There is a lot of enviable technology in the world of “Star Trek”. Spaceships move faster than the speed of light, diseases are cured with just a few sprays of hypospray, and food replicators can materialize delicious meals from energy in mere seconds.
But one of the handiest tools available to the Starfleet organization when traveling to extraterrestrial civilizations is the Universal Translator – a device that automatically translates speech into each person’s primary language. Essentially, an English speaker will hear everyone speaking to them in English, regardless of what language they actually speak.
What is machine translation?
Machine translation is a natural language processing application that trains computer models to translate texts between languages. Older methods relied on sentence-based techniques, but deep learning techniques now dominate the field of machine translation and produce much more accurate results.
But is it possible to build a universal translator in the real world? Natural language processing has made amazing progress in recent years, with projects like GPT-3 able to generate sentences and paragraphs that can fool even the experts In the field. In a world where this is possible, why does Netflix still face a shortage of translators to translate the subtitles of their TV shows?
Maite Taboada, a linguistics professor at Simon Fraser University in British Columbia, Canada, said the sticking point was context. It is in the subtleties of meaning that machines fall behind in their ability to make translations. And no matter how much data we throw at the problem, we may never get to the point where it’s accurate enough.
Machine translation is the ultimate problem in natural language processing
“Machine translation is the ultimate app,” Taboada said. “It’s like the Holy Grail.”
Taboada refers to the particular status of machine translation in the field of natural language processing, an area of computer science research that aims for computers to “understand” languages in the same way humans do. Machine translation relies on all the other knowledge in the field of natural language processing, such as grammar, language understanding and language generation. All of these underlying topics need to be mastered in order to build a good machine translation tool.
“So when we say, ‘What are the barriers to machine translation?’ — well, those are all the hurdles of natural language processing,” Taboada said.
The good news is that researchers have been studying natural language processing for over 50 years now. Many areas of this field are already well understood. For example, automating tasks like spell checking works almost perfectly: spell checking programs very rarely require human intervention and it is now, in practice, something for which humans can depend on machines, a said Taboada.
Today, machine translation is synonymous with deep learning
One of the first techniques used by researchers for machine translation was sentence-based machine translation, which uses supervised learning to translate known sentences. Supervised learning relies on humans to label training data before it is fed into the training model, creating bottlenecks around the data the algorithms have access to. The technique struggled with “long distance dependencies,” where the accuracy of translations falters on longer pieces of content.
For example, if a sentence mentions a car at the top of a paragraph and the last sentence of the paragraph refers to the car as “it”, the algorithm may be confused as to what that “it” is. reports.
Sentence-based machine translation techniques aren’t widely used anymore, and that’s mainly because in 2016, Google changed the algorithm taking their Google Translate tool from sentence-based translation to deep learning, a machine learning technique that relies on building large neural networks. Google said the new technique reduced errors by approximately 60 percent.
“And then the thing exploded, everybody wanted to use deep learning,” Taboada said. “I think there is no turning back.”
One of the main advantages of the deep learning technique is that it can be trained using mostly unsupervised learning – it doesn’t require as much human supervision to label the data for the process of training works.
Taboada said the lack of supervision is possible because deep learning can infer the meaning of words and phrases from context. Word meanings are mapped as “vectors” in multi-dimensional space, and when two are often observed together, the algorithm learns that their meanings are related. As a result, deep learning is able to use this vector-style understanding of word meanings to aid in the translation process.
Biases and lack of data for many languages hinder machine translation
Even with the deep learning model, many obstacles remain to build a universal translator. One is the problem of bias in training data. Because deep learning uses unsupervised methods, it learns while extracting data from the world and therefore inherits the same problems and biases that exist in the world.
Taboada illustrated the problem with the example of nouns that have genders. In some languages, such as Spanish, a translation into the language must include genders even when the original text does not specify genders. For example, if the word “doctor” is translated from English to Spanish, it must have a gender, and that gender can be determined by the predominant gender associated with doctors in the model training data.
“The data just reflects the way the world is, but it’s not necessarily the way we want the world to be, and it may not be appropriate.”
“So you go out into the world, what do you see? Perhaps “doctor” is 70% of the time translated as “el doctor” in Spanish and “nurse” is 85% of the time translated as “la enferra”, feminine,” Taboada said. “It’s actually not a fault of the data – the data just reflects the way the world is, but that’s not necessarily the way we want the world to be, and maybe that’s not appropriate.”
There are also other concerns. Some languages may simply not have enough data to create good training models. And algorithms may not be able to differentiate nuances like dialects, effectively flattening translations.
For streaming services like Netflix, part of the difficulty in translating subtitles for shows and movies is the physical constraints of screen time and space – sometimes translations are too long to fit. screen or be read fairly quickly. In these cases, humans are needed to make the tough decisions about what to cut so that the subtitles still look good.
The biggest obstacle to machine translation is the context
Another big hurdle for machine translations is the issue of context. Although there is plenty of data to train on for deep learning, such as content from Wikipedia, books, or academic papers, it can be difficult for algorithms to learn language differences between these. brackets.
“It’s just completely different the way I write an academic paper than the way I write a tweet, and it’s all rolled into one dataset,” Taboada said.
An disentanglement that still requires work on the part of humans. Taboada specializes in sentiment analysis, an area of natural language processing that analyzes the emotions behind sentences and expressions.
Although machine translation has come a long way, it still struggles to detect subtle positive and negative emotions. Deep learning algorithms are quite capable of translating texts like user manuals, which typically don’t contain emotional sentences or require a lot of cultural context to understand, Taboada said. And they’re also able to do a decent job of content moderation, allowing businesses to step up their content moderation and find inappropriate comments automatically.
But the algorithms are not yet very reliable for these tasks when they require a lot of nuances. A study 2016 examined words and phrases used by racist and anti-racist online communities and found many linguistic similarities. These similarities make it difficult to accurately detect hate speech, as it is easy to accidentally block anti-racist comments.
“I would never have a call between [President Joe] Biden and [Russian President Vladimir] Putin is automatically translated.
“Moderating hate speech is really difficult because the words overlap,” Taboada said. “So you need to know more about the context and how they are used to it to understand whether it’s something that needs to be removed or not.”
When dealing with sensitive issues and situations where nuance is important, such as legal contracts or political matters, machine translation is not an appropriate application.
“I would never have a call between [President Joe] Biden and [Russian President Vladimir] Putin is automatically translated,” she said.
Would machine translation be accurate and reliable enough to translate sensitive conversations, if deep learning models had unlimited data on which to train their translation models?
“I don’t think so,” Taboada said. “With a complex issue like machine translation, right now I don’t see how it’s ever going to be good enough that I can click a button and walk away and assume the translation is going to be great and I don’t have no need to do anything about it.