Training data only includes regular parallel data. At testing, model performances on unseen idioms are measured. 


University of California, Berkeley

There are three idiom testing conditions.


Idiom Testing Conditions


Zero Testing Condition


In neural machine translation experiments, the joint testing condition uses a training dataset containing both regular parallel data and a specific subset of idiom data (idiom-train data). During testing, the model's performance is measured on translating these same idioms, but within sentence contexts different from those seen in the training data.

Joint Testing Condition in Idiom Translation

In neural machine translation experiments, the upsampling testing condition is similar to the joint testing condition—which trains the model on a dataset containing both regular parallel data and a specific subset of idiom data (idiom-train data)—but the idiom-train data is upsampled $$N$$ times. This condition evaluates whether upsampling targeted training data is necessary or effective to achieve better idiom translation quality.

Learn Before

Related