Missing Phrase Recovering by Combining Forward and Backward Phrase Translation Tables

摘要

We propose a method to recover missing phrases dropped in the phrase extraction algorithm. Those phrases, therefore, are not translated even though we tested the system with the training data. On the other hand, in native-to-foreign, or backward training, some missing phrases can be recovered. In this paper, we combined two phrase translation tables extracted by the source-to-target and target-to-source training for the sake of more complete phrase translation table. We re-estimated the lexical weights and phrase translation probabilities for each phrase pair. Additional combining weights were applied to both tables. We assessed our method on different combining weights by counting the missing phrases and calculating the BLEU scores and NIST scores. Approximately 7% of missing phrases are recovered and 1.3% of BLEU score is increased.