Research Article

Learning to Translate: A Statistical and Computational Analysis

Figure 7

Unlearning curves. These results have been obtained using a fixed training set size equal to 629,957 pairs of sentences and Moses as translation system. In (a), “words swap TM’’ has been obtained by swapping the target phrases inside the TM. In (b), two unlearning curves have been compared. “Numerical swap LM’’ has been obtained applying numerical swaps only to the LM and “numerical swap TM’’ applying numerical swaps only to the LM.
484580.fig.007a
(a)
484580.fig.007b
(b)