Japan e-book distributor Media Do to build ‘AI translator’

From Nikkei Asian Review:

Japanese electronic book distributor Media Do will develop an artificial intelligence-based automatic translation system to make its e-books available for English-speaking readers.

The company hopes to reach a broader market and promote digitization at a time when Japan’s book market is shrinking.

Media Do has teamed up with two Tokyo-based AI startups for the project — Internet Research Institute and A.I. Squared. Media Do will invest about 1.1 billion yen ($9.82 million) to acquire about 20% of each company through a third-party share allotment at the end of this month.

Both startups have developed unique technology for summarizing and translating text. The summarization technology analyzes the relationships between words and sentences in a given piece of text and extracts key sentences to create a summary.

The translation technology “learns” set phrases in both Japanese and English — on top of vocabulary and grammar — to enhance the quality of its translations, a process known as deep learning.

Link to the rest at Nikkei Asian Review

5 thoughts on “Japan e-book distributor Media Do to build ‘AI translator’”

  1. We get enough laughs out of mis-translations between other languages. And they’re going to trust an AI?

    This should be good.

  2. Different languages are not simple word or phrase substitution codes of each other. Japanese and English have very different common ambiguities and elisions. A line by line translation can very quickly lead you astray. Referents might be paragraphs away. Until so-called AIs can understand what they read, the technology cannot produce robust translations.

    • This is the problem with the Bayesian constructs that are part and parcel to deep learning. They are naive, top down analyses of information. Operant conditioning functions this way and it is the crudest form of learning that organic systems use to respond to their environment. Humans use both top down and bottom up analysis to understand the world around us.

      An analogy is a comparison of Eastern medicine and ancient technologies like Wootz/Damascus Steel to the modern scientific method. In both cases, there is empirical evaluation of the results. The power of Western science is that there is a theoretical framework to build an understanding of underlying concepts. With a bottom up framework, it is often but not always possible to distinguish what steps led to the result and why, and when there was an element of dumb luck. The result is that we can quickly build on past learning to drive deeper understanding and accelerate our progress.

      The failure of deep learning to understand meaning is evident when Watson was playing Jeopardy! and was flummoxed by a simple Harry Potter question. It could not tell the difference between Harry Potter and Voldemort. Even those of us that have not read one word of Harry Potter know that Voldemort is the villain. Watson had no idea.

Comments are closed.