Comparing greedy and optimal coverage strategies for shallow-transfer machine translation
Jernej Vičič, Mikel L. Forcada · 2008
Using the open-source Apertium platform as an example of a shallow-transfer machine translation system and Spanish and Catalan data, we compare greedy (left-to-right, longest-match) and optimal-coverage strategies when chunking the input sentence using the left-hand-side patterns of shallow transfer rules. We find that, when rules are reasonably correct, there is almost no difference between both strategies. The selection of unreasonable rules (such as rules which delete or insert content words) may be curtailed by using a length modification penalty term.