Many syntactic machine translation decoders, including Moses, cdec, and Joshua, implement bottom-up dynamic programming to integrate N-gram language model proba- bilities into hypothesis scoring. These decoders concatenate hypotheses according to grammar rules, yielding larger hy- potheses and eventually complete translations. When hy- potheses are concatenated, the language model score is ad- justed to account for boundary-crossing n-grams. Words on the boundary of each hypothesis are encoded in state, con- sisting of left state (the first few words) and right state (the last few words). We speed concatenation by encoding left state using data structure pointers in lieu of vocabulary in- dices and by avoiding unnecessary queries. To increase the decoder’s opportunities to recombine hypothesis, we mini- mize the number of words encoded by left state. This has the effect of reducing search errors made by the decoder. The resulting gain in model score is smaller than for right state minimization, which we explain by observing a relationship between state minimization and language model probability. With a fixed cube pruning pop limit, we show a 3-6% re- duction in CPU time and improved model scores. Reducing the pop limit to the point where model scores tie the baseline yields a net 11% reduction in CPU time.
Left Language Model State for Syntactic Machine Translation
Federico, Marcello
2011-01-01
Abstract
Many syntactic machine translation decoders, including Moses, cdec, and Joshua, implement bottom-up dynamic programming to integrate N-gram language model proba- bilities into hypothesis scoring. These decoders concatenate hypotheses according to grammar rules, yielding larger hy- potheses and eventually complete translations. When hy- potheses are concatenated, the language model score is ad- justed to account for boundary-crossing n-grams. Words on the boundary of each hypothesis are encoded in state, con- sisting of left state (the first few words) and right state (the last few words). We speed concatenation by encoding left state using data structure pointers in lieu of vocabulary in- dices and by avoiding unnecessary queries. To increase the decoder’s opportunities to recombine hypothesis, we mini- mize the number of words encoded by left state. This has the effect of reducing search errors made by the decoder. The resulting gain in model score is smaller than for right state minimization, which we explain by observing a relationship between state minimization and language model probability. With a fixed cube pruning pop limit, we show a 3-6% re- duction in CPU time and improved model scores. Reducing the pop limit to the point where model scores tie the baseline yields a net 11% reduction in CPU time.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
