Note: This blog entry was originally written in Japanese in 2019 for our Japanese website. We used our machine translation platform Translation Designer to translate the entry into English and to post-edit the output for readers outside of Japan. The original Japanese entry can be found here.
Machine translation (MT) is one of the application fields of artificial intelligence (AI), and has developed in line with the evolution of AI. AI has been repeating a boom and winter era, and now, it is said that we are seeing the third AI boom. In this boom, the practicality of MT has dramatically increased. In this blog entry, let's look at the history of AI and MT and how they have developed throughout the years
The first AI boom: MT was just a toy?
The term "artificial intelligence" was coined at the 1956 Dartmouth Workshop. The world's first AI program Logic Theorist presented at this workshop was able to prove the theorem of mathematics. It was surprisingly welcomed that computers, which were supposed to be just calculators, could perform human-like intellectual activities, and the first boom in AI took place.
Attempts to make computers perform intellectual activities began in the early 1950s, and their application to translation has been considered from the beginning. As early as 1954, an MT system jointly developed by Georgetown University and IBM succeeded in translating more than 60 Russian sentences into English. However, this system was far from practical. Looking at it now, it was like a toy, with a vocabulary of only 250 words and only being able to translate a specific syntax.
At that time, AI was based on logical reasoning and brute-force search, and could only solve limited problems. Therefore, in the 1970s, AI enters the winter era. MT was no exception. Especially in the United States, in 1966, ALPAC (Automatic Language Processing Advisory Committee) compiled a report stating that the progress expected for MT was not seen, resulting in a significant reduction in the government budget.
The second AI boom: RBMT
Research on AI continued even in the winter era. In particular, the expert system, which teaches computers the knowledge possessed by experts in a specific field as a rule, reached the practical stage in the 1980s and was adopted by many companies. As a result, AI has once again boomed.
In the second boom of AI, IBM's chess computer Deep Blue, which its development began from a project at Carnegie Mellon University in 1985, finally defeated the world chess champion Garry Kasparov in 1997.
In translation, rule-based MT (RBMT), which teaches machines the knowledge of dictionaries and grammar, was commercialized at this time. The company SYSTRAN, founded in 1968, has provided Russian-English MT services to the US military during the US-Soviet Cold War. From 1978, it provided a translation software to Xerox, which became the basis of the automatic translation website AltaVista Babel Fish launched in 1997.
In rule-based AI, the number of rules that need to be taught to improve performance becomes enormous in general. It is often difficult to turn expert knowledge into such a huge number of machine-understandable rules. As a result, rule-based AI, including RBMT, has limited performance. Even if AI could play chess, it was thought that it would be difficult to have AI beat humans in the game of go and shogi. In 1987, the AI market shrank significantly, and the winter era came again.
Machine learning and SMT
Since the late 1980s, several MT systems have been proposed to replace RBMT. One of the most successful ones is statistical MT (SMT), which applies machine learning to translation.
In machine learning, instead of teaching human knowledge to machines, the machines learn by statistically processing large-scale amounts of data. In SMT, a mathematical model of a language is created from a large-scale amount of source and translated text pairs called corpus. Then, when you enter the original text of the source language (e.g. English), a text of the target language (e.g. Japanese), which is calculated to have the highest probability of being the translation for it, is output.
Machine learning, including SMT, requires high computing power and large-scale storage capacity. The term "machine learning" was coined in 1959, but it was in the 1990s, as hardware performance improved, that statistics-based machine learning replaced rule-based AI and became the leading role. SMT was also adopted for Google Translate, which was launched in 2006.
The third AI boom: NMT
Normal machine learning requires humans to teach how to model data. On the other hand, in deep learning, the machine itself creates an optimal mathematical model. After the 2000s, deep learning using neural network (NN) made large-scale strides in the field of image recognition, and the AI boom that continues to this day has occurred.
NN uses computers to mimic the network of nerve cells in an animal's brain. The idea has existed since the 1940s, but from the 1980s, improvements in computing power and parallel processing technology have led to the creation of deep neural network (DNN) that places multiple hidden layers between the input and output layers, which became to be used for deep learning.
Deep learning has been applied mainly in speech recognition in the 1990s, and in various fields from the 2000s to the 2010s. Above all, the fact that the Go program AlphaGo Master, developed by Google's group company DeepMind, defeated the world's top professional Go players one after another from 2016 to 2017 became a big news story.
MT using DNN and deep learning is called neural MT (NMT), and its practical application began in 2015. From 2016 to 2017, companies such as Google, Microsoft, SYSTRAN, and Globalese have launched services using NMT. In Japan, NMT has been available since 2017 with an automatic translation service developed by the National Institute of Information and Communications Technology (NICT).
The service by NICT is limited to non-commercial use, but we have launched a service called Kawamura NMT that allows businesses to use NICT's automatic translation service for commercial purposes, and it has become one of our popular services. For those interested in the service, please see our MT Solutions page.
Conclusion
We have rushed through the history of AI and MT. AI has surpassed humans in areas with fixed rules such as chess and Go, but just as self-driving cars have not yet been put to practical use, it has not fully replaced human intellectual activities including translation.
The AI boom of the past has come to an end because AI has failed to meet excessive expectations. Today's MT is also not a complete technology, and it can be disillusioning if you over-expect. On the other hand, MT has progressed over the past 60 years or so and has become useful as your "assistant" depending on how it is used. At the moment, it is still necessary to figure out the optimal use of MT while identifying the advantages and challenges of MT.
