Lost in Translation – Ask the Machine
There is a war going on that rarely makes it into the press as the battlefield is not on the ground, in the air, or in space, but inside the guts of massive processing nodes that are used to understand the nuance of language. The algorithms that direct these processor are based on a subset of Ai that deals directly with syntax, expression, and a host of other language variables that make understanding other languages a challenge for humans and a monumental task for digital entities. The opponents here are companies rather than political entities, and well-known ones at that, with Microsoft (MSFT), Google (GOOG) and Meta (FB) all pitted against each other with the focus of being the dominant force in the digital translation market.
There are no casualties or battle lines in this war, as it is fought with processing metrics and advertising, and that makes it hard to know who is winning, but the participants seem to have taken on a particular metric to make the public aware of who might be in the lead. That metric is the number of languages a system can translate, which seems to be used a s a gauge toward whose system is the most advanced. While the number of languages that a system can translate is certainly important, especially to those languages that might be considered secondary, but by far the most important metric for translation services is accuracy.
Recently Meta indicated that its NLLB-200 Ai model has increased the number of languages it can translate to 200, which it accomplished in two years, while Google Translate is only able to work with 133 languages and Microsoft’s system translates only 111 languages, although that includes two Klingon dialects, and is clearly staking a claim as the world’s most advanced translation tool. While the number of languages a system is able to detect is certainly easily understood by the general public, the quality of the translation, which is based on the algorithm and the sample base, is far more important and there are two ways in which that can be evaluated, by humans or by machines. Using humans to evaluate translation quality throws subjectivity into the mix, while automated machine evaluation does not, but a machine evaluates a translation by averaging words and sentences against a human evaluation, with the idea that the closer the machine score is to a human score, the better the translation is.
With all of this being beyond the scope or desire of the general population, translation giants will continue to use the simplest metrics to give credence to their systems, but will have little correlation to real world results, as the ability of the AI to understand nuance and what to do with that understanding is really the key. All three companies mentioned here have access to vast stores of speech, which certainly goes toward the ability of an AI to learn, but the algorithm is the key and that is something that not only grows with more resources but must change as humans better learn how to convert those subjective views into language that a machine can understand. So the question is, does the AI need more language to read or does it need more human evaluations in order to take the subjectivity out of their evaluation? The only way to know is…
E ninau i ka Mīkini - Hawaiian
Spurðu vélina - Icelandic
Faighnich dhan Inneal – Scottish Gaelic
quaeritur machina – Latin
Ask the Machine - English
There are no casualties or battle lines in this war, as it is fought with processing metrics and advertising, and that makes it hard to know who is winning, but the participants seem to have taken on a particular metric to make the public aware of who might be in the lead. That metric is the number of languages a system can translate, which seems to be used a s a gauge toward whose system is the most advanced. While the number of languages that a system can translate is certainly important, especially to those languages that might be considered secondary, but by far the most important metric for translation services is accuracy.
Recently Meta indicated that its NLLB-200 Ai model has increased the number of languages it can translate to 200, which it accomplished in two years, while Google Translate is only able to work with 133 languages and Microsoft’s system translates only 111 languages, although that includes two Klingon dialects, and is clearly staking a claim as the world’s most advanced translation tool. While the number of languages a system is able to detect is certainly easily understood by the general public, the quality of the translation, which is based on the algorithm and the sample base, is far more important and there are two ways in which that can be evaluated, by humans or by machines. Using humans to evaluate translation quality throws subjectivity into the mix, while automated machine evaluation does not, but a machine evaluates a translation by averaging words and sentences against a human evaluation, with the idea that the closer the machine score is to a human score, the better the translation is.
With all of this being beyond the scope or desire of the general population, translation giants will continue to use the simplest metrics to give credence to their systems, but will have little correlation to real world results, as the ability of the AI to understand nuance and what to do with that understanding is really the key. All three companies mentioned here have access to vast stores of speech, which certainly goes toward the ability of an AI to learn, but the algorithm is the key and that is something that not only grows with more resources but must change as humans better learn how to convert those subjective views into language that a machine can understand. So the question is, does the AI need more language to read or does it need more human evaluations in order to take the subjectivity out of their evaluation? The only way to know is…
E ninau i ka Mīkini - Hawaiian
Spurðu vélina - Icelandic
Faighnich dhan Inneal – Scottish Gaelic
quaeritur machina – Latin
Ask the Machine - English
RSS Feed