Language barriers and Artificial Intelligence

Vocal interactions play a key role in the technological advances concerning the so-called Artificial Intelligence, in particular in the case of the understanding of spoken language (NLU – natural language understanding), or in an extremely adjacent field such as that of linguistic translation.
Online language translation began in the 1990s with two main products: AltaVista’s Babelfish and Xerox’s Systran: these first web translation tools could only handle short pieces of text using statistical rules.
But the further one has gone, the more it has been necessary to manage huge volumes of data at a fast pace, which has led to a different approach by the language-translation tools that have established themselves in recent years, such as Google Translate, probably the better known, or Microsoft’s Translator, its most significant competitor.
Current AI language translation tools exploit a technique called neural machine translation (NMT), which appears to use artificially constructed neural networks, in which essentially the entire sentence (the set of words present) is analyzed rather than just single words, while still making the translation faster and more accurate.
But deep learning, despite all its perceived appeal, still has many limitations.
Google researchers spoke openly about some of these limitations in an interview with Wired Magazine. In particular, they pointed out that simply scaling up the neural network and adding more data doesn’t necessarily mean that human capabilities can be replicated.
In the same article, NYU (New York University) Professor Gary Marcus described deep learning as “greedy, fragile, dull and superficial.” By “greed” Marcus means that neural networks require huge data sets for training.
The training factor is even more evident when dealing with the specific needs of a sector, a profession, or with the specific knowledge base of a company, thanks to the terminologies used, technical or related to products.