I had a somewhat strange request last week, as someone wanted me to translate a short sample of machine translated text without looking at the original to show a client that it really could not be used. She gave me a choice between the versions generated by babelfish and google translate to pick the one that was a better translation. They seemed to be at about the same level of gibberish, and I ended up picking the babelfish one. Just as I mentioned in my last post about having trouble with picking the right defiinition for ‘linglong’, the machines clearly have difficulty in determining which definition to choose when there are several possibilities. For example, the topic of the excerpt was chemistry, and uniform became  制服, as in a type of clothing, and thermal bonding became 熱量保稅, using the tax word for bond. Also calling a phone number became 叫, and solid (as opposed to liquid) became 堅實. The translation of ‘flux’ into 漲潮, or rising tide, was another puzzle.

It was also interesting that the paragraph section fared much worse in the machine translation than the bullets. The shorter bullet points were actually somewhat comprehensible, while the longer paragraphs became nonsensical babble. And babble is pretty hard to translate! A lot of times technical paragraphs written by engineers will be quite poorly written, but this takes it to a whole other level. The best way to translate I figured was just word-by-word, as when there is no sensical sentence structure to make out, that’s really all you can do. 

It seems to me that even though machine translation has come a long way, there are just too many variables, including non-textual ones, to consider for it to ever become 100% accurate.

2 Comments »