so far.
The door to the room of artificial intelligence is already intriguingly open, and a lot of virtual spaces have been created. Are we ready to enter this room without getting lost? It depends on ourselves.