Machine Translation: Why is Japanese-English the only pair where rules are still used?

Question

The Microsoft Research Machine Translation system (MSR-MT) is actually a hybrid system, using both rules and statistical methods (1) .

According to Peter Norvig (2) (emphasis mine):

Machine translation: 100% of top competitors in competitions such as NIST use statistical methods. Some commercial systems use a hybrid of trained and rule-based approaches. Of the 4000 language pairs covered by machine translation systems, a statistical system is by far the best for every pair except Japanese-English, where the top statistical system is roughly equal to the top hybrid system.

Is this because Japanese is a very regular language (agglutinative), which makes it easier to use rules? Or is it only because of sociopolitical reasons (Microsoft had some sort of relationship with Japanese people, etc.)? If the first, why aren't other agglutinative languages (Turkish, Korean) also using statistical/rules hybrid systems?

Answer 1

As another user wrote in a comment:

Japanese-English does not used rules-based machine translation.

The Norvig quote must be interpreted in the historical context of 2011. The other facts mentioned in it are also out of date (eg 4000 language pairs, statistical methods as opposed to neural given that usually NMT is distinguished from SMT even though NMT is also statistical not rules-based).

If you re-read the Norvig quote, you will notice that he did not write that a rules-based system was being used for Japanese-English in production, he wrote only that for that one pair the results of the statistical system were only equal to those of the rules-based system.

That being so at that time had more to do with the available tools and data for the pair than some inherent property of the structure of one of the languages in the pair.

If you imagine some very obscure language like Chechen for which there are no researchers who know the rules and parsers, stemmers etc and no tools, it may be more clear to you for which languages a statistical approach will beat a rules- and tools-based approach.

To prove to yourself that this has nothing to do with typology, consider the case of a language like Swiss German, which has tools like Chechen but typology like English.

(As an aside, I do not remember if this was the case for Japanese-English only, English-Japanese only or both. As a rule, x-English systems perform better than their complements, again for reasons that have little to do with the inherent structure of the languages.)