Translation Rules. Of course, before any translation rules can be discovered at all, it’s necessary to decide what form they will take. Clearly, it would not be particularly effective to simply memorize translations of entire sentences, as it’s vanishingly unlikely that all novel source sentences presented to the system will have been seen before. Thus, we need to operate at the level of smaller linguistic units. The earliest, simplest statistical MT systems learned translation rules for individual source language words, as in Figure 1.1a (Xxxxx et al., 1994). However, due to syntactic divergence, it’s rare for word-to-word translations to be accurate in even closely related languages, so word-level systems have since been supplanted by phrase- based ones, which operate at the level of multi-word sequences, as in Figure 1.1b (Xxxxx et al., 2003). Phrase-based systems often achieve state-of-the-art performance for pairs of closely related languages, but don’t do as well when the source and target languages have more divergent structure and long-range reorderings are required. One simple way to add more 1If the decoder is parameterized, there is often also an intermediate stage where these parameters are tuned. S NP VP VBD NP PRP DT JJ NN I saw a white cat ví un gato xxxxxx
