Artificial neural networks trained on big data have really improved the quality of machine translation so much that it can be seen with the naked eye. If you search the Internet for screenshots with Google translator curiosities and test them today, the difference is obvious.
However, is neural machine translation (NMT) really getting close to human? The answer is unequivocal: not at all! So far, machine translation systems are not comparable to the brain of a human translator. They make mistakes that a person would never make – and which indicate that talk of “artificial intelligence” is premature.
What’s wrong with neural machine translation?
NRM has many disadvantages that are not at all similar to the problems of human translations. The disadvantages of neural translators can be divided into 3 categories: reliability, memory, and common sense.
Reliability: Perhaps the greatest concern is that the NRM may be unreliable, defiantly erroneous and completely incomprehensible. NMT systems do not guarantee the accuracy of the translation and often skip negations, individual words or entire phrases.
Memory: The LMT system is also characterized by loss of short-term memory. Systems are designed to translate one sentence. As a result, they forget the information learned from previous sentences. It turns out worse than the entertainment accepted at parties, in which each participant writes the next line of the story, while seeing only the previous one.
Common Sense: NWO systems do not have common sense in the human sense – that is, the external context and knowledge of the world. Being able to discern which contexts are appropriate for certain translations is important for our understanding of situations, but these contexts are often difficult to capture fully.
NMT systems are not armed with methods for determining the reliability of facts in the text of the translation. Worse, such inaccuracies are unpredictable and inconsistent, making it difficult to automatically detect and correct them. For example, IMT systems can confuse negations and omit entire pieces of information. What are the consequences of such mistakes?
Skewed data, skewed translations
Although the original Malay text did not contain any gender information, the English translation assumes that the nurse is female and the programmer is male. The HMT system suggested this option because there were more examples of female nurses and more examples of male programmers in the machine learning data.
This phenomenon is an unfortunate side effect of how neural networks are trained. Using real world data (such as gender ratios in the nursing profession and programming), the NMT system introduces unsound information into its translations.
This example is especially telling, as the translation system exacerbates existing inequalities in the world. However, this kind of error can occur whenever the system remembers a trend in the training data and may misuse this pattern when translating.
NMT systems have another notable flaw: they are heavily geared towards translating single sentences. Neural networks in modern translators do not remember well what happened before the sentence they are translating.
Why do NMT systems learn to translate one sentence at a time, and not the entire document at once? The reasons are technical. First, it is difficult for a neural system to read a long document, store all this information compactly, and memorize it efficiently. Secondly, these systems need more time with a large amount of initial data.
The inability to use the broader context is a major obstacle to the success of the NRM. Almost any translation requires understanding a few sentences, but in some cases, such as translating a story, this is critical.
Storytelling is a human activity like any other, requiring a combination of creativity, intelligence, and communication skills that make us different from animals. If AI translation systems can’t translate text coherently, let alone beautifully, then can we really say they have human-level translation?