+49-431-72004160 moin@otus2.com
Seite wählen

Why do we talk about Artificial Intelligence (AI) in a blog about translation?

The AI was always tightly connected to machine translation (MT). In the beginning, the AI was supposed to mimic human logic. For this purpose, applications were mostly written in special programming languages, such as LISP or Prolog.

A well-known example of an early AI application was called ELIZA, a program able to „talk“ to a person entering questions or statements into a terminal. Essentially, it provided general phrases or counterquestions intended to seduce people to believe ELIZA would „understand“ them. Nowadays, you would call ELIZA a chatbot.

A bit later, the first MT application appeared in the market, e.g. METAL developed by the University of Texas at Austin and acquired and distributed by SIEMENS. It used the same approach. The MT system was supposed to understand the text first and translate it afterwards, just like a human being. Hence it needed to know grammar rules both in the source language as well as in the target language. Thus, these MT systems are called RBMT systems (Rule-Based Machine Translation).

ELIZA and METAL are examples for AI applications that used to be very promising, but became more and more complex over time. Thius eventually lead to high complexity which made them unmanageable. This led to a stagnating quality level and, as a consquence, to harsh criticism and an almost complete abandonment of this approach. AI was sitting in the back corner for many years.

A quickly rising computer performance enable the AI to be revived on a completely new basis in the new millenium. The first approach was a bottom-up approach. In other words, many rules were used to create on overall structure. But now, a top-down approach is used: you start with existing results (data) and try to break them down and put them back together based on mathematical models.

Again, machine translation is one of the more prominent applications. It started with the statistical machine translation (SMT), which has nothing to do with a classical human translation. An SMT system does not know any linguistic information, neither grammar rujles nor the meaning of words. Using bilingual texts (existing translations), it forms patterns in the source language and tries to find corresponding patterns in the target language in order to match them.

A further development of SMT is NMT (Neural Machine Translation). Neural machine translation is a daring term, since it suggests that the system acts like the neurons in the human brain. But it still has a long way to go! For details on MT, see a different blog article.

Since AI covers a vast field, you will find statements like „… thanks to our AI-driven solutions …“ in the language industry. More often than not, this „AI“ turns out to be quite banal, e.g. a chatbot, which brings us back to ELIZA.

So what is the new AI approach? In simple words, it consists of collecting large volumes of data (Big Data) and writing rules analyzing these data. This easily reveals two weak spots. First of all, the quality of the rules determines the exactness of the results. It the rules are programmed erroneously, the systems delivers doubtful results.

But even more important: logic fights back! If your data are inconsistent, you start with a faulty assumption. And that means your result does not mean anything! If you e. g. start with the assumption 1 = 10, than you can „prove“ that Mount Rushmore is higher than Mount Everest.

Todays AI follows a swarm approach: if many individuals start in the same direction, this cannot be false. If you see that many US citizens still believe everything Trump says, although it has been demonstrated that he lies more often than not, you might get some doubts on this approach.

The current Corona crisis provides a good example for the caution you should use regarding the results of big data analyses. If you compare the numbers of infections, convalescenses, and fatalities in various countries, you will find striking differences. And it is hard to explain them just by better medical facilities and similar factors. If you then study how data are created, you get closer to the truth. For instance, considering the inhabitants of the homes for elderly citizens changes the numbers dramatically. And testing is done ad hoc, by chance, and in different ways and frequencies in the different countries. Who tests the people in Brazil’s favelas or in rural areas in Turkey? Considering the fact that a large percentage of US citizens has no health insurance or just lost it together with their jobs, you can ask yourself, who will test them? Hence, all cross-country comparisons are misleading and invalid.

There is an old saying of mathematicians: “I only believe in statistics I falsified myself!” The problem with the Big Data approach is: nobody knows (not even the programmers!), how the AI system got to its results. It is a black box. And if you consider how strongly people tend to believe in numbers, you get scared.

A physics student once told me the distance between earth an moon was 36 km (22.4 miles)! And he could not understand why I doubted that. He had calculated it, thus it needed to be correct! (Who cares about a few missing zeros.)