The issue with the chinese room thought experiment is that it does not take algorithmic complexity into account. The experiment describes a dumb lookup table, the table size would grow exponentially with the number of input output combinations. While NNs encodings allowing them to store the same information in a sub-exponential manner.
The most efficient manner to translate chinese is to understand chinese including every-day nuances that might be relevant to the translations. Just like a human translator can benefit from knowledge in the field about the text which he is translating.
http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.136...