MIT Technology Review article on machine translation

Thought some of the people here might be interested in this MIT Technology Review article on machine translation and a study done by NIST to see which services are the most effective:

http://www.technologyreview.com/BizTech/17793/

The results of the NIST study are available here:

http://www.nist.gov/speech/tests/mt/mt06eval_official_results.html

http://www.nist.gov/speech/tests/mt/

Let me know if you can't access the pages, I will mirror them or post the text here.

On the same general topic, this an interesting article too:

http://www.wired.com/wired/archive/14.12/translate.html

The company here is taking a different approach to translation. It is using a rules-based system the output of which is analysed using various SMT approaches that try to pick the most natural of various translations "proposed" by the system.

I'm not really that impressed by all of the writing in Wired and Nature about MT otherwise. This piece is a case in point -- look at how the author completely bags Systran for having low BLEU scores. He obviously doesn't know that this is because of the way their system works (BLEU scores of SMT systems tend to be higher than those of rules-based systems, but the two approaches can generate texts with radically differently levels of fluency when evaluated by human readers). The author clearly does not know this, which is a pity because it isn't exactly a secret and if he had interviewed anyone at Systran they probably would have told him that straight away.)

The idea isn't really that novel either (I seem to remember Language Weaver getting a grant for hundreds of thousands of dollars by the US govt to develop something that would do something similar to its own output [ahh.... how it pays to have funding connections.]) but I would put my money on this being a smarter approach to machine translation than SMT-exclusive approaches like those backed by Google with focus exclusively on parallel texts. It is so much easier to add SMT language smoothing to a rules-based system than to add rules-based manipulation (name recognition, whatever...) to an exclusively SMT approach, and you don't need to worry about having a large volume of parallel text if you only need to worry about statistical patterns in your target language.

I am sorry that I will have to mention the name of the company but I am too lazy to search for online articles about it. The company is Interlecta and they use and combine several MT methods and mostly use the statistical one. I don't know how they are rated by BLEU but I like the output I get.