Using decision tree to hybrid morphology generation of Persian verb for English-Persian translation

摘要

Languages such as English need to be morphologically analyzed in translation into morphologically rich languages such as Persian. Analyzing the output of English to Persian machine translation systems illustrates that Persian morphology comes with many challenges especially in the verb conjugation. In this paper, we investigate three ways to deal with the morphology of Persian verb in machine translation (MT): no morphology generation in statistical MT, rule-based morphology generation in rule-based MT and a hybrid-model-independent morphology generation. By model-independent we mean that it is not based on statistical or rule-based MT and could be applied to any English to Persian MT as a post-processor. We select Google translator (translate.google.com) to show the performance of a statistical MT without any morphology generation component for the verb conjugation. Rule-based morphology generation is implemented as a part of a rule-based MT. Finally, we enrich the rule-based approach by statistical methods and information to present a hybrid model. A set of linguistically motivated features are defined using both English and Persian linguistic knowledge obtained from a parallel corpus. Then we make a model to predict six morphological features of the verb in Persian using decision tree classifier and generate an inflected verb form. In a real translation process, by applying our model to the output of Google translator and a rule-based MT as a post-processor, we achieve an improvement of about 0.7% absolute BLEU score in the best case. When we are given the gold lemma in our reference experiments, using the most common feature values as a baseline shows an improvement of almost 2.8% absolute BLEU score on a test set containing 15K sentences.