This blog post explores how deep learning-based GNMT technology has revolutionized the accuracy and naturalness of Google Translate.
Most people who have used Google Translate have likely encountered its errors at least once in the past. Before using Google Translate, you might have expected the service to solve everything, even if you didn’t know the foreign language. However, after using Google Translate, that expectation likely vanished completely. Google Translate, which just a few years ago delivered absurd results, has recently undergone a remarkable transformation. The translated context has become much more natural, and it can even interpret poetic expressions.
Google Translate has historically struggled with handling the subtle nuances between languages. This often forced users to go through the hassle of revising or reinterpreting the results. However, this hassle has significantly decreased recently. Thanks to the introduction of new technology, Google Translate now provides much more sophisticated and accurate translations than before, marking a major advancement in global communication.
The reason Google Translate, which had been underperforming, was able to make such a dramatic leap forward is precisely because the Neural Machine Translation System (GNMT) was introduced into the Google Translate service. Before the introduction of GNMT technology, the system supporting the existing Google Translate service was a phrase-based machine translation system. This system relied on a database of grammatical rules and meanings input by humans. It translated each word or phrase within a given sentence individually, then pieced them together like a puzzle to form the translated sentence. Consequently, the sentence structure often felt disjointed, and the resulting translation’s word order and context were highly unnatural to read. Naturally, it also failed to grasp the author’s underlying meaning and intent within the sentence. Because Google provided translation services based on this system, many users experienced the frustration of receiving perplexing translation results.
However, the newly introduced GNMT technology is a translation technique that utilizes deep learning, a core technology of artificial intelligence. GNMT technology recognizes the flow of the entire sentence and even understands the author’s purpose within the sentence to provide translation, resulting in noticeably smoother interpretations. The introduction of GNMT technology is a major innovation in itself, but to understand the changes this technology has brought, it is necessary to grasp the basic concepts of deep learning.
So, what exactly is deep learning, the core AI technology that has enabled us to reach this stage? Deep learning is the general-purpose AI algorithm used in AlphaGo, the computer Go program that competed against Lee Sedol. It’s a term you’ve likely heard before, given its frequent mention in the public eye. You may also have heard that AlphaGo consumed an enormous amount of power during its matches against Lee Sedol. It reportedly used 1,200 CPUs, which is roughly equivalent to the processing power of 300 computers. These massive computing resources study like humans before facing Lee Sedol. They learn which trends lead to victory and which lead to defeat. AlphaGo analyzes existing game records, uses them to generate new ones, and learns patterns within these records. After completing a certain amount of learning, it then plays Go against Lee Sedol. During the game, AlphaGo searches among the previously studied records for patterns most similar to the current game. AlphaGo directs its learning toward winning game records, much like how humans learn. Programs trained this way can perform calculations hundreds or millions of times faster than humans, enabling them to sometimes achieve superior results.
As described above, deep learning is an advanced form of artificial intelligence developed from artificial neural networks. It utilizes information input/output layers similar to brain neurons to learn from data. Here, artificial neural networks—the precursor to deep learning—are algorithms modeled after the human brain to process various information in a manner similar to the human brain. The human brain is composed of structural units called neurons and can learn specific functions like pattern recognition and cognition through experience. These artificial neural networks became feasible as computer performance improved. However, this neural network technology also has several drawbacks. These include slow learning times and a tendency to lose direction as the number of layers increases. Deep learning technology addresses these shortcomings of artificial neural networks.
The introduction of deep learning has not only improved the performance of Google Translate but also opened up possibilities for application across diverse fields. For instance, in the medical field, diagnostic systems utilizing deep learning can play a crucial role in detecting specific diseases early and suggesting treatment methods. In this way, deep learning is increasingly permeating various aspects of our lives, and its potential is limitless.
GNMT technology, which applies deep learning as described above, does not match words or sentences 1:1 like humans do. GNMT technology recognizes entire sentences as the unit of translation, enabling it to grasp context and reflect it in the results. Furthermore, GNMT technology analyzes and learns from existing translations. In this process, GNMT technology can improve its performance by modifying the connections between artificial neural networks.
To evaluate GNMT’s performance, Google researchers selected sentences from Wikipedia explanatory texts and news articles and translated them into several languages. They then placed these translations side-by-side with those produced by Google’s existing system and by human translators. Human evaluators were then asked to assess the quality of the translations. The evaluation results showed that translations from notoriously difficult Chinese to English scored significantly higher than the existing system. Translations between certain other language pairs also achieved scores close to human translations in terms of accuracy. However, it lagged behind translations between Indian and European languages. The paper’s authors emphasized, “It should be noted that the selected sentences were well-constructed short sentences.”
Thus, Google Translate incorporates deep learning, a core AI technology. By digitizing internet content to accumulate big data, Google Translate can now treat entire sentences as a single unit for translation. Because the translation unit expanded from words and phrases to entire sentences, translation results have made a noticeable leap compared to the past. Regarding this, Barak Turovsky, Head of Product Management for Google Translate, stated: “Neural machine translation technology reduces the possibility of errors by up to 85%. A greater evolution has occurred than the achievements of the past decade.”
However, language translation involves immense variability. Even within a single language, individual differences and dialects exist, and language evolves over time. Furthermore, interpreting wordplay or poetic expressions remains challenging and often results in inadequate translations. Some analyses suggest neural machine translation technology is less accurate than phrase-based machine translation systems. Nevertheless, the core technology behind Google Translate is deep learning, meaning translation data is accumulating in the machine’s brain even as you read this. Over time, the machine can autonomously gather and learn from the newly accumulated data. As this knowledge builds, it will eventually enable translations that account for current shortcomings. Indeed, Google has admitted that this translation still falls short of human translation and contains a significant number of errors. However, it emphasized that with the accumulation of learning experiences by deep learning-based AI and advancements in related technologies, it will evolve toward near-perfection. We can reasonably expect that in the not-too-distant future, Google Translate, enhanced with neural network translation technology, will break down language barriers.
